1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This tablegen backend emits a target specifier matcher for converting parsed 11 // assembly operands in the MCInst structures. It also emits a matcher for 12 // custom operand parsing. 13 // 14 // Converting assembly operands into MCInst structures 15 // --------------------------------------------------- 16 // 17 // The input to the target specific matcher is a list of literal tokens and 18 // operands. The target specific parser should generally eliminate any syntax 19 // which is not relevant for matching; for example, comma tokens should have 20 // already been consumed and eliminated by the parser. Most instructions will 21 // end up with a single literal token (the instruction name) and some number of 22 // operands. 23 // 24 // Some example inputs, for X86: 25 // 'addl' (immediate ...) (register ...) 26 // 'add' (immediate ...) (memory ...) 27 // 'call' '*' %epc 28 // 29 // The assembly matcher is responsible for converting this input into a precise 30 // machine instruction (i.e., an instruction with a well defined encoding). This 31 // mapping has several properties which complicate matching: 32 // 33 // - It may be ambiguous; many architectures can legally encode particular 34 // variants of an instruction in different ways (for example, using a smaller 35 // encoding for small immediates). Such ambiguities should never be 36 // arbitrarily resolved by the assembler, the assembler is always responsible 37 // for choosing the "best" available instruction. 38 // 39 // - It may depend on the subtarget or the assembler context. Instructions 40 // which are invalid for the current mode, but otherwise unambiguous (e.g., 41 // an SSE instruction in a file being assembled for i486) should be accepted 42 // and rejected by the assembler front end. However, if the proper encoding 43 // for an instruction is dependent on the assembler context then the matcher 44 // is responsible for selecting the correct machine instruction for the 45 // current mode. 46 // 47 // The core matching algorithm attempts to exploit the regularity in most 48 // instruction sets to quickly determine the set of possibly matching 49 // instructions, and the simplify the generated code. Additionally, this helps 50 // to ensure that the ambiguities are intentionally resolved by the user. 51 // 52 // The matching is divided into two distinct phases: 53 // 54 // 1. Classification: Each operand is mapped to the unique set which (a) 55 // contains it, and (b) is the largest such subset for which a single 56 // instruction could match all members. 57 // 58 // For register classes, we can generate these subgroups automatically. For 59 // arbitrary operands, we expect the user to define the classes and their 60 // relations to one another (for example, 8-bit signed immediates as a 61 // subset of 32-bit immediates). 62 // 63 // By partitioning the operands in this way, we guarantee that for any 64 // tuple of classes, any single instruction must match either all or none 65 // of the sets of operands which could classify to that tuple. 66 // 67 // In addition, the subset relation amongst classes induces a partial order 68 // on such tuples, which we use to resolve ambiguities. 69 // 70 // 2. The input can now be treated as a tuple of classes (static tokens are 71 // simple singleton sets). Each such tuple should generally map to a single 72 // instruction (we currently ignore cases where this isn't true, whee!!!), 73 // which we can emit a simple matcher for. 74 // 75 // Custom Operand Parsing 76 // ---------------------- 77 // 78 // Some targets need a custom way to parse operands, some specific instructions 79 // can contain arguments that can represent processor flags and other kinds of 80 // identifiers that need to be mapped to specific values in the final encoded 81 // instructions. The target specific custom operand parsing works in the 82 // following way: 83 // 84 // 1. A operand match table is built, each entry contains a mnemonic, an 85 // operand class, a mask for all operand positions for that same 86 // class/mnemonic and target features to be checked while trying to match. 87 // 88 // 2. The operand matcher will try every possible entry with the same 89 // mnemonic and will check if the target feature for this mnemonic also 90 // matches. After that, if the operand to be matched has its index 91 // present in the mask, a successful match occurs. Otherwise, fallback 92 // to the regular operand parsing. 93 // 94 // 3. For a match success, each operand class that has a 'ParserMethod' 95 // becomes part of a switch from where the custom method is called. 96 // 97 //===----------------------------------------------------------------------===// 98 99 #include "CodeGenTarget.h" 100 #include "llvm/ADT/PointerUnion.h" 101 #include "llvm/ADT/STLExtras.h" 102 #include "llvm/ADT/SmallPtrSet.h" 103 #include "llvm/ADT/SmallVector.h" 104 #include "llvm/ADT/StringExtras.h" 105 #include "llvm/Support/CommandLine.h" 106 #include "llvm/Support/Debug.h" 107 #include "llvm/Support/ErrorHandling.h" 108 #include "llvm/TableGen/Error.h" 109 #include "llvm/TableGen/Record.h" 110 #include "llvm/TableGen/StringMatcher.h" 111 #include "llvm/TableGen/StringToOffsetTable.h" 112 #include "llvm/TableGen/TableGenBackend.h" 113 #include <cassert> 114 #include <cctype> 115 #include <forward_list> 116 #include <map> 117 #include <set> 118 119 using namespace llvm; 120 121 #define DEBUG_TYPE "asm-matcher-emitter" 122 123 static cl::opt<std::string> 124 MatchPrefix("match-prefix", cl::init(""), 125 cl::desc("Only match instructions with the given prefix")); 126 127 namespace { 128 class AsmMatcherInfo; 129 struct SubtargetFeatureInfo; 130 131 // Register sets are used as keys in some second-order sets TableGen creates 132 // when generating its data structures. This means that the order of two 133 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and 134 // can even affect compiler output (at least seen in diagnostics produced when 135 // all matches fail). So we use a type that sorts them consistently. 136 typedef std::set<Record*, LessRecordByID> RegisterSet; 137 138 class AsmMatcherEmitter { 139 RecordKeeper &Records; 140 public: 141 AsmMatcherEmitter(RecordKeeper &R) : Records(R) {} 142 143 void run(raw_ostream &o); 144 }; 145 146 /// ClassInfo - Helper class for storing the information about a particular 147 /// class of operands which can be matched. 148 struct ClassInfo { 149 enum ClassInfoKind { 150 /// Invalid kind, for use as a sentinel value. 151 Invalid = 0, 152 153 /// The class for a particular token. 154 Token, 155 156 /// The (first) register class, subsequent register classes are 157 /// RegisterClass0+1, and so on. 158 RegisterClass0, 159 160 /// The (first) user defined class, subsequent user defined classes are 161 /// UserClass0+1, and so on. 162 UserClass0 = 1<<16 163 }; 164 165 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + 166 /// N) for the Nth user defined class. 167 unsigned Kind; 168 169 /// SuperClasses - The super classes of this class. Note that for simplicities 170 /// sake user operands only record their immediate super class, while register 171 /// operands include all superclasses. 172 std::vector<ClassInfo*> SuperClasses; 173 174 /// Name - The full class name, suitable for use in an enum. 175 std::string Name; 176 177 /// ClassName - The unadorned generic name for this class (e.g., Token). 178 std::string ClassName; 179 180 /// ValueName - The name of the value this class represents; for a token this 181 /// is the literal token string, for an operand it is the TableGen class (or 182 /// empty if this is a derived class). 183 std::string ValueName; 184 185 /// PredicateMethod - The name of the operand method to test whether the 186 /// operand matches this class; this is not valid for Token or register kinds. 187 std::string PredicateMethod; 188 189 /// RenderMethod - The name of the operand method to add this operand to an 190 /// MCInst; this is not valid for Token or register kinds. 191 std::string RenderMethod; 192 193 /// ParserMethod - The name of the operand method to do a target specific 194 /// parsing on the operand. 195 std::string ParserMethod; 196 197 /// For register classes: the records for all the registers in this class. 198 RegisterSet Registers; 199 200 /// For custom match classes: the diagnostic kind for when the predicate fails. 201 std::string DiagnosticType; 202 203 /// Is this operand optional and not always required. 204 bool IsOptional; 205 206 /// DefaultMethod - The name of the method that returns the default operand 207 /// for optional operand 208 std::string DefaultMethod; 209 210 public: 211 /// isRegisterClass() - Check if this is a register class. 212 bool isRegisterClass() const { 213 return Kind >= RegisterClass0 && Kind < UserClass0; 214 } 215 216 /// isUserClass() - Check if this is a user defined class. 217 bool isUserClass() const { 218 return Kind >= UserClass0; 219 } 220 221 /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes 222 /// are related if they are in the same class hierarchy. 223 bool isRelatedTo(const ClassInfo &RHS) const { 224 // Tokens are only related to tokens. 225 if (Kind == Token || RHS.Kind == Token) 226 return Kind == Token && RHS.Kind == Token; 227 228 // Registers classes are only related to registers classes, and only if 229 // their intersection is non-empty. 230 if (isRegisterClass() || RHS.isRegisterClass()) { 231 if (!isRegisterClass() || !RHS.isRegisterClass()) 232 return false; 233 234 RegisterSet Tmp; 235 std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin()); 236 std::set_intersection(Registers.begin(), Registers.end(), 237 RHS.Registers.begin(), RHS.Registers.end(), 238 II, LessRecordByID()); 239 240 return !Tmp.empty(); 241 } 242 243 // Otherwise we have two users operands; they are related if they are in the 244 // same class hierarchy. 245 // 246 // FIXME: This is an oversimplification, they should only be related if they 247 // intersect, however we don't have that information. 248 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!"); 249 const ClassInfo *Root = this; 250 while (!Root->SuperClasses.empty()) 251 Root = Root->SuperClasses.front(); 252 253 const ClassInfo *RHSRoot = &RHS; 254 while (!RHSRoot->SuperClasses.empty()) 255 RHSRoot = RHSRoot->SuperClasses.front(); 256 257 return Root == RHSRoot; 258 } 259 260 /// isSubsetOf - Test whether this class is a subset of \p RHS. 261 bool isSubsetOf(const ClassInfo &RHS) const { 262 // This is a subset of RHS if it is the same class... 263 if (this == &RHS) 264 return true; 265 266 // ... or if any of its super classes are a subset of RHS. 267 for (const ClassInfo *CI : SuperClasses) 268 if (CI->isSubsetOf(RHS)) 269 return true; 270 271 return false; 272 } 273 274 int getTreeDepth() const { 275 int Depth = 0; 276 const ClassInfo *Root = this; 277 while (!Root->SuperClasses.empty()) { 278 Depth++; 279 Root = Root->SuperClasses.front(); 280 } 281 return Depth; 282 } 283 284 const ClassInfo *findRoot() const { 285 const ClassInfo *Root = this; 286 while (!Root->SuperClasses.empty()) 287 Root = Root->SuperClasses.front(); 288 return Root; 289 } 290 291 /// Compare two classes. This does not produce a total ordering, but does 292 /// guarantee that subclasses are sorted before their parents, and that the 293 /// ordering is transitive. 294 bool operator<(const ClassInfo &RHS) const { 295 if (this == &RHS) 296 return false; 297 298 // First, enforce the ordering between the three different types of class. 299 // Tokens sort before registers, which sort before user classes. 300 if (Kind == Token) { 301 if (RHS.Kind != Token) 302 return true; 303 assert(RHS.Kind == Token); 304 } else if (isRegisterClass()) { 305 if (RHS.Kind == Token) 306 return false; 307 else if (RHS.isUserClass()) 308 return true; 309 assert(RHS.isRegisterClass()); 310 } else if (isUserClass()) { 311 if (!RHS.isUserClass()) 312 return false; 313 assert(RHS.isUserClass()); 314 } else { 315 llvm_unreachable("Unknown ClassInfoKind"); 316 } 317 318 if (Kind == Token || isUserClass()) { 319 // Related tokens and user classes get sorted by depth in the inheritence 320 // tree (so that subclasses are before their parents). 321 if (isRelatedTo(RHS)) { 322 if (getTreeDepth() > RHS.getTreeDepth()) 323 return true; 324 if (getTreeDepth() < RHS.getTreeDepth()) 325 return false; 326 } else { 327 // Unrelated tokens and user classes are ordered by the name of their 328 // root nodes, so that there is a consistent ordering between 329 // unconnected trees. 330 return findRoot()->ValueName < RHS.findRoot()->ValueName; 331 } 332 } else if (isRegisterClass()) { 333 // For register sets, sort by number of registers. This guarantees that 334 // a set will always sort before all of it's strict supersets. 335 if (Registers.size() != RHS.Registers.size()) 336 return Registers.size() < RHS.Registers.size(); 337 } else { 338 llvm_unreachable("Unknown ClassInfoKind"); 339 } 340 341 // FIXME: We should be able to just return false here, as we only need a 342 // partial order (we use stable sorts, so this is deterministic) and the 343 // name of a class shouldn't be significant. However, some of the backends 344 // accidentally rely on this behaviour, so it will have to stay like this 345 // until they are fixed. 346 return ValueName < RHS.ValueName; 347 } 348 }; 349 350 class AsmVariantInfo { 351 public: 352 std::string RegisterPrefix; 353 std::string TokenizingCharacters; 354 std::string SeparatorCharacters; 355 std::string BreakCharacters; 356 int AsmVariantNo; 357 }; 358 359 /// MatchableInfo - Helper class for storing the necessary information for an 360 /// instruction or alias which is capable of being matched. 361 struct MatchableInfo { 362 struct AsmOperand { 363 /// Token - This is the token that the operand came from. 364 StringRef Token; 365 366 /// The unique class instance this operand should match. 367 ClassInfo *Class; 368 369 /// The operand name this is, if anything. 370 StringRef SrcOpName; 371 372 /// The suboperand index within SrcOpName, or -1 for the entire operand. 373 int SubOpIdx; 374 375 /// Whether the token is "isolated", i.e., it is preceded and followed 376 /// by separators. 377 bool IsIsolatedToken; 378 379 /// Register record if this token is singleton register. 380 Record *SingletonReg; 381 382 explicit AsmOperand(bool IsIsolatedToken, StringRef T) 383 : Token(T), Class(nullptr), SubOpIdx(-1), 384 IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {} 385 }; 386 387 /// ResOperand - This represents a single operand in the result instruction 388 /// generated by the match. In cases (like addressing modes) where a single 389 /// assembler operand expands to multiple MCOperands, this represents the 390 /// single assembler operand, not the MCOperand. 391 struct ResOperand { 392 enum { 393 /// RenderAsmOperand - This represents an operand result that is 394 /// generated by calling the render method on the assembly operand. The 395 /// corresponding AsmOperand is specified by AsmOperandNum. 396 RenderAsmOperand, 397 398 /// TiedOperand - This represents a result operand that is a duplicate of 399 /// a previous result operand. 