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/CachedHashString.h" 101 #include "llvm/ADT/PointerUnion.h" 102 #include "llvm/ADT/STLExtras.h" 103 #include "llvm/ADT/SmallPtrSet.h" 104 #include "llvm/ADT/SmallVector.h" 105 #include "llvm/ADT/StringExtras.h" 106 #include "llvm/Support/CommandLine.h" 107 #include "llvm/Support/Debug.h" 108 #include "llvm/Support/ErrorHandling.h" 109 #include "llvm/TableGen/Error.h" 110 #include "llvm/TableGen/Record.h" 111 #include "llvm/TableGen/StringMatcher.h" 112 #include "llvm/TableGen/StringToOffsetTable.h" 113 #include "llvm/TableGen/TableGenBackend.h" 114 #include <cassert> 115 #include <cctype> 116 #include <forward_list> 117 #include <map> 118 #include <set> 119 120 using namespace llvm; 121 122 #define DEBUG_TYPE "asm-matcher-emitter" 123 124 static cl::opt<std::string> 125 MatchPrefix("match-prefix", cl::init(""), 126 cl::desc("Only match instructions with the given prefix")); 127 128 namespace { 129 class AsmMatcherInfo; 130 struct SubtargetFeatureInfo; 131 132 // Register sets are used as keys in some second-order sets TableGen creates 133 // when generating its data structures. This means that the order of two 134 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and 135 // can even affect compiler output (at least seen in diagnostics produced when 136 // all matches fail). So we use a type that sorts them consistently. 137 typedef std::set<Record*, LessRecordByID> RegisterSet; 138 139 class AsmMatcherEmitter { 140 RecordKeeper &Records; 141 public: 142 AsmMatcherEmitter(RecordKeeper &R) : Records(R) {} 143 144 void run(raw_ostream &o); 145 }; 146 147 /// ClassInfo - Helper class for storing the information about a particular 148 /// class of operands which can be matched. 149 struct ClassInfo { 150 enum ClassInfoKind { 151 /// Invalid kind, for use as a sentinel value. 152 Invalid = 0, 153 154 /// The class for a particular token. 155 Token, 156 157 /// The (first) register class, subsequent register classes are 158 /// RegisterClass0+1, and so on. 159 RegisterClass0, 160 161 /// The (first) user defined class, subsequent user defined classes are 162 /// UserClass0+1, and so on. 163 UserClass0 = 1<<16 164 }; 165 166 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + 167 /// N) for the Nth user defined class. 168 unsigned Kind; 169 170 /// SuperClasses - The super classes of this class. Note that for simplicities 171 /// sake user operands only record their immediate super class, while register 172 /// operands include all superclasses. 173 std::vector<ClassInfo*> SuperClasses; 174 175 /// Name - The full class name, suitable for use in an enum. 176 std::string Name; 177 178 /// ClassName - The unadorned generic name for this class (e.g., Token). 179 std::string ClassName; 180 181 /// ValueName - The name of the value this class represents; for a token this 182 /// is the literal token string, for an operand it is the TableGen class (or 183 /// empty if this is a derived class). 184 std::string ValueName; 185 186 /// PredicateMethod - The name of the operand method to test whether the 187 /// operand matches this class; this is not valid for Token or register kinds. 188 std::string PredicateMethod; 189 190 /// RenderMethod - The name of the operand method to add this operand to an 191 /// MCInst; this is not valid for Token or register kinds. 192 std::string RenderMethod; 193 194 /// ParserMethod - The name of the operand method to do a target specific 195 /// parsing on the operand. 196 std::string ParserMethod; 197 198 /// For register classes: the records for all the registers in this class. 199 RegisterSet Registers; 200 201 /// For custom match classes: the diagnostic kind for when the predicate fails. 202 std::string DiagnosticType; 203 204 /// Is this operand optional and not always required. 205 bool IsOptional; 206 207 /// DefaultMethod - The name of the method that returns the default operand 208 /// for optional operand 209 std::string DefaultMethod; 210 211 public: 212 /// isRegisterClass() - Check if this is a register class. 213 bool isRegisterClass() const { 214 return Kind >= RegisterClass0 && Kind < UserClass0; 215 } 216 217 /// isUserClass() - Check if this is a user defined class. 218 bool isUserClass() const { 219 return Kind >= UserClass0; 220 } 221 222 /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes 223 /// are related if they are in the same class hierarchy. 224 bool isRelatedTo(const ClassInfo &RHS) const { 225 // Tokens are only related to tokens. 226 if (Kind == Token || RHS.Kind == Token) 227 return Kind == Token && RHS.Kind == Token; 228 229 // Registers classes are only related to registers classes, and only if 230 // their intersection is non-empty. 231 if (isRegisterClass() || RHS.isRegisterClass()) { 232 if (!isRegisterClass() || !RHS.isRegisterClass()) 233 return false; 234 235 RegisterSet Tmp; 236 std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin()); 237 std::set_intersection(Registers.begin(), Registers.end(), 238 RHS.Registers.begin(), RHS.Registers.end(), 239 II, LessRecordByID()); 240 241 return !Tmp.empty(); 242 } 243 244 // Otherwise we have two users operands; they are related if they are in the 245 // same class hierarchy. 246 // 247 // FIXME: This is an oversimplification, they should only be related if they 248 // intersect, however we don't have that information. 249 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!"); 250 const ClassInfo *Root = this; 251 while (!Root->SuperClasses.empty()) 252 Root = Root->SuperClasses.front(); 253 254 const ClassInfo *RHSRoot = &RHS; 255 while (!RHSRoot->SuperClasses.empty()) 256 RHSRoot = RHSRoot->SuperClasses.front(); 257 258 return Root == RHSRoot; 259 } 260 261 /// isSubsetOf - Test whether this class is a subset of \p RHS. 262 bool isSubsetOf(const ClassInfo &RHS) const { 263 // This is a subset of RHS if it is the same class... 264 if (this == &RHS) 265 return true; 266 267 // ... or if any of its super classes are a subset of RHS. 268 for (const ClassInfo *CI : SuperClasses) 269 if (CI->isSubsetOf(RHS)) 270 return true; 271 272 return false; 273 } 274 275 int getTreeDepth() const { 276 int Depth = 0; 277 const ClassInfo *Root = this; 278 while (!Root->SuperClasses.empty()) { 279 Depth++; 280 Root = Root->SuperClasses.front(); 281 } 282 return Depth; 283 } 284 285 const ClassInfo *findRoot() const { 286 const ClassInfo *Root = this; 287 while (!Root->SuperClasses.empty()) 288 Root = Root->SuperClasses.front(); 289 return Root; 290 } 291 292 /// Compare two classes. This does not produce a total ordering, but does 293 /// guarantee that subclasses are sorted before their parents, and that the 294 /// ordering is transitive. 295 bool operator<(const ClassInfo &RHS) const { 296 if (this == &RHS) 297 return false; 298 299 // First, enforce the ordering between the three different types of class. 300 // Tokens sort before registers, which sort before user classes. 301 if (Kind == Token) { 302 if (RHS.Kind != Token) 303 return true; 304 assert(RHS.Kind == Token); 305 } else if (isRegisterClass()) { 306 if (RHS.Kind == Token) 307 return false; 308 else if (RHS.isUserClass()) 309 return true; 310 assert(RHS.isRegisterClass()); 311 } else if (isUserClass()) { 312 if (!RHS.isUserClass()) 313 return false; 314 assert(RHS.isUserClass()); 315 } else { 316 llvm_unreachable("Unknown ClassInfoKind"); 317 } 318 319 if (Kind == Token || isUserClass()) { 320 // Related tokens and user classes get sorted by depth in the inheritence 321 // tree (so that subclasses are before their parents). 322 if (isRelatedTo(RHS)) { 323 if (getTreeDepth() > RHS.getTreeDepth()) 324 return true; 325 if (getTreeDepth() < RHS.getTreeDepth()) 326 return false; 327 } else { 328 // Unrelated tokens and user classes are ordered by the name of their 329 // root nodes, so that there is a consistent ordering between 330 // unconnected trees. 331 return findRoot()->ValueName < RHS.findRoot()->ValueName; 332 } 333 } else if (isRegisterClass()) { 334 // For register sets, sort by number of registers. This guarantees that 335 // a set will always sort before all of it's strict supersets. 336 if (Registers.size() != RHS.Registers.size()) 337 return Registers.size() < RHS.Registers.size(); 338 } else { 339 llvm_unreachable("Unknown ClassInfoKind"); 340 } 341 342 // FIXME: We should be able to just return false here, as we only need a 343 // partial order (we use stable sorts, so this is deterministic) and the 344 // name of a class shouldn't be significant. However, some of the backends 345 // accidentally rely on this behaviour, so it will have to stay like this 346 // until they are fixed. 347 return ValueName < RHS.ValueName; 348 } 349 }; 350 351 class AsmVariantInfo { 352 public: 353 std::string RegisterPrefix; 354 std::string TokenizingCharacters; 355 std::string SeparatorCharacters; 356 std::string BreakCharacters; 357 std::string Name; 358 int AsmVariantNo; 359 }; 360 361 /// MatchableInfo - Helper class for storing the necessary information for an 362 /// instruction or alias which is capable of being matched. 363 struct MatchableInfo { 364 struct AsmOperand { 365 /// Token - This is the token that the operand came from. 366 StringRef Token; 367 368 /// The unique class instance this operand should match. 369 ClassInfo *Class; 370 371 /// The operand name this is, if anything. 372 StringRef SrcOpName; 373 374 /// The suboperand index within SrcOpName, or -1 for the entire operand. 375 int SubOpIdx; 376 377 /// Whether the token is "isolated", i.e., it is preceded and followed 378 /// by separators. 379 bool IsIsolatedToken; 380 381 /// Register record if this token is singleton register. 382 Record *SingletonReg; 383 384 explicit AsmOperand(bool IsIsolatedToken, StringRef T) 385 : Token(T), Class(nullptr), SubOpIdx(-1), 386 IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {} 387 }; 388 389 /// ResOperand - This represents a single operand in the result instruction 390 /// generated by the match. In cases (like addressing modes) where a single 391 /// assembler operand expands to multiple MCOperands, this represents the 392 /// single assembler operand, not the MCOperand. 393 struct ResOperand { 394 enum { 395 /// RenderAsmOperand - This represents an operand result that is 396 /// generated by calling the render method on the assembly operand. The 397 /// corresponding AsmOperand is specified by AsmOperandNum. 398 RenderAsmOperand, 399 400 /// TiedOperand - This represents a result operand that is a duplicate of 401 /// a previous result operand. 402 TiedOperand, 403 404 /// ImmOperand - This represents an immediate value that is dumped into 405 /// the operand. 406 ImmOperand, 407 408 /// RegOperand - This represents a fixed register that is dumped in. 409 RegOperand 410 } Kind; 411 412 union { 413 /// This is the operand # in the AsmOperands list that this should be 414 /// copied from. 415 unsigned AsmOperandNum; 416 417 /// TiedOperandNum - This is the (earlier) result operand that should be 418 /// copied from. 419 unsigned TiedOperandNum; 420 421 /// ImmVal - This is the immediate value added to the instruction. 422 int64_t ImmVal; 423 424 /// Register - This is the register record. 425 Record *Register; 426 }; 427 428 /// MINumOperands - The number of MCInst operands populated by this 429 /// operand. 430 unsigned MINumOperands; 431 432 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) { 433 ResOperand X; 434 X.Kind = RenderAsmOperand; 435 X.AsmOperandNum = AsmOpNum; 436 X.MINumOperands = NumOperands; 437 return X; 438 } 439 440 static ResOperand getTiedOp(unsigned TiedOperandNum) { 441 ResOperand X; 442 X.Kind = TiedOperand; 443 X.TiedOperandNum = TiedOperandNum; 444 X.MINumOperands = 1; 445 return X; 446 } 447 448 static ResOperand getImmOp(int64_t Val) { 449 ResOperand X; 450 X.Kind = ImmOperand; 451 X.ImmVal = Val; 452 X.MINumOperands = 1; 453 return X; 454 } 455 456 static ResOperand getRegOp(Record *Reg) { 457 ResOperand X; 458 X.Kind = RegOperand; 459 X.Register = Reg; 460 X.MINumOperands = 1; 461 return X; 462 } 463 }; 464 465 /// AsmVariantID - Target's assembly syntax variant no. 466 int AsmVariantID; 467 468 /// AsmString - The assembly string for this instruction (with variants 469 /// removed), e.g. "movsx $src, $dst". 470 std::string AsmString; 471 472 /// TheDef - This is the definition of the instruction or InstAlias that this 473 /// matchable came from. 474 Record *const TheDef; 475 476 /// DefRec - This is the definition that it came from. 477 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec; 478 479 const CodeGenInstruction *getResultInst() const { 480 if (DefRec.is<const CodeGenInstruction*>()) 481 return DefRec.get<const CodeGenInstruction*>(); 482 return DefRec.get<const CodeGenInstAlias*>()->ResultInst; 483 } 484 485 /// ResOperands - This is the operand list that should be built for the result 486 /// MCInst. 487 SmallVector<ResOperand, 8> ResOperands; 488 489 /// Mnemonic - This is the first token of the matched instruction, its 490 /// mnemonic. 491 StringRef Mnemonic; 492 493 /// AsmOperands - The textual operands that this instruction matches, 494 /// annotated with a class and where in the OperandList they were defined. 495 /// This directly corresponds to the tokenized AsmString after the mnemonic is 496 /// removed. 497 SmallVector<AsmOperand, 8> AsmOperands; 498 499 /// Predicates - The required subtarget features to match this instruction. 500 SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures; 501 502 /// ConversionFnKind - The enum value which is passed to the generated 503 /// convertToMCInst to convert parsed operands into an MCInst for this 504 /// function. 505 std::string ConversionFnKind; 506 507 /// If this instruction is deprecated in some form. 508 bool HasDeprecation; 509 510 /// If this is an alias, this is use to determine whether or not to using 511 /// the conversion function defined by the instruction's AsmMatchConverter 512 /// or to use the function generated by the alias. 513 bool UseInstAsmMatchConverter; 514 515 MatchableInfo(const CodeGenInstruction &CGI) 516 : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI), 517 UseInstAsmMatchConverter(true) { 518 } 519 520 MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias) 521 : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef), 522 DefRec(Alias.release()), 523 UseInstAsmMatchConverter( 524 TheDef->getValueAsBit("UseInstAsmMatchConverter")) { 525 } 526 527 // Could remove this and the dtor if PointerUnion supported unique_ptr 528 // elements with a dynamic failure/assertion (like the one below) in the case 529 // where it was copied while being in an owning state. 