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 valeus 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 "AsmMatcherEmitter.h" 100 #include "CodeGenTarget.h" 101 #include "StringMatcher.h" 102 #include "StringToOffsetTable.h" 103 #include "llvm/ADT/OwningPtr.h" 104 #include "llvm/ADT/PointerUnion.h" 105 #include "llvm/ADT/SmallPtrSet.h" 106 #include "llvm/ADT/SmallVector.h" 107 #include "llvm/ADT/STLExtras.h" 108 #include "llvm/ADT/StringExtras.h" 109 #include "llvm/Support/CommandLine.h" 110 #include "llvm/Support/Debug.h" 111 #include "llvm/Support/ErrorHandling.h" 112 #include "llvm/TableGen/Error.h" 113 #include "llvm/TableGen/Record.h" 114 #include <map> 115 #include <set> 116 using namespace llvm; 117 118 static cl::opt<std::string> 119 MatchPrefix("match-prefix", cl::init(""), 120 cl::desc("Only match instructions with the given prefix")); 121 122 namespace { 123 class AsmMatcherInfo; 124 struct SubtargetFeatureInfo; 125 126 /// ClassInfo - Helper class for storing the information about a particular 127 /// class of operands which can be matched. 128 struct ClassInfo { 129 enum ClassInfoKind { 130 /// Invalid kind, for use as a sentinel value. 131 Invalid = 0, 132 133 /// The class for a particular token. 134 Token, 135 136 /// The (first) register class, subsequent register classes are 137 /// RegisterClass0+1, and so on. 138 RegisterClass0, 139 140 /// The (first) user defined class, subsequent user defined classes are 141 /// UserClass0+1, and so on. 142 UserClass0 = 1<<16 143 }; 144 145 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + 146 /// N) for the Nth user defined class. 147 unsigned Kind; 148 149 /// SuperClasses - The super classes of this class. Note that for simplicities 150 /// sake user operands only record their immediate super class, while register 151 /// operands include all superclasses. 152 std::vector<ClassInfo*> SuperClasses; 153 154 /// Name - The full class name, suitable for use in an enum. 155 std::string Name; 156 157 /// ClassName - The unadorned generic name for this class (e.g., Token). 158 std::string ClassName; 159 160 /// ValueName - The name of the value this class represents; for a token this 161 /// is the literal token string, for an operand it is the TableGen class (or 162 /// empty if this is a derived class). 163 std::string ValueName; 164 165 /// PredicateMethod - The name of the operand method to test whether the 166 /// operand matches this class; this is not valid for Token or register kinds. 167 std::string PredicateMethod; 168 169 /// RenderMethod - The name of the operand method to add this operand to an 170 /// MCInst; this is not valid for Token or register kinds. 171 std::string RenderMethod; 172 173 /// ParserMethod - The name of the operand method to do a target specific 174 /// parsing on the operand. 175 std::string ParserMethod; 176 177 /// For register classes, the records for all the registers in this class. 178 std::set<Record*> Registers; 179 180 public: 181 /// isRegisterClass() - Check if this is a register class. 182 bool isRegisterClass() const { 183 return Kind >= RegisterClass0 && Kind < UserClass0; 184 } 185 186 /// isUserClass() - Check if this is a user defined class. 187 bool isUserClass() const { 188 return Kind >= UserClass0; 189 } 190 191 /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes 192 /// are related if they are in the same class hierarchy. 193 bool isRelatedTo(const ClassInfo &RHS) const { 194 // Tokens are only related to tokens. 195 if (Kind == Token || RHS.Kind == Token) 196 return Kind == Token && RHS.Kind == Token; 197 198 // Registers classes are only related to registers classes, and only if 199 // their intersection is non-empty. 200 if (isRegisterClass() || RHS.isRegisterClass()) { 201 if (!isRegisterClass() || !RHS.isRegisterClass()) 202 return false; 203 204 std::set<Record*> Tmp; 205 std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin()); 206 std::set_intersection(Registers.begin(), Registers.end(), 207 RHS.Registers.begin(), RHS.Registers.end(), 208 II); 209 210 return !Tmp.empty(); 211 } 212 213 // Otherwise we have two users operands; they are related if they are in the 214 // same class hierarchy. 215 // 216 // FIXME: This is an oversimplification, they should only be related if they 217 // intersect, however we don't have that information. 218 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!"); 219 const ClassInfo *Root = this; 220 while (!Root->SuperClasses.empty()) 221 Root = Root->SuperClasses.front(); 222 223 const ClassInfo *RHSRoot = &RHS; 224 while (!RHSRoot->SuperClasses.empty()) 225 RHSRoot = RHSRoot->SuperClasses.front(); 226 227 return Root == RHSRoot; 228 } 229 230 /// isSubsetOf - Test whether this class is a subset of \arg RHS; 231 bool isSubsetOf(const ClassInfo &RHS) const { 232 // This is a subset of RHS if it is the same class... 233 if (this == &RHS) 234 return true; 235 236 // ... or if any of its super classes are a subset of RHS. 237 for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(), 238 ie = SuperClasses.end(); it != ie; ++it) 239 if ((*it)->isSubsetOf(RHS)) 240 return true; 241 242 return false; 243 } 244 245 /// operator< - Compare two classes. 246 bool operator<(const ClassInfo &RHS) const { 247 if (this == &RHS) 248 return false; 249 250 // Unrelated classes can be ordered by kind. 251 if (!isRelatedTo(RHS)) 252 return Kind < RHS.Kind; 253 254 switch (Kind) { 255 case Invalid: 256 llvm_unreachable("Invalid kind!"); 257 258 default: 259 // This class precedes the RHS if it is a proper subset of the RHS. 260 if (isSubsetOf(RHS)) 261 return true; 262 if (RHS.isSubsetOf(*this)) 263 return false; 264 265 // Otherwise, order by name to ensure we have a total ordering. 266 return ValueName < RHS.ValueName; 267 } 268 } 269 }; 270 271 /// MatchableInfo - Helper class for storing the necessary information for an 272 /// instruction or alias which is capable of being matched. 273 struct MatchableInfo { 274 struct AsmOperand { 275 /// Token - This is the token that the operand came from. 276 StringRef Token; 277 278 /// The unique class instance this operand should match. 279 ClassInfo *Class; 280 281 /// The operand name this is, if anything. 282 StringRef SrcOpName; 283 284 /// The suboperand index within SrcOpName, or -1 for the entire operand. 285 int SubOpIdx; 286 287 /// Register record if this token is singleton register. 288 Record *SingletonReg; 289 290 explicit AsmOperand(StringRef T) : Token(T), Class(0), SubOpIdx(-1), 291 SingletonReg(0) {} 292 }; 293 294 /// ResOperand - This represents a single operand in the result instruction 295 /// generated by the match. In cases (like addressing modes) where a single 296 /// assembler operand expands to multiple MCOperands, this represents the 297 /// single assembler operand, not the MCOperand. 298 struct ResOperand { 299 enum { 300 /// RenderAsmOperand - This represents an operand result that is 301 /// generated by calling the render method on the assembly operand. The 302 /// corresponding AsmOperand is specified by AsmOperandNum. 303 RenderAsmOperand, 304 305 /// TiedOperand - This represents a result operand that is a duplicate of 306 /// a previous result operand. 307 TiedOperand, 308 309 /// ImmOperand - This represents an immediate value that is dumped into 310 /// the operand. 311 ImmOperand, 312 313 /// RegOperand - This represents a fixed register that is dumped in. 314 RegOperand 315 } Kind; 316 317 union { 318 /// This is the operand # in the AsmOperands list that this should be 319 /// copied from. 320 unsigned AsmOperandNum; 321 322 /// TiedOperandNum - This is the (earlier) result operand that should be 323 /// copied from. 324 unsigned TiedOperandNum; 325 326 /// ImmVal - This is the immediate value added to the instruction. 327 int64_t ImmVal; 328 329 /// Register - This is the register record. 330 Record *Register; 331 }; 332 333 /// MINumOperands - The number of MCInst operands populated by this 334 /// operand. 335 unsigned MINumOperands; 336 337 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) { 338 ResOperand X; 339 X.Kind = RenderAsmOperand; 340 X.AsmOperandNum = AsmOpNum; 341 X.MINumOperands = NumOperands; 342 return X; 343 } 344 345 static ResOperand getTiedOp(unsigned TiedOperandNum) { 346 ResOperand X; 347 X.Kind = TiedOperand; 348 X.TiedOperandNum = TiedOperandNum; 349 X.MINumOperands = 1; 350 return X; 351 } 352 353 static ResOperand getImmOp(int64_t Val) { 354 ResOperand X; 355 X.Kind = ImmOperand; 356 X.ImmVal = Val; 357 X.MINumOperands = 1; 358 return X; 359 } 360 361 static ResOperand getRegOp(Record *Reg) { 362 ResOperand X; 363 X.Kind = RegOperand; 364 X.Register = Reg; 365 X.MINumOperands = 1; 366 return X; 367 } 368 }; 369 370 /// AsmVariantID - Target's assembly syntax variant no. 371 int AsmVariantID; 372 373 /// TheDef - This is the definition of the instruction or InstAlias that this 374 /// matchable came from. 375 Record *const TheDef; 376 377 /// DefRec - This is the definition that it came from. 378 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec; 379 380 const CodeGenInstruction *getResultInst() const { 381 if (DefRec.is<const CodeGenInstruction*>()) 382 return DefRec.get<const CodeGenInstruction*>(); 383 return DefRec.get<const CodeGenInstAlias*>()->ResultInst; 384 } 385 386 /// ResOperands - This is the operand list that should be built for the result 387 /// MCInst. 388 SmallVector<ResOperand, 8> ResOperands; 389 390 /// AsmString - The assembly string for this instruction (with variants 391 /// removed), e.g. "movsx $src, $dst". 392 std::string AsmString; 393 394 /// Mnemonic - This is the first token of the matched instruction, its 395 /// mnemonic. 396 StringRef Mnemonic; 397 398 /// AsmOperands - The textual operands that this instruction matches, 399 /// annotated with a class and where in the OperandList they were defined. 400 /// This directly corresponds to the tokenized AsmString after the mnemonic is 401 /// removed. 402 SmallVector<AsmOperand, 8> AsmOperands; 403 404 /// Predicates - The required subtarget features to match this instruction. 405 SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures; 406 407 /// ConversionFnKind - The enum value which is passed to the generated 408 /// ConvertToMCInst to convert parsed operands into an MCInst for this 409 /// function. 410 std::string ConversionFnKind; 411 412 MatchableInfo(const CodeGenInstruction &CGI) 413 : AsmVariantID(0), TheDef(CGI.TheDef), DefRec(&CGI), 414 AsmString(CGI.AsmString) { 415 } 416 417 MatchableInfo(const CodeGenInstAlias *Alias) 418 : AsmVariantID(0), TheDef(Alias->TheDef), DefRec(Alias), 419 AsmString(Alias->AsmString) { 420 } 421 422 // Two-operand aliases clone from the main matchable, but mark the second 423 // operand as a tied operand of the first for purposes of the assembler. 424 void formTwoOperandAlias(StringRef Constraint); 425 426 void initialize(const AsmMatcherInfo &Info, 427 SmallPtrSet<Record*, 16> &SingletonRegisters, 428 int AsmVariantNo, std::string &RegisterPrefix); 429 430 /// validate - Return true if this matchable is a valid thing to match against 431 /// and perform a bunch of validity checking. 