400 TiedOperand, 401 402 /// ImmOperand - This represents an immediate value that is dumped into 403 /// the operand. 404 ImmOperand, 405 406 /// RegOperand - This represents a fixed register that is dumped in. 407 RegOperand 408 } Kind; 409 410 union { 411 /// This is the operand # in the AsmOperands list that this should be 412 /// copied from. 413 unsigned AsmOperandNum; 414 415 /// TiedOperandNum - This is the (earlier) result operand that should be 416 /// copied from. 417 unsigned TiedOperandNum; 418 419 /// ImmVal - This is the immediate value added to the instruction. 420 int64_t ImmVal; 421 422 /// Register - This is the register record. 423 Record *Register; 424 }; 425 426 /// MINumOperands - The number of MCInst operands populated by this 427 /// operand. 428 unsigned MINumOperands; 429 430 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) { 431 ResOperand X; 432 X.Kind = RenderAsmOperand; 433 X.AsmOperandNum = AsmOpNum; 434 X.MINumOperands = NumOperands; 435 return X; 436 } 437 438 static ResOperand getTiedOp(unsigned TiedOperandNum) { 439 ResOperand X; 440 X.Kind = TiedOperand; 441 X.TiedOperandNum = TiedOperandNum; 442 X.MINumOperands = 1; 443 return X; 444 } 445 446 static ResOperand getImmOp(int64_t Val) { 447 ResOperand X; 448 X.Kind = ImmOperand; 449 X.ImmVal = Val; 450 X.MINumOperands = 1; 451 return X; 452 } 453 454 static ResOperand getRegOp(Record *Reg) { 455 ResOperand X; 456 X.Kind = RegOperand; 457 X.Register = Reg; 458 X.MINumOperands = 1; 459 return X; 460 } 461 }; 462 463 /// AsmVariantID - Target's assembly syntax variant no. 464 int AsmVariantID; 465 466 /// AsmString - The assembly string for this instruction (with variants 467 /// removed), e.g. "movsx $src, $dst". 468 std::string AsmString; 469 470 /// TheDef - This is the definition of the instruction or InstAlias that this 471 /// matchable came from. 472 Record *const TheDef; 473 474 /// DefRec - This is the definition that it came from. 475 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec; 476 477 const CodeGenInstruction *getResultInst() const { 478 if (DefRec.is<const CodeGenInstruction*>()) 479 return DefRec.get<const CodeGenInstruction*>(); 480 return DefRec.get<const CodeGenInstAlias*>()->ResultInst; 481 } 482 483 /// ResOperands - This is the operand list that should be built for the result 484 /// MCInst. 485 SmallVector<ResOperand, 8> ResOperands; 486 487 /// Mnemonic - This is the first token of the matched instruction, its 488 /// mnemonic. 489 StringRef Mnemonic; 490 491 /// AsmOperands - The textual operands that this instruction matches, 492 /// annotated with a class and where in the OperandList they were defined. 493 /// This directly corresponds to the tokenized AsmString after the mnemonic is 494 /// removed. 495 SmallVector<AsmOperand, 8> AsmOperands; 496 497 /// Predicates - The required subtarget features to match this instruction. 498 SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures; 499 500 /// ConversionFnKind - The enum value which is passed to the generated 501 /// convertToMCInst to convert parsed operands into an MCInst for this 502 /// function. 503 std::string ConversionFnKind; 504 505 /// If this instruction is deprecated in some form. 506 bool HasDeprecation; 507 508 /// If this is an alias, this is use to determine whether or not to using 509 /// the conversion function defined by the instruction's AsmMatchConverter 510 /// or to use the function generated by the alias. 511 bool UseInstAsmMatchConverter; 512 513 MatchableInfo(const CodeGenInstruction &CGI) 514 : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI), 515 UseInstAsmMatchConverter(true) { 516 } 517 518 MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias) 519 : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef), 520 DefRec(Alias.release()), 521 UseInstAsmMatchConverter( 522 TheDef->getValueAsBit("UseInstAsmMatchConverter")) { 523 } 524 525 // Could remove this and the dtor if PointerUnion supported unique_ptr 526 // elements with a dynamic failure/assertion (like the one below) in the case 527 // where it was copied while being in an owning state. 528 MatchableInfo(const MatchableInfo &RHS) 529 : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString), 530 TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands), 531 Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands), 532 RequiredFeatures(RHS.RequiredFeatures), 533 ConversionFnKind(RHS.ConversionFnKind), 534 HasDeprecation(RHS.HasDeprecation), 535 UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) { 536 assert(!DefRec.is<const CodeGenInstAlias *>()); 537 } 538 539 ~MatchableInfo() { 540 delete DefRec.dyn_cast<const CodeGenInstAlias*>(); 541 } 542 543 // Two-operand aliases clone from the main matchable, but mark the second 544 // operand as a tied operand of the first for purposes of the assembler. 545 void formTwoOperandAlias(StringRef Constraint); 546 547 void initialize(const AsmMatcherInfo &Info, 548 SmallPtrSetImpl<Record*> &SingletonRegisters, 549 AsmVariantInfo const &Variant, 550 bool HasMnemonicFirst); 551 552 /// validate - Return true if this matchable is a valid thing to match against 553 /// and perform a bunch of validity checking. 554 bool validate(StringRef CommentDelimiter, bool Hack) const; 555 556 /// findAsmOperand - Find the AsmOperand with the specified name and 557 /// suboperand index. 558 int findAsmOperand(StringRef N, int SubOpIdx) const { 559 auto I = std::find_if(AsmOperands.begin(), AsmOperands.end(), 560 [&](const AsmOperand &Op) { 561 return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx; 562 }); 563 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; 564 } 565 566 /// findAsmOperandNamed - Find the first AsmOperand with the specified name. 567 /// This does not check the suboperand index. 568 int findAsmOperandNamed(StringRef N) const { 569 auto I = std::find_if(AsmOperands.begin(), AsmOperands.end(), 570 [&](const AsmOperand &Op) { 571 return Op.SrcOpName == N; 572 }); 573 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; 574 } 575 576 void buildInstructionResultOperands(); 577 void buildAliasResultOperands(); 578 579 /// operator< - Compare two matchables. 580 bool operator<(const MatchableInfo &RHS) const { 581 // The primary comparator is the instruction mnemonic. 582 if (int Cmp = Mnemonic.compare(RHS.Mnemonic)) 583 return Cmp == -1; 584 585 if (AsmOperands.size() != RHS.AsmOperands.size()) 586 return AsmOperands.size() < RHS.AsmOperands.size(); 587 588 // Compare lexicographically by operand. The matcher validates that other 589 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith(). 590 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 591 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 592 return true; 593 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 594 return false; 595 } 596 597 // Give matches that require more features higher precedence. This is useful 598 // because we cannot define AssemblerPredicates with the negation of 599 // processor features. For example, ARM v6 "nop" may be either a HINT or 600 // MOV. With v6, we want to match HINT. The assembler has no way to 601 // predicate MOV under "NoV6", but HINT will always match first because it 602 // requires V6 while MOV does not. 603 if (RequiredFeatures.size() != RHS.RequiredFeatures.size()) 604 return RequiredFeatures.size() > RHS.RequiredFeatures.size(); 605 606 return false; 607 } 608 609 /// couldMatchAmbiguouslyWith - Check whether this matchable could 610 /// ambiguously match the same set of operands as \p RHS (without being a 611 /// strictly superior match). 612 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const { 613 // The primary comparator is the instruction mnemonic. 614 if (Mnemonic != RHS.Mnemonic) 615 return false; 616 617 // The number of operands is unambiguous. 618 if (AsmOperands.size() != RHS.AsmOperands.size()) 619 return false; 620 621 // Otherwise, make sure the ordering of the two instructions is unambiguous 622 // by checking that either (a) a token or operand kind discriminates them, 623 // or (b) the ordering among equivalent kinds is consistent. 624 625 // Tokens and operand kinds are unambiguous (assuming a correct target 626 // specific parser). 627 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) 628 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind || 629 AsmOperands[i].Class->Kind == ClassInfo::Token) 630 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class || 631 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 632 return false; 633 634 // Otherwise, this operand could commute if all operands are equivalent, or 635 // there is a pair of operands that compare less than and a pair that 636 // compare greater than. 637 bool HasLT = false, HasGT = false; 638 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 639 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 640 HasLT = true; 641 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 642 HasGT = true; 643 } 644 645 return HasLT == HasGT; 646 } 647 648 void dump() const; 649 650 private: 651 void tokenizeAsmString(AsmMatcherInfo const &Info, 652 AsmVariantInfo const &Variant); 653 void addAsmOperand(StringRef Token, bool IsIsolatedToken = false); 654 }; 655 656 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget 657 /// feature which participates in instruction matching. 658 struct SubtargetFeatureInfo { 659 /// \brief The predicate record for this feature. 660 Record *TheDef; 661 662 /// \brief An unique index assigned to represent this feature. 663 uint64_t Index; 664 665 SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {} 666 667 /// \brief The name of the enumerated constant identifying this feature. 668 std::string getEnumName() const { 669 return "Feature_" + TheDef->getName(); 670 } 671 672 void dump() const { 673 errs() << getEnumName() << " " << Index << "\n"; 674 TheDef->dump(); 675 } 676 }; 677 678 struct OperandMatchEntry { 679 unsigned OperandMask; 680 const MatchableInfo* MI; 681 ClassInfo *CI; 682 683 static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci, 684 unsigned opMask) { 685 OperandMatchEntry X; 686 X.OperandMask = opMask; 687 X.CI = ci; 688 X.MI = mi; 689 return X; 690 } 691 }; 692 693 class AsmMatcherInfo { 694 public: 695 /// Tracked Records 696 RecordKeeper &Records; 697 698 /// The tablegen AsmParser record. 699 Record *AsmParser; 700 701 /// Target - The target information. 702 CodeGenTarget &Target; 703 704 /// The classes which are needed for matching. 705 std::forward_list<ClassInfo> Classes; 706 707 /// The information on the matchables to match. 708 std::vector<std::unique_ptr<MatchableInfo>> Matchables; 709 710 /// Info for custom matching operands by user defined methods. 711 std::vector<OperandMatchEntry> OperandMatchInfo; 712 713 /// Map of Register records to their class information. 714 typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy; 715 RegisterClassesTy RegisterClasses; 716 717 /// Map of Predicate records to their subtarget information. 718 std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures; 719 720 /// Map of AsmOperandClass records to their class information. 721 std::map<Record*, ClassInfo*> AsmOperandClasses; 722 723 private: 724 /// Map of token to class information which has already been constructed. 725 std::map<std::string, ClassInfo*> TokenClasses; 726 727 /// Map of RegisterClass records to their class information. 728 std::map<Record*, ClassInfo*> RegisterClassClasses; 729 730 private: 731 /// getTokenClass - Lookup or create the class for the given token. 732 ClassInfo *getTokenClass(StringRef Token); 733 734 /// getOperandClass - Lookup or create the class for the given operand. 735 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI, 736 int SubOpIdx); 737 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx); 738 739 /// buildRegisterClasses - Build the ClassInfo* instances for register 740 /// classes. 741 void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters); 742 743 /// buildOperandClasses - Build the ClassInfo* instances for user defined 744 /// operand classes. 745 void buildOperandClasses(); 746 747 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName, 748 unsigned AsmOpIdx); 749 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName, 750 MatchableInfo::AsmOperand &Op); 751 752 public: 753 AsmMatcherInfo(Record *AsmParser, 754 CodeGenTarget &Target, 755 RecordKeeper &Records); 756 757 /// buildInfo - Construct the various tables used during matching. 758 void buildInfo(); 759 760 /// buildOperandMatchInfo - Build the necessary information to handle user 761 /// defined operand parsing methods. 762 void buildOperandMatchInfo(); 763 764 /// getSubtargetFeature - Lookup or create the subtarget feature info for the 765 /// given operand. 766 const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const { 767 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!"); 768 const auto &I = SubtargetFeatures.find(Def); 769 return I == SubtargetFeatures.end() ? nullptr : &I->second; 770 } 771 772 RecordKeeper &getRecords() const { 773 return Records; 774 } 775 776 bool hasOptionalOperands() const { 777 return std::find_if(Classes.begin(), Classes.end(), 778 [](const ClassInfo& Class){ return Class.IsOptional; }) 779 != Classes.end(); 780 } 781 }; 782 783 } // end anonymous namespace 784 785 void MatchableInfo::dump() const { 786 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n"; 787 788 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 789 const AsmOperand &Op = AsmOperands[i]; 790 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - "; 791 errs() << '\"' << Op.Token << "\"\n"; 792 } 793 } 794 795 static std::pair<StringRef, StringRef> 796 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) { 797 // Split via the '='. 798 std::pair<StringRef, StringRef> Ops = S.split('='); 799 if (Ops.second == "") 800 PrintFatalError(Loc, "missing '=' in two-operand alias constraint"); 801 // Trim whitespace and the leading '$' on the operand names. 802 size_t start = Ops.first.find_first_of('$'); 803 if (start == std::string::npos) 804 PrintFatalError(Loc, "expected '$' prefix on asm operand name"); 805 Ops.first = Ops.first.slice(start + 1, std::string::npos); 806 size_t end = Ops.first.find_last_of(" \t"); 807 Ops.first = Ops.first.slice(0, end); 808 // Now the second operand. 809 start = Ops.second.find_first_of('$'); 810 if (start == std::string::npos) 811 PrintFatalError(Loc, "expected '$' prefix on asm operand name"); 812 Ops.second = Ops.second.slice(start + 1, std::string::npos); 813 end = Ops.second.find_last_of(" \t"); 814 Ops.first = Ops.first.slice(0, end); 815 return Ops; 816 } 817 818 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) { 819 // Figure out which operands are aliased and mark them as tied. 820 std::pair<StringRef, StringRef> Ops = 821 parseTwoOperandConstraint(Constraint, TheDef->getLoc()); 822 823 // Find the AsmOperands that refer to the operands we're aliasing. 