530 MatchableInfo(const MatchableInfo &RHS) 531 : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString), 532 TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands), 533 Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands), 534 RequiredFeatures(RHS.RequiredFeatures), 535 ConversionFnKind(RHS.ConversionFnKind), 536 HasDeprecation(RHS.HasDeprecation), 537 UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) { 538 assert(!DefRec.is<const CodeGenInstAlias *>()); 539 } 540 541 ~MatchableInfo() { 542 delete DefRec.dyn_cast<const CodeGenInstAlias*>(); 543 } 544 545 // Two-operand aliases clone from the main matchable, but mark the second 546 // operand as a tied operand of the first for purposes of the assembler. 547 void formTwoOperandAlias(StringRef Constraint); 548 549 void initialize(const AsmMatcherInfo &Info, 550 SmallPtrSetImpl<Record*> &SingletonRegisters, 551 AsmVariantInfo const &Variant, 552 bool HasMnemonicFirst); 553 554 /// validate - Return true if this matchable is a valid thing to match against 555 /// and perform a bunch of validity checking. 556 bool validate(StringRef CommentDelimiter, bool Hack) const; 557 558 /// findAsmOperand - Find the AsmOperand with the specified name and 559 /// suboperand index. 560 int findAsmOperand(StringRef N, int SubOpIdx) const { 561 auto I = find_if(AsmOperands, [&](const AsmOperand &Op) { 562 return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx; 563 }); 564 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; 565 } 566 567 /// findAsmOperandNamed - Find the first AsmOperand with the specified name. 568 /// This does not check the suboperand index. 569 int findAsmOperandNamed(StringRef N) const { 570 auto I = find_if(AsmOperands, 571 [&](const AsmOperand &Op) { return Op.SrcOpName == N; }); 572 return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; 573 } 574 575 void buildInstructionResultOperands(); 576 void buildAliasResultOperands(); 577 578 /// operator< - Compare two matchables. 579 bool operator<(const MatchableInfo &RHS) const { 580 // The primary comparator is the instruction mnemonic. 581 if (int Cmp = Mnemonic.compare(RHS.Mnemonic)) 582 return Cmp == -1; 583 584 if (AsmOperands.size() != RHS.AsmOperands.size()) 585 return AsmOperands.size() < RHS.AsmOperands.size(); 586 587 // Compare lexicographically by operand. The matcher validates that other 588 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith(). 589 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 590 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 591 return true; 592 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 593 return false; 594 } 595 596 // Give matches that require more features higher precedence. This is useful 597 // because we cannot define AssemblerPredicates with the negation of 598 // processor features. For example, ARM v6 "nop" may be either a HINT or 599 // MOV. With v6, we want to match HINT. The assembler has no way to 600 // predicate MOV under "NoV6", but HINT will always match first because it 601 // requires V6 while MOV does not. 602 if (RequiredFeatures.size() != RHS.RequiredFeatures.size()) 603 return RequiredFeatures.size() > RHS.RequiredFeatures.size(); 604 605 return false; 606 } 607 608 /// couldMatchAmbiguouslyWith - Check whether this matchable could 609 /// ambiguously match the same set of operands as \p RHS (without being a 610 /// strictly superior match). 611 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const { 612 // The primary comparator is the instruction mnemonic. 613 if (Mnemonic != RHS.Mnemonic) 614 return false; 615 616 // The number of operands is unambiguous. 617 if (AsmOperands.size() != RHS.AsmOperands.size()) 618 return false; 619 620 // Otherwise, make sure the ordering of the two instructions is unambiguous 621 // by checking that either (a) a token or operand kind discriminates them, 622 // or (b) the ordering among equivalent kinds is consistent. 623 624 // Tokens and operand kinds are unambiguous (assuming a correct target 625 // specific parser). 626 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) 627 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind || 628 AsmOperands[i].Class->Kind == ClassInfo::Token) 629 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class || 630 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 631 return false; 632 633 // Otherwise, this operand could commute if all operands are equivalent, or 634 // there is a pair of operands that compare less than and a pair that 635 // compare greater than. 636 bool HasLT = false, HasGT = false; 637 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 638 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 639 HasLT = true; 640 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 641 HasGT = true; 642 } 643 644 return HasLT == HasGT; 645 } 646 647 void dump() const; 648 649 private: 650 void tokenizeAsmString(AsmMatcherInfo const &Info, 651 AsmVariantInfo const &Variant); 652 void addAsmOperand(StringRef Token, bool IsIsolatedToken = false); 653 }; 654 655 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget 656 /// feature which participates in instruction matching. 657 struct SubtargetFeatureInfo { 658 /// \brief The predicate record for this feature. 659 Record *TheDef; 660 661 /// \brief An unique index assigned to represent this feature. 662 uint64_t Index; 663 664 SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {} 665 666 /// \brief The name of the enumerated constant identifying this feature. 667 std::string getEnumName() const { 668 return "Feature_" + TheDef->getName(); 669 } 670 671 void dump() const { 672 errs() << getEnumName() << " " << Index << "\n"; 673 TheDef->dump(); 674 } 675 }; 676 677 struct OperandMatchEntry { 678 unsigned OperandMask; 679 const MatchableInfo* MI; 680 ClassInfo *CI; 681 682 static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci, 683 unsigned opMask) { 684 OperandMatchEntry X; 685 X.OperandMask = opMask; 686 X.CI = ci; 687 X.MI = mi; 688 return X; 689 } 690 }; 691 692 class AsmMatcherInfo { 693 public: 694 /// Tracked Records 695 RecordKeeper &Records; 696 697 /// The tablegen AsmParser record. 698 Record *AsmParser; 699 700 /// Target - The target information. 701 CodeGenTarget &Target; 702 703 /// The classes which are needed for matching. 704 std::forward_list<ClassInfo> Classes; 705 706 /// The information on the matchables to match. 707 std::vector<std::unique_ptr<MatchableInfo>> Matchables; 708 709 /// Info for custom matching operands by user defined methods. 710 std::vector<OperandMatchEntry> OperandMatchInfo; 711 712 /// Map of Register records to their class information. 713 typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy; 714 RegisterClassesTy RegisterClasses; 715 716 /// Map of Predicate records to their subtarget information. 717 std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures; 718 719 /// Map of AsmOperandClass records to their class information. 720 std::map<Record*, ClassInfo*> AsmOperandClasses; 721 722 private: 723 /// Map of token to class information which has already been constructed. 724 std::map<std::string, ClassInfo*> TokenClasses; 725 726 /// Map of RegisterClass records to their class information. 727 std::map<Record*, ClassInfo*> RegisterClassClasses; 728 729 private: 730 /// getTokenClass - Lookup or create the class for the given token. 731 ClassInfo *getTokenClass(StringRef Token); 732 733 /// getOperandClass - Lookup or create the class for the given operand. 734 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI, 735 int SubOpIdx); 736 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx); 737 738 /// buildRegisterClasses - Build the ClassInfo* instances for register 739 /// classes. 740 void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters); 741 742 /// buildOperandClasses - Build the ClassInfo* instances for user defined 743 /// operand classes. 744 void buildOperandClasses(); 745 746 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName, 747 unsigned AsmOpIdx); 748 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName, 749 MatchableInfo::AsmOperand &Op); 750 751 public: 752 AsmMatcherInfo(Record *AsmParser, 753 CodeGenTarget &Target, 754 RecordKeeper &Records); 755 756 /// buildInfo - Construct the various tables used during matching. 757 void buildInfo(); 758 759 /// buildOperandMatchInfo - Build the necessary information to handle user 760 /// defined operand parsing methods. 761 void buildOperandMatchInfo(); 762 763 /// getSubtargetFeature - Lookup or create the subtarget feature info for the 764 /// given operand. 765 const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const { 766 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!"); 767 const auto &I = SubtargetFeatures.find(Def); 768 return I == SubtargetFeatures.end() ? nullptr : &I->second; 769 } 770 771 RecordKeeper &getRecords() const { 772 return Records; 773 } 774 775 bool hasOptionalOperands() const { 776 return find_if(Classes, [](const ClassInfo &Class) { 777 return Class.IsOptional; 778 }) != Classes.end(); 779 } 780 }; 781 782 } // end anonymous namespace 783 784 void MatchableInfo::dump() const { 785 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n"; 786 787 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 788 const AsmOperand &Op = AsmOperands[i]; 789 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - "; 790 errs() << '\"' << Op.Token << "\"\n"; 791 } 792 } 793 794 static std::pair<StringRef, StringRef> 795 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) { 796 // Split via the '='. 797 std::pair<StringRef, StringRef> Ops = S.split('='); 798 if (Ops.second == "") 799 PrintFatalError(Loc, "missing '=' in two-operand alias constraint"); 800 // Trim whitespace and the leading '$' on the operand names. 801 size_t start = Ops.first.find_first_of('$'); 802 if (start == std::string::npos) 803 PrintFatalError(Loc, "expected '$' prefix on asm operand name"); 804 Ops.first = Ops.first.slice(start + 1, std::string::npos); 805 size_t end = Ops.first.find_last_of(" \t"); 806 Ops.first = Ops.first.slice(0, end); 807 // Now the second operand. 808 start = Ops.second.find_first_of('$'); 809 if (start == std::string::npos) 810 PrintFatalError(Loc, "expected '$' prefix on asm operand name"); 811 Ops.second = Ops.second.slice(start + 1, std::string::npos); 812 end = Ops.second.find_last_of(" \t"); 813 Ops.first = Ops.first.slice(0, end); 814 return Ops; 815 } 816 817 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) { 818 // Figure out which operands are aliased and mark them as tied. 819 std::pair<StringRef, StringRef> Ops = 820 parseTwoOperandConstraint(Constraint, TheDef->getLoc()); 821 822 // Find the AsmOperands that refer to the operands we're aliasing. 823 int SrcAsmOperand = findAsmOperandNamed(Ops.first); 824 int DstAsmOperand = findAsmOperandNamed(Ops.second); 825 if (SrcAsmOperand == -1) 826 PrintFatalError(TheDef->getLoc(), 827 "unknown source two-operand alias operand '" + Ops.first + 828 "'."); 829 if (DstAsmOperand == -1) 830 PrintFatalError(TheDef->getLoc(), 831 "unknown destination two-operand alias operand '" + 832 Ops.second + "'."); 833 834 // Find the ResOperand that refers to the operand we're aliasing away 835 // and update it to refer to the combined operand instead. 836 for (ResOperand &Op : ResOperands) { 837 if (Op.Kind == ResOperand::RenderAsmOperand && 838 Op.AsmOperandNum == (unsigned)SrcAsmOperand) { 839 Op.AsmOperandNum = DstAsmOperand; 840 break; 841 } 842 } 843 // Remove the AsmOperand for the alias operand. 844 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand); 845 // Adjust the ResOperand references to any AsmOperands that followed 846 // the one we just deleted. 847 for (ResOperand &Op : ResOperands) { 848 switch(Op.Kind) { 849 default: 850 // Nothing to do for operands that don't reference AsmOperands. 851 break; 852 case ResOperand::RenderAsmOperand: 853 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand) 854 --Op.AsmOperandNum; 855 break; 856 case ResOperand::TiedOperand: 857 if (Op.TiedOperandNum > (unsigned)SrcAsmOperand) 858 --Op.TiedOperandNum; 859 break; 860 } 861 } 862 } 863 864 /// extractSingletonRegisterForAsmOperand - Extract singleton register, 865 /// if present, from specified token. 866 static void 867 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op, 868 const AsmMatcherInfo &Info, 869 StringRef RegisterPrefix) { 870 StringRef Tok = Op.Token; 871 872 // If this token is not an isolated token, i.e., it isn't separated from 873 // other tokens (e.g. with whitespace), don't interpret it as a register name. 874 if (!Op.IsIsolatedToken) 875 return; 876 877 if (RegisterPrefix.empty()) { 878 std::string LoweredTok = Tok.lower(); 879 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok)) 880 Op.SingletonReg = Reg->TheDef; 881 return; 882 } 883 884 if (!Tok.startswith(RegisterPrefix)) 885 return; 886 887 StringRef RegName = Tok.substr(RegisterPrefix.size()); 888 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName)) 889 Op.SingletonReg = Reg->TheDef; 890 891 // If there is no register prefix (i.e. "%" in "%eax"), then this may 892 // be some random non-register token, just ignore it. 893 } 894 895 void MatchableInfo::initialize(const AsmMatcherInfo &Info, 896 SmallPtrSetImpl<Record*> &SingletonRegisters, 897 AsmVariantInfo const &Variant, 898 bool HasMnemonicFirst) { 899 AsmVariantID = Variant.AsmVariantNo; 900 AsmString = 901 CodeGenInstruction::FlattenAsmStringVariants(AsmString, 902 Variant.AsmVariantNo); 903 904 tokenizeAsmString(Info, Variant); 905 906 // The first token of the instruction is the mnemonic, which must be a 907 // simple string, not a $foo variable or a singleton register. 908 if (AsmOperands.empty()) 909 PrintFatalError(TheDef->getLoc(), 910 "Instruction '" + TheDef->getName() + "' has no tokens"); 911 912 assert(!AsmOperands[0].Token.empty()); 913 if (HasMnemonicFirst) { 914 Mnemonic = AsmOperands[0].Token; 915 if (Mnemonic[0] == '$') 916 PrintFatalError(TheDef->getLoc(), 917 "Invalid instruction mnemonic '" + Mnemonic + "'!"); 918 919 // Remove the first operand, it is tracked in the mnemonic field. 920 AsmOperands.erase(AsmOperands.begin()); 921 } else if (AsmOperands[0].Token[0] != '$') 922 Mnemonic = AsmOperands[0].Token; 923 924 // Compute the require features. 925 for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates")) 926 if (const SubtargetFeatureInfo *Feature = 927 Info.getSubtargetFeature(Predicate)) 928 RequiredFeatures.push_back(Feature); 929 930 // Collect singleton registers, if used. 931 for (MatchableInfo::AsmOperand &Op : AsmOperands) { 932 extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix); 933 if (Record *Reg = Op.SingletonReg) 934 SingletonRegisters.insert(Reg); 935 } 936 937 const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask"); 938 if (!DepMask) 939 DepMask = TheDef->getValue("ComplexDeprecationPredicate"); 940 941 HasDeprecation = 942 DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false; 943 } 944 945 /// Append an AsmOperand for the given substring of AsmString. 