432 bool validate(StringRef CommentDelimiter, bool Hack) const; 433 434 /// extractSingletonRegisterForAsmOperand - Extract singleton register, 435 /// if present, from specified token. 436 void 437 extractSingletonRegisterForAsmOperand(unsigned i, const AsmMatcherInfo &Info, 438 std::string &RegisterPrefix); 439 440 /// findAsmOperand - Find the AsmOperand with the specified name and 441 /// suboperand index. 442 int findAsmOperand(StringRef N, int SubOpIdx) const { 443 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) 444 if (N == AsmOperands[i].SrcOpName && 445 SubOpIdx == AsmOperands[i].SubOpIdx) 446 return i; 447 return -1; 448 } 449 450 /// findAsmOperandNamed - Find the first AsmOperand with the specified name. 451 /// This does not check the suboperand index. 452 int findAsmOperandNamed(StringRef N) const { 453 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) 454 if (N == AsmOperands[i].SrcOpName) 455 return i; 456 return -1; 457 } 458 459 void buildInstructionResultOperands(); 460 void buildAliasResultOperands(); 461 462 /// operator< - Compare two matchables. 463 bool operator<(const MatchableInfo &RHS) const { 464 // The primary comparator is the instruction mnemonic. 465 if (Mnemonic != RHS.Mnemonic) 466 return Mnemonic < RHS.Mnemonic; 467 468 if (AsmOperands.size() != RHS.AsmOperands.size()) 469 return AsmOperands.size() < RHS.AsmOperands.size(); 470 471 // Compare lexicographically by operand. The matcher validates that other 472 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith(). 473 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 474 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 475 return true; 476 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 477 return false; 478 } 479 480 return false; 481 } 482 483 /// couldMatchAmbiguouslyWith - Check whether this matchable could 484 /// ambiguously match the same set of operands as \arg RHS (without being a 485 /// strictly superior match). 486 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) { 487 // The primary comparator is the instruction mnemonic. 488 if (Mnemonic != RHS.Mnemonic) 489 return false; 490 491 // The number of operands is unambiguous. 492 if (AsmOperands.size() != RHS.AsmOperands.size()) 493 return false; 494 495 // Otherwise, make sure the ordering of the two instructions is unambiguous 496 // by checking that either (a) a token or operand kind discriminates them, 497 // or (b) the ordering among equivalent kinds is consistent. 498 499 // Tokens and operand kinds are unambiguous (assuming a correct target 500 // specific parser). 501 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) 502 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind || 503 AsmOperands[i].Class->Kind == ClassInfo::Token) 504 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class || 505 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 506 return false; 507 508 // Otherwise, this operand could commute if all operands are equivalent, or 509 // there is a pair of operands that compare less than and a pair that 510 // compare greater than. 511 bool HasLT = false, HasGT = false; 512 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 513 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class) 514 HasLT = true; 515 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class) 516 HasGT = true; 517 } 518 519 return !(HasLT ^ HasGT); 520 } 521 522 void dump(); 523 524 private: 525 void tokenizeAsmString(const AsmMatcherInfo &Info); 526 }; 527 528 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget 529 /// feature which participates in instruction matching. 530 struct SubtargetFeatureInfo { 531 /// \brief The predicate record for this feature. 532 Record *TheDef; 533 534 /// \brief An unique index assigned to represent this feature. 535 unsigned Index; 536 537 SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {} 538 539 /// \brief The name of the enumerated constant identifying this feature. 540 std::string getEnumName() const { 541 return "Feature_" + TheDef->getName(); 542 } 543 }; 544 545 struct OperandMatchEntry { 546 unsigned OperandMask; 547 MatchableInfo* MI; 548 ClassInfo *CI; 549 550 static OperandMatchEntry create(MatchableInfo* mi, ClassInfo *ci, 551 unsigned opMask) { 552 OperandMatchEntry X; 553 X.OperandMask = opMask; 554 X.CI = ci; 555 X.MI = mi; 556 return X; 557 } 558 }; 559 560 561 class AsmMatcherInfo { 562 public: 563 /// Tracked Records 564 RecordKeeper &Records; 565 566 /// The tablegen AsmParser record. 567 Record *AsmParser; 568 569 /// Target - The target information. 570 CodeGenTarget &Target; 571 572 /// The classes which are needed for matching. 573 std::vector<ClassInfo*> Classes; 574 575 /// The information on the matchables to match. 576 std::vector<MatchableInfo*> Matchables; 577 578 /// Info for custom matching operands by user defined methods. 579 std::vector<OperandMatchEntry> OperandMatchInfo; 580 581 /// Map of Register records to their class information. 582 std::map<Record*, ClassInfo*> RegisterClasses; 583 584 /// Map of Predicate records to their subtarget information. 585 std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures; 586 587 private: 588 /// Map of token to class information which has already been constructed. 589 std::map<std::string, ClassInfo*> TokenClasses; 590 591 /// Map of RegisterClass records to their class information. 592 std::map<Record*, ClassInfo*> RegisterClassClasses; 593 594 /// Map of AsmOperandClass records to their class information. 595 std::map<Record*, ClassInfo*> AsmOperandClasses; 596 597 private: 598 /// getTokenClass - Lookup or create the class for the given token. 599 ClassInfo *getTokenClass(StringRef Token); 600 601 /// getOperandClass - Lookup or create the class for the given operand. 602 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI, 603 int SubOpIdx); 604 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx); 605 606 /// buildRegisterClasses - Build the ClassInfo* instances for register 607 /// classes. 608 void buildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters); 609 610 /// buildOperandClasses - Build the ClassInfo* instances for user defined 611 /// operand classes. 612 void buildOperandClasses(); 613 614 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName, 615 unsigned AsmOpIdx); 616 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName, 617 MatchableInfo::AsmOperand &Op); 618 619 public: 620 AsmMatcherInfo(Record *AsmParser, 621 CodeGenTarget &Target, 622 RecordKeeper &Records); 623 624 /// buildInfo - Construct the various tables used during matching. 625 void buildInfo(); 626 627 /// buildOperandMatchInfo - Build the necessary information to handle user 628 /// defined operand parsing methods. 629 void buildOperandMatchInfo(); 630 631 /// getSubtargetFeature - Lookup or create the subtarget feature info for the 632 /// given operand. 633 SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const { 634 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!"); 635 std::map<Record*, SubtargetFeatureInfo*>::const_iterator I = 636 SubtargetFeatures.find(Def); 637 return I == SubtargetFeatures.end() ? 0 : I->second; 638 } 639 640 RecordKeeper &getRecords() const { 641 return Records; 642 } 643 }; 644 645 } 646 647 void MatchableInfo::dump() { 648 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n"; 649 650 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 651 AsmOperand &Op = AsmOperands[i]; 652 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - "; 653 errs() << '\"' << Op.Token << "\"\n"; 654 } 655 } 656 657 static std::pair<StringRef, StringRef> 658 parseTwoOperandConstraint(StringRef S, SMLoc Loc) { 659 // Split via the '='. 660 std::pair<StringRef, StringRef> Ops = S.split('='); 661 if (Ops.second == "") 662 throw TGError(Loc, "missing '=' in two-operand alias constraint"); 663 // Trim whitespace and the leading '$' on the operand names. 664 size_t start = Ops.first.find_first_of('$'); 665 if (start == std::string::npos) 666 throw TGError(Loc, "expected '$' prefix on asm operand name"); 667 Ops.first = Ops.first.slice(start + 1, std::string::npos); 668 size_t end = Ops.first.find_last_of(" \t"); 669 Ops.first = Ops.first.slice(0, end); 670 // Now the second operand. 671 start = Ops.second.find_first_of('$'); 672 if (start == std::string::npos) 673 throw TGError(Loc, "expected '$' prefix on asm operand name"); 674 Ops.second = Ops.second.slice(start + 1, std::string::npos); 675 end = Ops.second.find_last_of(" \t"); 676 Ops.first = Ops.first.slice(0, end); 677 return Ops; 678 } 679 680 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) { 681 // Figure out which operands are aliased and mark them as tied. 682 std::pair<StringRef, StringRef> Ops = 683 parseTwoOperandConstraint(Constraint, TheDef->getLoc()); 684 685 // Find the AsmOperands that refer to the operands we're aliasing. 686 int SrcAsmOperand = findAsmOperandNamed(Ops.first); 687 int DstAsmOperand = findAsmOperandNamed(Ops.second); 688 if (SrcAsmOperand == -1) 689 throw TGError(TheDef->getLoc(), 690 "unknown source two-operand alias operand '" + 691 Ops.first.str() + "'."); 692 if (DstAsmOperand == -1) 693 throw TGError(TheDef->getLoc(), 694 "unknown destination two-operand alias operand '" + 695 Ops.second.str() + "'."); 696 697 // Find the ResOperand that refers to the operand we're aliasing away 698 // and update it to refer to the combined operand instead. 699 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) { 700 ResOperand &Op = ResOperands[i]; 701 if (Op.Kind == ResOperand::RenderAsmOperand && 702 Op.AsmOperandNum == (unsigned)SrcAsmOperand) { 703 Op.AsmOperandNum = DstAsmOperand; 704 break; 705 } 706 } 707 // Remove the AsmOperand for the alias operand. 708 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand); 709 // Adjust the ResOperand references to any AsmOperands that followed 710 // the one we just deleted. 711 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) { 712 ResOperand &Op = ResOperands[i]; 713 switch(Op.Kind) { 714 default: 715 // Nothing to do for operands that don't reference AsmOperands. 716 break; 717 case ResOperand::RenderAsmOperand: 718 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand) 719 --Op.AsmOperandNum; 720 break; 721 case ResOperand::TiedOperand: 722 if (Op.TiedOperandNum > (unsigned)SrcAsmOperand) 723 --Op.TiedOperandNum; 724 break; 725 } 726 } 727 } 728 729 void MatchableInfo::initialize(const AsmMatcherInfo &Info, 730 SmallPtrSet<Record*, 16> &SingletonRegisters, 731 int AsmVariantNo, std::string &RegisterPrefix) { 732 AsmVariantID = AsmVariantNo; 733 AsmString = 734 CodeGenInstruction::FlattenAsmStringVariants(AsmString, AsmVariantNo); 735 736 tokenizeAsmString(Info); 737 738 // Compute the require features. 739 std::vector<Record*> Predicates =TheDef->getValueAsListOfDefs("Predicates"); 740 for (unsigned i = 0, e = Predicates.size(); i != e; ++i) 741 if (SubtargetFeatureInfo *Feature = 742 Info.getSubtargetFeature(Predicates[i])) 743 RequiredFeatures.push_back(Feature); 744 745 // Collect singleton registers, if used. 746 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 747 extractSingletonRegisterForAsmOperand(i, Info, RegisterPrefix); 748 if (Record *Reg = AsmOperands[i].SingletonReg) 749 SingletonRegisters.insert(Reg); 750 } 751 } 752 753 /// tokenizeAsmString - Tokenize a simplified assembly string. 