824 int SrcAsmOperand = findAsmOperandNamed(Ops.first); 825 int DstAsmOperand = findAsmOperandNamed(Ops.second); 826 if (SrcAsmOperand == -1) 827 PrintFatalError(TheDef->getLoc(), 828 "unknown source two-operand alias operand '" + Ops.first + 829 "'."); 830 if (DstAsmOperand == -1) 831 PrintFatalError(TheDef->getLoc(), 832 "unknown destination two-operand alias operand '" + 833 Ops.second + "'."); 834 835 // Find the ResOperand that refers to the operand we're aliasing away 836 // and update it to refer to the combined operand instead. 837 for (ResOperand &Op : ResOperands) { 838 if (Op.Kind == ResOperand::RenderAsmOperand && 839 Op.AsmOperandNum == (unsigned)SrcAsmOperand) { 840 Op.AsmOperandNum = DstAsmOperand; 841 break; 842 } 843 } 844 // Remove the AsmOperand for the alias operand. 845 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand); 846 // Adjust the ResOperand references to any AsmOperands that followed 847 // the one we just deleted. 848 for (ResOperand &Op : ResOperands) { 849 switch(Op.Kind) { 850 default: 851 // Nothing to do for operands that don't reference AsmOperands. 852 break; 853 case ResOperand::RenderAsmOperand: 854 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand) 855 --Op.AsmOperandNum; 856 break; 857 case ResOperand::TiedOperand: 858 if (Op.TiedOperandNum > (unsigned)SrcAsmOperand) 859 --Op.TiedOperandNum; 860 break; 861 } 862 } 863 } 864 865 /// extractSingletonRegisterForAsmOperand - Extract singleton register, 866 /// if present, from specified token. 867 static void 868 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op, 869 const AsmMatcherInfo &Info, 870 StringRef RegisterPrefix) { 871 StringRef Tok = Op.Token; 872 873 // If this token is not an isolated token, i.e., it isn't separated from 874 // other tokens (e.g. with whitespace), don't interpret it as a register name. 875 if (!Op.IsIsolatedToken) 876 return; 877 878 if (RegisterPrefix.empty()) { 879 std::string LoweredTok = Tok.lower(); 880 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok)) 881 Op.SingletonReg = Reg->TheDef; 882 return; 883 } 884 885 if (!Tok.startswith(RegisterPrefix)) 886 return; 887 888 StringRef RegName = Tok.substr(RegisterPrefix.size()); 889 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName)) 890 Op.SingletonReg = Reg->TheDef; 891 892 // If there is no register prefix (i.e. "%" in "%eax"), then this may 893 // be some random non-register token, just ignore it. 894 } 895 896 void MatchableInfo::initialize(const AsmMatcherInfo &Info, 897 SmallPtrSetImpl<Record*> &SingletonRegisters, 898 AsmVariantInfo const &Variant, 899 bool HasMnemonicFirst) { 900 AsmVariantID = Variant.AsmVariantNo; 901 AsmString = 902 CodeGenInstruction::FlattenAsmStringVariants(AsmString, 903 Variant.AsmVariantNo); 904 905 tokenizeAsmString(Info, Variant); 906 907 // The first token of the instruction is the mnemonic, which must be a 908 // simple string, not a $foo variable or a singleton register. 909 if (AsmOperands.empty()) 910 PrintFatalError(TheDef->getLoc(), 911 "Instruction '" + TheDef->getName() + "' has no tokens"); 912 913 assert(!AsmOperands[0].Token.empty()); 914 if (HasMnemonicFirst) { 915 Mnemonic = AsmOperands[0].Token; 916 if (Mnemonic[0] == '$') 917 PrintFatalError(TheDef->getLoc(), 918 "Invalid instruction mnemonic '" + Mnemonic + "'!"); 919 920 // Remove the first operand, it is tracked in the mnemonic field. 921 AsmOperands.erase(AsmOperands.begin()); 922 } else if (AsmOperands[0].Token[0] != '$') 923 Mnemonic = AsmOperands[0].Token; 924 925 // Compute the require features. 926 for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates")) 927 if (const SubtargetFeatureInfo *Feature = 928 Info.getSubtargetFeature(Predicate)) 929 RequiredFeatures.push_back(Feature); 930 931 // Collect singleton registers, if used. 932 for (MatchableInfo::AsmOperand &Op : AsmOperands) { 933 extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix); 934 if (Record *Reg = Op.SingletonReg) 935 SingletonRegisters.insert(Reg); 936 } 937 938 const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask"); 939 if (!DepMask) 940 DepMask = TheDef->getValue("ComplexDeprecationPredicate"); 941 942 HasDeprecation = 943 DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false; 944 } 945 946 /// Append an AsmOperand for the given substring of AsmString. 947 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) { 948 AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token)); 949 } 950 951 /// tokenizeAsmString - Tokenize a simplified assembly string. 952 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info, 953 AsmVariantInfo const &Variant) { 954 StringRef String = AsmString; 955 size_t Prev = 0; 956 bool InTok = false; 957 bool IsIsolatedToken = true; 958 for (size_t i = 0, e = String.size(); i != e; ++i) { 959 char Char = String[i]; 960 if (Variant.BreakCharacters.find(Char) != std::string::npos) { 961 if (InTok) { 962 addAsmOperand(String.slice(Prev, i), false); 963 Prev = i; 964 IsIsolatedToken = false; 965 } 966 InTok = true; 967 continue; 968 } 969 if (Variant.TokenizingCharacters.find(Char) != std::string::npos) { 970 if (InTok) { 971 addAsmOperand(String.slice(Prev, i), IsIsolatedToken); 972 InTok = false; 973 IsIsolatedToken = false; 974 } 975 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken); 976 Prev = i + 1; 977 IsIsolatedToken = true; 978 continue; 979 } 980 if (Variant.SeparatorCharacters.find(Char) != std::string::npos) { 981 if (InTok) { 982 addAsmOperand(String.slice(Prev, i), IsIsolatedToken); 983 InTok = false; 984 } 985 Prev = i + 1; 986 IsIsolatedToken = true; 987 continue; 988 } 989 990 switch (Char) { 991 case '\\': 992 if (InTok) { 993 addAsmOperand(String.slice(Prev, i), false); 994 InTok = false; 995 IsIsolatedToken = false; 996 } 997 ++i; 998 assert(i != String.size() && "Invalid quoted character"); 999 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken); 1000 Prev = i + 1; 1001 IsIsolatedToken = false; 1002 break; 1003 1004 case '$': { 1005 if (InTok) { 1006 addAsmOperand(String.slice(Prev, i), false); 1007 InTok = false; 1008 IsIsolatedToken = false; 1009 } 1010 1011 // If this isn't "${", start new identifier looking like "$xxx" 1012 if (i + 1 == String.size() || String[i + 1] != '{') { 1013 Prev = i; 1014 break; 1015 } 1016 1017 size_t EndPos = String.find('}', i); 1018 assert(EndPos != StringRef::npos && 1019 "Missing brace in operand reference!"); 1020 addAsmOperand(String.slice(i, EndPos+1), IsIsolatedToken); 1021 Prev = EndPos + 1; 1022 i = EndPos; 1023 IsIsolatedToken = false; 1024 break; 1025 } 1026 1027 default: 1028 InTok = true; 1029 break; 1030 } 1031 } 1032 if (InTok && Prev != String.size()) 1033 addAsmOperand(String.substr(Prev), IsIsolatedToken); 1034 } 1035 1036 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const { 1037 // Reject matchables with no .s string. 1038 if (AsmString.empty()) 1039 PrintFatalError(TheDef->getLoc(), "instruction with empty asm string"); 1040 1041 // Reject any matchables with a newline in them, they should be marked 1042 // isCodeGenOnly if they are pseudo instructions. 1043 if (AsmString.find('\n') != std::string::npos) 1044 PrintFatalError(TheDef->getLoc(), 1045 "multiline instruction is not valid for the asmparser, " 1046 "mark it isCodeGenOnly"); 1047 1048 // Remove comments from the asm string. We know that the asmstring only 1049 // has one line. 1050 if (!CommentDelimiter.empty() && 1051 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos) 1052 PrintFatalError(TheDef->getLoc(), 1053 "asmstring for instruction has comment character in it, " 1054 "mark it isCodeGenOnly"); 1055 1056 // Reject matchables with operand modifiers, these aren't something we can 1057 // handle, the target should be refactored to use operands instead of 1058 // modifiers. 1059 // 1060 // Also, check for instructions which reference the operand multiple times; 1061 // this implies a constraint we would not honor. 1062 std::set<std::string> OperandNames; 1063 for (const AsmOperand &Op : AsmOperands) { 1064 StringRef Tok = Op.Token; 1065 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos) 1066 PrintFatalError(TheDef->getLoc(), 1067 "matchable with operand modifier '" + Tok + 1068 "' not supported by asm matcher. Mark isCodeGenOnly!"); 1069 1070 // Verify that any operand is only mentioned once. 1071 // We reject aliases and ignore instructions for now. 1072 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) { 1073 if (!Hack) 1074 PrintFatalError(TheDef->getLoc(), 1075 "ERROR: matchable with tied operand '" + Tok + 1076 "' can never be matched!"); 1077 // FIXME: Should reject these. The ARM backend hits this with $lane in a 1078 // bunch of instructions. It is unclear what the right answer is. 1079 DEBUG({ 1080 errs() << "warning: '" << TheDef->getName() << "': " 1081 << "ignoring instruction with tied operand '" 1082 << Tok << "'\n"; 1083 }); 1084 return false; 1085 } 1086 } 1087 1088 return true; 1089 } 1090 1091 static std::string getEnumNameForToken(StringRef Str) { 1092 std::string Res; 1093 1094 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) { 1095 switch (*it) { 1096 case '*': Res += "_STAR_"; break; 1097 case '%': Res += "_PCT_"; break; 1098 case ':': Res += "_COLON_"; break; 1099 case '!': Res += "_EXCLAIM_"; break; 1100 case '.': Res += "_DOT_"; break; 1101 case '<': Res += "_LT_"; break; 1102 case '>': Res += "_GT_"; break; 1103 case '-': Res += "_MINUS_"; break; 1104 default: 1105 if ((*it >= 'A' && *it <= 'Z') || 1106 (*it >= 'a' && *it <= 'z') || 1107 (*it >= '0' && *it <= '9')) 1108 Res += *it; 1109 else 1110 Res += "_" + utostr((unsigned) *it) + "_"; 1111 } 1112 } 1113 1114 return Res; 1115 } 1116 1117 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) { 1118 ClassInfo *&Entry = TokenClasses[Token]; 1119 1120 if (!Entry) { 1121 Classes.emplace_front(); 1122 Entry = &Classes.front(); 1123 Entry->Kind = ClassInfo::Token; 1124 Entry->ClassName = "Token"; 1125 Entry->Name = "MCK_" + getEnumNameForToken(Token); 1126 Entry->ValueName = Token; 1127 Entry->PredicateMethod = "<invalid>"; 1128 Entry->RenderMethod = "<invalid>"; 1129 Entry->ParserMethod = ""; 1130 Entry->DiagnosticType = ""; 1131 Entry->IsOptional = false; 1132 Entry->DefaultMethod = "<invalid>"; 1133 } 1134 1135 return Entry; 1136 } 1137 1138 ClassInfo * 1139 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI, 1140 int SubOpIdx) { 1141 Record *Rec = OI.Rec; 1142 if (SubOpIdx != -1) 1143 Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef(); 1144 return getOperandClass(Rec, SubOpIdx); 1145 } 1146 1147 ClassInfo * 1148 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) { 1149 if (Rec->isSubClassOf("RegisterOperand")) { 1150 // RegisterOperand may have an associated ParserMatchClass. If it does, 1151 // use it, else just fall back to the underlying register class. 1152 const RecordVal *R = Rec->getValue("ParserMatchClass"); 1153 if (!R || !R->getValue()) 1154 PrintFatalError("Record `" + Rec->getName() + 1155 "' does not have a ParserMatchClass!\n"); 1156 1157 if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) { 1158 Record *MatchClass = DI->getDef(); 1159 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 1160 return CI; 1161 } 1162 1163 // No custom match class. Just use the register class. 1164 Record *ClassRec = Rec->getValueAsDef("RegClass"); 1165 if (!ClassRec) 1166 PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() + 1167 "' has no associated register class!\n"); 1168 if (ClassInfo *CI = RegisterClassClasses[ClassRec]) 1169 return CI; 1170 PrintFatalError(Rec->getLoc(), "register class has no class info!"); 1171 } 1172 1173 if (Rec->isSubClassOf("RegisterClass")) { 1174 if (ClassInfo *CI = RegisterClassClasses[Rec]) 1175 return CI; 1176 PrintFatalError(Rec->getLoc(), "register class has no class info!"); 1177 } 1178 1179 if (!Rec->isSubClassOf("Operand")) 1180 PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() + 1181 "' does not derive from class Operand!\n"); 1182 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1183 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 1184 return CI; 1185 1186 PrintFatalError(Rec->getLoc(), "operand has no match class!"); 1187 } 1188 1189 struct LessRegisterSet { 1190 bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const { 1191 // std::set<T> defines its own compariso "operator<", but it 1192 // performs a lexicographical comparison by T's innate comparison 1193 // for some reason. We don't want non-deterministic pointer 1194 // comparisons so use this instead. 1195 return std::lexicographical_compare(LHS.begin(), LHS.end(), 1196 RHS.begin(), RHS.end(), 1197 LessRecordByID()); 1198 } 1199 }; 1200 1201 void AsmMatcherInfo:: 1202 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) { 1203 const auto &Registers = Target.getRegBank().getRegisters(); 1204 auto &RegClassList = Target.getRegBank().getRegClasses(); 1205 1206 typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet; 1207 1208 // The register sets used for matching. 1209 RegisterSetSet RegisterSets; 1210 1211 // Gather the defined sets. 1212 for (const CodeGenRegisterClass &RC : RegClassList) 1213 RegisterSets.insert( 1214 RegisterSet(RC.getOrder().begin(), RC.getOrder().end())); 1215 1216 // Add any required singleton sets. 1217 for (Record *Rec : SingletonRegisters) { 1218 RegisterSets.insert(RegisterSet(&Rec, &Rec + 1)); 1219 } 1220 1221 // Introduce derived sets where necessary (when a register does not determine 1222 // a unique register set class), and build the mapping of registers to the set 1223 // they should classify to. 1224 std::map<Record*, RegisterSet> RegisterMap; 1225 for (const CodeGenRegister &CGR : Registers) { 1226 // Compute the intersection of all sets containing this register. 1227 RegisterSet ContainingSet; 1228 1229 for (const RegisterSet &RS : RegisterSets) { 1230 if (!RS.count(CGR.TheDef)) 1231 continue; 1232 1233 if (ContainingSet.empty()) { 1234 ContainingSet = RS; 1235 continue; 1236 } 1237 1238 RegisterSet Tmp; 1239 std::swap(Tmp, ContainingSet); 1240 std::insert_iterator<RegisterSet> II(ContainingSet, 1241 ContainingSet.begin()); 1242 std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II, 1243 LessRecordByID()); 1244 } 1245 1246 if (!ContainingSet.empty()) { 1247 RegisterSets.insert(ContainingSet); 1248 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet)); 1249 } 1250 } 1251 1252 // Construct the register classes. 1253 std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses; 1254 unsigned Index = 0; 1255 for (const RegisterSet &RS : RegisterSets) { 1256 Classes.emplace_front(); 1257 ClassInfo *CI = &Classes.front(); 1258 CI->Kind = ClassInfo::RegisterClass0 + Index; 1259 CI->ClassName = "Reg" + utostr(Index); 1260 CI->Name = "MCK_Reg" + utostr(Index); 1261 CI->ValueName = ""; 1262 CI->PredicateMethod = ""; // unused 1263 CI->RenderMethod = "addRegOperands"; 1264 CI->Registers = RS; 1265 // FIXME: diagnostic type. 1266 CI->DiagnosticType = ""; 1267 CI->IsOptional = false; 1268 CI->DefaultMethod = ""; // unused 1269 RegisterSetClasses.insert(std::make_pair(RS, CI)); 1270 ++Index; 1271 } 1272 1273 // Find the superclasses; we could compute only the subgroup lattice edges, 1274 // but there isn't really a point. 