946 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) { 947 AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token)); 948 } 949 950 /// tokenizeAsmString - Tokenize a simplified assembly string. 951 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info, 952 AsmVariantInfo const &Variant) { 953 StringRef String = AsmString; 954 size_t Prev = 0; 955 bool InTok = false; 956 bool IsIsolatedToken = true; 957 for (size_t i = 0, e = String.size(); i != e; ++i) { 958 char Char = String[i]; 959 if (Variant.BreakCharacters.find(Char) != std::string::npos) { 960 if (InTok) { 961 addAsmOperand(String.slice(Prev, i), false); 962 Prev = i; 963 IsIsolatedToken = false; 964 } 965 InTok = true; 966 continue; 967 } 968 if (Variant.TokenizingCharacters.find(Char) != std::string::npos) { 969 if (InTok) { 970 addAsmOperand(String.slice(Prev, i), IsIsolatedToken); 971 InTok = false; 972 IsIsolatedToken = false; 973 } 974 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken); 975 Prev = i + 1; 976 IsIsolatedToken = true; 977 continue; 978 } 979 if (Variant.SeparatorCharacters.find(Char) != std::string::npos) { 980 if (InTok) { 981 addAsmOperand(String.slice(Prev, i), IsIsolatedToken); 982 InTok = false; 983 } 984 Prev = i + 1; 985 IsIsolatedToken = true; 986 continue; 987 } 988 989 switch (Char) { 990 case '\\': 991 if (InTok) { 992 addAsmOperand(String.slice(Prev, i), false); 993 InTok = false; 994 IsIsolatedToken = false; 995 } 996 ++i; 997 assert(i != String.size() && "Invalid quoted character"); 998 addAsmOperand(String.slice(i, i + 1), IsIsolatedToken); 999 Prev = i + 1; 1000 IsIsolatedToken = false; 1001 break; 1002 1003 case '$': { 1004 if (InTok) { 1005 addAsmOperand(String.slice(Prev, i), false); 1006 InTok = false; 1007 IsIsolatedToken = false; 1008 } 1009 1010 // If this isn't "${", start new identifier looking like "$xxx" 1011 if (i + 1 == String.size() || String[i + 1] != '{') { 1012 Prev = i; 1013 break; 1014 } 1015 1016 size_t EndPos = String.find('}', i); 1017 assert(EndPos != StringRef::npos && 1018 "Missing brace in operand reference!"); 1019 addAsmOperand(String.slice(i, EndPos+1), IsIsolatedToken); 1020 Prev = EndPos + 1; 1021 i = EndPos; 1022 IsIsolatedToken = false; 1023 break; 1024 } 1025 1026 default: 1027 InTok = true; 1028 break; 1029 } 1030 } 1031 if (InTok && Prev != String.size()) 1032 addAsmOperand(String.substr(Prev), IsIsolatedToken); 1033 } 1034 1035 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const { 1036 // Reject matchables with no .s string. 1037 if (AsmString.empty()) 1038 PrintFatalError(TheDef->getLoc(), "instruction with empty asm string"); 1039 1040 // Reject any matchables with a newline in them, they should be marked 1041 // isCodeGenOnly if they are pseudo instructions. 1042 if (AsmString.find('\n') != std::string::npos) 1043 PrintFatalError(TheDef->getLoc(), 1044 "multiline instruction is not valid for the asmparser, " 1045 "mark it isCodeGenOnly"); 1046 1047 // Remove comments from the asm string. We know that the asmstring only 1048 // has one line. 1049 if (!CommentDelimiter.empty() && 1050 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos) 1051 PrintFatalError(TheDef->getLoc(), 1052 "asmstring for instruction has comment character in it, " 1053 "mark it isCodeGenOnly"); 1054 1055 // Reject matchables with operand modifiers, these aren't something we can 1056 // handle, the target should be refactored to use operands instead of 1057 // modifiers. 1058 // 1059 // Also, check for instructions which reference the operand multiple times; 1060 // this implies a constraint we would not honor. 1061 std::set<std::string> OperandNames; 1062 for (const AsmOperand &Op : AsmOperands) { 1063 StringRef Tok = Op.Token; 1064 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos) 1065 PrintFatalError(TheDef->getLoc(), 1066 "matchable with operand modifier '" + Tok + 1067 "' not supported by asm matcher. Mark isCodeGenOnly!"); 1068 1069 // Verify that any operand is only mentioned once. 1070 // We reject aliases and ignore instructions for now. 1071 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) { 1072 if (!Hack) 1073 PrintFatalError(TheDef->getLoc(), 1074 "ERROR: matchable with tied operand '" + Tok + 1075 "' can never be matched!"); 1076 // FIXME: Should reject these. The ARM backend hits this with $lane in a 1077 // bunch of instructions. It is unclear what the right answer is. 1078 DEBUG({ 1079 errs() << "warning: '" << TheDef->getName() << "': " 1080 << "ignoring instruction with tied operand '" 1081 << Tok << "'\n"; 1082 }); 1083 return false; 1084 } 1085 } 1086 1087 return true; 1088 } 1089 1090 static std::string getEnumNameForToken(StringRef Str) { 1091 std::string Res; 1092 1093 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) { 1094 switch (*it) { 1095 case '*': Res += "_STAR_"; break; 1096 case '%': Res += "_PCT_"; break; 1097 case ':': Res += "_COLON_"; break; 1098 case '!': Res += "_EXCLAIM_"; break; 1099 case '.': Res += "_DOT_"; break; 1100 case '<': Res += "_LT_"; break; 1101 case '>': Res += "_GT_"; break; 1102 case '-': Res += "_MINUS_"; break; 1103 default: 1104 if ((*it >= 'A' && *it <= 'Z') || 1105 (*it >= 'a' && *it <= 'z') || 1106 (*it >= '0' && *it <= '9')) 1107 Res += *it; 1108 else 1109 Res += "_" + utostr((unsigned) *it) + "_"; 1110 } 1111 } 1112 1113 return Res; 1114 } 1115 1116 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) { 1117 ClassInfo *&Entry = TokenClasses[Token]; 1118 1119 if (!Entry) { 1120 Classes.emplace_front(); 1121 Entry = &Classes.front(); 1122 Entry->Kind = ClassInfo::Token; 1123 Entry->ClassName = "Token"; 1124 Entry->Name = "MCK_" + getEnumNameForToken(Token); 1125 Entry->ValueName = Token; 1126 Entry->PredicateMethod = "<invalid>"; 1127 Entry->RenderMethod = "<invalid>"; 1128 Entry->ParserMethod = ""; 1129 Entry->DiagnosticType = ""; 1130 Entry->IsOptional = false; 1131 Entry->DefaultMethod = "<invalid>"; 1132 } 1133 1134 return Entry; 1135 } 1136 1137 ClassInfo * 1138 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI, 1139 int SubOpIdx) { 1140 Record *Rec = OI.Rec; 1141 if (SubOpIdx != -1) 1142 Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef(); 1143 return getOperandClass(Rec, SubOpIdx); 1144 } 1145 1146 ClassInfo * 1147 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) { 1148 if (Rec->isSubClassOf("RegisterOperand")) { 1149 // RegisterOperand may have an associated ParserMatchClass. If it does, 1150 // use it, else just fall back to the underlying register class. 1151 const RecordVal *R = Rec->getValue("ParserMatchClass"); 1152 if (!R || !R->getValue()) 1153 PrintFatalError("Record `" + Rec->getName() + 1154 "' does not have a ParserMatchClass!\n"); 1155 1156 if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) { 1157 Record *MatchClass = DI->getDef(); 1158 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 1159 return CI; 1160 } 1161 1162 // No custom match class. Just use the register class. 1163 Record *ClassRec = Rec->getValueAsDef("RegClass"); 1164 if (!ClassRec) 1165 PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() + 1166 "' has no associated register class!\n"); 1167 if (ClassInfo *CI = RegisterClassClasses[ClassRec]) 1168 return CI; 1169 PrintFatalError(Rec->getLoc(), "register class has no class info!"); 1170 } 1171 1172 if (Rec->isSubClassOf("RegisterClass")) { 1173 if (ClassInfo *CI = RegisterClassClasses[Rec]) 1174 return CI; 1175 PrintFatalError(Rec->getLoc(), "register class has no class info!"); 1176 } 1177 1178 if (!Rec->isSubClassOf("Operand")) 1179 PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() + 1180 "' does not derive from class Operand!\n"); 1181 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1182 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 1183 return CI; 1184 1185 PrintFatalError(Rec->getLoc(), "operand has no match class!"); 1186 } 1187 1188 struct LessRegisterSet { 1189 bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const { 1190 // std::set<T> defines its own compariso "operator<", but it 1191 // performs a lexicographical comparison by T's innate comparison 1192 // for some reason. We don't want non-deterministic pointer 1193 // comparisons so use this instead. 1194 return std::lexicographical_compare(LHS.begin(), LHS.end(), 1195 RHS.begin(), RHS.end(), 1196 LessRecordByID()); 1197 } 1198 }; 1199 1200 void AsmMatcherInfo:: 1201 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) { 1202 const auto &Registers = Target.getRegBank().getRegisters(); 1203 auto &RegClassList = Target.getRegBank().getRegClasses(); 1204 1205 typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet; 1206 1207 // The register sets used for matching. 1208 RegisterSetSet RegisterSets; 1209 1210 // Gather the defined sets. 1211 for (const CodeGenRegisterClass &RC : RegClassList) 1212 RegisterSets.insert( 1213 RegisterSet(RC.getOrder().begin(), RC.getOrder().end())); 1214 1215 // Add any required singleton sets. 1216 for (Record *Rec : SingletonRegisters) { 1217 RegisterSets.insert(RegisterSet(&Rec, &Rec + 1)); 1218 } 1219 1220 // Introduce derived sets where necessary (when a register does not determine 1221 // a unique register set class), and build the mapping of registers to the set 1222 // they should classify to. 1223 std::map<Record*, RegisterSet> RegisterMap; 1224 for (const CodeGenRegister &CGR : Registers) { 1225 // Compute the intersection of all sets containing this register. 1226 RegisterSet ContainingSet; 1227 1228 for (const RegisterSet &RS : RegisterSets) { 1229 if (!RS.count(CGR.TheDef)) 1230 continue; 1231 1232 if (ContainingSet.empty()) { 1233 ContainingSet = RS; 1234 continue; 1235 } 1236 1237 RegisterSet Tmp; 1238 std::swap(Tmp, ContainingSet); 1239 std::insert_iterator<RegisterSet> II(ContainingSet, 1240 ContainingSet.begin()); 1241 std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II, 1242 LessRecordByID()); 1243 } 1244 1245 if (!ContainingSet.empty()) { 1246 RegisterSets.insert(ContainingSet); 1247 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet)); 1248 } 1249 } 1250 1251 // Construct the register classes. 1252 std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses; 1253 unsigned Index = 0; 1254 for (const RegisterSet &RS : RegisterSets) { 1255 Classes.emplace_front(); 1256 ClassInfo *CI = &Classes.front(); 1257 CI->Kind = ClassInfo::RegisterClass0 + Index; 1258 CI->ClassName = "Reg" + utostr(Index); 1259 CI->Name = "MCK_Reg" + utostr(Index); 1260 CI->ValueName = ""; 1261 CI->PredicateMethod = ""; // unused 1262 CI->RenderMethod = "addRegOperands"; 1263 CI->Registers = RS; 1264 // FIXME: diagnostic type. 1265 CI->DiagnosticType = ""; 1266 CI->IsOptional = false; 1267 CI->DefaultMethod = ""; // unused 1268 RegisterSetClasses.insert(std::make_pair(RS, CI)); 1269 ++Index; 1270 } 1271 1272 // Find the superclasses; we could compute only the subgroup lattice edges, 1273 // but there isn't really a point. 1274 for (const RegisterSet &RS : RegisterSets) { 1275 ClassInfo *CI = RegisterSetClasses[RS]; 1276 for (const RegisterSet &RS2 : RegisterSets) 1277 if (RS != RS2 && 1278 std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(), 1279 LessRecordByID())) 1280 CI->SuperClasses.push_back(RegisterSetClasses[RS2]); 1281 } 1282 1283 // Name the register classes which correspond to a user defined RegisterClass. 1284 for (const CodeGenRegisterClass &RC : RegClassList) { 1285 // Def will be NULL for non-user defined register classes. 1286 Record *Def = RC.getDef(); 1287 if (!Def) 1288 continue; 1289 ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(), 1290 RC.getOrder().end())]; 1291 if (CI->ValueName.empty()) { 1292 CI->ClassName = RC.getName(); 1293 CI->Name = "MCK_" + RC.getName(); 1294 CI->ValueName = RC.getName(); 1295 } else 1296 CI->ValueName = CI->ValueName + "," + RC.getName(); 1297 1298 RegisterClassClasses.insert(std::make_pair(Def, CI)); 1299 } 1300 1301 // Populate the map for individual registers. 1302 for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(), 1303 ie = RegisterMap.end(); it != ie; ++it) 1304 RegisterClasses[it->first] = RegisterSetClasses[it->second]; 1305 1306 // Name the register classes which correspond to singleton registers. 1307 for (Record *Rec : SingletonRegisters) { 1308 ClassInfo *CI = RegisterClasses[Rec]; 1309 assert(CI && "Missing singleton register class info!"); 1310 1311 if (CI->ValueName.empty()) { 1312 CI->ClassName = Rec->getName(); 1313 CI->Name = "MCK_" + Rec->getName(); 1314 CI->ValueName = Rec->getName(); 1315 } else 1316 CI->ValueName = CI->ValueName + "," + Rec->getName(); 1317 } 1318 } 1319 1320 void AsmMatcherInfo::buildOperandClasses() { 1321 std::vector<Record*> AsmOperands = 1322 Records.getAllDerivedDefinitions("AsmOperandClass"); 1323 1324 // Pre-populate AsmOperandClasses map. 1325 for (Record *Rec : AsmOperands) { 1326 Classes.emplace_front(); 1327 AsmOperandClasses[Rec] = &Classes.front(); 1328 } 1329 1330 unsigned Index = 0; 1331 for (Record *Rec : AsmOperands) { 1332 ClassInfo *CI = AsmOperandClasses[Rec]; 1333 CI->Kind = ClassInfo::UserClass0 + Index; 1334 1335 ListInit *Supers = Rec->getValueAsListInit("SuperClasses"); 1336 for (Init *I : Supers->getValues()) { 1337 DefInit *DI = dyn_cast<DefInit>(I); 1338 if (!DI) { 1339 PrintError(Rec->getLoc(), "Invalid super class reference!"); 1340 continue; 1341 } 1342 1343 ClassInfo *SC = AsmOperandClasses[DI->getDef()]; 1344 if (!SC) 1345 PrintError(Rec->getLoc(), "Invalid super class reference!"); 1346 else 1347 CI->SuperClasses.push_back(SC); 1348 } 1349 CI->ClassName = Rec->getValueAsString("Name"); 1350 CI->Name = "MCK_" + CI->ClassName; 1351 CI->ValueName = Rec->getName(); 1352 1353 // Get or construct the predicate method name. 1354 Init *PMName = Rec->getValueInit("PredicateMethod"); 1355 if (StringInit *SI = dyn_cast<StringInit>(PMName)) { 1356 CI->PredicateMethod = SI->getValue(); 1357 } else { 1358 assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!"); 1359 CI->PredicateMethod = "is" + CI->ClassName; 1360 } 1361 1362 // Get or construct the render method name. 1363 Init *RMName = Rec->getValueInit("RenderMethod"); 1364 if (StringInit *SI = dyn_cast<StringInit>(RMName)) { 1365 CI->RenderMethod = SI->getValue(); 1366 } else { 1367 assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!"); 1368 CI->RenderMethod = "add" + CI->ClassName + "Operands"; 1369 } 1370 1371 // Get the parse method name or leave it as empty. 