754 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info) { 755 StringRef String = AsmString; 756 unsigned Prev = 0; 757 bool InTok = true; 758 for (unsigned i = 0, e = String.size(); i != e; ++i) { 759 switch (String[i]) { 760 case '[': 761 case ']': 762 case '*': 763 case '!': 764 case ' ': 765 case '\t': 766 case ',': 767 if (InTok) { 768 AsmOperands.push_back(AsmOperand(String.slice(Prev, i))); 769 InTok = false; 770 } 771 if (!isspace(String[i]) && String[i] != ',') 772 AsmOperands.push_back(AsmOperand(String.substr(i, 1))); 773 Prev = i + 1; 774 break; 775 776 case '\\': 777 if (InTok) { 778 AsmOperands.push_back(AsmOperand(String.slice(Prev, i))); 779 InTok = false; 780 } 781 ++i; 782 assert(i != String.size() && "Invalid quoted character"); 783 AsmOperands.push_back(AsmOperand(String.substr(i, 1))); 784 Prev = i + 1; 785 break; 786 787 case '$': { 788 if (InTok) { 789 AsmOperands.push_back(AsmOperand(String.slice(Prev, i))); 790 InTok = false; 791 } 792 793 // If this isn't "${", treat like a normal token. 794 if (i + 1 == String.size() || String[i + 1] != '{') { 795 Prev = i; 796 break; 797 } 798 799 StringRef::iterator End = std::find(String.begin() + i, String.end(),'}'); 800 assert(End != String.end() && "Missing brace in operand reference!"); 801 size_t EndPos = End - String.begin(); 802 AsmOperands.push_back(AsmOperand(String.slice(i, EndPos+1))); 803 Prev = EndPos + 1; 804 i = EndPos; 805 break; 806 } 807 808 case '.': 809 if (InTok) 810 AsmOperands.push_back(AsmOperand(String.slice(Prev, i))); 811 Prev = i; 812 InTok = true; 813 break; 814 815 default: 816 InTok = true; 817 } 818 } 819 if (InTok && Prev != String.size()) 820 AsmOperands.push_back(AsmOperand(String.substr(Prev))); 821 822 // The first token of the instruction is the mnemonic, which must be a 823 // simple string, not a $foo variable or a singleton register. 824 if (AsmOperands.empty()) 825 throw TGError(TheDef->getLoc(), 826 "Instruction '" + TheDef->getName() + "' has no tokens"); 827 Mnemonic = AsmOperands[0].Token; 828 // FIXME : Check and raise an error if it is a register. 829 if (Mnemonic[0] == '$') 830 throw TGError(TheDef->getLoc(), 831 "Invalid instruction mnemonic '" + Mnemonic.str() + "'!"); 832 833 // Remove the first operand, it is tracked in the mnemonic field. 834 AsmOperands.erase(AsmOperands.begin()); 835 } 836 837 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const { 838 // Reject matchables with no .s string. 839 if (AsmString.empty()) 840 throw TGError(TheDef->getLoc(), "instruction with empty asm string"); 841 842 // Reject any matchables with a newline in them, they should be marked 843 // isCodeGenOnly if they are pseudo instructions. 844 if (AsmString.find('\n') != std::string::npos) 845 throw TGError(TheDef->getLoc(), 846 "multiline instruction is not valid for the asmparser, " 847 "mark it isCodeGenOnly"); 848 849 // Remove comments from the asm string. We know that the asmstring only 850 // has one line. 851 if (!CommentDelimiter.empty() && 852 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos) 853 throw TGError(TheDef->getLoc(), 854 "asmstring for instruction has comment character in it, " 855 "mark it isCodeGenOnly"); 856 857 // Reject matchables with operand modifiers, these aren't something we can 858 // handle, the target should be refactored to use operands instead of 859 // modifiers. 860 // 861 // Also, check for instructions which reference the operand multiple times; 862 // this implies a constraint we would not honor. 863 std::set<std::string> OperandNames; 864 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) { 865 StringRef Tok = AsmOperands[i].Token; 866 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos) 867 throw TGError(TheDef->getLoc(), 868 "matchable with operand modifier '" + Tok.str() + 869 "' not supported by asm matcher. Mark isCodeGenOnly!"); 870 871 // Verify that any operand is only mentioned once. 872 // We reject aliases and ignore instructions for now. 873 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) { 874 if (!Hack) 875 throw TGError(TheDef->getLoc(), 876 "ERROR: matchable with tied operand '" + Tok.str() + 877 "' can never be matched!"); 878 // FIXME: Should reject these. The ARM backend hits this with $lane in a 879 // bunch of instructions. It is unclear what the right answer is. 880 DEBUG({ 881 errs() << "warning: '" << TheDef->getName() << "': " 882 << "ignoring instruction with tied operand '" 883 << Tok.str() << "'\n"; 884 }); 885 return false; 886 } 887 } 888 889 return true; 890 } 891 892 /// extractSingletonRegisterForAsmOperand - Extract singleton register, 893 /// if present, from specified token. 894 void MatchableInfo:: 895 extractSingletonRegisterForAsmOperand(unsigned OperandNo, 896 const AsmMatcherInfo &Info, 897 std::string &RegisterPrefix) { 898 StringRef Tok = AsmOperands[OperandNo].Token; 899 if (RegisterPrefix.empty()) { 900 std::string LoweredTok = Tok.lower(); 901 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok)) 902 AsmOperands[OperandNo].SingletonReg = Reg->TheDef; 903 return; 904 } 905 906 if (!Tok.startswith(RegisterPrefix)) 907 return; 908 909 StringRef RegName = Tok.substr(RegisterPrefix.size()); 910 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName)) 911 AsmOperands[OperandNo].SingletonReg = Reg->TheDef; 912 913 // If there is no register prefix (i.e. "%" in "%eax"), then this may 914 // be some random non-register token, just ignore it. 915 return; 916 } 917 918 static std::string getEnumNameForToken(StringRef Str) { 919 std::string Res; 920 921 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) { 922 switch (*it) { 923 case '*': Res += "_STAR_"; break; 924 case '%': Res += "_PCT_"; break; 925 case ':': Res += "_COLON_"; break; 926 case '!': Res += "_EXCLAIM_"; break; 927 case '.': Res += "_DOT_"; break; 928 default: 929 if (isalnum(*it)) 930 Res += *it; 931 else 932 Res += "_" + utostr((unsigned) *it) + "_"; 933 } 934 } 935 936 return Res; 937 } 938 939 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) { 940 ClassInfo *&Entry = TokenClasses[Token]; 941 942 if (!Entry) { 943 Entry = new ClassInfo(); 944 Entry->Kind = ClassInfo::Token; 945 Entry->ClassName = "Token"; 946 Entry->Name = "MCK_" + getEnumNameForToken(Token); 947 Entry->ValueName = Token; 948 Entry->PredicateMethod = "<invalid>"; 949 Entry->RenderMethod = "<invalid>"; 950 Entry->ParserMethod = ""; 951 Classes.push_back(Entry); 952 } 953 954 return Entry; 955 } 956 957 ClassInfo * 958 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI, 959 int SubOpIdx) { 960 Record *Rec = OI.Rec; 961 if (SubOpIdx != -1) 962 Rec = dynamic_cast<DefInit*>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef(); 963 return getOperandClass(Rec, SubOpIdx); 964 } 965 966 ClassInfo * 967 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) { 968 if (Rec->isSubClassOf("RegisterOperand")) { 969 // RegisterOperand may have an associated ParserMatchClass. If it does, 970 // use it, else just fall back to the underlying register class. 971 const RecordVal *R = Rec->getValue("ParserMatchClass"); 972 if (R == 0 || R->getValue() == 0) 973 throw "Record `" + Rec->getName() + 974 "' does not have a ParserMatchClass!\n"; 975 976 if (DefInit *DI= dynamic_cast<DefInit*>(R->getValue())) { 977 Record *MatchClass = DI->getDef(); 978 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 979 return CI; 980 } 981 982 // No custom match class. Just use the register class. 983 Record *ClassRec = Rec->getValueAsDef("RegClass"); 984 if (!ClassRec) 985 throw TGError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() + 986 "' has no associated register class!\n"); 987 if (ClassInfo *CI = RegisterClassClasses[ClassRec]) 988 return CI; 989 throw TGError(Rec->getLoc(), "register class has no class info!"); 990 } 991 992 993 if (Rec->isSubClassOf("RegisterClass")) { 994 if (ClassInfo *CI = RegisterClassClasses[Rec]) 995 return CI; 996 throw TGError(Rec->getLoc(), "register class has no class info!"); 997 } 998 999 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!"); 1000 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1001 if (ClassInfo *CI = AsmOperandClasses[MatchClass]) 1002 return CI; 1003 1004 throw TGError(Rec->getLoc(), "operand has no match class!"); 1005 } 1006 1007 void AsmMatcherInfo:: 1008 buildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters) { 1009 const std::vector<CodeGenRegister*> &Registers = 1010 Target.getRegBank().getRegisters(); 1011 ArrayRef<CodeGenRegisterClass*> RegClassList = 1012 Target.getRegBank().getRegClasses(); 1013 1014 // The register sets used for matching. 1015 std::set< std::set<Record*> > RegisterSets; 1016 1017 // Gather the defined sets. 1018 for (ArrayRef<CodeGenRegisterClass*>::const_iterator it = 1019 RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it) 1020 RegisterSets.insert(std::set<Record*>( 1021 (*it)->getOrder().begin(), (*it)->getOrder().end())); 1022 1023 // Add any required singleton sets. 1024 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(), 1025 ie = SingletonRegisters.end(); it != ie; ++it) { 1026 Record *Rec = *it; 1027 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1)); 1028 } 1029 1030 // Introduce derived sets where necessary (when a register does not determine 1031 // a unique register set class), and build the mapping of registers to the set 1032 // they should classify to. 1033 std::map<Record*, std::set<Record*> > RegisterMap; 1034 for (std::vector<CodeGenRegister*>::const_iterator it = Registers.begin(), 1035 ie = Registers.end(); it != ie; ++it) { 1036 const CodeGenRegister &CGR = **it; 1037 // Compute the intersection of all sets containing this register. 1038 std::set<Record*> ContainingSet; 1039 1040 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 1041 ie = RegisterSets.end(); it != ie; ++it) { 1042 if (!it->count(CGR.TheDef)) 1043 continue; 1044 1045 if (ContainingSet.empty()) { 1046 ContainingSet = *it; 1047 continue; 1048 } 1049 1050 std::set<Record*> Tmp; 1051 std::swap(Tmp, ContainingSet); 1052 std::insert_iterator< std::set<Record*> > II(ContainingSet, 1053 ContainingSet.begin()); 1054 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(), II); 1055 } 1056 1057 if (!ContainingSet.empty()) { 1058 RegisterSets.insert(ContainingSet); 1059 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet)); 1060 } 1061 } 1062 1063 // Construct the register classes. 1064 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses; 1065 unsigned Index = 0; 1066 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 1067 ie = RegisterSets.end(); it != ie; ++it, ++Index) { 1068 ClassInfo *CI = new ClassInfo(); 1069 CI->Kind = ClassInfo::RegisterClass0 + Index; 1070 CI->ClassName = "Reg" + utostr(Index); 1071 CI->Name = "MCK_Reg" + utostr(Index); 1072 CI->ValueName = ""; 1073 CI->PredicateMethod = ""; // unused 1074 CI->RenderMethod = "addRegOperands"; 1075 CI->Registers = *it; 1076 Classes.push_back(CI); 1077 RegisterSetClasses.insert(std::make_pair(*it, CI)); 1078 } 1079 1080 // Find the superclasses; we could compute only the subgroup lattice edges, 1081 // but there isn't really a point. 