1275 for (const RegisterSet &RS : RegisterSets) { 1276 ClassInfo *CI = RegisterSetClasses[RS]; 1277 for (const RegisterSet &RS2 : RegisterSets) 1278 if (RS != RS2 && 1279 std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(), 1280 LessRecordByID())) 1281 CI->SuperClasses.push_back(RegisterSetClasses[RS2]); 1282 } 1283 1284 // Name the register classes which correspond to a user defined RegisterClass. 1285 for (const CodeGenRegisterClass &RC : RegClassList) { 1286 // Def will be NULL for non-user defined register classes. 1287 Record *Def = RC.getDef(); 1288 if (!Def) 1289 continue; 1290 ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(), 1291 RC.getOrder().end())]; 1292 if (CI->ValueName.empty()) { 1293 CI->ClassName = RC.getName(); 1294 CI->Name = "MCK_" + RC.getName(); 1295 CI->ValueName = RC.getName(); 1296 } else 1297 CI->ValueName = CI->ValueName + "," + RC.getName(); 1298 1299 RegisterClassClasses.insert(std::make_pair(Def, CI)); 1300 } 1301 1302 // Populate the map for individual registers. 1303 for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(), 1304 ie = RegisterMap.end(); it != ie; ++it) 1305 RegisterClasses[it->first] = RegisterSetClasses[it->second]; 1306 1307 // Name the register classes which correspond to singleton registers. 1308 for (Record *Rec : SingletonRegisters) { 1309 ClassInfo *CI = RegisterClasses[Rec]; 1310 assert(CI && "Missing singleton register class info!"); 1311 1312 if (CI->ValueName.empty()) { 1313 CI->ClassName = Rec->getName(); 1314 CI->Name = "MCK_" + Rec->getName(); 1315 CI->ValueName = Rec->getName(); 1316 } else 1317 CI->ValueName = CI->ValueName + "," + Rec->getName(); 1318 } 1319 } 1320 1321 void AsmMatcherInfo::buildOperandClasses() { 1322 std::vector<Record*> AsmOperands = 1323 Records.getAllDerivedDefinitions("AsmOperandClass"); 1324 1325 // Pre-populate AsmOperandClasses map. 1326 for (Record *Rec : AsmOperands) { 1327 Classes.emplace_front(); 1328 AsmOperandClasses[Rec] = &Classes.front(); 1329 } 1330 1331 unsigned Index = 0; 1332 for (Record *Rec : AsmOperands) { 1333 ClassInfo *CI = AsmOperandClasses[Rec]; 1334 CI->Kind = ClassInfo::UserClass0 + Index; 1335 1336 ListInit *Supers = Rec->getValueAsListInit("SuperClasses"); 1337 for (Init *I : Supers->getValues()) { 1338 DefInit *DI = dyn_cast<DefInit>(I); 1339 if (!DI) { 1340 PrintError(Rec->getLoc(), "Invalid super class reference!"); 1341 continue; 1342 } 1343 1344 ClassInfo *SC = AsmOperandClasses[DI->getDef()]; 1345 if (!SC) 1346 PrintError(Rec->getLoc(), "Invalid super class reference!"); 1347 else 1348 CI->SuperClasses.push_back(SC); 1349 } 1350 CI->ClassName = Rec->getValueAsString("Name"); 1351 CI->Name = "MCK_" + CI->ClassName; 1352 CI->ValueName = Rec->getName(); 1353 1354 // Get or construct the predicate method name. 1355 Init *PMName = Rec->getValueInit("PredicateMethod"); 1356 if (StringInit *SI = dyn_cast<StringInit>(PMName)) { 1357 CI->PredicateMethod = SI->getValue(); 1358 } else { 1359 assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!"); 1360 CI->PredicateMethod = "is" + CI->ClassName; 1361 } 1362 1363 // Get or construct the render method name. 1364 Init *RMName = Rec->getValueInit("RenderMethod"); 1365 if (StringInit *SI = dyn_cast<StringInit>(RMName)) { 1366 CI->RenderMethod = SI->getValue(); 1367 } else { 1368 assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!"); 1369 CI->RenderMethod = "add" + CI->ClassName + "Operands"; 1370 } 1371 1372 // Get the parse method name or leave it as empty. 1373 Init *PRMName = Rec->getValueInit("ParserMethod"); 1374 if (StringInit *SI = dyn_cast<StringInit>(PRMName)) 1375 CI->ParserMethod = SI->getValue(); 1376 1377 // Get the diagnostic type or leave it as empty. 1378 // Get the parse method name or leave it as empty. 1379 Init *DiagnosticType = Rec->getValueInit("DiagnosticType"); 1380 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType)) 1381 CI->DiagnosticType = SI->getValue(); 1382 1383 Init *IsOptional = Rec->getValueInit("IsOptional"); 1384 if (BitInit *BI = dyn_cast<BitInit>(IsOptional)) 1385 CI->IsOptional = BI->getValue(); 1386 1387 // Get or construct the default method name. 1388 Init *DMName = Rec->getValueInit("DefaultMethod"); 1389 if (StringInit *SI = dyn_cast<StringInit>(DMName)) { 1390 CI->DefaultMethod = SI->getValue(); 1391 } else { 1392 assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!"); 1393 CI->DefaultMethod = "default" + CI->ClassName + "Operands"; 1394 } 1395 1396 ++Index; 1397 } 1398 } 1399 1400 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, 1401 CodeGenTarget &target, 1402 RecordKeeper &records) 1403 : Records(records), AsmParser(asmParser), Target(target) { 1404 } 1405 1406 /// buildOperandMatchInfo - Build the necessary information to handle user 1407 /// defined operand parsing methods. 1408 void AsmMatcherInfo::buildOperandMatchInfo() { 1409 1410 /// Map containing a mask with all operands indices that can be found for 1411 /// that class inside a instruction. 1412 typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy; 1413 OpClassMaskTy OpClassMask; 1414 1415 for (const auto &MI : Matchables) { 1416 OpClassMask.clear(); 1417 1418 // Keep track of all operands of this instructions which belong to the 1419 // same class. 1420 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) { 1421 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i]; 1422 if (Op.Class->ParserMethod.empty()) 1423 continue; 1424 unsigned &OperandMask = OpClassMask[Op.Class]; 1425 OperandMask |= (1 << i); 1426 } 1427 1428 // Generate operand match info for each mnemonic/operand class pair. 1429 for (const auto &OCM : OpClassMask) { 1430 unsigned OpMask = OCM.second; 1431 ClassInfo *CI = OCM.first; 1432 OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI, 1433 OpMask)); 1434 } 1435 } 1436 } 1437 1438 void AsmMatcherInfo::buildInfo() { 1439 // Build information about all of the AssemblerPredicates. 1440 std::vector<Record*> AllPredicates = 1441 Records.getAllDerivedDefinitions("Predicate"); 1442 for (Record *Pred : AllPredicates) { 1443 // Ignore predicates that are not intended for the assembler. 1444 if (!Pred->getValueAsBit("AssemblerMatcherPredicate")) 1445 continue; 1446 1447 if (Pred->getName().empty()) 1448 PrintFatalError(Pred->getLoc(), "Predicate has no name!"); 1449 1450 SubtargetFeatures.insert(std::make_pair( 1451 Pred, SubtargetFeatureInfo(Pred, SubtargetFeatures.size()))); 1452 DEBUG(SubtargetFeatures.find(Pred)->second.dump()); 1453 assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!"); 1454 } 1455 1456 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst"); 1457 1458 // Parse the instructions; we need to do this first so that we can gather the 1459 // singleton register classes. 1460 SmallPtrSet<Record*, 16> SingletonRegisters; 1461 unsigned VariantCount = Target.getAsmParserVariantCount(); 1462 for (unsigned VC = 0; VC != VariantCount; ++VC) { 1463 Record *AsmVariant = Target.getAsmParserVariant(VC); 1464 std::string CommentDelimiter = 1465 AsmVariant->getValueAsString("CommentDelimiter"); 1466 AsmVariantInfo Variant; 1467 Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix"); 1468 Variant.TokenizingCharacters = 1469 AsmVariant->getValueAsString("TokenizingCharacters"); 1470 Variant.SeparatorCharacters = 1471 AsmVariant->getValueAsString("SeparatorCharacters"); 1472 Variant.BreakCharacters = 1473 AsmVariant->getValueAsString("BreakCharacters"); 1474 Variant.AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 1475 1476 for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) { 1477 1478 // If the tblgen -match-prefix option is specified (for tblgen hackers), 1479 // filter the set of instructions we consider. 1480 if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix)) 1481 continue; 1482 1483 // Ignore "codegen only" instructions. 1484 if (CGI->TheDef->getValueAsBit("isCodeGenOnly")) 1485 continue; 1486 1487 auto II = llvm::make_unique<MatchableInfo>(*CGI); 1488 1489 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst); 1490 1491 // Ignore instructions which shouldn't be matched and diagnose invalid 1492 // instruction definitions with an error. 1493 if (!II->validate(CommentDelimiter, true)) 1494 continue; 1495 1496 Matchables.push_back(std::move(II)); 1497 } 1498 1499 // Parse all of the InstAlias definitions and stick them in the list of 1500 // matchables. 1501 std::vector<Record*> AllInstAliases = 1502 Records.getAllDerivedDefinitions("InstAlias"); 1503 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) { 1504 auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i], 1505 Variant.AsmVariantNo, 1506 Target); 1507 1508 // If the tblgen -match-prefix option is specified (for tblgen hackers), 1509 // filter the set of instruction aliases we consider, based on the target 1510 // instruction. 1511 if (!StringRef(Alias->ResultInst->TheDef->getName()) 1512 .startswith( MatchPrefix)) 1513 continue; 1514 1515 auto II = llvm::make_unique<MatchableInfo>(std::move(Alias)); 1516 1517 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst); 1518 1519 // Validate the alias definitions. 1520 II->validate(CommentDelimiter, false); 1521 1522 Matchables.push_back(std::move(II)); 1523 } 1524 } 1525 1526 // Build info for the register classes. 1527 buildRegisterClasses(SingletonRegisters); 1528 1529 // Build info for the user defined assembly operand classes. 1530 buildOperandClasses(); 1531 1532 // Build the information about matchables, now that we have fully formed 1533 // classes. 1534 std::vector<std::unique_ptr<MatchableInfo>> NewMatchables; 1535 for (auto &II : Matchables) { 1536 // Parse the tokens after the mnemonic. 1537 // Note: buildInstructionOperandReference may insert new AsmOperands, so 1538 // don't precompute the loop bound. 1539 for (unsigned i = 0; i != II->AsmOperands.size(); ++i) { 1540 MatchableInfo::AsmOperand &Op = II->AsmOperands[i]; 1541 StringRef Token = Op.Token; 1542 1543 // Check for singleton registers. 1544 if (Record *RegRecord = Op.SingletonReg) { 1545 Op.Class = RegisterClasses[RegRecord]; 1546 assert(Op.Class && Op.Class->Registers.size() == 1 && 1547 "Unexpected class for singleton register"); 1548 continue; 1549 } 1550 1551 // Check for simple tokens. 1552 if (Token[0] != '$') { 1553 Op.Class = getTokenClass(Token); 1554 continue; 1555 } 1556 1557 if (Token.size() > 1 && isdigit(Token[1])) { 1558 Op.Class = getTokenClass(Token); 1559 continue; 1560 } 1561 1562 // Otherwise this is an operand reference. 1563 StringRef OperandName; 1564 if (Token[1] == '{') 1565 OperandName = Token.substr(2, Token.size() - 3); 1566 else 1567 OperandName = Token.substr(1); 1568 1569 if (II->DefRec.is<const CodeGenInstruction*>()) 1570 buildInstructionOperandReference(II.get(), OperandName, i); 1571 else 1572 buildAliasOperandReference(II.get(), OperandName, Op); 1573 } 1574 1575 if (II->DefRec.is<const CodeGenInstruction*>()) { 1576 II->buildInstructionResultOperands(); 1577 // If the instruction has a two-operand alias, build up the 1578 // matchable here. We'll add them in bulk at the end to avoid 1579 // confusing this loop. 1580 std::string Constraint = 1581 II->TheDef->getValueAsString("TwoOperandAliasConstraint"); 1582 if (Constraint != "") { 1583 // Start by making a copy of the original matchable. 1584 auto AliasII = llvm::make_unique<MatchableInfo>(*II); 1585 1586 // Adjust it to be a two-operand alias. 1587 AliasII->formTwoOperandAlias(Constraint); 1588 1589 // Add the alias to the matchables list. 1590 NewMatchables.push_back(std::move(AliasII)); 1591 } 1592 } else 1593 II->buildAliasResultOperands(); 1594 } 1595 if (!NewMatchables.empty()) 1596 Matchables.insert(Matchables.end(), 1597 std::make_move_iterator(NewMatchables.begin()), 1598 std::make_move_iterator(NewMatchables.end())); 1599 1600 // Process token alias definitions and set up the associated superclass 1601 // information. 1602 std::vector<Record*> AllTokenAliases = 1603 Records.getAllDerivedDefinitions("TokenAlias"); 1604 for (Record *Rec : AllTokenAliases) { 1605 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken")); 1606 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken")); 1607 if (FromClass == ToClass) 1608 PrintFatalError(Rec->getLoc(), 1609 "error: Destination value identical to source value."); 1610 FromClass->SuperClasses.push_back(ToClass); 1611 } 1612 1613 // Reorder classes so that classes precede super classes. 1614 Classes.sort(); 1615 1616 #ifndef NDEBUG 1617 // Verify that the table is now sorted 1618 for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) { 1619 for (auto J = I; J != E; ++J) { 1620 assert(!(*J < *I)); 1621 assert(I == J || !J->isSubsetOf(*I)); 1622 } 1623 } 1624 #endif // NDEBUG 1625 } 1626 1627 /// buildInstructionOperandReference - The specified operand is a reference to a 1628 /// named operand such as $src. Resolve the Class and OperandInfo pointers. 1629 void AsmMatcherInfo:: 1630 buildInstructionOperandReference(MatchableInfo *II, 1631 StringRef OperandName, 1632 unsigned AsmOpIdx) { 1633 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>(); 1634 const CGIOperandList &Operands = CGI.Operands; 1635 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx]; 1636 1637 // Map this token to an operand. 1638 unsigned Idx; 1639 if (!Operands.hasOperandNamed(OperandName, Idx)) 1640 PrintFatalError(II->TheDef->getLoc(), 1641 "error: unable to find operand: '" + OperandName + "'"); 1642 1643 // If the instruction operand has multiple suboperands, but the parser 1644 // match class for the asm operand is still the default "ImmAsmOperand", 1645 // then handle each suboperand separately. 1646 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) { 1647 Record *Rec = Operands[Idx].Rec; 1648 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!"); 1649 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1650 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") { 1651 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands. 1652 StringRef Token = Op->Token; // save this in case Op gets moved 1653 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) { 1654 MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token); 1655 NewAsmOp.SubOpIdx = SI; 1656 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp); 1657 } 1658 // Replace Op with first suboperand. 1659 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved 1660 Op->SubOpIdx = 0; 1661 } 1662 } 1663 1664 // Set up the operand class. 1665 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx); 1666 1667 // If the named operand is tied, canonicalize it to the untied operand. 1668 // For example, something like: 1669 // (outs GPR:$dst), (ins GPR:$src) 1670 // with an asmstring of 1671 // "inc $src" 1672 // we want to canonicalize to: 1673 // "inc $dst" 1674 // so that we know how to provide the $dst operand when filling in the result. 1675 int OITied = -1; 1676 if (Operands[Idx].MINumOperands == 1) 1677 OITied = Operands[Idx].getTiedRegister(); 1678 if (OITied != -1) { 1679 // The tied operand index is an MIOperand index, find the operand that 1680 // contains it. 1681 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied); 1682 OperandName = Operands[Idx.first].Name; 1683 Op->SubOpIdx = Idx.second; 1684 } 1685 1686 Op->SrcOpName = OperandName; 1687 } 1688 1689 /// buildAliasOperandReference - When parsing an operand reference out of the 1690 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the 1691 /// operand reference is by looking it up in the result pattern definition. 