1372 Init *PRMName = Rec->getValueInit("ParserMethod"); 1373 if (StringInit *SI = dyn_cast<StringInit>(PRMName)) 1374 CI->ParserMethod = SI->getValue(); 1375 1376 // Get the diagnostic type or leave it as empty. 1377 // Get the parse method name or leave it as empty. 1378 Init *DiagnosticType = Rec->getValueInit("DiagnosticType"); 1379 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType)) 1380 CI->DiagnosticType = SI->getValue(); 1381 1382 Init *IsOptional = Rec->getValueInit("IsOptional"); 1383 if (BitInit *BI = dyn_cast<BitInit>(IsOptional)) 1384 CI->IsOptional = BI->getValue(); 1385 1386 // Get or construct the default method name. 1387 Init *DMName = Rec->getValueInit("DefaultMethod"); 1388 if (StringInit *SI = dyn_cast<StringInit>(DMName)) { 1389 CI->DefaultMethod = SI->getValue(); 1390 } else { 1391 assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!"); 1392 CI->DefaultMethod = "default" + CI->ClassName + "Operands"; 1393 } 1394 1395 ++Index; 1396 } 1397 } 1398 1399 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, 1400 CodeGenTarget &target, 1401 RecordKeeper &records) 1402 : Records(records), AsmParser(asmParser), Target(target) { 1403 } 1404 1405 /// buildOperandMatchInfo - Build the necessary information to handle user 1406 /// defined operand parsing methods. 1407 void AsmMatcherInfo::buildOperandMatchInfo() { 1408 1409 /// Map containing a mask with all operands indices that can be found for 1410 /// that class inside a instruction. 1411 typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy; 1412 OpClassMaskTy OpClassMask; 1413 1414 for (const auto &MI : Matchables) { 1415 OpClassMask.clear(); 1416 1417 // Keep track of all operands of this instructions which belong to the 1418 // same class. 1419 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) { 1420 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i]; 1421 if (Op.Class->ParserMethod.empty()) 1422 continue; 1423 unsigned &OperandMask = OpClassMask[Op.Class]; 1424 OperandMask |= (1 << i); 1425 } 1426 1427 // Generate operand match info for each mnemonic/operand class pair. 1428 for (const auto &OCM : OpClassMask) { 1429 unsigned OpMask = OCM.second; 1430 ClassInfo *CI = OCM.first; 1431 OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI, 1432 OpMask)); 1433 } 1434 } 1435 } 1436 1437 void AsmMatcherInfo::buildInfo() { 1438 // Build information about all of the AssemblerPredicates. 1439 std::vector<Record*> AllPredicates = 1440 Records.getAllDerivedDefinitions("Predicate"); 1441 for (Record *Pred : AllPredicates) { 1442 // Ignore predicates that are not intended for the assembler. 1443 if (!Pred->getValueAsBit("AssemblerMatcherPredicate")) 1444 continue; 1445 1446 if (Pred->getName().empty()) 1447 PrintFatalError(Pred->getLoc(), "Predicate has no name!"); 1448 1449 SubtargetFeatures.insert(std::make_pair( 1450 Pred, SubtargetFeatureInfo(Pred, SubtargetFeatures.size()))); 1451 DEBUG(SubtargetFeatures.find(Pred)->second.dump()); 1452 assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!"); 1453 } 1454 1455 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst"); 1456 1457 // Parse the instructions; we need to do this first so that we can gather the 1458 // singleton register classes. 1459 SmallPtrSet<Record*, 16> SingletonRegisters; 1460 unsigned VariantCount = Target.getAsmParserVariantCount(); 1461 for (unsigned VC = 0; VC != VariantCount; ++VC) { 1462 Record *AsmVariant = Target.getAsmParserVariant(VC); 1463 std::string CommentDelimiter = 1464 AsmVariant->getValueAsString("CommentDelimiter"); 1465 AsmVariantInfo Variant; 1466 Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix"); 1467 Variant.TokenizingCharacters = 1468 AsmVariant->getValueAsString("TokenizingCharacters"); 1469 Variant.SeparatorCharacters = 1470 AsmVariant->getValueAsString("SeparatorCharacters"); 1471 Variant.BreakCharacters = 1472 AsmVariant->getValueAsString("BreakCharacters"); 1473 Variant.Name = AsmVariant->getValueAsString("Name"); 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 // Ignore instructions for different instructions 1488 const std::string V = CGI->TheDef->getValueAsString("AsmVariantName"); 1489 if (!V.empty() && V != Variant.Name) 1490 continue; 1491 1492 auto II = llvm::make_unique<MatchableInfo>(*CGI); 1493 1494 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst); 1495 1496 // Ignore instructions which shouldn't be matched and diagnose invalid 1497 // instruction definitions with an error. 1498 if (!II->validate(CommentDelimiter, true)) 1499 continue; 1500 1501 Matchables.push_back(std::move(II)); 1502 } 1503 1504 // Parse all of the InstAlias definitions and stick them in the list of 1505 // matchables. 1506 std::vector<Record*> AllInstAliases = 1507 Records.getAllDerivedDefinitions("InstAlias"); 1508 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) { 1509 auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i], 1510 Variant.AsmVariantNo, 1511 Target); 1512 1513 // If the tblgen -match-prefix option is specified (for tblgen hackers), 1514 // filter the set of instruction aliases we consider, based on the target 1515 // instruction. 1516 if (!StringRef(Alias->ResultInst->TheDef->getName()) 1517 .startswith( MatchPrefix)) 1518 continue; 1519 1520 const std::string V = Alias->TheDef->getValueAsString("AsmVariantName"); 1521 if (!V.empty() && V != Variant.Name) 1522 continue; 1523 1524 auto II = llvm::make_unique<MatchableInfo>(std::move(Alias)); 1525 1526 II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst); 1527 1528 // Validate the alias definitions. 1529 II->validate(CommentDelimiter, false); 1530 1531 Matchables.push_back(std::move(II)); 1532 } 1533 } 1534 1535 // Build info for the register classes. 1536 buildRegisterClasses(SingletonRegisters); 1537 1538 // Build info for the user defined assembly operand classes. 1539 buildOperandClasses(); 1540 1541 // Build the information about matchables, now that we have fully formed 1542 // classes. 1543 std::vector<std::unique_ptr<MatchableInfo>> NewMatchables; 1544 for (auto &II : Matchables) { 1545 // Parse the tokens after the mnemonic. 1546 // Note: buildInstructionOperandReference may insert new AsmOperands, so 1547 // don't precompute the loop bound. 1548 for (unsigned i = 0; i != II->AsmOperands.size(); ++i) { 1549 MatchableInfo::AsmOperand &Op = II->AsmOperands[i]; 1550 StringRef Token = Op.Token; 1551 1552 // Check for singleton registers. 1553 if (Record *RegRecord = Op.SingletonReg) { 1554 Op.Class = RegisterClasses[RegRecord]; 1555 assert(Op.Class && Op.Class->Registers.size() == 1 && 1556 "Unexpected class for singleton register"); 1557 continue; 1558 } 1559 1560 // Check for simple tokens. 1561 if (Token[0] != '$') { 1562 Op.Class = getTokenClass(Token); 1563 continue; 1564 } 1565 1566 if (Token.size() > 1 && isdigit(Token[1])) { 1567 Op.Class = getTokenClass(Token); 1568 continue; 1569 } 1570 1571 // Otherwise this is an operand reference. 1572 StringRef OperandName; 1573 if (Token[1] == '{') 1574 OperandName = Token.substr(2, Token.size() - 3); 1575 else 1576 OperandName = Token.substr(1); 1577 1578 if (II->DefRec.is<const CodeGenInstruction*>()) 1579 buildInstructionOperandReference(II.get(), OperandName, i); 1580 else 1581 buildAliasOperandReference(II.get(), OperandName, Op); 1582 } 1583 1584 if (II->DefRec.is<const CodeGenInstruction*>()) { 1585 II->buildInstructionResultOperands(); 1586 // If the instruction has a two-operand alias, build up the 1587 // matchable here. We'll add them in bulk at the end to avoid 1588 // confusing this loop. 1589 std::string Constraint = 1590 II->TheDef->getValueAsString("TwoOperandAliasConstraint"); 1591 if (Constraint != "") { 1592 // Start by making a copy of the original matchable. 1593 auto AliasII = llvm::make_unique<MatchableInfo>(*II); 1594 1595 // Adjust it to be a two-operand alias. 1596 AliasII->formTwoOperandAlias(Constraint); 1597 1598 // Add the alias to the matchables list. 1599 NewMatchables.push_back(std::move(AliasII)); 1600 } 1601 } else 1602 II->buildAliasResultOperands(); 1603 } 1604 if (!NewMatchables.empty()) 1605 Matchables.insert(Matchables.end(), 1606 std::make_move_iterator(NewMatchables.begin()), 1607 std::make_move_iterator(NewMatchables.end())); 1608 1609 // Process token alias definitions and set up the associated superclass 1610 // information. 1611 std::vector<Record*> AllTokenAliases = 1612 Records.getAllDerivedDefinitions("TokenAlias"); 1613 for (Record *Rec : AllTokenAliases) { 1614 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken")); 1615 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken")); 1616 if (FromClass == ToClass) 1617 PrintFatalError(Rec->getLoc(), 1618 "error: Destination value identical to source value."); 1619 FromClass->SuperClasses.push_back(ToClass); 1620 } 1621 1622 // Reorder classes so that classes precede super classes. 1623 Classes.sort(); 1624 1625 #ifndef NDEBUG 1626 // Verify that the table is now sorted 1627 for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) { 1628 for (auto J = I; J != E; ++J) { 1629 assert(!(*J < *I)); 1630 assert(I == J || !J->isSubsetOf(*I)); 1631 } 1632 } 1633 #endif // NDEBUG 1634 } 1635 1636 /// buildInstructionOperandReference - The specified operand is a reference to a 1637 /// named operand such as $src. Resolve the Class and OperandInfo pointers. 1638 void AsmMatcherInfo:: 1639 buildInstructionOperandReference(MatchableInfo *II, 1640 StringRef OperandName, 1641 unsigned AsmOpIdx) { 1642 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>(); 1643 const CGIOperandList &Operands = CGI.Operands; 1644 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx]; 1645 1646 // Map this token to an operand. 1647 unsigned Idx; 1648 if (!Operands.hasOperandNamed(OperandName, Idx)) 1649 PrintFatalError(II->TheDef->getLoc(), 1650 "error: unable to find operand: '" + OperandName + "'"); 1651 1652 // If the instruction operand has multiple suboperands, but the parser 1653 // match class for the asm operand is still the default "ImmAsmOperand", 1654 // then handle each suboperand separately. 1655 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) { 1656 Record *Rec = Operands[Idx].Rec; 1657 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!"); 1658 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1659 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") { 1660 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands. 1661 StringRef Token = Op->Token; // save this in case Op gets moved 1662 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) { 1663 MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token); 1664 NewAsmOp.SubOpIdx = SI; 1665 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp); 1666 } 1667 // Replace Op with first suboperand. 1668 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved 1669 Op->SubOpIdx = 0; 1670 } 1671 } 1672 1673 // Set up the operand class. 1674 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx); 1675 1676 // If the named operand is tied, canonicalize it to the untied operand. 1677 // For example, something like: 1678 // (outs GPR:$dst), (ins GPR:$src) 1679 // with an asmstring of 1680 // "inc $src" 1681 // we want to canonicalize to: 1682 // "inc $dst" 1683 // so that we know how to provide the $dst operand when filling in the result. 1684 int OITied = -1; 1685 if (Operands[Idx].MINumOperands == 1) 1686 OITied = Operands[Idx].getTiedRegister(); 1687 if (OITied != -1) { 1688 // The tied operand index is an MIOperand index, find the operand that 1689 // contains it. 1690 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied); 1691 OperandName = Operands[Idx.first].Name; 1692 Op->SubOpIdx = Idx.second; 1693 } 1694 1695 Op->SrcOpName = OperandName; 1696 } 1697 1698 /// buildAliasOperandReference - When parsing an operand reference out of the 1699 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the 1700 /// operand reference is by looking it up in the result pattern definition. 1701 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II, 1702 StringRef OperandName, 1703 MatchableInfo::AsmOperand &Op) { 1704 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>(); 1705 1706 // Set up the operand class. 1707 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i) 1708 if (CGA.ResultOperands[i].isRecord() && 1709 CGA.ResultOperands[i].getName() == OperandName) { 1710 // It's safe to go with the first one we find, because CodeGenInstAlias 1711 // validates that all operands with the same name have the same record. 1712 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second; 1713 // Use the match class from the Alias definition, not the 1714 // destination instruction, as we may have an immediate that's 1715 // being munged by the match class. 1716 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(), 1717 Op.SubOpIdx); 1718 Op.SrcOpName = OperandName; 1719 return; 1720 } 1721 1722 PrintFatalError(II->TheDef->getLoc(), 1723 "error: unable to find operand: '" + OperandName + "'"); 1724 } 1725 1726 void MatchableInfo::buildInstructionResultOperands() { 1727 const CodeGenInstruction *ResultInst = getResultInst(); 1728 1729 // Loop over all operands of the result instruction, determining how to 1730 // populate them. 1731 for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) { 1732 // If this is a tied operand, just copy from the previously handled operand. 1733 int TiedOp = -1; 1734 if (OpInfo.MINumOperands == 1) 1735 TiedOp = OpInfo.getTiedRegister(); 1736 if (TiedOp != -1) { 1737 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1738 continue; 1739 } 1740 1741 // Find out what operand from the asmparser this MCInst operand comes from. 1742 int SrcOperand = findAsmOperandNamed(OpInfo.Name); 1743 if (OpInfo.Name.empty() || SrcOperand == -1) { 1744 // This may happen for operands that are tied to a suboperand of a 1745 // complex operand. Simply use a dummy value here; nobody should 1746 // use this operand slot. 1747 // FIXME: The long term goal is for the MCOperand list to not contain 1748 // tied operands at all. 1749 ResOperands.push_back(ResOperand::getImmOp(0)); 1750 continue; 1751 } 1752 1753 // Check if the one AsmOperand populates the entire operand. 1754 unsigned NumOperands = OpInfo.MINumOperands; 1755 if (AsmOperands[SrcOperand].SubOpIdx == -1) { 1756 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands)); 1757 continue; 1758 } 1759 1760 // Add a separate ResOperand for each suboperand. 