1082 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 1083 ie = RegisterSets.end(); it != ie; ++it) { 1084 ClassInfo *CI = RegisterSetClasses[*it]; 1085 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(), 1086 ie2 = RegisterSets.end(); it2 != ie2; ++it2) 1087 if (*it != *it2 && 1088 std::includes(it2->begin(), it2->end(), it->begin(), it->end())) 1089 CI->SuperClasses.push_back(RegisterSetClasses[*it2]); 1090 } 1091 1092 // Name the register classes which correspond to a user defined RegisterClass. 1093 for (ArrayRef<CodeGenRegisterClass*>::const_iterator 1094 it = RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it) { 1095 const CodeGenRegisterClass &RC = **it; 1096 // Def will be NULL for non-user defined register classes. 1097 Record *Def = RC.getDef(); 1098 if (!Def) 1099 continue; 1100 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(RC.getOrder().begin(), 1101 RC.getOrder().end())]; 1102 if (CI->ValueName.empty()) { 1103 CI->ClassName = RC.getName(); 1104 CI->Name = "MCK_" + RC.getName(); 1105 CI->ValueName = RC.getName(); 1106 } else 1107 CI->ValueName = CI->ValueName + "," + RC.getName(); 1108 1109 RegisterClassClasses.insert(std::make_pair(Def, CI)); 1110 } 1111 1112 // Populate the map for individual registers. 1113 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(), 1114 ie = RegisterMap.end(); it != ie; ++it) 1115 RegisterClasses[it->first] = RegisterSetClasses[it->second]; 1116 1117 // Name the register classes which correspond to singleton registers. 1118 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(), 1119 ie = SingletonRegisters.end(); it != ie; ++it) { 1120 Record *Rec = *it; 1121 ClassInfo *CI = RegisterClasses[Rec]; 1122 assert(CI && "Missing singleton register class info!"); 1123 1124 if (CI->ValueName.empty()) { 1125 CI->ClassName = Rec->getName(); 1126 CI->Name = "MCK_" + Rec->getName(); 1127 CI->ValueName = Rec->getName(); 1128 } else 1129 CI->ValueName = CI->ValueName + "," + Rec->getName(); 1130 } 1131 } 1132 1133 void AsmMatcherInfo::buildOperandClasses() { 1134 std::vector<Record*> AsmOperands = 1135 Records.getAllDerivedDefinitions("AsmOperandClass"); 1136 1137 // Pre-populate AsmOperandClasses map. 1138 for (std::vector<Record*>::iterator it = AsmOperands.begin(), 1139 ie = AsmOperands.end(); it != ie; ++it) 1140 AsmOperandClasses[*it] = new ClassInfo(); 1141 1142 unsigned Index = 0; 1143 for (std::vector<Record*>::iterator it = AsmOperands.begin(), 1144 ie = AsmOperands.end(); it != ie; ++it, ++Index) { 1145 ClassInfo *CI = AsmOperandClasses[*it]; 1146 CI->Kind = ClassInfo::UserClass0 + Index; 1147 1148 ListInit *Supers = (*it)->getValueAsListInit("SuperClasses"); 1149 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) { 1150 DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i)); 1151 if (!DI) { 1152 PrintError((*it)->getLoc(), "Invalid super class reference!"); 1153 continue; 1154 } 1155 1156 ClassInfo *SC = AsmOperandClasses[DI->getDef()]; 1157 if (!SC) 1158 PrintError((*it)->getLoc(), "Invalid super class reference!"); 1159 else 1160 CI->SuperClasses.push_back(SC); 1161 } 1162 CI->ClassName = (*it)->getValueAsString("Name"); 1163 CI->Name = "MCK_" + CI->ClassName; 1164 CI->ValueName = (*it)->getName(); 1165 1166 // Get or construct the predicate method name. 1167 Init *PMName = (*it)->getValueInit("PredicateMethod"); 1168 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) { 1169 CI->PredicateMethod = SI->getValue(); 1170 } else { 1171 assert(dynamic_cast<UnsetInit*>(PMName) && 1172 "Unexpected PredicateMethod field!"); 1173 CI->PredicateMethod = "is" + CI->ClassName; 1174 } 1175 1176 // Get or construct the render method name. 1177 Init *RMName = (*it)->getValueInit("RenderMethod"); 1178 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) { 1179 CI->RenderMethod = SI->getValue(); 1180 } else { 1181 assert(dynamic_cast<UnsetInit*>(RMName) && 1182 "Unexpected RenderMethod field!"); 1183 CI->RenderMethod = "add" + CI->ClassName + "Operands"; 1184 } 1185 1186 // Get the parse method name or leave it as empty. 1187 Init *PRMName = (*it)->getValueInit("ParserMethod"); 1188 if (StringInit *SI = dynamic_cast<StringInit*>(PRMName)) 1189 CI->ParserMethod = SI->getValue(); 1190 1191 AsmOperandClasses[*it] = CI; 1192 Classes.push_back(CI); 1193 } 1194 } 1195 1196 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, 1197 CodeGenTarget &target, 1198 RecordKeeper &records) 1199 : Records(records), AsmParser(asmParser), Target(target) { 1200 } 1201 1202 /// buildOperandMatchInfo - Build the necessary information to handle user 1203 /// defined operand parsing methods. 1204 void AsmMatcherInfo::buildOperandMatchInfo() { 1205 1206 /// Map containing a mask with all operands indices that can be found for 1207 /// that class inside a instruction. 1208 std::map<ClassInfo*, unsigned> OpClassMask; 1209 1210 for (std::vector<MatchableInfo*>::const_iterator it = 1211 Matchables.begin(), ie = Matchables.end(); 1212 it != ie; ++it) { 1213 MatchableInfo &II = **it; 1214 OpClassMask.clear(); 1215 1216 // Keep track of all operands of this instructions which belong to the 1217 // same class. 1218 for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) { 1219 MatchableInfo::AsmOperand &Op = II.AsmOperands[i]; 1220 if (Op.Class->ParserMethod.empty()) 1221 continue; 1222 unsigned &OperandMask = OpClassMask[Op.Class]; 1223 OperandMask |= (1 << i); 1224 } 1225 1226 // Generate operand match info for each mnemonic/operand class pair. 1227 for (std::map<ClassInfo*, unsigned>::iterator iit = OpClassMask.begin(), 1228 iie = OpClassMask.end(); iit != iie; ++iit) { 1229 unsigned OpMask = iit->second; 1230 ClassInfo *CI = iit->first; 1231 OperandMatchInfo.push_back(OperandMatchEntry::create(&II, CI, OpMask)); 1232 } 1233 } 1234 } 1235 1236 void AsmMatcherInfo::buildInfo() { 1237 // Build information about all of the AssemblerPredicates. 1238 std::vector<Record*> AllPredicates = 1239 Records.getAllDerivedDefinitions("Predicate"); 1240 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) { 1241 Record *Pred = AllPredicates[i]; 1242 // Ignore predicates that are not intended for the assembler. 1243 if (!Pred->getValueAsBit("AssemblerMatcherPredicate")) 1244 continue; 1245 1246 if (Pred->getName().empty()) 1247 throw TGError(Pred->getLoc(), "Predicate has no name!"); 1248 1249 unsigned FeatureNo = SubtargetFeatures.size(); 1250 SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo); 1251 assert(FeatureNo < 32 && "Too many subtarget features!"); 1252 } 1253 1254 // Parse the instructions; we need to do this first so that we can gather the 1255 // singleton register classes. 1256 SmallPtrSet<Record*, 16> SingletonRegisters; 1257 unsigned VariantCount = Target.getAsmParserVariantCount(); 1258 for (unsigned VC = 0; VC != VariantCount; ++VC) { 1259 Record *AsmVariant = Target.getAsmParserVariant(VC); 1260 std::string CommentDelimiter = 1261 AsmVariant->getValueAsString("CommentDelimiter"); 1262 std::string RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix"); 1263 int AsmVariantNo = AsmVariant->getValueAsInt("Variant"); 1264 1265 for (CodeGenTarget::inst_iterator I = Target.inst_begin(), 1266 E = Target.inst_end(); I != E; ++I) { 1267 const CodeGenInstruction &CGI = **I; 1268 1269 // If the tblgen -match-prefix option is specified (for tblgen hackers), 1270 // filter the set of instructions we consider. 1271 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix)) 1272 continue; 1273 1274 // Ignore "codegen only" instructions. 1275 if (CGI.TheDef->getValueAsBit("isCodeGenOnly")) 1276 continue; 1277 1278 // Validate the operand list to ensure we can handle this instruction. 1279 for (unsigned i = 0, e = CGI.Operands.size(); i != e; ++i) { 1280 const CGIOperandList::OperandInfo &OI = CGI.Operands[i]; 1281 1282 // Validate tied operands. 1283 if (OI.getTiedRegister() != -1) { 1284 // If we have a tied operand that consists of multiple MCOperands, 1285 // reject it. We reject aliases and ignore instructions for now. 1286 if (OI.MINumOperands != 1) { 1287 // FIXME: Should reject these. The ARM backend hits this with $lane 1288 // in a bunch of instructions. The right answer is unclear. 1289 DEBUG({ 1290 errs() << "warning: '" << CGI.TheDef->getName() << "': " 1291 << "ignoring instruction with multi-operand tied operand '" 1292 << OI.Name << "'\n"; 1293 }); 1294 continue; 1295 } 1296 } 1297 } 1298 1299 OwningPtr<MatchableInfo> II(new MatchableInfo(CGI)); 1300 1301 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix); 1302 1303 // Ignore instructions which shouldn't be matched and diagnose invalid 1304 // instruction definitions with an error. 1305 if (!II->validate(CommentDelimiter, true)) 1306 continue; 1307 1308 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases. 1309 // 1310 // FIXME: This is a total hack. 1311 if (StringRef(II->TheDef->getName()).startswith("Int_") || 1312 StringRef(II->TheDef->getName()).endswith("_Int")) 1313 continue; 1314 1315 Matchables.push_back(II.take()); 1316 } 1317 1318 // Parse all of the InstAlias definitions and stick them in the list of 1319 // matchables. 1320 std::vector<Record*> AllInstAliases = 1321 Records.getAllDerivedDefinitions("InstAlias"); 1322 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) { 1323 CodeGenInstAlias *Alias = new CodeGenInstAlias(AllInstAliases[i], Target); 1324 1325 // If the tblgen -match-prefix option is specified (for tblgen hackers), 1326 // filter the set of instruction aliases we consider, based on the target 1327 // instruction. 1328 if (!StringRef(Alias->ResultInst->TheDef->getName()) 1329 .startswith( MatchPrefix)) 1330 continue; 1331 1332 OwningPtr<MatchableInfo> II(new MatchableInfo(Alias)); 1333 1334 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix); 1335 1336 // Validate the alias definitions. 1337 II->validate(CommentDelimiter, false); 1338 1339 Matchables.push_back(II.take()); 1340 } 1341 } 1342 1343 // Build info for the register classes. 1344 buildRegisterClasses(SingletonRegisters); 1345 1346 // Build info for the user defined assembly operand classes. 1347 buildOperandClasses(); 1348 1349 // Build the information about matchables, now that we have fully formed 1350 // classes. 1351 std::vector<MatchableInfo*> NewMatchables; 1352 for (std::vector<MatchableInfo*>::iterator it = Matchables.begin(), 1353 ie = Matchables.end(); it != ie; ++it) { 1354 MatchableInfo *II = *it; 1355 1356 // Parse the tokens after the mnemonic. 1357 // Note: buildInstructionOperandReference may insert new AsmOperands, so 1358 // don't precompute the loop bound. 1359 for (unsigned i = 0; i != II->AsmOperands.size(); ++i) { 1360 MatchableInfo::AsmOperand &Op = II->AsmOperands[i]; 1361 StringRef Token = Op.Token; 1362 1363 // Check for singleton registers. 1364 if (Record *RegRecord = II->AsmOperands[i].SingletonReg) { 1365 Op.Class = RegisterClasses[RegRecord]; 1366 assert(Op.Class && Op.Class->Registers.size() == 1 && 1367 "Unexpected class for singleton register"); 1368 continue; 1369 } 1370 1371 // Check for simple tokens. 