1692 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II, 1693 StringRef OperandName, 1694 MatchableInfo::AsmOperand &Op) { 1695 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>(); 1696 1697 // Set up the operand class. 1698 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i) 1699 if (CGA.ResultOperands[i].isRecord() && 1700 CGA.ResultOperands[i].getName() == OperandName) { 1701 // It's safe to go with the first one we find, because CodeGenInstAlias 1702 // validates that all operands with the same name have the same record. 1703 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second; 1704 // Use the match class from the Alias definition, not the 1705 // destination instruction, as we may have an immediate that's 1706 // being munged by the match class. 1707 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(), 1708 Op.SubOpIdx); 1709 Op.SrcOpName = OperandName; 1710 return; 1711 } 1712 1713 PrintFatalError(II->TheDef->getLoc(), 1714 "error: unable to find operand: '" + OperandName + "'"); 1715 } 1716 1717 void MatchableInfo::buildInstructionResultOperands() { 1718 const CodeGenInstruction *ResultInst = getResultInst(); 1719 1720 // Loop over all operands of the result instruction, determining how to 1721 // populate them. 1722 for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) { 1723 // If this is a tied operand, just copy from the previously handled operand. 1724 int TiedOp = -1; 1725 if (OpInfo.MINumOperands == 1) 1726 TiedOp = OpInfo.getTiedRegister(); 1727 if (TiedOp != -1) { 1728 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1729 continue; 1730 } 1731 1732 // Find out what operand from the asmparser this MCInst operand comes from. 1733 int SrcOperand = findAsmOperandNamed(OpInfo.Name); 1734 if (OpInfo.Name.empty() || SrcOperand == -1) { 1735 // This may happen for operands that are tied to a suboperand of a 1736 // complex operand. Simply use a dummy value here; nobody should 1737 // use this operand slot. 1738 // FIXME: The long term goal is for the MCOperand list to not contain 1739 // tied operands at all. 1740 ResOperands.push_back(ResOperand::getImmOp(0)); 1741 continue; 1742 } 1743 1744 // Check if the one AsmOperand populates the entire operand. 1745 unsigned NumOperands = OpInfo.MINumOperands; 1746 if (AsmOperands[SrcOperand].SubOpIdx == -1) { 1747 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands)); 1748 continue; 1749 } 1750 1751 // Add a separate ResOperand for each suboperand. 1752 for (unsigned AI = 0; AI < NumOperands; ++AI) { 1753 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI && 1754 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name && 1755 "unexpected AsmOperands for suboperands"); 1756 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1)); 1757 } 1758 } 1759 } 1760 1761 void MatchableInfo::buildAliasResultOperands() { 1762 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>(); 1763 const CodeGenInstruction *ResultInst = getResultInst(); 1764 1765 // Loop over all operands of the result instruction, determining how to 1766 // populate them. 1767 unsigned AliasOpNo = 0; 1768 unsigned LastOpNo = CGA.ResultInstOperandIndex.size(); 1769 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) { 1770 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i]; 1771 1772 // If this is a tied operand, just copy from the previously handled operand. 1773 int TiedOp = -1; 1774 if (OpInfo->MINumOperands == 1) 1775 TiedOp = OpInfo->getTiedRegister(); 1776 if (TiedOp != -1) { 1777 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1778 continue; 1779 } 1780 1781 // Handle all the suboperands for this operand. 1782 const std::string &OpName = OpInfo->Name; 1783 for ( ; AliasOpNo < LastOpNo && 1784 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) { 1785 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second; 1786 1787 // Find out what operand from the asmparser that this MCInst operand 1788 // comes from. 1789 switch (CGA.ResultOperands[AliasOpNo].Kind) { 1790 case CodeGenInstAlias::ResultOperand::K_Record: { 1791 StringRef Name = CGA.ResultOperands[AliasOpNo].getName(); 1792 int SrcOperand = findAsmOperand(Name, SubIdx); 1793 if (SrcOperand == -1) 1794 PrintFatalError(TheDef->getLoc(), "Instruction '" + 1795 TheDef->getName() + "' has operand '" + OpName + 1796 "' that doesn't appear in asm string!"); 1797 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1); 1798 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, 1799 NumOperands)); 1800 break; 1801 } 1802 case CodeGenInstAlias::ResultOperand::K_Imm: { 1803 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm(); 1804 ResOperands.push_back(ResOperand::getImmOp(ImmVal)); 1805 break; 1806 } 1807 case CodeGenInstAlias::ResultOperand::K_Reg: { 1808 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister(); 1809 ResOperands.push_back(ResOperand::getRegOp(Reg)); 1810 break; 1811 } 1812 } 1813 } 1814 } 1815 } 1816 1817 static unsigned getConverterOperandID(const std::string &Name, 1818 SmallSetVector<std::string, 16> &Table, 1819 bool &IsNew) { 1820 IsNew = Table.insert(Name); 1821 1822 unsigned ID = IsNew ? Table.size() - 1 : 1823 std::find(Table.begin(), Table.end(), Name) - Table.begin(); 1824 1825 assert(ID < Table.size()); 1826 1827 return ID; 1828 } 1829 1830 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName, 1831 std::vector<std::unique_ptr<MatchableInfo>> &Infos, 1832 bool HasMnemonicFirst, bool HasOptionalOperands, 1833 raw_ostream &OS) { 1834 SmallSetVector<std::string, 16> OperandConversionKinds; 1835 SmallSetVector<std::string, 16> InstructionConversionKinds; 1836 std::vector<std::vector<uint8_t> > ConversionTable; 1837 size_t MaxRowLength = 2; // minimum is custom converter plus terminator. 1838 1839 // TargetOperandClass - This is the target's operand class, like X86Operand. 1840 std::string TargetOperandClass = Target.getName() + "Operand"; 1841 1842 // Write the convert function to a separate stream, so we can drop it after 1843 // the enum. We'll build up the conversion handlers for the individual 1844 // operand types opportunistically as we encounter them. 1845 std::string ConvertFnBody; 1846 raw_string_ostream CvtOS(ConvertFnBody); 1847 // Start the unified conversion function. 1848 if (HasOptionalOperands) { 1849 CvtOS << "void " << Target.getName() << ClassName << "::\n" 1850 << "convertToMCInst(unsigned Kind, MCInst &Inst, " 1851 << "unsigned Opcode,\n" 1852 << " const OperandVector &Operands,\n" 1853 << " const SmallBitVector &OptionalOperandsMask) {\n"; 1854 } else { 1855 CvtOS << "void " << Target.getName() << ClassName << "::\n" 1856 << "convertToMCInst(unsigned Kind, MCInst &Inst, " 1857 << "unsigned Opcode,\n" 1858 << " const OperandVector &Operands) {\n"; 1859 } 1860 CvtOS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"; 1861 CvtOS << " const uint8_t *Converter = ConversionTable[Kind];\n"; 1862 if (HasOptionalOperands) { 1863 CvtOS << " unsigned NumDefaults = 0;\n"; 1864 } 1865 CvtOS << " unsigned OpIdx;\n"; 1866 CvtOS << " Inst.setOpcode(Opcode);\n"; 1867 CvtOS << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"; 1868 if (HasOptionalOperands) { 1869 CvtOS << " OpIdx = *(p + 1) - NumDefaults;\n"; 1870 } else { 1871 CvtOS << " OpIdx = *(p + 1);\n"; 1872 } 1873 CvtOS << " switch (*p) {\n"; 1874 CvtOS << " default: llvm_unreachable(\"invalid conversion entry!\");\n"; 1875 CvtOS << " case CVT_Reg:\n"; 1876 CvtOS << " static_cast<" << TargetOperandClass 1877 << "&>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n"; 1878 CvtOS << " break;\n"; 1879 CvtOS << " case CVT_Tied:\n"; 1880 CvtOS << " Inst.addOperand(Inst.getOperand(OpIdx));\n"; 1881 CvtOS << " break;\n"; 1882 1883 std::string OperandFnBody; 1884 raw_string_ostream OpOS(OperandFnBody); 1885 // Start the operand number lookup function. 1886 OpOS << "void " << Target.getName() << ClassName << "::\n" 1887 << "convertToMapAndConstraints(unsigned Kind,\n"; 1888 OpOS.indent(27); 1889 OpOS << "const OperandVector &Operands) {\n" 1890 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n" 1891 << " unsigned NumMCOperands = 0;\n" 1892 << " const uint8_t *Converter = ConversionTable[Kind];\n" 1893 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n" 1894 << " switch (*p) {\n" 1895 << " default: llvm_unreachable(\"invalid conversion entry!\");\n" 1896 << " case CVT_Reg:\n" 1897 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 1898 << " Operands[*(p + 1)]->setConstraint(\"r\");\n" 1899 << " ++NumMCOperands;\n" 1900 << " break;\n" 1901 << " case CVT_Tied:\n" 1902 << " ++NumMCOperands;\n" 1903 << " break;\n"; 1904 1905 // Pre-populate the operand conversion kinds with the standard always 1906 // available entries. 1907 OperandConversionKinds.insert("CVT_Done"); 1908 OperandConversionKinds.insert("CVT_Reg"); 1909 OperandConversionKinds.insert("CVT_Tied"); 1910 enum { CVT_Done, CVT_Reg, CVT_Tied }; 1911 1912 for (auto &II : Infos) { 1913 // Check if we have a custom match function. 1914 std::string AsmMatchConverter = 1915 II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter"); 1916 if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) { 1917 std::string Signature = "ConvertCustom_" + AsmMatchConverter; 1918 II->ConversionFnKind = Signature; 1919 1920 // Check if we have already generated this signature. 1921 if (!InstructionConversionKinds.insert(Signature)) 1922 continue; 1923 1924 // Remember this converter for the kind enum. 1925 unsigned KindID = OperandConversionKinds.size(); 1926 OperandConversionKinds.insert("CVT_" + 1927 getEnumNameForToken(AsmMatchConverter)); 1928 1929 // Add the converter row for this instruction. 1930 ConversionTable.emplace_back(); 1931 ConversionTable.back().push_back(KindID); 1932 ConversionTable.back().push_back(CVT_Done); 1933 1934 // Add the handler to the conversion driver function. 1935 CvtOS << " case CVT_" 1936 << getEnumNameForToken(AsmMatchConverter) << ":\n" 1937 << " " << AsmMatchConverter << "(Inst, Operands);\n" 1938 << " break;\n"; 1939 1940 // FIXME: Handle the operand number lookup for custom match functions. 1941 continue; 1942 } 1943 1944 // Build the conversion function signature. 1945 std::string Signature = "Convert"; 1946 1947 std::vector<uint8_t> ConversionRow; 1948 1949 // Compute the convert enum and the case body. 1950 MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 ); 1951 1952 for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) { 1953 const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i]; 1954 1955 // Generate code to populate each result operand. 1956 switch (OpInfo.Kind) { 1957 case MatchableInfo::ResOperand::RenderAsmOperand: { 1958 // This comes from something we parsed. 1959 const MatchableInfo::AsmOperand &Op = 1960 II->AsmOperands[OpInfo.AsmOperandNum]; 1961 1962 // Registers are always converted the same, don't duplicate the 1963 // conversion function based on them. 1964 Signature += "__"; 1965 std::string Class; 1966 Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName; 1967 Signature += Class; 1968 Signature += utostr(OpInfo.MINumOperands); 1969 Signature += "_" + itostr(OpInfo.AsmOperandNum); 1970 1971 // Add the conversion kind, if necessary, and get the associated ID 1972 // the index of its entry in the vector). 1973 std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" : 1974 Op.Class->RenderMethod); 1975 if (Op.Class->IsOptional) { 1976 // For optional operands we must also care about DefaultMethod 1977 assert(HasOptionalOperands); 1978 Name += "_" + Op.Class->DefaultMethod; 1979 } 1980 Name = getEnumNameForToken(Name); 1981 1982 bool IsNewConverter = false; 1983 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 1984 IsNewConverter); 1985 1986 // Add the operand entry to the instruction kind conversion row. 1987 ConversionRow.push_back(ID); 1988 ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst); 1989 1990 if (!IsNewConverter) 1991 break; 1992 1993 // This is a new operand kind. Add a handler for it to the 1994 // converter driver. 1995 CvtOS << " case " << Name << ":\n"; 1996 if (Op.Class->IsOptional) { 1997 // If optional operand is not present in actual instruction then we 1998 // should call its DefaultMethod before RenderMethod 1999 assert(HasOptionalOperands); 2000 CvtOS << " if (OptionalOperandsMask[*(p + 1) - 1]) {\n" 2001 << " " << Op.Class->DefaultMethod << "()" 2002 << "->" << Op.Class->RenderMethod << "(Inst, " 2003 << OpInfo.MINumOperands << ");\n" 2004 << " ++NumDefaults;\n" 2005 << " } else {\n" 2006 << " static_cast<" << TargetOperandClass 2007 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod 2008 << "(Inst, " << OpInfo.MINumOperands << ");\n" 2009 << " }\n"; 2010 } else { 2011 CvtOS << " static_cast<" << TargetOperandClass 2012 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod 2013 << "(Inst, " << OpInfo.MINumOperands << ");\n"; 2014 } 2015 CvtOS << " break;\n"; 2016 2017 // Add a handler for the operand number lookup. 2018 OpOS << " case " << Name << ":\n" 2019 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"; 2020 2021 if (Op.Class->isRegisterClass()) 2022 OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n"; 2023 else 2024 OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n"; 2025 OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n" 2026 << " break;\n"; 2027 break; 2028 } 2029 case MatchableInfo::ResOperand::TiedOperand: { 2030 // If this operand is tied to a previous one, just copy the MCInst 2031 // operand from the earlier one.We can only tie single MCOperand values. 2032 assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand"); 2033 unsigned TiedOp = OpInfo.TiedOperandNum; 2034 assert(i > TiedOp && "Tied operand precedes its target!"); 2035 Signature += "__Tie" + utostr(TiedOp); 2036 ConversionRow.push_back(CVT_Tied); 2037 ConversionRow.push_back(TiedOp); 2038 break; 2039 } 2040 case MatchableInfo::ResOperand::ImmOperand: { 2041 int64_t Val = OpInfo.ImmVal; 2042 std::string Ty = "imm_" + itostr(Val); 2043 Ty = getEnumNameForToken(Ty); 2044 Signature += "__" + Ty; 2045 2046 std::string Name = "CVT_" + Ty; 2047 bool IsNewConverter = false; 2048 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 2049 IsNewConverter); 2050 // Add the operand entry to the instruction kind conversion row. 2051 ConversionRow.push_back(ID); 2052 ConversionRow.push_back(0); 2053 2054 if (!IsNewConverter) 2055 break; 2056 2057 CvtOS << " case " << Name << ":\n" 2058 << " Inst.addOperand(MCOperand::createImm(" << Val << "));\n" 2059 << " break;\n"; 2060 2061 OpOS << " case " << Name << ":\n" 2062 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 2063 << " Operands[*(p + 1)]->setConstraint(\"\");\n" 2064 << " ++NumMCOperands;\n" 2065 << " break;\n"; 2066 break; 2067 } 2068 case MatchableInfo::ResOperand::RegOperand: { 2069 std::string Reg, Name; 2070 if (!OpInfo.Register) { 2071 Name = "reg0"; 2072 Reg = "0"; 2073 } else { 2074 Reg = getQualifiedName(OpInfo.Register); 2075 Name = "reg" + OpInfo.Register->getName(); 2076 } 2077 Signature += "__" + Name; 2078 Name = "CVT_" + Name; 2079 bool IsNewConverter = false; 2080 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 2081 IsNewConverter); 2082 // Add the operand entry to the instruction kind conversion row. 2083 ConversionRow.push_back(ID); 2084 ConversionRow.push_back(0); 2085 2086 if (!IsNewConverter) 2087 break; 2088 CvtOS << " case " << Name << ":\n" 2089 << " Inst.addOperand(MCOperand::createReg(" << Reg << "));\n" 2090 << " break;\n"; 2091 2092 OpOS << " case " << Name << ":\n" 2093 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 2094 << " Operands[*(p + 1)]->setConstraint(\"m\");\n" 2095 << " ++NumMCOperands;\n" 2096 << " break;\n"; 2097 } 2098 } 2099 } 2100 2101 // If there were no operands, add to the signature to that effect 2102 if (Signature == "Convert") 2103 Signature += "_NoOperands"; 2104 2105 II->ConversionFnKind = Signature; 2106 2107 // Save the signature. If we already have it, don't add a new row 2108 // to the table. 