1761 for (unsigned AI = 0; AI < NumOperands; ++AI) { 1762 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI && 1763 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name && 1764 "unexpected AsmOperands for suboperands"); 1765 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1)); 1766 } 1767 } 1768 } 1769 1770 void MatchableInfo::buildAliasResultOperands() { 1771 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>(); 1772 const CodeGenInstruction *ResultInst = getResultInst(); 1773 1774 // Loop over all operands of the result instruction, determining how to 1775 // populate them. 1776 unsigned AliasOpNo = 0; 1777 unsigned LastOpNo = CGA.ResultInstOperandIndex.size(); 1778 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) { 1779 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i]; 1780 1781 // If this is a tied operand, just copy from the previously handled operand. 1782 int TiedOp = -1; 1783 if (OpInfo->MINumOperands == 1) 1784 TiedOp = OpInfo->getTiedRegister(); 1785 if (TiedOp != -1) { 1786 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1787 continue; 1788 } 1789 1790 // Handle all the suboperands for this operand. 1791 const std::string &OpName = OpInfo->Name; 1792 for ( ; AliasOpNo < LastOpNo && 1793 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) { 1794 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second; 1795 1796 // Find out what operand from the asmparser that this MCInst operand 1797 // comes from. 1798 switch (CGA.ResultOperands[AliasOpNo].Kind) { 1799 case CodeGenInstAlias::ResultOperand::K_Record: { 1800 StringRef Name = CGA.ResultOperands[AliasOpNo].getName(); 1801 int SrcOperand = findAsmOperand(Name, SubIdx); 1802 if (SrcOperand == -1) 1803 PrintFatalError(TheDef->getLoc(), "Instruction '" + 1804 TheDef->getName() + "' has operand '" + OpName + 1805 "' that doesn't appear in asm string!"); 1806 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1); 1807 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, 1808 NumOperands)); 1809 break; 1810 } 1811 case CodeGenInstAlias::ResultOperand::K_Imm: { 1812 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm(); 1813 ResOperands.push_back(ResOperand::getImmOp(ImmVal)); 1814 break; 1815 } 1816 case CodeGenInstAlias::ResultOperand::K_Reg: { 1817 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister(); 1818 ResOperands.push_back(ResOperand::getRegOp(Reg)); 1819 break; 1820 } 1821 } 1822 } 1823 } 1824 } 1825 1826 static unsigned 1827 getConverterOperandID(const std::string &Name, 1828 SmallSetVector<CachedHashString, 16> &Table, 1829 bool &IsNew) { 1830 IsNew = Table.insert(CachedHashString(Name)); 1831 1832 unsigned ID = IsNew ? Table.size() - 1 : find(Table, Name) - Table.begin(); 1833 1834 assert(ID < Table.size()); 1835 1836 return ID; 1837 } 1838 1839 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName, 1840 std::vector<std::unique_ptr<MatchableInfo>> &Infos, 1841 bool HasMnemonicFirst, bool HasOptionalOperands, 1842 raw_ostream &OS) { 1843 SmallSetVector<CachedHashString, 16> OperandConversionKinds; 1844 SmallSetVector<CachedHashString, 16> InstructionConversionKinds; 1845 std::vector<std::vector<uint8_t> > ConversionTable; 1846 size_t MaxRowLength = 2; // minimum is custom converter plus terminator. 1847 1848 // TargetOperandClass - This is the target's operand class, like X86Operand. 1849 std::string TargetOperandClass = Target.getName() + "Operand"; 1850 1851 // Write the convert function to a separate stream, so we can drop it after 1852 // the enum. We'll build up the conversion handlers for the individual 1853 // operand types opportunistically as we encounter them. 1854 std::string ConvertFnBody; 1855 raw_string_ostream CvtOS(ConvertFnBody); 1856 // Start the unified conversion function. 1857 if (HasOptionalOperands) { 1858 CvtOS << "void " << Target.getName() << ClassName << "::\n" 1859 << "convertToMCInst(unsigned Kind, MCInst &Inst, " 1860 << "unsigned Opcode,\n" 1861 << " const OperandVector &Operands,\n" 1862 << " const SmallBitVector &OptionalOperandsMask) {\n"; 1863 } else { 1864 CvtOS << "void " << Target.getName() << ClassName << "::\n" 1865 << "convertToMCInst(unsigned Kind, MCInst &Inst, " 1866 << "unsigned Opcode,\n" 1867 << " const OperandVector &Operands) {\n"; 1868 } 1869 CvtOS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"; 1870 CvtOS << " const uint8_t *Converter = ConversionTable[Kind];\n"; 1871 if (HasOptionalOperands) { 1872 CvtOS << " unsigned NumDefaults = 0;\n"; 1873 } 1874 CvtOS << " unsigned OpIdx;\n"; 1875 CvtOS << " Inst.setOpcode(Opcode);\n"; 1876 CvtOS << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"; 1877 if (HasOptionalOperands) { 1878 CvtOS << " OpIdx = *(p + 1) - NumDefaults;\n"; 1879 } else { 1880 CvtOS << " OpIdx = *(p + 1);\n"; 1881 } 1882 CvtOS << " switch (*p) {\n"; 1883 CvtOS << " default: llvm_unreachable(\"invalid conversion entry!\");\n"; 1884 CvtOS << " case CVT_Reg:\n"; 1885 CvtOS << " static_cast<" << TargetOperandClass 1886 << "&>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n"; 1887 CvtOS << " break;\n"; 1888 CvtOS << " case CVT_Tied:\n"; 1889 CvtOS << " Inst.addOperand(Inst.getOperand(OpIdx));\n"; 1890 CvtOS << " break;\n"; 1891 1892 std::string OperandFnBody; 1893 raw_string_ostream OpOS(OperandFnBody); 1894 // Start the operand number lookup function. 1895 OpOS << "void " << Target.getName() << ClassName << "::\n" 1896 << "convertToMapAndConstraints(unsigned Kind,\n"; 1897 OpOS.indent(27); 1898 OpOS << "const OperandVector &Operands) {\n" 1899 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n" 1900 << " unsigned NumMCOperands = 0;\n" 1901 << " const uint8_t *Converter = ConversionTable[Kind];\n" 1902 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n" 1903 << " switch (*p) {\n" 1904 << " default: llvm_unreachable(\"invalid conversion entry!\");\n" 1905 << " case CVT_Reg:\n" 1906 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 1907 << " Operands[*(p + 1)]->setConstraint(\"r\");\n" 1908 << " ++NumMCOperands;\n" 1909 << " break;\n" 1910 << " case CVT_Tied:\n" 1911 << " ++NumMCOperands;\n" 1912 << " break;\n"; 1913 1914 // Pre-populate the operand conversion kinds with the standard always 1915 // available entries. 1916 OperandConversionKinds.insert(CachedHashString("CVT_Done")); 1917 OperandConversionKinds.insert(CachedHashString("CVT_Reg")); 1918 OperandConversionKinds.insert(CachedHashString("CVT_Tied")); 1919 enum { CVT_Done, CVT_Reg, CVT_Tied }; 1920 1921 for (auto &II : Infos) { 1922 // Check if we have a custom match function. 1923 std::string AsmMatchConverter = 1924 II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter"); 1925 if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) { 1926 std::string Signature = "ConvertCustom_" + AsmMatchConverter; 1927 II->ConversionFnKind = Signature; 1928 1929 // Check if we have already generated this signature. 1930 if (!InstructionConversionKinds.insert(CachedHashString(Signature))) 1931 continue; 1932 1933 // Remember this converter for the kind enum. 1934 unsigned KindID = OperandConversionKinds.size(); 1935 OperandConversionKinds.insert( 1936 CachedHashString("CVT_" + getEnumNameForToken(AsmMatchConverter))); 1937 1938 // Add the converter row for this instruction. 1939 ConversionTable.emplace_back(); 1940 ConversionTable.back().push_back(KindID); 1941 ConversionTable.back().push_back(CVT_Done); 1942 1943 // Add the handler to the conversion driver function. 1944 CvtOS << " case CVT_" 1945 << getEnumNameForToken(AsmMatchConverter) << ":\n" 1946 << " " << AsmMatchConverter << "(Inst, Operands);\n" 1947 << " break;\n"; 1948 1949 // FIXME: Handle the operand number lookup for custom match functions. 1950 continue; 1951 } 1952 1953 // Build the conversion function signature. 1954 std::string Signature = "Convert"; 1955 1956 std::vector<uint8_t> ConversionRow; 1957 1958 // Compute the convert enum and the case body. 1959 MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 ); 1960 1961 for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) { 1962 const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i]; 1963 1964 // Generate code to populate each result operand. 1965 switch (OpInfo.Kind) { 1966 case MatchableInfo::ResOperand::RenderAsmOperand: { 1967 // This comes from something we parsed. 1968 const MatchableInfo::AsmOperand &Op = 1969 II->AsmOperands[OpInfo.AsmOperandNum]; 1970 1971 // Registers are always converted the same, don't duplicate the 1972 // conversion function based on them. 1973 Signature += "__"; 1974 std::string Class; 1975 Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName; 1976 Signature += Class; 1977 Signature += utostr(OpInfo.MINumOperands); 1978 Signature += "_" + itostr(OpInfo.AsmOperandNum); 1979 1980 // Add the conversion kind, if necessary, and get the associated ID 1981 // the index of its entry in the vector). 1982 std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" : 1983 Op.Class->RenderMethod); 1984 if (Op.Class->IsOptional) { 1985 // For optional operands we must also care about DefaultMethod 1986 assert(HasOptionalOperands); 1987 Name += "_" + Op.Class->DefaultMethod; 1988 } 1989 Name = getEnumNameForToken(Name); 1990 1991 bool IsNewConverter = false; 1992 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 1993 IsNewConverter); 1994 1995 // Add the operand entry to the instruction kind conversion row. 1996 ConversionRow.push_back(ID); 1997 ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst); 1998 1999 if (!IsNewConverter) 2000 break; 2001 2002 // This is a new operand kind. Add a handler for it to the 2003 // converter driver. 2004 CvtOS << " case " << Name << ":\n"; 2005 if (Op.Class->IsOptional) { 2006 // If optional operand is not present in actual instruction then we 2007 // should call its DefaultMethod before RenderMethod 2008 assert(HasOptionalOperands); 2009 CvtOS << " if (OptionalOperandsMask[*(p + 1) - 1]) {\n" 2010 << " " << Op.Class->DefaultMethod << "()" 2011 << "->" << Op.Class->RenderMethod << "(Inst, " 2012 << OpInfo.MINumOperands << ");\n" 2013 << " ++NumDefaults;\n" 2014 << " } else {\n" 2015 << " static_cast<" << TargetOperandClass 2016 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod 2017 << "(Inst, " << OpInfo.MINumOperands << ");\n" 2018 << " }\n"; 2019 } else { 2020 CvtOS << " static_cast<" << TargetOperandClass 2021 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod 2022 << "(Inst, " << OpInfo.MINumOperands << ");\n"; 2023 } 2024 CvtOS << " break;\n"; 2025 2026 // Add a handler for the operand number lookup. 2027 OpOS << " case " << Name << ":\n" 2028 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"; 2029 2030 if (Op.Class->isRegisterClass()) 2031 OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n"; 2032 else 2033 OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n"; 2034 OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n" 2035 << " break;\n"; 2036 break; 2037 } 2038 case MatchableInfo::ResOperand::TiedOperand: { 2039 // If this operand is tied to a previous one, just copy the MCInst 2040 // operand from the earlier one.We can only tie single MCOperand values. 2041 assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand"); 2042 unsigned TiedOp = OpInfo.TiedOperandNum; 2043 assert(i > TiedOp && "Tied operand precedes its target!"); 2044 Signature += "__Tie" + utostr(TiedOp); 2045 ConversionRow.push_back(CVT_Tied); 2046 ConversionRow.push_back(TiedOp); 2047 break; 2048 } 2049 case MatchableInfo::ResOperand::ImmOperand: { 2050 int64_t Val = OpInfo.ImmVal; 2051 std::string Ty = "imm_" + itostr(Val); 2052 Ty = getEnumNameForToken(Ty); 2053 Signature += "__" + Ty; 2054 2055 std::string Name = "CVT_" + Ty; 2056 bool IsNewConverter = false; 2057 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 2058 IsNewConverter); 2059 // Add the operand entry to the instruction kind conversion row. 2060 ConversionRow.push_back(ID); 2061 ConversionRow.push_back(0); 2062 2063 if (!IsNewConverter) 2064 break; 2065 2066 CvtOS << " case " << Name << ":\n" 2067 << " Inst.addOperand(MCOperand::createImm(" << Val << "));\n" 2068 << " break;\n"; 2069 2070 OpOS << " case " << Name << ":\n" 2071 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 2072 << " Operands[*(p + 1)]->setConstraint(\"\");\n" 2073 << " ++NumMCOperands;\n" 2074 << " break;\n"; 2075 break; 2076 } 2077 case MatchableInfo::ResOperand::RegOperand: { 2078 std::string Reg, Name; 2079 if (!OpInfo.Register) { 2080 Name = "reg0"; 2081 Reg = "0"; 2082 } else { 2083 Reg = getQualifiedName(OpInfo.Register); 2084 Name = "reg" + OpInfo.Register->getName(); 2085 } 2086 Signature += "__" + Name; 2087 Name = "CVT_" + Name; 2088 bool IsNewConverter = false; 2089 unsigned ID = getConverterOperandID(Name, OperandConversionKinds, 2090 IsNewConverter); 2091 // Add the operand entry to the instruction kind conversion row. 2092 ConversionRow.push_back(ID); 2093 ConversionRow.push_back(0); 2094 2095 if (!IsNewConverter) 2096 break; 2097 CvtOS << " case " << Name << ":\n" 2098 << " Inst.addOperand(MCOperand::createReg(" << Reg << "));\n" 2099 << " break;\n"; 2100 2101 OpOS << " case " << Name << ":\n" 2102 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" 2103 << " Operands[*(p + 1)]->setConstraint(\"m\");\n" 2104 << " ++NumMCOperands;\n" 2105 << " break;\n"; 2106 } 2107 } 2108 } 2109 2110 // If there were no operands, add to the signature to that effect 2111 if (Signature == "Convert") 2112 Signature += "_NoOperands"; 2113 2114 II->ConversionFnKind = Signature; 2115 2116 // Save the signature. If we already have it, don't add a new row 2117 // to the table. 2118 if (!InstructionConversionKinds.insert(CachedHashString(Signature))) 2119 continue; 2120 2121 // Add the row to the table. 2122 ConversionTable.push_back(std::move(ConversionRow)); 2123 } 2124 2125 // Finish up the converter driver function. 2126 CvtOS << " }\n }\n}\n\n"; 2127 2128 // Finish up the operand number lookup function. 2129 OpOS << " }\n }\n}\n\n"; 2130 2131 OS << "namespace {\n"; 2132 2133 // Output the operand conversion kind enum. 2134 OS << "enum OperatorConversionKind {\n"; 2135 for (const auto &Converter : OperandConversionKinds) 2136 OS << " " << Converter << ",\n"; 2137 OS << " CVT_NUM_CONVERTERS\n"; 2138 OS << "};\n\n"; 2139 2140 // Output the instruction conversion kind enum. 2141 OS << "enum InstructionConversionKind {\n"; 2142 for (const auto &Signature : InstructionConversionKinds) 2143 OS << " " << Signature << ",\n"; 2144 OS << " CVT_NUM_SIGNATURES\n"; 2145 OS << "};\n\n"; 2146 2147 OS << "} // end anonymous namespace\n\n"; 2148 2149 // Output the conversion table. 2150 OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES][" 2151 << MaxRowLength << "] = {\n"; 2152 2153 for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) { 2154 assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!"); 2155 OS << " // " << InstructionConversionKinds[Row] << "\n"; 2156 OS << " { "; 2157 for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2) 2158 OS << OperandConversionKinds[ConversionTable[Row][i]] << ", " 2159 << (unsigned)(ConversionTable[Row][i + 1]) << ", "; 2160 OS << "CVT_Done },\n"; 2161 } 2162 2163 OS << "};\n\n"; 2164 2165 // Spit out the conversion driver function. 