1372 if (Token[0] != '$') { 1373 Op.Class = getTokenClass(Token); 1374 continue; 1375 } 1376 1377 if (Token.size() > 1 && isdigit(Token[1])) { 1378 Op.Class = getTokenClass(Token); 1379 continue; 1380 } 1381 1382 // Otherwise this is an operand reference. 1383 StringRef OperandName; 1384 if (Token[1] == '{') 1385 OperandName = Token.substr(2, Token.size() - 3); 1386 else 1387 OperandName = Token.substr(1); 1388 1389 if (II->DefRec.is<const CodeGenInstruction*>()) 1390 buildInstructionOperandReference(II, OperandName, i); 1391 else 1392 buildAliasOperandReference(II, OperandName, Op); 1393 } 1394 1395 if (II->DefRec.is<const CodeGenInstruction*>()) { 1396 II->buildInstructionResultOperands(); 1397 // If the instruction has a two-operand alias, build up the 1398 // matchable here. We'll add them in bulk at the end to avoid 1399 // confusing this loop. 1400 std::string Constraint = 1401 II->TheDef->getValueAsString("TwoOperandAliasConstraint"); 1402 if (Constraint != "") { 1403 // Start by making a copy of the original matchable. 1404 OwningPtr<MatchableInfo> AliasII(new MatchableInfo(*II)); 1405 1406 // Adjust it to be a two-operand alias. 1407 AliasII->formTwoOperandAlias(Constraint); 1408 1409 // Add the alias to the matchables list. 1410 NewMatchables.push_back(AliasII.take()); 1411 } 1412 } else 1413 II->buildAliasResultOperands(); 1414 } 1415 if (!NewMatchables.empty()) 1416 Matchables.insert(Matchables.end(), NewMatchables.begin(), 1417 NewMatchables.end()); 1418 1419 // Process token alias definitions and set up the associated superclass 1420 // information. 1421 std::vector<Record*> AllTokenAliases = 1422 Records.getAllDerivedDefinitions("TokenAlias"); 1423 for (unsigned i = 0, e = AllTokenAliases.size(); i != e; ++i) { 1424 Record *Rec = AllTokenAliases[i]; 1425 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken")); 1426 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken")); 1427 if (FromClass == ToClass) 1428 throw TGError(Rec->getLoc(), 1429 "error: Destination value identical to source value."); 1430 FromClass->SuperClasses.push_back(ToClass); 1431 } 1432 1433 // Reorder classes so that classes precede super classes. 1434 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>()); 1435 } 1436 1437 /// buildInstructionOperandReference - The specified operand is a reference to a 1438 /// named operand such as $src. Resolve the Class and OperandInfo pointers. 1439 void AsmMatcherInfo:: 1440 buildInstructionOperandReference(MatchableInfo *II, 1441 StringRef OperandName, 1442 unsigned AsmOpIdx) { 1443 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>(); 1444 const CGIOperandList &Operands = CGI.Operands; 1445 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx]; 1446 1447 // Map this token to an operand. 1448 unsigned Idx; 1449 if (!Operands.hasOperandNamed(OperandName, Idx)) 1450 throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" + 1451 OperandName.str() + "'"); 1452 1453 // If the instruction operand has multiple suboperands, but the parser 1454 // match class for the asm operand is still the default "ImmAsmOperand", 1455 // then handle each suboperand separately. 1456 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) { 1457 Record *Rec = Operands[Idx].Rec; 1458 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!"); 1459 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass"); 1460 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") { 1461 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands. 1462 StringRef Token = Op->Token; // save this in case Op gets moved 1463 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) { 1464 MatchableInfo::AsmOperand NewAsmOp(Token); 1465 NewAsmOp.SubOpIdx = SI; 1466 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp); 1467 } 1468 // Replace Op with first suboperand. 1469 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved 1470 Op->SubOpIdx = 0; 1471 } 1472 } 1473 1474 // Set up the operand class. 1475 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx); 1476 1477 // If the named operand is tied, canonicalize it to the untied operand. 1478 // For example, something like: 1479 // (outs GPR:$dst), (ins GPR:$src) 1480 // with an asmstring of 1481 // "inc $src" 1482 // we want to canonicalize to: 1483 // "inc $dst" 1484 // so that we know how to provide the $dst operand when filling in the result. 1485 int OITied = Operands[Idx].getTiedRegister(); 1486 if (OITied != -1) { 1487 // The tied operand index is an MIOperand index, find the operand that 1488 // contains it. 1489 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied); 1490 OperandName = Operands[Idx.first].Name; 1491 Op->SubOpIdx = Idx.second; 1492 } 1493 1494 Op->SrcOpName = OperandName; 1495 } 1496 1497 /// buildAliasOperandReference - When parsing an operand reference out of the 1498 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the 1499 /// operand reference is by looking it up in the result pattern definition. 1500 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II, 1501 StringRef OperandName, 1502 MatchableInfo::AsmOperand &Op) { 1503 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>(); 1504 1505 // Set up the operand class. 1506 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i) 1507 if (CGA.ResultOperands[i].isRecord() && 1508 CGA.ResultOperands[i].getName() == OperandName) { 1509 // It's safe to go with the first one we find, because CodeGenInstAlias 1510 // validates that all operands with the same name have the same record. 1511 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second; 1512 // Use the match class from the Alias definition, not the 1513 // destination instruction, as we may have an immediate that's 1514 // being munged by the match class. 1515 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(), 1516 Op.SubOpIdx); 1517 Op.SrcOpName = OperandName; 1518 return; 1519 } 1520 1521 throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" + 1522 OperandName.str() + "'"); 1523 } 1524 1525 void MatchableInfo::buildInstructionResultOperands() { 1526 const CodeGenInstruction *ResultInst = getResultInst(); 1527 1528 // Loop over all operands of the result instruction, determining how to 1529 // populate them. 1530 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) { 1531 const CGIOperandList::OperandInfo &OpInfo = ResultInst->Operands[i]; 1532 1533 // If this is a tied operand, just copy from the previously handled operand. 1534 int TiedOp = OpInfo.getTiedRegister(); 1535 if (TiedOp != -1) { 1536 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1537 continue; 1538 } 1539 1540 // Find out what operand from the asmparser this MCInst operand comes from. 1541 int SrcOperand = findAsmOperandNamed(OpInfo.Name); 1542 if (OpInfo.Name.empty() || SrcOperand == -1) 1543 throw TGError(TheDef->getLoc(), "Instruction '" + 1544 TheDef->getName() + "' has operand '" + OpInfo.Name + 1545 "' that doesn't appear in asm string!"); 1546 1547 // Check if the one AsmOperand populates the entire operand. 1548 unsigned NumOperands = OpInfo.MINumOperands; 1549 if (AsmOperands[SrcOperand].SubOpIdx == -1) { 1550 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands)); 1551 continue; 1552 } 1553 1554 // Add a separate ResOperand for each suboperand. 1555 for (unsigned AI = 0; AI < NumOperands; ++AI) { 1556 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI && 1557 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name && 1558 "unexpected AsmOperands for suboperands"); 1559 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1)); 1560 } 1561 } 1562 } 1563 1564 void MatchableInfo::buildAliasResultOperands() { 1565 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>(); 1566 const CodeGenInstruction *ResultInst = getResultInst(); 1567 1568 // Loop over all operands of the result instruction, determining how to 1569 // populate them. 1570 unsigned AliasOpNo = 0; 1571 unsigned LastOpNo = CGA.ResultInstOperandIndex.size(); 1572 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) { 1573 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i]; 1574 1575 // If this is a tied operand, just copy from the previously handled operand. 1576 int TiedOp = OpInfo->getTiedRegister(); 1577 if (TiedOp != -1) { 1578 ResOperands.push_back(ResOperand::getTiedOp(TiedOp)); 1579 continue; 1580 } 1581 1582 // Handle all the suboperands for this operand. 1583 const std::string &OpName = OpInfo->Name; 1584 for ( ; AliasOpNo < LastOpNo && 1585 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) { 1586 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second; 1587 1588 // Find out what operand from the asmparser that this MCInst operand 1589 // comes from. 1590 switch (CGA.ResultOperands[AliasOpNo].Kind) { 1591 case CodeGenInstAlias::ResultOperand::K_Record: { 1592 StringRef Name = CGA.ResultOperands[AliasOpNo].getName(); 1593 int SrcOperand = findAsmOperand(Name, SubIdx); 1594 if (SrcOperand == -1) 1595 throw TGError(TheDef->getLoc(), "Instruction '" + 1596 TheDef->getName() + "' has operand '" + OpName + 1597 "' that doesn't appear in asm string!"); 1598 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1); 1599 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, 1600 NumOperands)); 1601 break; 1602 } 1603 case CodeGenInstAlias::ResultOperand::K_Imm: { 1604 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm(); 1605 ResOperands.push_back(ResOperand::getImmOp(ImmVal)); 1606 break; 1607 } 1608 case CodeGenInstAlias::ResultOperand::K_Reg: { 1609 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister(); 1610 ResOperands.push_back(ResOperand::getRegOp(Reg)); 1611 break; 1612 } 1613 } 1614 } 1615 } 1616 } 1617 1618 static void emitConvertToMCInst(CodeGenTarget &Target, StringRef ClassName, 1619 std::vector<MatchableInfo*> &Infos, 1620 raw_ostream &OS) { 1621 // Write the convert function to a separate stream, so we can drop it after 1622 // the enum. 1623 std::string ConvertFnBody; 1624 raw_string_ostream CvtOS(ConvertFnBody); 1625 1626 // Function we have already generated. 1627 std::set<std::string> GeneratedFns; 1628 1629 // Start the unified conversion function. 1630 CvtOS << "bool " << Target.getName() << ClassName << "::\n"; 1631 CvtOS << "ConvertToMCInst(unsigned Kind, MCInst &Inst, " 1632 << "unsigned Opcode,\n" 1633 << " const SmallVectorImpl<MCParsedAsmOperand*" 1634 << "> &Operands) {\n"; 1635 CvtOS << " Inst.setOpcode(Opcode);\n"; 1636 CvtOS << " switch (Kind) {\n"; 1637 CvtOS << " default:\n"; 1638 1639 // Start the enum, which we will generate inline. 1640 1641 OS << "// Unified function for converting operands to MCInst instances.\n\n"; 1642 OS << "enum ConversionKind {\n"; 1643 1644 // TargetOperandClass - This is the target's operand class, like X86Operand. 1645 std::string TargetOperandClass = Target.getName() + "Operand"; 1646 1647 for (std::vector<MatchableInfo*>::const_iterator it = Infos.begin(), 1648 ie = Infos.end(); it != ie; ++it) { 1649 MatchableInfo &II = **it; 1650 1651 // Check if we have a custom match function. 1652 std::string AsmMatchConverter = 1653 II.getResultInst()->TheDef->getValueAsString("AsmMatchConverter"); 1654 if (!AsmMatchConverter.empty()) { 1655 std::string Signature = "ConvertCustom_" + AsmMatchConverter; 1656 II.ConversionFnKind = Signature; 1657 1658 // Check if we have already generated this signature. 