2109 if (!InstructionConversionKinds.insert(Signature)) 2110 continue; 2111 2112 // Add the row to the table. 2113 ConversionTable.push_back(std::move(ConversionRow)); 2114 } 2115 2116 // Finish up the converter driver function. 2117 CvtOS << " }\n }\n}\n\n"; 2118 2119 // Finish up the operand number lookup function. 2120 OpOS << " }\n }\n}\n\n"; 2121 2122 OS << "namespace {\n"; 2123 2124 // Output the operand conversion kind enum. 2125 OS << "enum OperatorConversionKind {\n"; 2126 for (const std::string &Converter : OperandConversionKinds) 2127 OS << " " << Converter << ",\n"; 2128 OS << " CVT_NUM_CONVERTERS\n"; 2129 OS << "};\n\n"; 2130 2131 // Output the instruction conversion kind enum. 2132 OS << "enum InstructionConversionKind {\n"; 2133 for (const std::string &Signature : InstructionConversionKinds) 2134 OS << " " << Signature << ",\n"; 2135 OS << " CVT_NUM_SIGNATURES\n"; 2136 OS << "};\n\n"; 2137 2138 OS << "} // end anonymous namespace\n\n"; 2139 2140 // Output the conversion table. 2141 OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES][" 2142 << MaxRowLength << "] = {\n"; 2143 2144 for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) { 2145 assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!"); 2146 OS << " // " << InstructionConversionKinds[Row] << "\n"; 2147 OS << " { "; 2148 for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2) 2149 OS << OperandConversionKinds[ConversionTable[Row][i]] << ", " 2150 << (unsigned)(ConversionTable[Row][i + 1]) << ", "; 2151 OS << "CVT_Done },\n"; 2152 } 2153 2154 OS << "};\n\n"; 2155 2156 // Spit out the conversion driver function. 2157 OS << CvtOS.str(); 2158 2159 // Spit out the operand number lookup function. 2160 OS << OpOS.str(); 2161 } 2162 2163 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds. 2164 static void emitMatchClassEnumeration(CodeGenTarget &Target, 2165 std::forward_list<ClassInfo> &Infos, 2166 raw_ostream &OS) { 2167 OS << "namespace {\n\n"; 2168 2169 OS << "/// MatchClassKind - The kinds of classes which participate in\n" 2170 << "/// instruction matching.\n"; 2171 OS << "enum MatchClassKind {\n"; 2172 OS << " InvalidMatchClass = 0,\n"; 2173 OS << " OptionalMatchClass = 1,\n"; 2174 for (const auto &CI : Infos) { 2175 OS << " " << CI.Name << ", // "; 2176 if (CI.Kind == ClassInfo::Token) { 2177 OS << "'" << CI.ValueName << "'\n"; 2178 } else if (CI.isRegisterClass()) { 2179 if (!CI.ValueName.empty()) 2180 OS << "register class '" << CI.ValueName << "'\n"; 2181 else 2182 OS << "derived register class\n"; 2183 } else { 2184 OS << "user defined class '" << CI.ValueName << "'\n"; 2185 } 2186 } 2187 OS << " NumMatchClassKinds\n"; 2188 OS << "};\n\n"; 2189 2190 OS << "}\n\n"; 2191 } 2192 2193 /// emitValidateOperandClass - Emit the function to validate an operand class. 2194 static void emitValidateOperandClass(AsmMatcherInfo &Info, 2195 raw_ostream &OS) { 2196 OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, " 2197 << "MatchClassKind Kind) {\n"; 2198 OS << " " << Info.Target.getName() << "Operand &Operand = (" 2199 << Info.Target.getName() << "Operand&)GOp;\n"; 2200 2201 // The InvalidMatchClass is not to match any operand. 2202 OS << " if (Kind == InvalidMatchClass)\n"; 2203 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n"; 2204 2205 // Check for Token operands first. 2206 // FIXME: Use a more specific diagnostic type. 2207 OS << " if (Operand.isToken())\n"; 2208 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n" 2209 << " MCTargetAsmParser::Match_Success :\n" 2210 << " MCTargetAsmParser::Match_InvalidOperand;\n\n"; 2211 2212 // Check the user classes. We don't care what order since we're only 2213 // actually matching against one of them. 2214 OS << " switch (Kind) {\n" 2215 " default: break;\n"; 2216 for (const auto &CI : Info.Classes) { 2217 if (!CI.isUserClass()) 2218 continue; 2219 2220 OS << " // '" << CI.ClassName << "' class\n"; 2221 OS << " case " << CI.Name << ":\n"; 2222 OS << " if (Operand." << CI.PredicateMethod << "())\n"; 2223 OS << " return MCTargetAsmParser::Match_Success;\n"; 2224 if (!CI.DiagnosticType.empty()) 2225 OS << " return " << Info.Target.getName() << "AsmParser::Match_" 2226 << CI.DiagnosticType << ";\n"; 2227 else 2228 OS << " break;\n"; 2229 } 2230 OS << " } // end switch (Kind)\n\n"; 2231 2232 // Check for register operands, including sub-classes. 2233 OS << " if (Operand.isReg()) {\n"; 2234 OS << " MatchClassKind OpKind;\n"; 2235 OS << " switch (Operand.getReg()) {\n"; 2236 OS << " default: OpKind = InvalidMatchClass; break;\n"; 2237 for (const auto &RC : Info.RegisterClasses) 2238 OS << " case " << Info.Target.getName() << "::" 2239 << RC.first->getName() << ": OpKind = " << RC.second->Name 2240 << "; break;\n"; 2241 OS << " }\n"; 2242 OS << " return isSubclass(OpKind, Kind) ? " 2243 << "MCTargetAsmParser::Match_Success :\n " 2244 << " MCTargetAsmParser::Match_InvalidOperand;\n }\n\n"; 2245 2246 // Generic fallthrough match failure case for operands that don't have 2247 // specialized diagnostic types. 2248 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n"; 2249 OS << "}\n\n"; 2250 } 2251 2252 /// emitIsSubclass - Emit the subclass predicate function. 2253 static void emitIsSubclass(CodeGenTarget &Target, 2254 std::forward_list<ClassInfo> &Infos, 2255 raw_ostream &OS) { 2256 OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n"; 2257 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n"; 2258 OS << " if (A == B)\n"; 2259 OS << " return true;\n\n"; 2260 2261 bool EmittedSwitch = false; 2262 for (const auto &A : Infos) { 2263 std::vector<StringRef> SuperClasses; 2264 if (A.IsOptional) 2265 SuperClasses.push_back("OptionalMatchClass"); 2266 for (const auto &B : Infos) { 2267 if (&A != &B && A.isSubsetOf(B)) 2268 SuperClasses.push_back(B.Name); 2269 } 2270 2271 if (SuperClasses.empty()) 2272 continue; 2273 2274 // If this is the first SuperClass, emit the switch header. 2275 if (!EmittedSwitch) { 2276 OS << " switch (A) {\n"; 2277 OS << " default:\n"; 2278 OS << " return false;\n"; 2279 EmittedSwitch = true; 2280 } 2281 2282 OS << "\n case " << A.Name << ":\n"; 2283 2284 if (SuperClasses.size() == 1) { 2285 OS << " return B == " << SuperClasses.back() << ";\n"; 2286 continue; 2287 } 2288 2289 if (!SuperClasses.empty()) { 2290 OS << " switch (B) {\n"; 2291 OS << " default: return false;\n"; 2292 for (StringRef SC : SuperClasses) 2293 OS << " case " << SC << ": return true;\n"; 2294 OS << " }\n"; 2295 } else { 2296 // No case statement to emit 2297 OS << " return false;\n"; 2298 } 2299 } 2300 2301 // If there were case statements emitted into the string stream write the 2302 // default. 2303 if (EmittedSwitch) 2304 OS << " }\n"; 2305 else 2306 OS << " return false;\n"; 2307 2308 OS << "}\n\n"; 2309 } 2310 2311 /// emitMatchTokenString - Emit the function to match a token string to the 2312 /// appropriate match class value. 2313 static void emitMatchTokenString(CodeGenTarget &Target, 2314 std::forward_list<ClassInfo> &Infos, 2315 raw_ostream &OS) { 2316 // Construct the match list. 2317 std::vector<StringMatcher::StringPair> Matches; 2318 for (const auto &CI : Infos) { 2319 if (CI.Kind == ClassInfo::Token) 2320 Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";"); 2321 } 2322 2323 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n"; 2324 2325 StringMatcher("Name", Matches, OS).Emit(); 2326 2327 OS << " return InvalidMatchClass;\n"; 2328 OS << "}\n\n"; 2329 } 2330 2331 /// emitMatchRegisterName - Emit the function to match a string to the target 2332 /// specific register enum. 2333 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser, 2334 raw_ostream &OS) { 2335 // Construct the match list. 2336 std::vector<StringMatcher::StringPair> Matches; 2337 const auto &Regs = Target.getRegBank().getRegisters(); 2338 for (const CodeGenRegister &Reg : Regs) { 2339 if (Reg.TheDef->getValueAsString("AsmName").empty()) 2340 continue; 2341 2342 Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"), 2343 "return " + utostr(Reg.EnumValue) + ";"); 2344 } 2345 2346 OS << "static unsigned MatchRegisterName(StringRef Name) {\n"; 2347 2348 StringMatcher("Name", Matches, OS).Emit(); 2349 2350 OS << " return 0;\n"; 2351 OS << "}\n\n"; 2352 } 2353 2354 /// Emit the function to match a string to the target 2355 /// specific register enum. 2356 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser, 2357 raw_ostream &OS) { 2358 // Construct the match list. 2359 std::vector<StringMatcher::StringPair> Matches; 2360 const auto &Regs = Target.getRegBank().getRegisters(); 2361 for (const CodeGenRegister &Reg : Regs) { 2362 2363 auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames"); 2364 2365 for (auto AltName : AltNames) { 2366 AltName = StringRef(AltName).trim(); 2367 2368 // don't handle empty alternative names 2369 if (AltName.empty()) 2370 continue; 2371 2372 Matches.emplace_back(AltName, 2373 "return " + utostr(Reg.EnumValue) + ";"); 2374 } 2375 } 2376 2377 OS << "static unsigned MatchRegisterAltName(StringRef Name) {\n"; 2378 2379 StringMatcher("Name", Matches, OS).Emit(); 2380 2381 OS << " return 0;\n"; 2382 OS << "}\n\n"; 2383 } 2384 2385 static const char *getMinimalTypeForRange(uint64_t Range) { 2386 assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large"); 2387 if (Range > 0xFFFFFFFFULL) 2388 return "uint64_t"; 2389 if (Range > 0xFFFF) 2390 return "uint32_t"; 2391 if (Range > 0xFF) 2392 return "uint16_t"; 2393 return "uint8_t"; 2394 } 2395 2396 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) { 2397 uint64_t MaxIndex = Info.SubtargetFeatures.size(); 2398 if (MaxIndex > 0) 2399 MaxIndex--; 2400 return getMinimalTypeForRange(1ULL << MaxIndex); 2401 } 2402 2403 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag 2404 /// definitions. 2405 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info, 2406 raw_ostream &OS) { 2407 OS << "// Flags for subtarget features that participate in " 2408 << "instruction matching.\n"; 2409 OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info) 2410 << " {\n"; 2411 for (const auto &SF : Info.SubtargetFeatures) { 2412 const SubtargetFeatureInfo &SFI = SF.second; 2413 OS << " " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n"; 2414 } 2415 OS << " Feature_None = 0\n"; 2416 OS << "};\n\n"; 2417 } 2418 2419 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types. 2420 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) { 2421 // Get the set of diagnostic types from all of the operand classes. 2422 std::set<StringRef> Types; 2423 for (const auto &OpClassEntry : Info.AsmOperandClasses) { 2424 if (!OpClassEntry.second->DiagnosticType.empty()) 2425 Types.insert(OpClassEntry.second->DiagnosticType); 2426 } 2427 2428 if (Types.empty()) return; 2429 2430 // Now emit the enum entries. 2431 for (StringRef Type : Types) 2432 OS << " Match_" << Type << ",\n"; 2433 OS << " END_OPERAND_DIAGNOSTIC_TYPES\n"; 2434 } 2435 2436 /// emitGetSubtargetFeatureName - Emit the helper function to get the 2437 /// user-level name for a subtarget feature. 2438 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) { 2439 OS << "// User-level names for subtarget features that participate in\n" 2440 << "// instruction matching.\n" 2441 << "static const char *getSubtargetFeatureName(uint64_t Val) {\n"; 2442 if (!Info.SubtargetFeatures.empty()) { 2443 OS << " switch(Val) {\n"; 2444 for (const auto &SF : Info.SubtargetFeatures) { 2445 const SubtargetFeatureInfo &SFI = SF.second; 2446 // FIXME: Totally just a placeholder name to get the algorithm working. 2447 OS << " case " << SFI.getEnumName() << ": return \"" 2448 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n"; 2449 } 2450 OS << " default: return \"(unknown)\";\n"; 2451 OS << " }\n"; 2452 } else { 2453 // Nothing to emit, so skip the switch 2454 OS << " return \"(unknown)\";\n"; 2455 } 2456 OS << "}\n\n"; 2457 } 2458 2459 /// emitComputeAvailableFeatures - Emit the function to compute the list of 2460 /// available features given a subtarget. 2461 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info, 2462 raw_ostream &OS) { 2463 std::string ClassName = 2464 Info.AsmParser->getValueAsString("AsmParserClassName"); 2465 2466 OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n" 2467 << "ComputeAvailableFeatures(const FeatureBitset& FB) const {\n"; 2468 OS << " uint64_t Features = 0;\n"; 2469 for (const auto &SF : Info.SubtargetFeatures) { 2470 const SubtargetFeatureInfo &SFI = SF.second; 2471 2472 OS << " if ("; 2473 std::string CondStorage = 2474 SFI.TheDef->getValueAsString("AssemblerCondString"); 2475 StringRef Conds = CondStorage; 2476 std::pair<StringRef,StringRef> Comma = Conds.split(','); 2477 bool First = true; 2478 do { 2479 if (!First) 2480 OS << " && "; 2481 2482 bool Neg = false; 2483 StringRef Cond = Comma.first; 2484 if (Cond[0] == '!') { 2485 Neg = true; 2486 Cond = Cond.substr(1); 2487 } 2488 2489 OS << "("; 2490 if (Neg) 2491 OS << "!"; 2492 OS << "FB[" << Info.Target.getName() << "::" << Cond << "])"; 2493 2494 if (Comma.second.empty()) 2495 break; 2496 2497 First = false; 2498 Comma = Comma.second.split(','); 2499 } while (true); 2500 2501 OS << ")\n"; 2502 OS << " Features |= " << SFI.getEnumName() << ";\n"; 2503 } 2504 OS << " return Features;\n"; 2505 OS << "}\n\n"; 2506 } 2507 2508 static std::string GetAliasRequiredFeatures(Record *R, 2509 const AsmMatcherInfo &Info) { 2510 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates"); 2511 std::string Result; 2512 unsigned NumFeatures = 0; 2513 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) { 2514 const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]); 2515 2516 if (!F) 2517 PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() + 2518 "' is not marked as an AssemblerPredicate!"); 2519 2520 if (NumFeatures) 2521 Result += '|'; 2522 2523 Result += F->getEnumName(); 2524 ++NumFeatures; 2525 } 2526 2527 if (NumFeatures > 1) 2528 Result = '(' + Result + ')'; 2529 return Result; 2530 } 2531 2532 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info, 2533 std::vector<Record*> &Aliases, 2534 unsigned Indent = 0, 2535 StringRef AsmParserVariantName = StringRef()){ 2536 // Keep track of all the aliases from a mnemonic. Use an std::map so that the 2537 // iteration order of the map is stable. 