2166 OS << CvtOS.str(); 2167 2168 // Spit out the operand number lookup function. 2169 OS << OpOS.str(); 2170 } 2171 2172 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds. 2173 static void emitMatchClassEnumeration(CodeGenTarget &Target, 2174 std::forward_list<ClassInfo> &Infos, 2175 raw_ostream &OS) { 2176 OS << "namespace {\n\n"; 2177 2178 OS << "/// MatchClassKind - The kinds of classes which participate in\n" 2179 << "/// instruction matching.\n"; 2180 OS << "enum MatchClassKind {\n"; 2181 OS << " InvalidMatchClass = 0,\n"; 2182 OS << " OptionalMatchClass = 1,\n"; 2183 for (const auto &CI : Infos) { 2184 OS << " " << CI.Name << ", // "; 2185 if (CI.Kind == ClassInfo::Token) { 2186 OS << "'" << CI.ValueName << "'\n"; 2187 } else if (CI.isRegisterClass()) { 2188 if (!CI.ValueName.empty()) 2189 OS << "register class '" << CI.ValueName << "'\n"; 2190 else 2191 OS << "derived register class\n"; 2192 } else { 2193 OS << "user defined class '" << CI.ValueName << "'\n"; 2194 } 2195 } 2196 OS << " NumMatchClassKinds\n"; 2197 OS << "};\n\n"; 2198 2199 OS << "}\n\n"; 2200 } 2201 2202 /// emitValidateOperandClass - Emit the function to validate an operand class. 2203 static void emitValidateOperandClass(AsmMatcherInfo &Info, 2204 raw_ostream &OS) { 2205 OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, " 2206 << "MatchClassKind Kind) {\n"; 2207 OS << " " << Info.Target.getName() << "Operand &Operand = (" 2208 << Info.Target.getName() << "Operand&)GOp;\n"; 2209 2210 // The InvalidMatchClass is not to match any operand. 2211 OS << " if (Kind == InvalidMatchClass)\n"; 2212 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n"; 2213 2214 // Check for Token operands first. 2215 // FIXME: Use a more specific diagnostic type. 2216 OS << " if (Operand.isToken())\n"; 2217 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n" 2218 << " MCTargetAsmParser::Match_Success :\n" 2219 << " MCTargetAsmParser::Match_InvalidOperand;\n\n"; 2220 2221 // Check the user classes. We don't care what order since we're only 2222 // actually matching against one of them. 2223 OS << " switch (Kind) {\n" 2224 " default: break;\n"; 2225 for (const auto &CI : Info.Classes) { 2226 if (!CI.isUserClass()) 2227 continue; 2228 2229 OS << " // '" << CI.ClassName << "' class\n"; 2230 OS << " case " << CI.Name << ":\n"; 2231 OS << " if (Operand." << CI.PredicateMethod << "())\n"; 2232 OS << " return MCTargetAsmParser::Match_Success;\n"; 2233 if (!CI.DiagnosticType.empty()) 2234 OS << " return " << Info.Target.getName() << "AsmParser::Match_" 2235 << CI.DiagnosticType << ";\n"; 2236 else 2237 OS << " break;\n"; 2238 } 2239 OS << " } // end switch (Kind)\n\n"; 2240 2241 // Check for register operands, including sub-classes. 2242 OS << " if (Operand.isReg()) {\n"; 2243 OS << " MatchClassKind OpKind;\n"; 2244 OS << " switch (Operand.getReg()) {\n"; 2245 OS << " default: OpKind = InvalidMatchClass; break;\n"; 2246 for (const auto &RC : Info.RegisterClasses) 2247 OS << " case " << Info.Target.getName() << "::" 2248 << RC.first->getName() << ": OpKind = " << RC.second->Name 2249 << "; break;\n"; 2250 OS << " }\n"; 2251 OS << " return isSubclass(OpKind, Kind) ? " 2252 << "MCTargetAsmParser::Match_Success :\n " 2253 << " MCTargetAsmParser::Match_InvalidOperand;\n }\n\n"; 2254 2255 // Generic fallthrough match failure case for operands that don't have 2256 // specialized diagnostic types. 2257 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n"; 2258 OS << "}\n\n"; 2259 } 2260 2261 /// emitIsSubclass - Emit the subclass predicate function. 2262 static void emitIsSubclass(CodeGenTarget &Target, 2263 std::forward_list<ClassInfo> &Infos, 2264 raw_ostream &OS) { 2265 OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n"; 2266 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n"; 2267 OS << " if (A == B)\n"; 2268 OS << " return true;\n\n"; 2269 2270 bool EmittedSwitch = false; 2271 for (const auto &A : Infos) { 2272 std::vector<StringRef> SuperClasses; 2273 if (A.IsOptional) 2274 SuperClasses.push_back("OptionalMatchClass"); 2275 for (const auto &B : Infos) { 2276 if (&A != &B && A.isSubsetOf(B)) 2277 SuperClasses.push_back(B.Name); 2278 } 2279 2280 if (SuperClasses.empty()) 2281 continue; 2282 2283 // If this is the first SuperClass, emit the switch header. 2284 if (!EmittedSwitch) { 2285 OS << " switch (A) {\n"; 2286 OS << " default:\n"; 2287 OS << " return false;\n"; 2288 EmittedSwitch = true; 2289 } 2290 2291 OS << "\n case " << A.Name << ":\n"; 2292 2293 if (SuperClasses.size() == 1) { 2294 OS << " return B == " << SuperClasses.back() << ";\n"; 2295 continue; 2296 } 2297 2298 if (!SuperClasses.empty()) { 2299 OS << " switch (B) {\n"; 2300 OS << " default: return false;\n"; 2301 for (StringRef SC : SuperClasses) 2302 OS << " case " << SC << ": return true;\n"; 2303 OS << " }\n"; 2304 } else { 2305 // No case statement to emit 2306 OS << " return false;\n"; 2307 } 2308 } 2309 2310 // If there were case statements emitted into the string stream write the 2311 // default. 2312 if (EmittedSwitch) 2313 OS << " }\n"; 2314 else 2315 OS << " return false;\n"; 2316 2317 OS << "}\n\n"; 2318 } 2319 2320 /// emitMatchTokenString - Emit the function to match a token string to the 2321 /// appropriate match class value. 2322 static void emitMatchTokenString(CodeGenTarget &Target, 2323 std::forward_list<ClassInfo> &Infos, 2324 raw_ostream &OS) { 2325 // Construct the match list. 2326 std::vector<StringMatcher::StringPair> Matches; 2327 for (const auto &CI : Infos) { 2328 if (CI.Kind == ClassInfo::Token) 2329 Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";"); 2330 } 2331 2332 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n"; 2333 2334 StringMatcher("Name", Matches, OS).Emit(); 2335 2336 OS << " return InvalidMatchClass;\n"; 2337 OS << "}\n\n"; 2338 } 2339 2340 /// emitMatchRegisterName - Emit the function to match a string to the target 2341 /// specific register enum. 2342 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser, 2343 raw_ostream &OS) { 2344 // Construct the match list. 2345 std::vector<StringMatcher::StringPair> Matches; 2346 const auto &Regs = Target.getRegBank().getRegisters(); 2347 for (const CodeGenRegister &Reg : Regs) { 2348 if (Reg.TheDef->getValueAsString("AsmName").empty()) 2349 continue; 2350 2351 Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"), 2352 "return " + utostr(Reg.EnumValue) + ";"); 2353 } 2354 2355 OS << "static unsigned MatchRegisterName(StringRef Name) {\n"; 2356 2357 StringMatcher("Name", Matches, OS).Emit(); 2358 2359 OS << " return 0;\n"; 2360 OS << "}\n\n"; 2361 } 2362 2363 /// Emit the function to match a string to the target 2364 /// specific register enum. 2365 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser, 2366 raw_ostream &OS) { 2367 // Construct the match list. 2368 std::vector<StringMatcher::StringPair> Matches; 2369 const auto &Regs = Target.getRegBank().getRegisters(); 2370 for (const CodeGenRegister &Reg : Regs) { 2371 2372 auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames"); 2373 2374 for (auto AltName : AltNames) { 2375 AltName = StringRef(AltName).trim(); 2376 2377 // don't handle empty alternative names 2378 if (AltName.empty()) 2379 continue; 2380 2381 Matches.emplace_back(AltName, 2382 "return " + utostr(Reg.EnumValue) + ";"); 2383 } 2384 } 2385 2386 OS << "static unsigned MatchRegisterAltName(StringRef Name) {\n"; 2387 2388 StringMatcher("Name", Matches, OS).Emit(); 2389 2390 OS << " return 0;\n"; 2391 OS << "}\n\n"; 2392 } 2393 2394 static const char *getMinimalTypeForRange(uint64_t Range) { 2395 assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large"); 2396 if (Range > 0xFFFFFFFFULL) 2397 return "uint64_t"; 2398 if (Range > 0xFFFF) 2399 return "uint32_t"; 2400 if (Range > 0xFF) 2401 return "uint16_t"; 2402 return "uint8_t"; 2403 } 2404 2405 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) { 2406 uint64_t MaxIndex = Info.SubtargetFeatures.size(); 2407 if (MaxIndex > 0) 2408 MaxIndex--; 2409 return getMinimalTypeForRange(1ULL << MaxIndex); 2410 } 2411 2412 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag 2413 /// definitions. 2414 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info, 2415 raw_ostream &OS) { 2416 OS << "// Flags for subtarget features that participate in " 2417 << "instruction matching.\n"; 2418 OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info) 2419 << " {\n"; 2420 for (const auto &SF : Info.SubtargetFeatures) { 2421 const SubtargetFeatureInfo &SFI = SF.second; 2422 OS << " " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n"; 2423 } 2424 OS << " Feature_None = 0\n"; 2425 OS << "};\n\n"; 2426 } 2427 2428 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types. 2429 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) { 2430 // Get the set of diagnostic types from all of the operand classes. 2431 std::set<StringRef> Types; 2432 for (const auto &OpClassEntry : Info.AsmOperandClasses) { 2433 if (!OpClassEntry.second->DiagnosticType.empty()) 2434 Types.insert(OpClassEntry.second->DiagnosticType); 2435 } 2436 2437 if (Types.empty()) return; 2438 2439 // Now emit the enum entries. 2440 for (StringRef Type : Types) 2441 OS << " Match_" << Type << ",\n"; 2442 OS << " END_OPERAND_DIAGNOSTIC_TYPES\n"; 2443 } 2444 2445 /// emitGetSubtargetFeatureName - Emit the helper function to get the 2446 /// user-level name for a subtarget feature. 2447 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) { 2448 OS << "// User-level names for subtarget features that participate in\n" 2449 << "// instruction matching.\n" 2450 << "static const char *getSubtargetFeatureName(uint64_t Val) {\n"; 2451 if (!Info.SubtargetFeatures.empty()) { 2452 OS << " switch(Val) {\n"; 2453 for (const auto &SF : Info.SubtargetFeatures) { 2454 const SubtargetFeatureInfo &SFI = SF.second; 2455 // FIXME: Totally just a placeholder name to get the algorithm working. 2456 OS << " case " << SFI.getEnumName() << ": return \"" 2457 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n"; 2458 } 2459 OS << " default: return \"(unknown)\";\n"; 2460 OS << " }\n"; 2461 } else { 2462 // Nothing to emit, so skip the switch 2463 OS << " return \"(unknown)\";\n"; 2464 } 2465 OS << "}\n\n"; 2466 } 2467 2468 /// emitComputeAvailableFeatures - Emit the function to compute the list of 2469 /// available features given a subtarget. 2470 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info, 2471 raw_ostream &OS) { 2472 std::string ClassName = 2473 Info.AsmParser->getValueAsString("AsmParserClassName"); 2474 2475 OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n" 2476 << "ComputeAvailableFeatures(const FeatureBitset& FB) const {\n"; 2477 OS << " uint64_t Features = 0;\n"; 2478 for (const auto &SF : Info.SubtargetFeatures) { 2479 const SubtargetFeatureInfo &SFI = SF.second; 2480 2481 OS << " if ("; 2482 std::string CondStorage = 2483 SFI.TheDef->getValueAsString("AssemblerCondString"); 2484 StringRef Conds = CondStorage; 2485 std::pair<StringRef,StringRef> Comma = Conds.split(','); 2486 bool First = true; 2487 do { 2488 if (!First) 2489 OS << " && "; 2490 2491 bool Neg = false; 2492 StringRef Cond = Comma.first; 2493 if (Cond[0] == '!') { 2494 Neg = true; 2495 Cond = Cond.substr(1); 2496 } 2497 2498 OS << "("; 2499 if (Neg) 2500 OS << "!"; 2501 OS << "FB[" << Info.Target.getName() << "::" << Cond << "])"; 2502 2503 if (Comma.second.empty()) 2504 break; 2505 2506 First = false; 2507 Comma = Comma.second.split(','); 2508 } while (true); 2509 2510 OS << ")\n"; 2511 OS << " Features |= " << SFI.getEnumName() << ";\n"; 2512 } 2513 OS << " return Features;\n"; 2514 OS << "}\n\n"; 2515 } 2516 2517 static std::string GetAliasRequiredFeatures(Record *R, 2518 const AsmMatcherInfo &Info) { 2519 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates"); 2520 std::string Result; 2521 unsigned NumFeatures = 0; 2522 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) { 2523 const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]); 2524 2525 if (!F) 2526 PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() + 2527 "' is not marked as an AssemblerPredicate!"); 2528 2529 if (NumFeatures) 2530 Result += '|'; 2531 2532 Result += F->getEnumName(); 2533 ++NumFeatures; 2534 } 2535 2536 if (NumFeatures > 1) 2537 Result = '(' + Result + ')'; 2538 return Result; 2539 } 2540 2541 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info, 2542 std::vector<Record*> &Aliases, 2543 unsigned Indent = 0, 2544 StringRef AsmParserVariantName = StringRef()){ 2545 // Keep track of all the aliases from a mnemonic. Use an std::map so that the 2546 // iteration order of the map is stable. 2547 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic; 2548 2549 for (Record *R : Aliases) { 2550 // FIXME: Allow AssemblerVariantName to be a comma separated list. 2551 std::string AsmVariantName = R->getValueAsString("AsmVariantName"); 2552 if (AsmVariantName != AsmParserVariantName) 2553 continue; 2554 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R); 2555 } 2556 if (AliasesFromMnemonic.empty()) 2557 return; 2558 2559 // Process each alias a "from" mnemonic at a time, building the code executed 2560 // by the string remapper. 2561 std::vector<StringMatcher::StringPair> Cases; 2562 for (const auto &AliasEntry : AliasesFromMnemonic) { 2563 const std::vector<Record*> &ToVec = AliasEntry.second; 2564 2565 // Loop through each alias and emit code that handles each case. If there 2566 // are two instructions without predicates, emit an error. If there is one, 2567 // emit it last. 2568 std::string MatchCode; 2569 int AliasWithNoPredicate = -1; 2570 2571 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) { 2572 Record *R = ToVec[i]; 2573 std::string FeatureMask = GetAliasRequiredFeatures(R, Info); 2574 2575 // If this unconditionally matches, remember it for later and diagnose 2576 // duplicates. 2577 if (FeatureMask.empty()) { 2578 if (AliasWithNoPredicate != -1) { 2579 // We can't have two aliases from the same mnemonic with no predicate. 