1659 if (!GeneratedFns.insert(Signature).second) 1660 continue; 1661 1662 // If not, emit it now. Add to the enum list. 1663 OS << " " << Signature << ",\n"; 1664 1665 CvtOS << " case " << Signature << ":\n"; 1666 CvtOS << " return " << AsmMatchConverter 1667 << "(Inst, Opcode, Operands);\n"; 1668 continue; 1669 } 1670 1671 // Build the conversion function signature. 1672 std::string Signature = "Convert"; 1673 std::string CaseBody; 1674 raw_string_ostream CaseOS(CaseBody); 1675 1676 // Compute the convert enum and the case body. 1677 for (unsigned i = 0, e = II.ResOperands.size(); i != e; ++i) { 1678 const MatchableInfo::ResOperand &OpInfo = II.ResOperands[i]; 1679 1680 // Generate code to populate each result operand. 1681 switch (OpInfo.Kind) { 1682 case MatchableInfo::ResOperand::RenderAsmOperand: { 1683 // This comes from something we parsed. 1684 MatchableInfo::AsmOperand &Op = II.AsmOperands[OpInfo.AsmOperandNum]; 1685 1686 // Registers are always converted the same, don't duplicate the 1687 // conversion function based on them. 1688 Signature += "__"; 1689 if (Op.Class->isRegisterClass()) 1690 Signature += "Reg"; 1691 else 1692 Signature += Op.Class->ClassName; 1693 Signature += utostr(OpInfo.MINumOperands); 1694 Signature += "_" + itostr(OpInfo.AsmOperandNum); 1695 1696 CaseOS << " ((" << TargetOperandClass << "*)Operands[" 1697 << (OpInfo.AsmOperandNum+1) << "])->" << Op.Class->RenderMethod 1698 << "(Inst, " << OpInfo.MINumOperands << ");\n"; 1699 break; 1700 } 1701 1702 case MatchableInfo::ResOperand::TiedOperand: { 1703 // If this operand is tied to a previous one, just copy the MCInst 1704 // operand from the earlier one.We can only tie single MCOperand values. 1705 //assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand"); 1706 unsigned TiedOp = OpInfo.TiedOperandNum; 1707 assert(i > TiedOp && "Tied operand precedes its target!"); 1708 CaseOS << " Inst.addOperand(Inst.getOperand(" << TiedOp << "));\n"; 1709 Signature += "__Tie" + utostr(TiedOp); 1710 break; 1711 } 1712 case MatchableInfo::ResOperand::ImmOperand: { 1713 int64_t Val = OpInfo.ImmVal; 1714 CaseOS << " Inst.addOperand(MCOperand::CreateImm(" << Val << "));\n"; 1715 Signature += "__imm" + itostr(Val); 1716 break; 1717 } 1718 case MatchableInfo::ResOperand::RegOperand: { 1719 if (OpInfo.Register == 0) { 1720 CaseOS << " Inst.addOperand(MCOperand::CreateReg(0));\n"; 1721 Signature += "__reg0"; 1722 } else { 1723 std::string N = getQualifiedName(OpInfo.Register); 1724 CaseOS << " Inst.addOperand(MCOperand::CreateReg(" << N << "));\n"; 1725 Signature += "__reg" + OpInfo.Register->getName(); 1726 } 1727 } 1728 } 1729 } 1730 1731 II.ConversionFnKind = Signature; 1732 1733 // Check if we have already generated this signature. 1734 if (!GeneratedFns.insert(Signature).second) 1735 continue; 1736 1737 // If not, emit it now. Add to the enum list. 1738 OS << " " << Signature << ",\n"; 1739 1740 CvtOS << " case " << Signature << ":\n"; 1741 CvtOS << CaseOS.str(); 1742 CvtOS << " return true;\n"; 1743 } 1744 1745 // Finish the convert function. 1746 1747 CvtOS << " }\n"; 1748 CvtOS << " return false;\n"; 1749 CvtOS << "}\n\n"; 1750 1751 // Finish the enum, and drop the convert function after it. 1752 1753 OS << " NumConversionVariants\n"; 1754 OS << "};\n\n"; 1755 1756 OS << CvtOS.str(); 1757 } 1758 1759 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds. 1760 static void emitMatchClassEnumeration(CodeGenTarget &Target, 1761 std::vector<ClassInfo*> &Infos, 1762 raw_ostream &OS) { 1763 OS << "namespace {\n\n"; 1764 1765 OS << "/// MatchClassKind - The kinds of classes which participate in\n" 1766 << "/// instruction matching.\n"; 1767 OS << "enum MatchClassKind {\n"; 1768 OS << " InvalidMatchClass = 0,\n"; 1769 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1770 ie = Infos.end(); it != ie; ++it) { 1771 ClassInfo &CI = **it; 1772 OS << " " << CI.Name << ", // "; 1773 if (CI.Kind == ClassInfo::Token) { 1774 OS << "'" << CI.ValueName << "'\n"; 1775 } else if (CI.isRegisterClass()) { 1776 if (!CI.ValueName.empty()) 1777 OS << "register class '" << CI.ValueName << "'\n"; 1778 else 1779 OS << "derived register class\n"; 1780 } else { 1781 OS << "user defined class '" << CI.ValueName << "'\n"; 1782 } 1783 } 1784 OS << " NumMatchClassKinds\n"; 1785 OS << "};\n\n"; 1786 1787 OS << "}\n\n"; 1788 } 1789 1790 /// emitValidateOperandClass - Emit the function to validate an operand class. 1791 static void emitValidateOperandClass(AsmMatcherInfo &Info, 1792 raw_ostream &OS) { 1793 OS << "static bool validateOperandClass(MCParsedAsmOperand *GOp, " 1794 << "MatchClassKind Kind) {\n"; 1795 OS << " " << Info.Target.getName() << "Operand &Operand = *(" 1796 << Info.Target.getName() << "Operand*)GOp;\n"; 1797 1798 // The InvalidMatchClass is not to match any operand. 1799 OS << " if (Kind == InvalidMatchClass)\n"; 1800 OS << " return false;\n\n"; 1801 1802 // Check for Token operands first. 1803 OS << " if (Operand.isToken())\n"; 1804 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind);" 1805 << "\n\n"; 1806 1807 // Check for register operands, including sub-classes. 1808 OS << " if (Operand.isReg()) {\n"; 1809 OS << " MatchClassKind OpKind;\n"; 1810 OS << " switch (Operand.getReg()) {\n"; 1811 OS << " default: OpKind = InvalidMatchClass; break;\n"; 1812 for (std::map<Record*, ClassInfo*>::iterator 1813 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end(); 1814 it != ie; ++it) 1815 OS << " case " << Info.Target.getName() << "::" 1816 << it->first->getName() << ": OpKind = " << it->second->Name 1817 << "; break;\n"; 1818 OS << " }\n"; 1819 OS << " return isSubclass(OpKind, Kind);\n"; 1820 OS << " }\n\n"; 1821 1822 // Check the user classes. We don't care what order since we're only 1823 // actually matching against one of them. 1824 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(), 1825 ie = Info.Classes.end(); it != ie; ++it) { 1826 ClassInfo &CI = **it; 1827 1828 if (!CI.isUserClass()) 1829 continue; 1830 1831 OS << " // '" << CI.ClassName << "' class\n"; 1832 OS << " if (Kind == " << CI.Name 1833 << " && Operand." << CI.PredicateMethod << "()) {\n"; 1834 OS << " return true;\n"; 1835 OS << " }\n\n"; 1836 } 1837 1838 OS << " return false;\n"; 1839 OS << "}\n\n"; 1840 } 1841 1842 /// emitIsSubclass - Emit the subclass predicate function. 1843 static void emitIsSubclass(CodeGenTarget &Target, 1844 std::vector<ClassInfo*> &Infos, 1845 raw_ostream &OS) { 1846 OS << "/// isSubclass - Compute whether \\arg A is a subclass of \\arg B.\n"; 1847 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n"; 1848 OS << " if (A == B)\n"; 1849 OS << " return true;\n\n"; 1850 1851 OS << " switch (A) {\n"; 1852 OS << " default:\n"; 1853 OS << " return false;\n"; 1854 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1855 ie = Infos.end(); it != ie; ++it) { 1856 ClassInfo &A = **it; 1857 1858 std::vector<StringRef> SuperClasses; 1859 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1860 ie = Infos.end(); it != ie; ++it) { 1861 ClassInfo &B = **it; 1862 1863 if (&A != &B && A.isSubsetOf(B)) 1864 SuperClasses.push_back(B.Name); 1865 } 1866 1867 if (SuperClasses.empty()) 1868 continue; 1869 1870 OS << "\n case " << A.Name << ":\n"; 1871 1872 if (SuperClasses.size() == 1) { 1873 OS << " return B == " << SuperClasses.back() << ";\n"; 1874 continue; 1875 } 1876 1877 OS << " switch (B) {\n"; 1878 OS << " default: return false;\n"; 1879 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i) 1880 OS << " case " << SuperClasses[i] << ": return true;\n"; 1881 OS << " }\n"; 1882 } 1883 OS << " }\n"; 1884 OS << "}\n\n"; 1885 } 1886 1887 /// emitMatchTokenString - Emit the function to match a token string to the 1888 /// appropriate match class value. 1889 static void emitMatchTokenString(CodeGenTarget &Target, 1890 std::vector<ClassInfo*> &Infos, 1891 raw_ostream &OS) { 1892 // Construct the match list. 1893 std::vector<StringMatcher::StringPair> Matches; 1894 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1895 ie = Infos.end(); it != ie; ++it) { 1896 ClassInfo &CI = **it; 1897 1898 if (CI.Kind == ClassInfo::Token) 1899 Matches.push_back(StringMatcher::StringPair(CI.ValueName, 1900 "return " + CI.Name + ";")); 1901 } 1902 1903 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n"; 1904 1905 StringMatcher("Name", Matches, OS).Emit(); 1906 1907 OS << " return InvalidMatchClass;\n"; 1908 OS << "}\n\n"; 1909 } 1910 1911 /// emitMatchRegisterName - Emit the function to match a string to the target 1912 /// specific register enum. 1913 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser, 1914 raw_ostream &OS) { 1915 // Construct the match list. 1916 std::vector<StringMatcher::StringPair> Matches; 1917 const std::vector<CodeGenRegister*> &Regs = 1918 Target.getRegBank().getRegisters(); 1919 for (unsigned i = 0, e = Regs.size(); i != e; ++i) { 1920 const CodeGenRegister *Reg = Regs[i]; 1921 if (Reg->TheDef->getValueAsString("AsmName").empty()) 1922 continue; 1923 1924 Matches.push_back(StringMatcher::StringPair( 1925 Reg->TheDef->getValueAsString("AsmName"), 1926 "return " + utostr(Reg->EnumValue) + ";")); 1927 } 1928 1929 OS << "static unsigned MatchRegisterName(StringRef Name) {\n"; 1930 1931 StringMatcher("Name", Matches, OS).Emit(); 1932 1933 OS << " return 0;\n"; 1934 OS << "}\n\n"; 1935 } 1936 1937 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag 1938 /// definitions. 1939 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info, 1940 raw_ostream &OS) { 1941 OS << "// Flags for subtarget features that participate in " 1942 << "instruction matching.\n"; 1943 OS << "enum SubtargetFeatureFlag {\n"; 1944 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator 1945 it = Info.SubtargetFeatures.begin(), 1946 ie = Info.SubtargetFeatures.end(); it != ie; ++it) { 1947 SubtargetFeatureInfo &SFI = *it->second; 1948 OS << " " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n"; 1949 } 1950 OS << " Feature_None = 0\n"; 1951 OS << "};\n\n"; 1952 } 1953 1954 /// emitGetSubtargetFeatureName - Emit the helper function to get the 1955 /// user-level name for a subtarget feature. 1956 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) { 1957 OS << "// User-level names for subtarget features that participate in\n" 1958 << "// instruction matching.\n" 1959 << "static const char *getSubtargetFeatureName(unsigned Val) {\n" 1960 << " switch(Val) {\n"; 1961 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator 1962 it = Info.SubtargetFeatures.begin(), 1963 ie = Info.SubtargetFeatures.end(); it != ie; ++it) { 1964 SubtargetFeatureInfo &SFI = *it->second; 1965 // FIXME: Totally just a placeholder name to get the algorithm working. 1966 OS << " case " << SFI.getEnumName() << ": return \"" 1967 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n"; 1968 } 1969 OS << " default: return \"(unknown)\";\n"; 1970 OS << " }\n}\n\n"; 1971 } 1972 1973 /// emitComputeAvailableFeatures - Emit the function to compute the list of 1974 /// available features given a subtarget. 