2538 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic; 2539 2540 for (Record *R : Aliases) { 2541 // FIXME: Allow AssemblerVariantName to be a comma separated list. 2542 std::string AsmVariantName = R->getValueAsString("AsmVariantName"); 2543 if (AsmVariantName != AsmParserVariantName) 2544 continue; 2545 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R); 2546 } 2547 if (AliasesFromMnemonic.empty()) 2548 return; 2549 2550 // Process each alias a "from" mnemonic at a time, building the code executed 2551 // by the string remapper. 2552 std::vector<StringMatcher::StringPair> Cases; 2553 for (const auto &AliasEntry : AliasesFromMnemonic) { 2554 const std::vector<Record*> &ToVec = AliasEntry.second; 2555 2556 // Loop through each alias and emit code that handles each case. If there 2557 // are two instructions without predicates, emit an error. If there is one, 2558 // emit it last. 2559 std::string MatchCode; 2560 int AliasWithNoPredicate = -1; 2561 2562 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) { 2563 Record *R = ToVec[i]; 2564 std::string FeatureMask = GetAliasRequiredFeatures(R, Info); 2565 2566 // If this unconditionally matches, remember it for later and diagnose 2567 // duplicates. 2568 if (FeatureMask.empty()) { 2569 if (AliasWithNoPredicate != -1) { 2570 // We can't have two aliases from the same mnemonic with no predicate. 2571 PrintError(ToVec[AliasWithNoPredicate]->getLoc(), 2572 "two MnemonicAliases with the same 'from' mnemonic!"); 2573 PrintFatalError(R->getLoc(), "this is the other MnemonicAlias."); 2574 } 2575 2576 AliasWithNoPredicate = i; 2577 continue; 2578 } 2579 if (R->getValueAsString("ToMnemonic") == AliasEntry.first) 2580 PrintFatalError(R->getLoc(), "MnemonicAlias to the same string"); 2581 2582 if (!MatchCode.empty()) 2583 MatchCode += "else "; 2584 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n"; 2585 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n"; 2586 } 2587 2588 if (AliasWithNoPredicate != -1) { 2589 Record *R = ToVec[AliasWithNoPredicate]; 2590 if (!MatchCode.empty()) 2591 MatchCode += "else\n "; 2592 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n"; 2593 } 2594 2595 MatchCode += "return;"; 2596 2597 Cases.push_back(std::make_pair(AliasEntry.first, MatchCode)); 2598 } 2599 StringMatcher("Mnemonic", Cases, OS).Emit(Indent); 2600 } 2601 2602 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions, 2603 /// emit a function for them and return true, otherwise return false. 2604 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info, 2605 CodeGenTarget &Target) { 2606 // Ignore aliases when match-prefix is set. 2607 if (!MatchPrefix.empty()) 2608 return false; 2609 2610 std::vector<Record*> Aliases = 2611 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias"); 2612 if (Aliases.empty()) return false; 2613 2614 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, " 2615 "uint64_t Features, unsigned VariantID) {\n"; 2616 OS << " switch (VariantID) {\n"; 2617 unsigned VariantCount = Target.getAsmParserVariantCount(); 2618 for (unsigned VC = 0; VC != VariantCount; ++VC) { 2619 Record *AsmVariant = Target.getAsmParserVariant(VC); 2620 int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant"); 2621 std::string AsmParserVariantName = AsmVariant->getValueAsString("Name"); 2622 OS << " case " << AsmParserVariantNo << ":\n"; 2623 emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2, 2624 AsmParserVariantName); 2625 OS << " break;\n"; 2626 } 2627 OS << " }\n"; 2628 2629 // Emit aliases that apply to all variants. 2630 emitMnemonicAliasVariant(OS, Info, Aliases); 2631 2632 OS << "}\n\n"; 2633 2634 return true; 2635 } 2636 2637 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target, 2638 const AsmMatcherInfo &Info, StringRef ClassName, 2639 StringToOffsetTable &StringTable, 2640 unsigned MaxMnemonicIndex, bool HasMnemonicFirst) { 2641 unsigned MaxMask = 0; 2642 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { 2643 MaxMask |= OMI.OperandMask; 2644 } 2645 2646 // Emit the static custom operand parsing table; 2647 OS << "namespace {\n"; 2648 OS << " struct OperandMatchEntry {\n"; 2649 OS << " " << getMinimalRequiredFeaturesType(Info) 2650 << " RequiredFeatures;\n"; 2651 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex) 2652 << " Mnemonic;\n"; 2653 OS << " " << getMinimalTypeForRange(std::distance( 2654 Info.Classes.begin(), Info.Classes.end())) << " Class;\n"; 2655 OS << " " << getMinimalTypeForRange(MaxMask) 2656 << " OperandMask;\n\n"; 2657 OS << " StringRef getMnemonic() const {\n"; 2658 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 2659 OS << " MnemonicTable[Mnemonic]);\n"; 2660 OS << " }\n"; 2661 OS << " };\n\n"; 2662 2663 OS << " // Predicate for searching for an opcode.\n"; 2664 OS << " struct LessOpcodeOperand {\n"; 2665 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n"; 2666 OS << " return LHS.getMnemonic() < RHS;\n"; 2667 OS << " }\n"; 2668 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n"; 2669 OS << " return LHS < RHS.getMnemonic();\n"; 2670 OS << " }\n"; 2671 OS << " bool operator()(const OperandMatchEntry &LHS,"; 2672 OS << " const OperandMatchEntry &RHS) {\n"; 2673 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 2674 OS << " }\n"; 2675 OS << " };\n"; 2676 2677 OS << "} // end anonymous namespace.\n\n"; 2678 2679 OS << "static const OperandMatchEntry OperandMatchTable[" 2680 << Info.OperandMatchInfo.size() << "] = {\n"; 2681 2682 OS << " /* Operand List Mask, Mnemonic, Operand Class, Features */\n"; 2683 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { 2684 const MatchableInfo &II = *OMI.MI; 2685 2686 OS << " { "; 2687 2688 // Write the required features mask. 2689 if (!II.RequiredFeatures.empty()) { 2690 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) { 2691 if (i) OS << "|"; 2692 OS << II.RequiredFeatures[i]->getEnumName(); 2693 } 2694 } else 2695 OS << "0"; 2696 2697 // Store a pascal-style length byte in the mnemonic. 2698 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str(); 2699 OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 2700 << " /* " << II.Mnemonic << " */, "; 2701 2702 OS << OMI.CI->Name; 2703 2704 OS << ", " << OMI.OperandMask; 2705 OS << " /* "; 2706 bool printComma = false; 2707 for (int i = 0, e = 31; i !=e; ++i) 2708 if (OMI.OperandMask & (1 << i)) { 2709 if (printComma) 2710 OS << ", "; 2711 OS << i; 2712 printComma = true; 2713 } 2714 OS << " */"; 2715 2716 OS << " },\n"; 2717 } 2718 OS << "};\n\n"; 2719 2720 // Emit the operand class switch to call the correct custom parser for 2721 // the found operand class. 2722 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2723 << Target.getName() << ClassName << "::\n" 2724 << "tryCustomParseOperand(OperandVector" 2725 << " &Operands,\n unsigned MCK) {\n\n" 2726 << " switch(MCK) {\n"; 2727 2728 for (const auto &CI : Info.Classes) { 2729 if (CI.ParserMethod.empty()) 2730 continue; 2731 OS << " case " << CI.Name << ":\n" 2732 << " return " << CI.ParserMethod << "(Operands);\n"; 2733 } 2734 2735 OS << " default:\n"; 2736 OS << " return MatchOperand_NoMatch;\n"; 2737 OS << " }\n"; 2738 OS << " return MatchOperand_NoMatch;\n"; 2739 OS << "}\n\n"; 2740 2741 // Emit the static custom operand parser. This code is very similar with 2742 // the other matcher. Also use MatchResultTy here just in case we go for 2743 // a better error handling. 2744 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2745 << Target.getName() << ClassName << "::\n" 2746 << "MatchOperandParserImpl(OperandVector" 2747 << " &Operands,\n StringRef Mnemonic) {\n"; 2748 2749 // Emit code to get the available features. 2750 OS << " // Get the current feature set.\n"; 2751 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n"; 2752 2753 OS << " // Get the next operand index.\n"; 2754 OS << " unsigned NextOpNum = Operands.size()" 2755 << (HasMnemonicFirst ? " - 1" : "") << ";\n"; 2756 2757 // Emit code to search the table. 2758 OS << " // Search the table.\n"; 2759 if (HasMnemonicFirst) { 2760 OS << " auto MnemonicRange =\n"; 2761 OS << " std::equal_range(std::begin(OperandMatchTable), " 2762 "std::end(OperandMatchTable),\n"; 2763 OS << " Mnemonic, LessOpcodeOperand());\n\n"; 2764 } else { 2765 OS << " auto MnemonicRange = std::make_pair(std::begin(OperandMatchTable)," 2766 " std::end(OperandMatchTable));\n"; 2767 OS << " if (!Mnemonic.empty())\n"; 2768 OS << " MnemonicRange =\n"; 2769 OS << " std::equal_range(std::begin(OperandMatchTable), " 2770 "std::end(OperandMatchTable),\n"; 2771 OS << " Mnemonic, LessOpcodeOperand());\n\n"; 2772 } 2773 2774 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 2775 OS << " return MatchOperand_NoMatch;\n\n"; 2776 2777 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n" 2778 << " *ie = MnemonicRange.second; it != ie; ++it) {\n"; 2779 2780 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 2781 OS << " assert(Mnemonic == it->getMnemonic());\n\n"; 2782 2783 // Emit check that the required features are available. 2784 OS << " // check if the available features match\n"; 2785 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 2786 << "!= it->RequiredFeatures) {\n"; 2787 OS << " continue;\n"; 2788 OS << " }\n\n"; 2789 2790 // Emit check to ensure the operand number matches. 2791 OS << " // check if the operand in question has a custom parser.\n"; 2792 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n"; 2793 OS << " continue;\n\n"; 2794 2795 // Emit call to the custom parser method 2796 OS << " // call custom parse method to handle the operand\n"; 2797 OS << " OperandMatchResultTy Result = "; 2798 OS << "tryCustomParseOperand(Operands, it->Class);\n"; 2799 OS << " if (Result != MatchOperand_NoMatch)\n"; 2800 OS << " return Result;\n"; 2801 OS << " }\n\n"; 2802 2803 OS << " // Okay, we had no match.\n"; 2804 OS << " return MatchOperand_NoMatch;\n"; 2805 OS << "}\n\n"; 2806 } 2807 2808 void AsmMatcherEmitter::run(raw_ostream &OS) { 2809 CodeGenTarget Target(Records); 2810 Record *AsmParser = Target.getAsmParser(); 2811 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName"); 2812 2813 // Compute the information on the instructions to match. 2814 AsmMatcherInfo Info(AsmParser, Target, Records); 2815 Info.buildInfo(); 2816 2817 // Sort the instruction table using the partial order on classes. We use 2818 // stable_sort to ensure that ambiguous instructions are still 2819 // deterministically ordered. 2820 std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(), 2821 [](const std::unique_ptr<MatchableInfo> &a, 2822 const std::unique_ptr<MatchableInfo> &b){ 2823 return *a < *b;}); 2824 2825 #ifndef NDEBUG 2826 // Verify that the table is now sorted 2827 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; 2828 ++I) { 2829 for (auto J = I; J != E; ++J) { 2830 assert(!(**J < **I)); 2831 } 2832 } 2833 #endif // NDEBUG 2834 2835 DEBUG_WITH_TYPE("instruction_info", { 2836 for (const auto &MI : Info.Matchables) 2837 MI->dump(); 2838 }); 2839 2840 // Check for ambiguous matchables. 2841 DEBUG_WITH_TYPE("ambiguous_instrs", { 2842 unsigned NumAmbiguous = 0; 2843 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; 2844 ++I) { 2845 for (auto J = std::next(I); J != E; ++J) { 2846 const MatchableInfo &A = **I; 2847 const MatchableInfo &B = **J; 2848 2849 if (A.couldMatchAmbiguouslyWith(B)) { 2850 errs() << "warning: ambiguous matchables:\n"; 2851 A.dump(); 2852 errs() << "\nis incomparable with:\n"; 2853 B.dump(); 2854 errs() << "\n\n"; 2855 ++NumAmbiguous; 2856 } 2857 } 2858 } 2859 if (NumAmbiguous) 2860 errs() << "warning: " << NumAmbiguous 2861 << " ambiguous matchables!\n"; 2862 }); 2863 2864 // Compute the information on the custom operand parsing. 2865 Info.buildOperandMatchInfo(); 2866 2867 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst"); 2868 bool HasOptionalOperands = Info.hasOptionalOperands(); 2869 2870 // Write the output. 2871 2872 // Information for the class declaration. 2873 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n"; 2874 OS << "#undef GET_ASSEMBLER_HEADER\n"; 2875 OS << " // This should be included into the middle of the declaration of\n"; 2876 OS << " // your subclasses implementation of MCTargetAsmParser.\n"; 2877 OS << " uint64_t ComputeAvailableFeatures(const FeatureBitset& FB) const;\n"; 2878 if (HasOptionalOperands) { 2879 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " 2880 << "unsigned Opcode,\n" 2881 << " const OperandVector &Operands,\n" 2882 << " const SmallBitVector &OptionalOperandsMask);\n"; 2883 } else { 2884 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " 2885 << "unsigned Opcode,\n" 2886 << " const OperandVector &Operands);\n"; 2887 } 2888 OS << " void convertToMapAndConstraints(unsigned Kind,\n "; 2889 OS << " const OperandVector &Operands) override;\n"; 2890 if (HasMnemonicFirst) 2891 OS << " bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID);\n"; 2892 OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n" 2893 << " MCInst &Inst,\n" 2894 << " uint64_t &ErrorInfo," 2895 << " bool matchingInlineAsm,\n" 2896 << " unsigned VariantID = 0);\n"; 2897 2898 if (!Info.OperandMatchInfo.empty()) { 2899 OS << "\n enum OperandMatchResultTy {\n"; 2900 OS << " MatchOperand_Success, // operand matched successfully\n"; 2901 OS << " MatchOperand_NoMatch, // operand did not match\n"; 2902 OS << " MatchOperand_ParseFail // operand matched but had errors\n"; 2903 OS << " };\n"; 2904 OS << " OperandMatchResultTy MatchOperandParserImpl(\n"; 2905 OS << " OperandVector &Operands,\n"; 2906 OS << " StringRef Mnemonic);\n"; 2907 2908 OS << " OperandMatchResultTy tryCustomParseOperand(\n"; 2909 OS << " OperandVector &Operands,\n"; 2910 OS << " unsigned MCK);\n\n"; 2911 } 2912 2913 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n"; 2914 2915 // Emit the operand match diagnostic enum names. 2916 OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n"; 2917 OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n"; 2918 emitOperandDiagnosticTypes(Info, OS); 2919 OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n"; 2920 2921 OS << "\n#ifdef GET_REGISTER_MATCHER\n"; 2922 OS << "#undef GET_REGISTER_MATCHER\n\n"; 2923 2924 // Emit the subtarget feature enumeration. 2925 emitSubtargetFeatureFlagEnumeration(Info, OS); 2926 2927 // Emit the function to match a register name to number. 2928 // This should be omitted for Mips target 2929 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName")) 2930 emitMatchRegisterName(Target, AsmParser, OS); 2931 2932 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName")) 2933 emitMatchRegisterAltName(Target, AsmParser, OS); 2934 2935 OS << "#endif // GET_REGISTER_MATCHER\n\n"; 2936 2937 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n"; 2938 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n"; 2939 2940 // Generate the helper function to get the names for subtarget features. 2941 emitGetSubtargetFeatureName(Info, OS); 2942 2943 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n"; 2944 2945 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n"; 2946 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n"; 2947 2948 // Generate the function that remaps for mnemonic aliases. 2949 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target); 2950 2951 // Generate the convertToMCInst function to convert operands into an MCInst. 