2580 PrintError(ToVec[AliasWithNoPredicate]->getLoc(), 2581 "two MnemonicAliases with the same 'from' mnemonic!"); 2582 PrintFatalError(R->getLoc(), "this is the other MnemonicAlias."); 2583 } 2584 2585 AliasWithNoPredicate = i; 2586 continue; 2587 } 2588 if (R->getValueAsString("ToMnemonic") == AliasEntry.first) 2589 PrintFatalError(R->getLoc(), "MnemonicAlias to the same string"); 2590 2591 if (!MatchCode.empty()) 2592 MatchCode += "else "; 2593 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n"; 2594 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n"; 2595 } 2596 2597 if (AliasWithNoPredicate != -1) { 2598 Record *R = ToVec[AliasWithNoPredicate]; 2599 if (!MatchCode.empty()) 2600 MatchCode += "else\n "; 2601 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n"; 2602 } 2603 2604 MatchCode += "return;"; 2605 2606 Cases.push_back(std::make_pair(AliasEntry.first, MatchCode)); 2607 } 2608 StringMatcher("Mnemonic", Cases, OS).Emit(Indent); 2609 } 2610 2611 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions, 2612 /// emit a function for them and return true, otherwise return false. 2613 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info, 2614 CodeGenTarget &Target) { 2615 // Ignore aliases when match-prefix is set. 2616 if (!MatchPrefix.empty()) 2617 return false; 2618 2619 std::vector<Record*> Aliases = 2620 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias"); 2621 if (Aliases.empty()) return false; 2622 2623 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, " 2624 "uint64_t Features, unsigned VariantID) {\n"; 2625 OS << " switch (VariantID) {\n"; 2626 unsigned VariantCount = Target.getAsmParserVariantCount(); 2627 for (unsigned VC = 0; VC != VariantCount; ++VC) { 2628 Record *AsmVariant = Target.getAsmParserVariant(VC); 2629 int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant"); 2630 std::string AsmParserVariantName = AsmVariant->getValueAsString("Name"); 2631 OS << " case " << AsmParserVariantNo << ":\n"; 2632 emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2, 2633 AsmParserVariantName); 2634 OS << " break;\n"; 2635 } 2636 OS << " }\n"; 2637 2638 // Emit aliases that apply to all variants. 2639 emitMnemonicAliasVariant(OS, Info, Aliases); 2640 2641 OS << "}\n\n"; 2642 2643 return true; 2644 } 2645 2646 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target, 2647 const AsmMatcherInfo &Info, StringRef ClassName, 2648 StringToOffsetTable &StringTable, 2649 unsigned MaxMnemonicIndex, bool HasMnemonicFirst) { 2650 unsigned MaxMask = 0; 2651 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { 2652 MaxMask |= OMI.OperandMask; 2653 } 2654 2655 // Emit the static custom operand parsing table; 2656 OS << "namespace {\n"; 2657 OS << " struct OperandMatchEntry {\n"; 2658 OS << " " << getMinimalRequiredFeaturesType(Info) 2659 << " RequiredFeatures;\n"; 2660 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex) 2661 << " Mnemonic;\n"; 2662 OS << " " << getMinimalTypeForRange(std::distance( 2663 Info.Classes.begin(), Info.Classes.end())) << " Class;\n"; 2664 OS << " " << getMinimalTypeForRange(MaxMask) 2665 << " OperandMask;\n\n"; 2666 OS << " StringRef getMnemonic() const {\n"; 2667 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 2668 OS << " MnemonicTable[Mnemonic]);\n"; 2669 OS << " }\n"; 2670 OS << " };\n\n"; 2671 2672 OS << " // Predicate for searching for an opcode.\n"; 2673 OS << " struct LessOpcodeOperand {\n"; 2674 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n"; 2675 OS << " return LHS.getMnemonic() < RHS;\n"; 2676 OS << " }\n"; 2677 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n"; 2678 OS << " return LHS < RHS.getMnemonic();\n"; 2679 OS << " }\n"; 2680 OS << " bool operator()(const OperandMatchEntry &LHS,"; 2681 OS << " const OperandMatchEntry &RHS) {\n"; 2682 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 2683 OS << " }\n"; 2684 OS << " };\n"; 2685 2686 OS << "} // end anonymous namespace.\n\n"; 2687 2688 OS << "static const OperandMatchEntry OperandMatchTable[" 2689 << Info.OperandMatchInfo.size() << "] = {\n"; 2690 2691 OS << " /* Operand List Mask, Mnemonic, Operand Class, Features */\n"; 2692 for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { 2693 const MatchableInfo &II = *OMI.MI; 2694 2695 OS << " { "; 2696 2697 // Write the required features mask. 2698 if (!II.RequiredFeatures.empty()) { 2699 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) { 2700 if (i) OS << "|"; 2701 OS << II.RequiredFeatures[i]->getEnumName(); 2702 } 2703 } else 2704 OS << "0"; 2705 2706 // Store a pascal-style length byte in the mnemonic. 2707 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str(); 2708 OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 2709 << " /* " << II.Mnemonic << " */, "; 2710 2711 OS << OMI.CI->Name; 2712 2713 OS << ", " << OMI.OperandMask; 2714 OS << " /* "; 2715 bool printComma = false; 2716 for (int i = 0, e = 31; i !=e; ++i) 2717 if (OMI.OperandMask & (1 << i)) { 2718 if (printComma) 2719 OS << ", "; 2720 OS << i; 2721 printComma = true; 2722 } 2723 OS << " */"; 2724 2725 OS << " },\n"; 2726 } 2727 OS << "};\n\n"; 2728 2729 // Emit the operand class switch to call the correct custom parser for 2730 // the found operand class. 2731 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2732 << Target.getName() << ClassName << "::\n" 2733 << "tryCustomParseOperand(OperandVector" 2734 << " &Operands,\n unsigned MCK) {\n\n" 2735 << " switch(MCK) {\n"; 2736 2737 for (const auto &CI : Info.Classes) { 2738 if (CI.ParserMethod.empty()) 2739 continue; 2740 OS << " case " << CI.Name << ":\n" 2741 << " return " << CI.ParserMethod << "(Operands);\n"; 2742 } 2743 2744 OS << " default:\n"; 2745 OS << " return MatchOperand_NoMatch;\n"; 2746 OS << " }\n"; 2747 OS << " return MatchOperand_NoMatch;\n"; 2748 OS << "}\n\n"; 2749 2750 // Emit the static custom operand parser. This code is very similar with 2751 // the other matcher. Also use MatchResultTy here just in case we go for 2752 // a better error handling. 2753 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2754 << Target.getName() << ClassName << "::\n" 2755 << "MatchOperandParserImpl(OperandVector" 2756 << " &Operands,\n StringRef Mnemonic) {\n"; 2757 2758 // Emit code to get the available features. 2759 OS << " // Get the current feature set.\n"; 2760 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n"; 2761 2762 OS << " // Get the next operand index.\n"; 2763 OS << " unsigned NextOpNum = Operands.size()" 2764 << (HasMnemonicFirst ? " - 1" : "") << ";\n"; 2765 2766 // Emit code to search the table. 2767 OS << " // Search the table.\n"; 2768 if (HasMnemonicFirst) { 2769 OS << " auto MnemonicRange =\n"; 2770 OS << " std::equal_range(std::begin(OperandMatchTable), " 2771 "std::end(OperandMatchTable),\n"; 2772 OS << " Mnemonic, LessOpcodeOperand());\n\n"; 2773 } else { 2774 OS << " auto MnemonicRange = std::make_pair(std::begin(OperandMatchTable)," 2775 " std::end(OperandMatchTable));\n"; 2776 OS << " if (!Mnemonic.empty())\n"; 2777 OS << " MnemonicRange =\n"; 2778 OS << " std::equal_range(std::begin(OperandMatchTable), " 2779 "std::end(OperandMatchTable),\n"; 2780 OS << " Mnemonic, LessOpcodeOperand());\n\n"; 2781 } 2782 2783 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 2784 OS << " return MatchOperand_NoMatch;\n\n"; 2785 2786 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n" 2787 << " *ie = MnemonicRange.second; it != ie; ++it) {\n"; 2788 2789 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 2790 OS << " assert(Mnemonic == it->getMnemonic());\n\n"; 2791 2792 // Emit check that the required features are available. 2793 OS << " // check if the available features match\n"; 2794 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 2795 << "!= it->RequiredFeatures) {\n"; 2796 OS << " continue;\n"; 2797 OS << " }\n\n"; 2798 2799 // Emit check to ensure the operand number matches. 2800 OS << " // check if the operand in question has a custom parser.\n"; 2801 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n"; 2802 OS << " continue;\n\n"; 2803 2804 // Emit call to the custom parser method 2805 OS << " // call custom parse method to handle the operand\n"; 2806 OS << " OperandMatchResultTy Result = "; 2807 OS << "tryCustomParseOperand(Operands, it->Class);\n"; 2808 OS << " if (Result != MatchOperand_NoMatch)\n"; 2809 OS << " return Result;\n"; 2810 OS << " }\n\n"; 2811 2812 OS << " // Okay, we had no match.\n"; 2813 OS << " return MatchOperand_NoMatch;\n"; 2814 OS << "}\n\n"; 2815 } 2816 2817 void AsmMatcherEmitter::run(raw_ostream &OS) { 2818 CodeGenTarget Target(Records); 2819 Record *AsmParser = Target.getAsmParser(); 2820 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName"); 2821 2822 // Compute the information on the instructions to match. 2823 AsmMatcherInfo Info(AsmParser, Target, Records); 2824 Info.buildInfo(); 2825 2826 // Sort the instruction table using the partial order on classes. We use 2827 // stable_sort to ensure that ambiguous instructions are still 2828 // deterministically ordered. 2829 std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(), 2830 [](const std::unique_ptr<MatchableInfo> &a, 2831 const std::unique_ptr<MatchableInfo> &b){ 2832 return *a < *b;}); 2833 2834 #ifndef NDEBUG 2835 // Verify that the table is now sorted 2836 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; 2837 ++I) { 2838 for (auto J = I; J != E; ++J) { 2839 assert(!(**J < **I)); 2840 } 2841 } 2842 #endif // NDEBUG 2843 2844 DEBUG_WITH_TYPE("instruction_info", { 2845 for (const auto &MI : Info.Matchables) 2846 MI->dump(); 2847 }); 2848 2849 // Check for ambiguous matchables. 2850 DEBUG_WITH_TYPE("ambiguous_instrs", { 2851 unsigned NumAmbiguous = 0; 2852 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; 2853 ++I) { 2854 for (auto J = std::next(I); J != E; ++J) { 2855 const MatchableInfo &A = **I; 2856 const MatchableInfo &B = **J; 2857 2858 if (A.couldMatchAmbiguouslyWith(B)) { 2859 errs() << "warning: ambiguous matchables:\n"; 2860 A.dump(); 2861 errs() << "\nis incomparable with:\n"; 2862 B.dump(); 2863 errs() << "\n\n"; 2864 ++NumAmbiguous; 2865 } 2866 } 2867 } 2868 if (NumAmbiguous) 2869 errs() << "warning: " << NumAmbiguous 2870 << " ambiguous matchables!\n"; 2871 }); 2872 2873 // Compute the information on the custom operand parsing. 2874 Info.buildOperandMatchInfo(); 2875 2876 bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst"); 2877 bool HasOptionalOperands = Info.hasOptionalOperands(); 2878 2879 // Write the output. 2880 2881 // Information for the class declaration. 2882 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n"; 2883 OS << "#undef GET_ASSEMBLER_HEADER\n"; 2884 OS << " // This should be included into the middle of the declaration of\n"; 2885 OS << " // your subclasses implementation of MCTargetAsmParser.\n"; 2886 OS << " uint64_t ComputeAvailableFeatures(const FeatureBitset& FB) const;\n"; 2887 if (HasOptionalOperands) { 2888 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " 2889 << "unsigned Opcode,\n" 2890 << " const OperandVector &Operands,\n" 2891 << " const SmallBitVector &OptionalOperandsMask);\n"; 2892 } else { 2893 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " 2894 << "unsigned Opcode,\n" 2895 << " const OperandVector &Operands);\n"; 2896 } 2897 OS << " void convertToMapAndConstraints(unsigned Kind,\n "; 2898 OS << " const OperandVector &Operands) override;\n"; 2899 if (HasMnemonicFirst) 2900 OS << " bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID);\n"; 2901 OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n" 2902 << " MCInst &Inst,\n" 2903 << " uint64_t &ErrorInfo," 2904 << " bool matchingInlineAsm,\n" 2905 << " unsigned VariantID = 0);\n"; 2906 2907 if (!Info.OperandMatchInfo.empty()) { 2908 OS << "\n enum OperandMatchResultTy {\n"; 2909 OS << " MatchOperand_Success, // operand matched successfully\n"; 2910 OS << " MatchOperand_NoMatch, // operand did not match\n"; 2911 OS << " MatchOperand_ParseFail // operand matched but had errors\n"; 2912 OS << " };\n"; 2913 OS << " OperandMatchResultTy MatchOperandParserImpl(\n"; 2914 OS << " OperandVector &Operands,\n"; 2915 OS << " StringRef Mnemonic);\n"; 2916 2917 OS << " OperandMatchResultTy tryCustomParseOperand(\n"; 2918 OS << " OperandVector &Operands,\n"; 2919 OS << " unsigned MCK);\n\n"; 2920 } 2921 2922 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n"; 2923 2924 // Emit the operand match diagnostic enum names. 2925 OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n"; 2926 OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n"; 2927 emitOperandDiagnosticTypes(Info, OS); 2928 OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n"; 2929 2930 OS << "\n#ifdef GET_REGISTER_MATCHER\n"; 2931 OS << "#undef GET_REGISTER_MATCHER\n\n"; 2932 2933 // Emit the subtarget feature enumeration. 2934 emitSubtargetFeatureFlagEnumeration(Info, OS); 2935 2936 // Emit the function to match a register name to number. 2937 // This should be omitted for Mips target 2938 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName")) 2939 emitMatchRegisterName(Target, AsmParser, OS); 2940 2941 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName")) 2942 emitMatchRegisterAltName(Target, AsmParser, OS); 2943 2944 OS << "#endif // GET_REGISTER_MATCHER\n\n"; 2945 2946 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n"; 2947 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n"; 2948 2949 // Generate the helper function to get the names for subtarget features. 2950 emitGetSubtargetFeatureName(Info, OS); 2951 2952 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n"; 2953 2954 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n"; 2955 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n"; 2956 2957 // Generate the function that remaps for mnemonic aliases. 2958 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target); 2959 2960 // Generate the convertToMCInst function to convert operands into an MCInst. 2961 // Also, generate the convertToMapAndConstraints function for MS-style inline 2962 // assembly. The latter doesn't actually generate a MCInst. 2963 emitConvertFuncs(Target, ClassName, Info.Matchables, HasMnemonicFirst, 2964 HasOptionalOperands, OS); 2965 2966 // Emit the enumeration for classes which participate in matching. 2967 emitMatchClassEnumeration(Target, Info.Classes, OS); 2968 2969 // Emit the routine to match token strings to their match class. 2970 emitMatchTokenString(Target, Info.Classes, OS); 2971 2972 // Emit the subclass predicate routine. 2973 emitIsSubclass(Target, Info.Classes, OS); 2974 2975 // Emit the routine to validate an operand against a match class. 2976 emitValidateOperandClass(Info, OS); 2977 2978 // Emit the available features compute function. 2979 emitComputeAvailableFeatures(Info, OS); 2980 2981 StringToOffsetTable StringTable; 2982 2983 size_t MaxNumOperands = 0; 2984 unsigned MaxMnemonicIndex = 0; 2985 bool HasDeprecation = false; 2986 for (const auto &MI : Info.Matchables) { 2987 MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size()); 2988 HasDeprecation |= MI->HasDeprecation; 2989 2990 // Store a pascal-style length byte in the mnemonic. 