1975 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info, 1976 raw_ostream &OS) { 1977 std::string ClassName = 1978 Info.AsmParser->getValueAsString("AsmParserClassName"); 1979 1980 OS << "unsigned " << Info.Target.getName() << ClassName << "::\n" 1981 << "ComputeAvailableFeatures(uint64_t FB) const {\n"; 1982 OS << " unsigned Features = 0;\n"; 1983 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator 1984 it = Info.SubtargetFeatures.begin(), 1985 ie = Info.SubtargetFeatures.end(); it != ie; ++it) { 1986 SubtargetFeatureInfo &SFI = *it->second; 1987 1988 OS << " if ("; 1989 std::string CondStorage = 1990 SFI.TheDef->getValueAsString("AssemblerCondString"); 1991 StringRef Conds = CondStorage; 1992 std::pair<StringRef,StringRef> Comma = Conds.split(','); 1993 bool First = true; 1994 do { 1995 if (!First) 1996 OS << " && "; 1997 1998 bool Neg = false; 1999 StringRef Cond = Comma.first; 2000 if (Cond[0] == '!') { 2001 Neg = true; 2002 Cond = Cond.substr(1); 2003 } 2004 2005 OS << "((FB & " << Info.Target.getName() << "::" << Cond << ")"; 2006 if (Neg) 2007 OS << " == 0"; 2008 else 2009 OS << " != 0"; 2010 OS << ")"; 2011 2012 if (Comma.second.empty()) 2013 break; 2014 2015 First = false; 2016 Comma = Comma.second.split(','); 2017 } while (true); 2018 2019 OS << ")\n"; 2020 OS << " Features |= " << SFI.getEnumName() << ";\n"; 2021 } 2022 OS << " return Features;\n"; 2023 OS << "}\n\n"; 2024 } 2025 2026 static std::string GetAliasRequiredFeatures(Record *R, 2027 const AsmMatcherInfo &Info) { 2028 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates"); 2029 std::string Result; 2030 unsigned NumFeatures = 0; 2031 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) { 2032 SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]); 2033 2034 if (F == 0) 2035 throw TGError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() + 2036 "' is not marked as an AssemblerPredicate!"); 2037 2038 if (NumFeatures) 2039 Result += '|'; 2040 2041 Result += F->getEnumName(); 2042 ++NumFeatures; 2043 } 2044 2045 if (NumFeatures > 1) 2046 Result = '(' + Result + ')'; 2047 return Result; 2048 } 2049 2050 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions, 2051 /// emit a function for them and return true, otherwise return false. 2052 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) { 2053 // Ignore aliases when match-prefix is set. 2054 if (!MatchPrefix.empty()) 2055 return false; 2056 2057 std::vector<Record*> Aliases = 2058 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias"); 2059 if (Aliases.empty()) return false; 2060 2061 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, " 2062 "unsigned Features) {\n"; 2063 2064 // Keep track of all the aliases from a mnemonic. Use an std::map so that the 2065 // iteration order of the map is stable. 2066 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic; 2067 2068 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) { 2069 Record *R = Aliases[i]; 2070 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R); 2071 } 2072 2073 // Process each alias a "from" mnemonic at a time, building the code executed 2074 // by the string remapper. 2075 std::vector<StringMatcher::StringPair> Cases; 2076 for (std::map<std::string, std::vector<Record*> >::iterator 2077 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end(); 2078 I != E; ++I) { 2079 const std::vector<Record*> &ToVec = I->second; 2080 2081 // Loop through each alias and emit code that handles each case. If there 2082 // are two instructions without predicates, emit an error. If there is one, 2083 // emit it last. 2084 std::string MatchCode; 2085 int AliasWithNoPredicate = -1; 2086 2087 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) { 2088 Record *R = ToVec[i]; 2089 std::string FeatureMask = GetAliasRequiredFeatures(R, Info); 2090 2091 // If this unconditionally matches, remember it for later and diagnose 2092 // duplicates. 2093 if (FeatureMask.empty()) { 2094 if (AliasWithNoPredicate != -1) { 2095 // We can't have two aliases from the same mnemonic with no predicate. 2096 PrintError(ToVec[AliasWithNoPredicate]->getLoc(), 2097 "two MnemonicAliases with the same 'from' mnemonic!"); 2098 throw TGError(R->getLoc(), "this is the other MnemonicAlias."); 2099 } 2100 2101 AliasWithNoPredicate = i; 2102 continue; 2103 } 2104 if (R->getValueAsString("ToMnemonic") == I->first) 2105 throw TGError(R->getLoc(), "MnemonicAlias to the same string"); 2106 2107 if (!MatchCode.empty()) 2108 MatchCode += "else "; 2109 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n"; 2110 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n"; 2111 } 2112 2113 if (AliasWithNoPredicate != -1) { 2114 Record *R = ToVec[AliasWithNoPredicate]; 2115 if (!MatchCode.empty()) 2116 MatchCode += "else\n "; 2117 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n"; 2118 } 2119 2120 MatchCode += "return;"; 2121 2122 Cases.push_back(std::make_pair(I->first, MatchCode)); 2123 } 2124 2125 StringMatcher("Mnemonic", Cases, OS).Emit(); 2126 OS << "}\n\n"; 2127 2128 return true; 2129 } 2130 2131 static const char *getMinimalTypeForRange(uint64_t Range) { 2132 assert(Range < 0xFFFFFFFFULL && "Enum too large"); 2133 if (Range > 0xFFFF) 2134 return "uint32_t"; 2135 if (Range > 0xFF) 2136 return "uint16_t"; 2137 return "uint8_t"; 2138 } 2139 2140 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target, 2141 const AsmMatcherInfo &Info, StringRef ClassName) { 2142 // Emit the static custom operand parsing table; 2143 OS << "namespace {\n"; 2144 OS << " struct OperandMatchEntry {\n"; 2145 OS << " static const char *const MnemonicTable;\n"; 2146 OS << " uint32_t OperandMask;\n"; 2147 OS << " uint32_t Mnemonic;\n"; 2148 OS << " " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size()) 2149 << " RequiredFeatures;\n"; 2150 OS << " " << getMinimalTypeForRange(Info.Classes.size()) 2151 << " Class;\n\n"; 2152 OS << " StringRef getMnemonic() const {\n"; 2153 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 2154 OS << " MnemonicTable[Mnemonic]);\n"; 2155 OS << " }\n"; 2156 OS << " };\n\n"; 2157 2158 OS << " // Predicate for searching for an opcode.\n"; 2159 OS << " struct LessOpcodeOperand {\n"; 2160 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n"; 2161 OS << " return LHS.getMnemonic() < RHS;\n"; 2162 OS << " }\n"; 2163 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n"; 2164 OS << " return LHS < RHS.getMnemonic();\n"; 2165 OS << " }\n"; 2166 OS << " bool operator()(const OperandMatchEntry &LHS,"; 2167 OS << " const OperandMatchEntry &RHS) {\n"; 2168 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 2169 OS << " }\n"; 2170 OS << " };\n"; 2171 2172 OS << "} // end anonymous namespace.\n\n"; 2173 2174 StringToOffsetTable StringTable; 2175 2176 OS << "static const OperandMatchEntry OperandMatchTable[" 2177 << Info.OperandMatchInfo.size() << "] = {\n"; 2178 2179 OS << " /* Operand List Mask, Mnemonic, Operand Class, Features */\n"; 2180 for (std::vector<OperandMatchEntry>::const_iterator it = 2181 Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end(); 2182 it != ie; ++it) { 2183 const OperandMatchEntry &OMI = *it; 2184 const MatchableInfo &II = *OMI.MI; 2185 2186 OS << " { " << OMI.OperandMask; 2187 2188 OS << " /* "; 2189 bool printComma = false; 2190 for (int i = 0, e = 31; i !=e; ++i) 2191 if (OMI.OperandMask & (1 << i)) { 2192 if (printComma) 2193 OS << ", "; 2194 OS << i; 2195 printComma = true; 2196 } 2197 OS << " */"; 2198 2199 // Store a pascal-style length byte in the mnemonic. 2200 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str(); 2201 OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 2202 << " /* " << II.Mnemonic << " */, "; 2203 2204 // Write the required features mask. 2205 if (!II.RequiredFeatures.empty()) { 2206 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) { 2207 if (i) OS << "|"; 2208 OS << II.RequiredFeatures[i]->getEnumName(); 2209 } 2210 } else 2211 OS << "0"; 2212 2213 OS << ", " << OMI.CI->Name; 2214 2215 OS << " },\n"; 2216 } 2217 OS << "};\n\n"; 2218 2219 OS << "const char *const OperandMatchEntry::MnemonicTable =\n"; 2220 StringTable.EmitString(OS); 2221 OS << ";\n\n"; 2222 2223 // Emit the operand class switch to call the correct custom parser for 2224 // the found operand class. 2225 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2226 << Target.getName() << ClassName << "::\n" 2227 << "tryCustomParseOperand(SmallVectorImpl<MCParsedAsmOperand*>" 2228 << " &Operands,\n unsigned MCK) {\n\n" 2229 << " switch(MCK) {\n"; 2230 2231 for (std::vector<ClassInfo*>::const_iterator it = Info.Classes.begin(), 2232 ie = Info.Classes.end(); it != ie; ++it) { 2233 ClassInfo *CI = *it; 2234 if (CI->ParserMethod.empty()) 2235 continue; 2236 OS << " case " << CI->Name << ":\n" 2237 << " return " << CI->ParserMethod << "(Operands);\n"; 2238 } 2239 2240 OS << " default:\n"; 2241 OS << " return MatchOperand_NoMatch;\n"; 2242 OS << " }\n"; 2243 OS << " return MatchOperand_NoMatch;\n"; 2244 OS << "}\n\n"; 2245 2246 // Emit the static custom operand parser. This code is very similar with 2247 // the other matcher. Also use MatchResultTy here just in case we go for 2248 // a better error handling. 2249 OS << Target.getName() << ClassName << "::OperandMatchResultTy " 2250 << Target.getName() << ClassName << "::\n" 2251 << "MatchOperandParserImpl(SmallVectorImpl<MCParsedAsmOperand*>" 2252 << " &Operands,\n StringRef Mnemonic) {\n"; 2253 2254 // Emit code to get the available features. 2255 OS << " // Get the current feature set.\n"; 2256 OS << " unsigned AvailableFeatures = getAvailableFeatures();\n\n"; 2257 2258 OS << " // Get the next operand index.\n"; 2259 OS << " unsigned NextOpNum = Operands.size()-1;\n"; 2260 2261 // Emit code to search the table. 2262 OS << " // Search the table.\n"; 2263 OS << " std::pair<const OperandMatchEntry*, const OperandMatchEntry*>"; 2264 OS << " MnemonicRange =\n"; 2265 OS << " std::equal_range(OperandMatchTable, OperandMatchTable+" 2266 << Info.OperandMatchInfo.size() << ", Mnemonic,\n" 2267 << " LessOpcodeOperand());\n\n"; 2268 2269 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 2270 OS << " return MatchOperand_NoMatch;\n\n"; 2271 2272 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n" 2273 << " *ie = MnemonicRange.second; it != ie; ++it) {\n"; 2274 2275 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 2276 OS << " assert(Mnemonic == it->getMnemonic());\n\n"; 2277 2278 // Emit check that the required features are available. 2279 OS << " // check if the available features match\n"; 2280 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 2281 << "!= it->RequiredFeatures) {\n"; 2282 OS << " continue;\n"; 2283 OS << " }\n\n"; 2284 2285 // Emit check to ensure the operand number matches. 2286 OS << " // check if the operand in question has a custom parser.\n"; 2287 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n"; 2288 OS << " continue;\n\n"; 2289 2290 // Emit call to the custom parser method 2291 OS << " // call custom parse method to handle the operand\n"; 2292 OS << " OperandMatchResultTy Result = "; 2293 OS << "tryCustomParseOperand(Operands, it->Class);\n"; 2294 OS << " if (Result != MatchOperand_NoMatch)\n"; 2295 OS << " return Result;\n"; 2296 OS << " }\n\n"; 2297 2298 OS << " // Okay, we had no match.