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