2991 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str(); 2992 MaxMnemonicIndex = std::max(MaxMnemonicIndex, 2993 StringTable.GetOrAddStringOffset(LenMnemonic, false)); 2994 } 2995 2996 OS << "static const char *const MnemonicTable =\n"; 2997 StringTable.EmitString(OS); 2998 OS << ";\n\n"; 2999 3000 // Emit the static match table; unused classes get initalized to 0 which is 3001 // guaranteed to be InvalidMatchClass. 3002 // 3003 // FIXME: We can reduce the size of this table very easily. First, we change 3004 // it so that store the kinds in separate bit-fields for each index, which 3005 // only needs to be the max width used for classes at that index (we also need 3006 // to reject based on this during classification). If we then make sure to 3007 // order the match kinds appropriately (putting mnemonics last), then we 3008 // should only end up using a few bits for each class, especially the ones 3009 // following the mnemonic. 3010 OS << "namespace {\n"; 3011 OS << " struct MatchEntry {\n"; 3012 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex) 3013 << " Mnemonic;\n"; 3014 OS << " uint16_t Opcode;\n"; 3015 OS << " " << getMinimalTypeForRange(Info.Matchables.size()) 3016 << " ConvertFn;\n"; 3017 OS << " " << getMinimalRequiredFeaturesType(Info) 3018 << " RequiredFeatures;\n"; 3019 OS << " " << getMinimalTypeForRange( 3020 std::distance(Info.Classes.begin(), Info.Classes.end())) 3021 << " Classes[" << MaxNumOperands << "];\n"; 3022 OS << " StringRef getMnemonic() const {\n"; 3023 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 3024 OS << " MnemonicTable[Mnemonic]);\n"; 3025 OS << " }\n"; 3026 OS << " };\n\n"; 3027 3028 OS << " // Predicate for searching for an opcode.\n"; 3029 OS << " struct LessOpcode {\n"; 3030 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n"; 3031 OS << " return LHS.getMnemonic() < RHS;\n"; 3032 OS << " }\n"; 3033 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n"; 3034 OS << " return LHS < RHS.getMnemonic();\n"; 3035 OS << " }\n"; 3036 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n"; 3037 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 3038 OS << " }\n"; 3039 OS << " };\n"; 3040 3041 OS << "} // end anonymous namespace.\n\n"; 3042 3043 unsigned VariantCount = Target.getAsmParserVariantCount(); 3044 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3045 Record *AsmVariant = Target.getAsmParserVariant(VC); 3046 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3047 3048 OS << "static const MatchEntry MatchTable" << VC << "[] = {\n"; 3049 3050 for (const auto &MI : Info.Matchables) { 3051 if (MI->AsmVariantID != AsmVariantNo) 3052 continue; 3053 3054 // Store a pascal-style length byte in the mnemonic. 3055 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str(); 3056 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 3057 << " /* " << MI->Mnemonic << " */, " 3058 << Target.getName() << "::" 3059 << MI->getResultInst()->TheDef->getName() << ", " 3060 << MI->ConversionFnKind << ", "; 3061 3062 // Write the required features mask. 3063 if (!MI->RequiredFeatures.empty()) { 3064 for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i) { 3065 if (i) OS << "|"; 3066 OS << MI->RequiredFeatures[i]->getEnumName(); 3067 } 3068 } else 3069 OS << "0"; 3070 3071 OS << ", { "; 3072 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) { 3073 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i]; 3074 3075 if (i) OS << ", "; 3076 OS << Op.Class->Name; 3077 } 3078 OS << " }, },\n"; 3079 } 3080 3081 OS << "};\n\n"; 3082 } 3083 3084 // A method to determine if a mnemonic is in the list. 3085 if (HasMnemonicFirst) { 3086 OS << "bool " << Target.getName() << ClassName << "::\n" 3087 << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n"; 3088 OS << " // Find the appropriate table for this asm variant.\n"; 3089 OS << " const MatchEntry *Start, *End;\n"; 3090 OS << " switch (VariantID) {\n"; 3091 OS << " default: llvm_unreachable(\"invalid variant!\");\n"; 3092 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3093 Record *AsmVariant = Target.getAsmParserVariant(VC); 3094 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3095 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC 3096 << "); End = std::end(MatchTable" << VC << "); break;\n"; 3097 } 3098 OS << " }\n"; 3099 OS << " // Search the table.\n"; 3100 OS << " auto MnemonicRange = "; 3101 OS << "std::equal_range(Start, End, Mnemonic, LessOpcode());\n"; 3102 OS << " return MnemonicRange.first != MnemonicRange.second;\n"; 3103 OS << "}\n\n"; 3104 } 3105 3106 // Finally, build the match function. 3107 OS << "unsigned " << Target.getName() << ClassName << "::\n" 3108 << "MatchInstructionImpl(const OperandVector &Operands,\n"; 3109 OS << " MCInst &Inst, uint64_t &ErrorInfo,\n" 3110 << " bool matchingInlineAsm, unsigned VariantID) {\n"; 3111 3112 OS << " // Eliminate obvious mismatches.\n"; 3113 OS << " if (Operands.size() > " 3114 << (MaxNumOperands + HasMnemonicFirst) << ") {\n"; 3115 OS << " ErrorInfo = " 3116 << (MaxNumOperands + HasMnemonicFirst) << ";\n"; 3117 OS << " return Match_InvalidOperand;\n"; 3118 OS << " }\n\n"; 3119 3120 // Emit code to get the available features. 3121 OS << " // Get the current feature set.\n"; 3122 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n"; 3123 3124 OS << " // Get the instruction mnemonic, which is the first token.\n"; 3125 if (HasMnemonicFirst) { 3126 OS << " StringRef Mnemonic = ((" << Target.getName() 3127 << "Operand&)*Operands[0]).getToken();\n\n"; 3128 } else { 3129 OS << " StringRef Mnemonic;\n"; 3130 OS << " if (Operands[0]->isToken())\n"; 3131 OS << " Mnemonic = ((" << Target.getName() 3132 << "Operand&)*Operands[0]).getToken();\n\n"; 3133 } 3134 3135 if (HasMnemonicAliases) { 3136 OS << " // Process all MnemonicAliases to remap the mnemonic.\n"; 3137 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n"; 3138 } 3139 3140 // Emit code to compute the class list for this operand vector. 3141 OS << " // Some state to try to produce better error messages.\n"; 3142 OS << " bool HadMatchOtherThanFeatures = false;\n"; 3143 OS << " bool HadMatchOtherThanPredicate = false;\n"; 3144 OS << " unsigned RetCode = Match_InvalidOperand;\n"; 3145 OS << " uint64_t MissingFeatures = ~0ULL;\n"; 3146 if (HasOptionalOperands) { 3147 OS << " SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n"; 3148 } 3149 OS << " // Set ErrorInfo to the operand that mismatches if it is\n"; 3150 OS << " // wrong for all instances of the instruction.\n"; 3151 OS << " ErrorInfo = ~0ULL;\n"; 3152 3153 // Emit code to search the table. 3154 OS << " // Find the appropriate table for this asm variant.\n"; 3155 OS << " const MatchEntry *Start, *End;\n"; 3156 OS << " switch (VariantID) {\n"; 3157 OS << " default: llvm_unreachable(\"invalid variant!\");\n"; 3158 for (unsigned VC = 0; VC != VariantCount; ++VC) { 3159 Record *AsmVariant = Target.getAsmParserVariant(VC); 3160 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 3161 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC 3162 << "); End = std::end(MatchTable" << VC << "); break;\n"; 3163 } 3164 OS << " }\n"; 3165 3166 OS << " // Search the table.\n"; 3167 if (HasMnemonicFirst) { 3168 OS << " auto MnemonicRange = " 3169 "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n"; 3170 } else { 3171 OS << " auto MnemonicRange = std::make_pair(Start, End);\n"; 3172 OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n"; 3173 OS << " if (!Mnemonic.empty())\n"; 3174 OS << " MnemonicRange = " 3175 "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n"; 3176 } 3177 3178 OS << " // Return a more specific error code if no mnemonics match.\n"; 3179 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 3180 OS << " return Match_MnemonicFail;\n\n"; 3181 3182 OS << " for (const MatchEntry *it = MnemonicRange.first, " 3183 << "*ie = MnemonicRange.second;\n"; 3184 OS << " it != ie; ++it) {\n"; 3185 3186 if (HasMnemonicFirst) { 3187 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 3188 OS << " assert(Mnemonic == it->getMnemonic());\n"; 3189 } 3190 3191 // Emit check that the subclasses match. 3192 OS << " bool OperandsValid = true;\n"; 3193 if (HasOptionalOperands) { 3194 OS << " OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n"; 3195 } 3196 OS << " for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex") 3197 << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex") 3198 << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n"; 3199 OS << " auto Formal = " 3200 << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n"; 3201 OS << " if (ActualIdx >= Operands.size()) {\n"; 3202 OS << " OperandsValid = (Formal == " <<"InvalidMatchClass) || " 3203 "isSubclass(Formal, OptionalMatchClass);\n"; 3204 OS << " if (!OperandsValid) ErrorInfo = ActualIdx;\n"; 3205 if (HasOptionalOperands) { 3206 OS << " OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands 3207 << ");\n"; 3208 } 3209 OS << " break;\n"; 3210 OS << " }\n"; 3211 OS << " MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n"; 3212 OS << " unsigned Diag = validateOperandClass(Actual, Formal);\n"; 3213 OS << " if (Diag == Match_Success) {\n"; 3214 OS << " ++ActualIdx;\n"; 3215 OS << " continue;\n"; 3216 OS << " }\n"; 3217 OS << " // If the generic handler indicates an invalid operand\n"; 3218 OS << " // failure, check for a special case.\n"; 3219 OS << " if (Diag == Match_InvalidOperand) {\n"; 3220 OS << " Diag = validateTargetOperandClass(Actual, Formal);\n"; 3221 OS << " if (Diag == Match_Success) {\n"; 3222 OS << " ++ActualIdx;\n"; 3223 OS << " continue;\n"; 3224 OS << " }\n"; 3225 OS << " }\n"; 3226 OS << " // If current formal operand wasn't matched and it is optional\n" 3227 << " // then try to match next formal operand\n"; 3228 OS << " if (Diag == Match_InvalidOperand " 3229 << "&& isSubclass(Formal, OptionalMatchClass)) {\n"; 3230 if (HasOptionalOperands) { 3231 OS << " OptionalOperandsMask.set(FormalIdx);\n"; 3232 } 3233 OS << " continue;\n"; 3234 OS << " }\n"; 3235 OS << " // If this operand is broken for all of the instances of this\n"; 3236 OS << " // mnemonic, keep track of it so we can report loc info.\n"; 3237 OS << " // If we already had a match that only failed due to a\n"; 3238 OS << " // target predicate, that diagnostic is preferred.\n"; 3239 OS << " if (!HadMatchOtherThanPredicate &&\n"; 3240 OS << " (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) {\n"; 3241 OS << " ErrorInfo = ActualIdx;\n"; 3242 OS << " // InvalidOperand is the default. Prefer specificity.\n"; 3243 OS << " if (Diag != Match_InvalidOperand)\n"; 3244 OS << " RetCode = Diag;\n"; 3245 OS << " }\n"; 3246 OS << " // Otherwise, just reject this instance of the mnemonic.\n"; 3247 OS << " OperandsValid = false;\n"; 3248 OS << " break;\n"; 3249 OS << " }\n\n"; 3250 3251 OS << " if (!OperandsValid) continue;\n"; 3252 3253 // Emit check that the required features are available. 3254 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 3255 << "!= it->RequiredFeatures) {\n"; 3256 OS << " HadMatchOtherThanFeatures = true;\n"; 3257 OS << " uint64_t NewMissingFeatures = it->RequiredFeatures & " 3258 "~AvailableFeatures;\n"; 3259 OS << " if (countPopulation(NewMissingFeatures) <=\n" 3260 " countPopulation(MissingFeatures))\n"; 3261 OS << " MissingFeatures = NewMissingFeatures;\n"; 3262 OS << " continue;\n"; 3263 OS << " }\n"; 3264 OS << "\n"; 3265 OS << " Inst.clear();\n\n"; 3266 OS << " Inst.setOpcode(it->Opcode);\n"; 3267 // Verify the instruction with the target-specific match predicate function. 3268 OS << " // We have a potential match but have not rendered the operands.\n" 3269 << " // Check the target predicate to handle any context sensitive\n" 3270 " // constraints.\n" 3271 << " // For example, Ties that are referenced multiple times must be\n" 3272 " // checked here to ensure the input is the same for each match\n" 3273 " // constraints. If we leave it any later the ties will have been\n" 3274 " // canonicalized\n" 3275 << " unsigned MatchResult;\n" 3276 << " if ((MatchResult = checkEarlyTargetMatchPredicate(Inst, " 3277 "Operands)) != Match_Success) {\n" 3278 << " Inst.clear();\n" 3279 << " RetCode = MatchResult;\n" 3280 << " HadMatchOtherThanPredicate = true;\n" 3281 << " continue;\n" 3282 << " }\n\n"; 3283 OS << " if (matchingInlineAsm) {\n"; 3284 OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n"; 3285 OS << " return Match_Success;\n"; 3286 OS << " }\n\n"; 3287 OS << " // We have selected a definite instruction, convert the parsed\n" 3288 << " // operands into the appropriate MCInst.\n"; 3289 if (HasOptionalOperands) { 3290 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n" 3291 << " OptionalOperandsMask);\n"; 3292 } else { 3293 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n"; 3294 } 3295 OS << "\n"; 3296 3297 // Verify the instruction with the target-specific match predicate function. 3298 OS << " // We have a potential match. Check the target predicate to\n" 3299 << " // handle any context sensitive constraints.\n" 3300 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !=" 3301 << " Match_Success) {\n" 3302 << " Inst.clear();\n" 3303 << " RetCode = MatchResult;\n" 3304 << " HadMatchOtherThanPredicate = true;\n" 3305 << " continue;\n" 3306 << " }\n\n"; 3307 3308 // Call the post-processing function, if used. 3309 std::string InsnCleanupFn = 3310 AsmParser->getValueAsString("AsmParserInstCleanup"); 3311 if (!InsnCleanupFn.empty()) 3312 OS << " " << InsnCleanupFn << "(Inst);\n"; 3313 3314 if (HasDeprecation) { 3315 OS << " std::string Info;\n"; 3316 OS << " if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n"; 3317 OS << " SMLoc Loc = ((" << Target.getName() 3318 << "Operand&)*Operands[0]).getStartLoc();\n"; 3319 OS << " getParser().Warning(Loc, Info, None);\n"; 3320 OS << " }\n"; 3321 } 3322 3323 OS << " return Match_Success;\n"; 3324 OS << " }\n\n"; 3325 3326 OS << " // Okay, we had no match. Try to return a useful error code.\n"; 3327 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n"; 3328 OS << " return RetCode;\n\n"; 3329 OS << " // Missing feature matches return which features were missing\n"; 3330 OS << " ErrorInfo = MissingFeatures;\n"; 3331 OS << " return Match_MissingFeature;\n"; 3332 OS << "}\n\n"; 3333 3334 if (!Info.OperandMatchInfo.empty()) 3335 emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable, 3336 MaxMnemonicIndex, HasMnemonicFirst); 3337 3338 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n"; 3339 } 3340 3341 namespace llvm { 3342 3343 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) { 3344 emitSourceFileHeader("Assembly Matcher Source Fragment", OS); 3345 AsmMatcherEmitter(RK).run(OS); 3346 } 3347 3348 } // end namespace llvm 3349