2952 // Also, generate the convertToMapAndConstraints function for MS-style inline 2953 // assembly. The latter doesn't actually generate a MCInst. 2954 emitConvertFuncs(Target, ClassName, Info.Matchables, HasMnemonicFirst, 2955 HasOptionalOperands, OS); 2956 2957 // Emit the enumeration for classes which participate in matching. 2958 emitMatchClassEnumeration(Target, Info.Classes, OS); 2959 2960 // Emit the routine to match token strings to their match class. 2961 emitMatchTokenString(Target, Info.Classes, OS); 2962 2963 // Emit the subclass predicate routine. 2964 emitIsSubclass(Target, Info.Classes, OS); 2965 2966 // Emit the routine to validate an operand against a match class. 2967 emitValidateOperandClass(Info, OS); 2968 2969 // Emit the available features compute function. 2970 emitComputeAvailableFeatures(Info, OS); 2971 2972 StringToOffsetTable StringTable; 2973 2974 size_t MaxNumOperands = 0; 2975 unsigned MaxMnemonicIndex = 0; 2976 bool HasDeprecation = false; 2977 for (const auto &MI : Info.Matchables) { 2978 MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size()); 2979 HasDeprecation |= MI->HasDeprecation; 2980 2981 // Store a pascal-style length byte in the mnemonic. 2982 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str(); 2983 MaxMnemonicIndex = std::max(MaxMnemonicIndex, 2984 StringTable.GetOrAddStringOffset(LenMnemonic, false)); 2985 } 2986 2987 OS << "static const char *const MnemonicTable =\n"; 2988 StringTable.EmitString(OS); 2989 OS << ";\n\n"; 2990 2991 // Emit the static match table; unused classes get initalized to 0 which is 2992 // guaranteed to be InvalidMatchClass. 2993 // 2994 // FIXME: We can reduce the size of this table very easily. First, we change 2995 // it so that store the kinds in separate bit-fields for each index, which 2996 // only needs to be the max width used for classes at that index (we also need 2997 // to reject based on this during classification). If we then make sure to 2998 // order the match kinds appropriately (putting mnemonics last), then we 2999 // should only end up using a few bits for each class, especially the ones 3000 // following the mnemonic. 3001 OS << "namespace {\n"; 3002 OS << " struct MatchEntry {\n"; 3003 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex) 3004 << " Mnemonic;\n"; 3005 OS << " uint16_t Opcode;\n"; 3006 OS << " " << getMinimalTypeForRange(Info.Matchables.size()) 3007 << " ConvertFn;\n"; 3008 OS << " " << getMinimalRequiredFeaturesType(Info) 3009 << " RequiredFeatures;\n"; 3010 OS << " " << getMinimalTypeForRange( 3011 std::distance(Info.Classes.begin(), Info.Classes.end())) 3012 << " Classes[" << MaxNumOperands << "];\n"; 3013 OS << " StringRef getMnemonic() const {\n"; 3014 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 3015 OS << " MnemonicTable[Mnemonic]);\n"; 3016 OS << " }\n"; 3017 OS << " };\n\n"; 3018 3019 OS << " // Predicate for searching for an opcode.\n"; 3020 OS << " struct LessOpcode {\n"; 3021 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n"; 3022 OS << " return LHS.getMnemonic() < RHS;\n"; 3023 OS << " }\n"; 3024 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n"; 3025 OS << " return LHS < RHS.getMnemonic();\n"; 3026 OS << " }\n"; 3027 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n"; 3028 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 3029 OS << " }\n"; 3030 OS << " };\n"; 3031 3032 OS << "} // end anonymous namespace.\n\n"; 3033 3034 unsigned VariantCount = Target.getAsmParserVariantCount(); 3035 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3036 Record *AsmVariant = Target.getAsmParserVariant(VC); 3037 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3038 3039 OS << "static const MatchEntry MatchTable" << VC << "[] = {\n"; 3040 3041 for (const auto &MI : Info.Matchables) { 3042 if (MI->AsmVariantID != AsmVariantNo) 3043 continue; 3044 3045 // Store a pascal-style length byte in the mnemonic. 3046 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str(); 3047 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 3048 << " /* " << MI->Mnemonic << " */, " 3049 << Target.getName() << "::" 3050 << MI->getResultInst()->TheDef->getName() << ", " 3051 << MI->ConversionFnKind << ", "; 3052 3053 // Write the required features mask. 3054 if (!MI->RequiredFeatures.empty()) { 3055 for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i) { 3056 if (i) OS << "|"; 3057 OS << MI->RequiredFeatures[i]->getEnumName(); 3058 } 3059 } else 3060 OS << "0"; 3061 3062 OS << ", { "; 3063 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) { 3064 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i]; 3065 3066 if (i) OS << ", "; 3067 OS << Op.Class->Name; 3068 } 3069 OS << " }, },\n"; 3070 } 3071 3072 OS << "};\n\n"; 3073 } 3074 3075 // A method to determine if a mnemonic is in the list. 3076 if (HasMnemonicFirst) { 3077 OS << "bool " << Target.getName() << ClassName << "::\n" 3078 << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n"; 3079 OS << " // Find the appropriate table for this asm variant.\n"; 3080 OS << " const MatchEntry *Start, *End;\n"; 3081 OS << " switch (VariantID) {\n"; 3082 OS << " default: llvm_unreachable(\"invalid variant!\");\n"; 3083 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3084 Record *AsmVariant = Target.getAsmParserVariant(VC); 3085 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3086 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC 3087 << "); End = std::end(MatchTable" << VC << "); break;\n"; 3088 } 3089 OS << " }\n"; 3090 OS << " // Search the table.\n"; 3091 OS << " auto MnemonicRange = "; 3092 OS << "std::equal_range(Start, End, Mnemonic, LessOpcode());\n"; 3093 OS << " return MnemonicRange.first != MnemonicRange.second;\n"; 3094 OS << "}\n\n"; 3095 } 3096 3097 // Finally, build the match function. 3098 OS << "unsigned " << Target.getName() << ClassName << "::\n" 3099 << "MatchInstructionImpl(const OperandVector &Operands,\n"; 3100 OS << " MCInst &Inst, uint64_t &ErrorInfo,\n" 3101 << " bool matchingInlineAsm, unsigned VariantID) {\n"; 3102 3103 OS << " // Eliminate obvious mismatches.\n"; 3104 OS << " if (Operands.size() > " 3105 << (MaxNumOperands + HasMnemonicFirst) << ") {\n"; 3106 OS << " ErrorInfo = " 3107 << (MaxNumOperands + HasMnemonicFirst) << ";\n"; 3108 OS << " return Match_InvalidOperand;\n"; 3109 OS << " }\n\n"; 3110 3111 // Emit code to get the available features. 3112 OS << " // Get the current feature set.\n"; 3113 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n"; 3114 3115 OS << " // Get the instruction mnemonic, which is the first token.\n"; 3116 if (HasMnemonicFirst) { 3117 OS << " StringRef Mnemonic = ((" << Target.getName() 3118 << "Operand&)*Operands[0]).getToken();\n\n"; 3119 } else { 3120 OS << " StringRef Mnemonic;\n"; 3121 OS << " if (Operands[0]->isToken())\n"; 3122 OS << " Mnemonic = ((" << Target.getName() 3123 << "Operand&)*Operands[0]).getToken();\n\n"; 3124 } 3125 3126 if (HasMnemonicAliases) { 3127 OS << " // Process all MnemonicAliases to remap the mnemonic.\n"; 3128 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n"; 3129 } 3130 3131 // Emit code to compute the class list for this operand vector. 3132 OS << " // Some state to try to produce better error messages.\n"; 3133 OS << " bool HadMatchOtherThanFeatures = false;\n"; 3134 OS << " bool HadMatchOtherThanPredicate = false;\n"; 3135 OS << " unsigned RetCode = Match_InvalidOperand;\n"; 3136 OS << " uint64_t MissingFeatures = ~0ULL;\n"; 3137 if (HasOptionalOperands) { 3138 OS << " SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n"; 3139 } 3140 OS << " // Set ErrorInfo to the operand that mismatches if it is\n"; 3141 OS << " // wrong for all instances of the instruction.\n"; 3142 OS << " ErrorInfo = ~0ULL;\n"; 3143 3144 // Emit code to search the table. 3145 OS << " // Find the appropriate table for this asm variant.\n"; 3146 OS << " const MatchEntry *Start, *End;\n"; 3147 OS << " switch (VariantID) {\n"; 3148 OS << " default: llvm_unreachable(\"invalid variant!\");\n"; 3149 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3150 Record *AsmVariant = Target.getAsmParserVariant(VC); 3151 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3152 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC 3153 << "); End = std::end(MatchTable" << VC << "); break;\n"; 3154 } 3155 OS << " }\n"; 3156 3157 OS << " // Search the table.\n"; 3158 if (HasMnemonicFirst) { 3159 OS << " auto MnemonicRange = " 3160 "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n"; 3161 } else { 3162 OS << " auto MnemonicRange = std::make_pair(Start, End);\n"; 3163 OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n"; 3164 OS << " if (!Mnemonic.empty())\n"; 3165 OS << " MnemonicRange = " 3166 "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n"; 3167 } 3168 3169 OS << " // Return a more specific error code if no mnemonics match.\n"; 3170 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 3171 OS << " return Match_MnemonicFail;\n\n"; 3172 3173 OS << " for (const MatchEntry *it = MnemonicRange.first, " 3174 << "*ie = MnemonicRange.second;\n"; 3175 OS << " it != ie; ++it) {\n"; 3176 3177 if (HasMnemonicFirst) { 3178 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 3179 OS << " assert(Mnemonic == it->getMnemonic());\n"; 3180 } 3181 3182 // Emit check that the subclasses match. 3183 OS << " bool OperandsValid = true;\n"; 3184 if (HasOptionalOperands) { 3185 OS << " OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n"; 3186 } 3187 OS << " for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex") 3188 << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex") 3189 << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n"; 3190 OS << " auto Formal = " 3191 << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n"; 3192 OS << " if (ActualIdx >= Operands.size()) {\n"; 3193 OS << " OperandsValid = (Formal == " <<"InvalidMatchClass) || " 3194 "isSubclass(Formal, OptionalMatchClass);\n"; 3195 OS << " if (!OperandsValid) ErrorInfo = ActualIdx;\n"; 3196 if (HasOptionalOperands) { 3197 OS << " OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands 3198 << ");\n"; 3199 } 3200 OS << " break;\n"; 3201 OS << " }\n"; 3202 OS << " MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n"; 3203 OS << " unsigned Diag = validateOperandClass(Actual, Formal);\n"; 3204 OS << " if (Diag == Match_Success) {\n"; 3205 OS << " ++ActualIdx;\n"; 3206 OS << " continue;\n"; 3207 OS << " }\n"; 3208 OS << " // If the generic handler indicates an invalid operand\n"; 3209 OS << " // failure, check for a special case.\n"; 3210 OS << " if (Diag == Match_InvalidOperand) {\n"; 3211 OS << " Diag = validateTargetOperandClass(Actual, Formal);\n"; 3212 OS << " if (Diag == Match_Success) {\n"; 3213 OS << " ++ActualIdx;\n"; 3214 OS << " continue;\n"; 3215 OS << " }\n"; 3216 OS << " }\n"; 3217 OS << " // If current formal operand wasn't matched and it is optional\n" 3218 << " // then try to match next formal operand\n"; 3219 OS << " if (Diag == Match_InvalidOperand " 3220 << "&& isSubclass(Formal, OptionalMatchClass)) {\n"; 3221 if (HasOptionalOperands) { 3222 OS << " OptionalOperandsMask.set(FormalIdx);\n"; 3223 } 3224 OS << " continue;\n"; 3225 OS << " }\n"; 3226 OS << " // If this operand is broken for all of the instances of this\n"; 3227 OS << " // mnemonic, keep track of it so we can report loc info.\n"; 3228 OS << " // If we already had a match that only failed due to a\n"; 3229 OS << " // target predicate, that diagnostic is preferred.\n"; 3230 OS << " if (!HadMatchOtherThanPredicate &&\n"; 3231 OS << " (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) {\n"; 3232 OS << " ErrorInfo = ActualIdx;\n"; 3233 OS << " // InvalidOperand is the default. Prefer specificity.\n"; 3234 OS << " if (Diag != Match_InvalidOperand)\n"; 3235 OS << " RetCode = Diag;\n"; 3236 OS << " }\n"; 3237 OS << " // Otherwise, just reject this instance of the mnemonic.\n"; 3238 OS << " OperandsValid = false;\n"; 3239 OS << " break;\n"; 3240 OS << " }\n\n"; 3241 3242 OS << " if (!OperandsValid) continue;\n"; 3243 3244 // Emit check that the required features are available. 3245 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 3246 << "!= it->RequiredFeatures) {\n"; 3247 OS << " HadMatchOtherThanFeatures = true;\n"; 3248 OS << " uint64_t NewMissingFeatures = it->RequiredFeatures & " 3249 "~AvailableFeatures;\n"; 3250 OS << " if (countPopulation(NewMissingFeatures) <=\n" 3251 " countPopulation(MissingFeatures))\n"; 3252 OS << " MissingFeatures = NewMissingFeatures;\n"; 3253 OS << " continue;\n"; 3254 OS << " }\n"; 3255 OS << "\n"; 3256 OS << " Inst.clear();\n\n"; 3257 OS << " if (matchingInlineAsm) {\n"; 3258 OS << " Inst.setOpcode(it->Opcode);\n"; 3259 OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n"; 3260 OS << " return Match_Success;\n"; 3261 OS << " }\n\n"; 3262 OS << " // We have selected a definite instruction, convert the parsed\n" 3263 << " // operands into the appropriate MCInst.\n"; 3264 if (HasOptionalOperands) { 3265 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n" 3266 << " OptionalOperandsMask);\n"; 3267 } else { 3268 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n"; 3269 } 3270 OS << "\n"; 3271 3272 // Verify the instruction with the target-specific match predicate function. 3273 OS << " // We have a potential match. Check the target predicate to\n" 3274 << " // handle any context sensitive constraints.\n" 3275 << " unsigned MatchResult;\n" 3276 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !=" 3277 << " Match_Success) {\n" 3278 << " Inst.clear();\n" 3279 << " RetCode = MatchResult;\n" 3280 << " HadMatchOtherThanPredicate = true;\n" 3281 << " continue;\n" 3282 << " }\n\n"; 3283 3284 // Call the post-processing function, if used. 3285 std::string InsnCleanupFn = 3286 AsmParser->getValueAsString("AsmParserInstCleanup"); 3287 if (!InsnCleanupFn.empty()) 3288 OS << " " << InsnCleanupFn << "(Inst);\n"; 3289 3290 if (HasDeprecation) { 3291 OS << " std::string Info;\n"; 3292 OS << " if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n"; 3293 OS << " SMLoc Loc = ((" << Target.getName() 3294 << "Operand&)*Operands[0]).getStartLoc();\n"; 3295 OS << " getParser().Warning(Loc, Info, None);\n"; 3296 OS << " }\n"; 3297 } 3298 3299 OS << " return Match_Success;\n"; 3300 OS << " }\n\n"; 3301 3302 OS << " // Okay, we had no match. Try to return a useful error code.\n"; 3303 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n"; 3304 OS << " return RetCode;\n\n"; 3305 OS << " // Missing feature matches return which features were missing\n"; 3306 OS << " ErrorInfo = MissingFeatures;\n"; 3307 OS << " return Match_MissingFeature;\n"; 3308 OS << "}\n\n"; 3309 3310 if (!Info.OperandMatchInfo.empty()) 3311 emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable, 3312 MaxMnemonicIndex, HasMnemonicFirst); 3313 3314 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n"; 3315 } 3316 3317 namespace llvm { 3318 3319 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) { 3320 emitSourceFileHeader("Assembly Matcher Source Fragment", OS); 3321 AsmMatcherEmitter(RK).run(OS); 3322 } 3323 3324 } // end namespace llvm 3325