\n"; 2299 OS << " return MatchOperand_NoMatch;\n"; 2300 OS << "}\n\n"; 2301 } 2302 2303 void AsmMatcherEmitter::run(raw_ostream &OS) { 2304 CodeGenTarget Target(Records); 2305 Record *AsmParser = Target.getAsmParser(); 2306 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName"); 2307 2308 // Compute the information on the instructions to match. 2309 AsmMatcherInfo Info(AsmParser, Target, Records); 2310 Info.buildInfo(); 2311 2312 // Sort the instruction table using the partial order on classes. We use 2313 // stable_sort to ensure that ambiguous instructions are still 2314 // deterministically ordered. 2315 std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(), 2316 less_ptr<MatchableInfo>()); 2317 2318 DEBUG_WITH_TYPE("instruction_info", { 2319 for (std::vector<MatchableInfo*>::iterator 2320 it = Info.Matchables.begin(), ie = Info.Matchables.end(); 2321 it != ie; ++it) 2322 (*it)->dump(); 2323 }); 2324 2325 // Check for ambiguous matchables. 2326 DEBUG_WITH_TYPE("ambiguous_instrs", { 2327 unsigned NumAmbiguous = 0; 2328 for (unsigned i = 0, e = Info.Matchables.size(); i != e; ++i) { 2329 for (unsigned j = i + 1; j != e; ++j) { 2330 MatchableInfo &A = *Info.Matchables[i]; 2331 MatchableInfo &B = *Info.Matchables[j]; 2332 2333 if (A.couldMatchAmbiguouslyWith(B)) { 2334 errs() << "warning: ambiguous matchables:\n"; 2335 A.dump(); 2336 errs() << "\nis incomparable with:\n"; 2337 B.dump(); 2338 errs() << "\n\n"; 2339 ++NumAmbiguous; 2340 } 2341 } 2342 } 2343 if (NumAmbiguous) 2344 errs() << "warning: " << NumAmbiguous 2345 << " ambiguous matchables!\n"; 2346 }); 2347 2348 // Compute the information on the custom operand parsing. 2349 Info.buildOperandMatchInfo(); 2350 2351 // Write the output. 2352 2353 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS); 2354 2355 // Information for the class declaration. 2356 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n"; 2357 OS << "#undef GET_ASSEMBLER_HEADER\n"; 2358 OS << " // This should be included into the middle of the declaration of\n"; 2359 OS << " // your subclasses implementation of MCTargetAsmParser.\n"; 2360 OS << " unsigned ComputeAvailableFeatures(uint64_t FeatureBits) const;\n"; 2361 OS << " bool ConvertToMCInst(unsigned Kind, MCInst &Inst, " 2362 << "unsigned Opcode,\n" 2363 << " const SmallVectorImpl<MCParsedAsmOperand*> " 2364 << "&Operands);\n"; 2365 OS << " bool MnemonicIsValid(StringRef Mnemonic);\n"; 2366 OS << " unsigned MatchInstructionImpl(\n"; 2367 OS << " const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n"; 2368 OS << " MCInst &Inst, unsigned &ErrorInfo, unsigned VariantID = 0);\n"; 2369 2370 if (Info.OperandMatchInfo.size()) { 2371 OS << "\n enum OperandMatchResultTy {\n"; 2372 OS << " MatchOperand_Success, // operand matched successfully\n"; 2373 OS << " MatchOperand_NoMatch, // operand did not match\n"; 2374 OS << " MatchOperand_ParseFail // operand matched but had errors\n"; 2375 OS << " };\n"; 2376 OS << " OperandMatchResultTy MatchOperandParserImpl(\n"; 2377 OS << " SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n"; 2378 OS << " StringRef Mnemonic);\n"; 2379 2380 OS << " OperandMatchResultTy tryCustomParseOperand(\n"; 2381 OS << " SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n"; 2382 OS << " unsigned MCK);\n\n"; 2383 } 2384 2385 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n"; 2386 2387 OS << "\n#ifdef GET_REGISTER_MATCHER\n"; 2388 OS << "#undef GET_REGISTER_MATCHER\n\n"; 2389 2390 // Emit the subtarget feature enumeration. 2391 emitSubtargetFeatureFlagEnumeration(Info, OS); 2392 2393 // Emit the function to match a register name to number. 2394 emitMatchRegisterName(Target, AsmParser, OS); 2395 2396 OS << "#endif // GET_REGISTER_MATCHER\n\n"; 2397 2398 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n"; 2399 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n"; 2400 2401 // Generate the helper function to get the names for subtarget features. 2402 emitGetSubtargetFeatureName(Info, OS); 2403 2404 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n"; 2405 2406 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n"; 2407 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n"; 2408 2409 // Generate the function that remaps for mnemonic aliases. 2410 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info); 2411 2412 // Generate the unified function to convert operands into an MCInst. 2413 emitConvertToMCInst(Target, ClassName, Info.Matchables, OS); 2414 2415 // Emit the enumeration for classes which participate in matching. 2416 emitMatchClassEnumeration(Target, Info.Classes, OS); 2417 2418 // Emit the routine to match token strings to their match class. 2419 emitMatchTokenString(Target, Info.Classes, OS); 2420 2421 // Emit the subclass predicate routine. 2422 emitIsSubclass(Target, Info.Classes, OS); 2423 2424 // Emit the routine to validate an operand against a match class. 2425 emitValidateOperandClass(Info, OS); 2426 2427 // Emit the available features compute function. 2428 emitComputeAvailableFeatures(Info, OS); 2429 2430 2431 size_t MaxNumOperands = 0; 2432 for (std::vector<MatchableInfo*>::const_iterator it = 2433 Info.Matchables.begin(), ie = Info.Matchables.end(); 2434 it != ie; ++it) 2435 MaxNumOperands = std::max(MaxNumOperands, (*it)->AsmOperands.size()); 2436 2437 // Emit the static match table; unused classes get initalized to 0 which is 2438 // guaranteed to be InvalidMatchClass. 2439 // 2440 // FIXME: We can reduce the size of this table very easily. First, we change 2441 // it so that store the kinds in separate bit-fields for each index, which 2442 // only needs to be the max width used for classes at that index (we also need 2443 // to reject based on this during classification). If we then make sure to 2444 // order the match kinds appropriately (putting mnemonics last), then we 2445 // should only end up using a few bits for each class, especially the ones 2446 // following the mnemonic. 2447 OS << "namespace {\n"; 2448 OS << " struct MatchEntry {\n"; 2449 OS << " static const char *const MnemonicTable;\n"; 2450 OS << " uint32_t Mnemonic;\n"; 2451 OS << " uint16_t Opcode;\n"; 2452 OS << " " << getMinimalTypeForRange(Info.Matchables.size()) 2453 << " ConvertFn;\n"; 2454 OS << " " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size()) 2455 << " RequiredFeatures;\n"; 2456 OS << " " << getMinimalTypeForRange(Info.Classes.size()) 2457 << " Classes[" << MaxNumOperands << "];\n"; 2458 OS << " uint8_t AsmVariantID;\n\n"; 2459 OS << " StringRef getMnemonic() const {\n"; 2460 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n"; 2461 OS << " MnemonicTable[Mnemonic]);\n"; 2462 OS << " }\n"; 2463 OS << " };\n\n"; 2464 2465 OS << " // Predicate for searching for an opcode.\n"; 2466 OS << " struct LessOpcode {\n"; 2467 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n"; 2468 OS << " return LHS.getMnemonic() < RHS;\n"; 2469 OS << " }\n"; 2470 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n"; 2471 OS << " return LHS < RHS.getMnemonic();\n"; 2472 OS << " }\n"; 2473 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n"; 2474 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n"; 2475 OS << " }\n"; 2476 OS << " };\n"; 2477 2478 OS << "} // end anonymous namespace.\n\n"; 2479 2480 StringToOffsetTable StringTable; 2481 2482 OS << "static const MatchEntry MatchTable[" 2483 << Info.Matchables.size() << "] = {\n"; 2484 2485 for (std::vector<MatchableInfo*>::const_iterator it = 2486 Info.Matchables.begin(), ie = Info.Matchables.end(); 2487 it != ie; ++it) { 2488 MatchableInfo &II = **it; 2489 2490 // Store a pascal-style length byte in the mnemonic. 2491 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str(); 2492 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false) 2493 << " /* " << II.Mnemonic << " */, " 2494 << Target.getName() << "::" 2495 << II.getResultInst()->TheDef->getName() << ", " 2496 << II.ConversionFnKind << ", "; 2497 2498 // Write the required features mask. 2499 if (!II.RequiredFeatures.empty()) { 2500 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) { 2501 if (i) OS << "|"; 2502 OS << II.RequiredFeatures[i]->getEnumName(); 2503 } 2504 } else 2505 OS << "0"; 2506 2507 OS << ", { "; 2508 for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) { 2509 MatchableInfo::AsmOperand &Op = II.AsmOperands[i]; 2510 2511 if (i) OS << ", "; 2512 OS << Op.Class->Name; 2513 } 2514 OS << " }, " << II.AsmVariantID; 2515 OS << "},\n"; 2516 } 2517 2518 OS << "};\n\n"; 2519 2520 OS << "const char *const MatchEntry::MnemonicTable =\n"; 2521 StringTable.EmitString(OS); 2522 OS << ";\n\n"; 2523 2524 // A method to determine if a mnemonic is in the list. 2525 OS << "bool " << Target.getName() << ClassName << "::\n" 2526 << "MnemonicIsValid(StringRef Mnemonic) {\n"; 2527 OS << " // Search the table.\n"; 2528 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n"; 2529 OS << " std::equal_range(MatchTable, MatchTable+" 2530 << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n"; 2531 OS << " return MnemonicRange.first != MnemonicRange.second;\n"; 2532 OS << "}\n\n"; 2533 2534 // Finally, build the match function. 2535 OS << "unsigned " 2536 << Target.getName() << ClassName << "::\n" 2537 << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>" 2538 << " &Operands,\n"; 2539 OS << " MCInst &Inst, unsigned &ErrorInfo, "; 2540 OS << "unsigned VariantID) {\n"; 2541 2542 // Emit code to get the available features. 2543 OS << " // Get the current feature set.\n"; 2544 OS << " unsigned AvailableFeatures = getAvailableFeatures();\n\n"; 2545 2546 OS << " // Get the instruction mnemonic, which is the first token.\n"; 2547 OS << " StringRef Mnemonic = ((" << Target.getName() 2548 << "Operand*)Operands[0])->getToken();\n\n"; 2549 2550 if (HasMnemonicAliases) { 2551 OS << " // Process all MnemonicAliases to remap the mnemonic.\n"; 2552 OS << " // FIXME : Add an entry in AsmParserVariant to check this.\n"; 2553 OS << " if (!VariantID)\n"; 2554 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n"; 2555 } 2556 2557 // Emit code to compute the class list for this operand vector. 2558 OS << " // Eliminate obvious mismatches.\n"; 2559 OS << " if (Operands.size() > " << (MaxNumOperands+1) << ") {\n"; 2560 OS << " ErrorInfo = " << (MaxNumOperands+1) << ";\n"; 2561 OS << " return Match_InvalidOperand;\n"; 2562 OS << " }\n\n"; 2563 2564 OS << " // Some state to try to produce better error messages.\n"; 2565 OS << " bool HadMatchOtherThanFeatures = false;\n"; 2566 OS << " bool HadMatchOtherThanPredicate = false;\n"; 2567 OS << " unsigned RetCode = Match_InvalidOperand;\n"; 2568 OS << " // Set ErrorInfo to the operand that mismatches if it is\n"; 2569 OS << " // wrong for all instances of the instruction.\n"; 2570 OS << " ErrorInfo = ~0U;\n"; 2571 2572 // Emit code to search the table. 2573 OS << " // Search the table.\n"; 2574 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n"; 2575 OS << " std::equal_range(MatchTable, MatchTable+" 2576 << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n\n"; 2577 2578 OS << " // Return a more specific error code if no mnemonics match.\n"; 2579 OS << " if (MnemonicRange.first == MnemonicRange.second)\n"; 2580 OS << " return Match_MnemonicFail;\n\n"; 2581 2582 OS << " for (const MatchEntry *it = MnemonicRange.first, " 2583 << "*ie = MnemonicRange.second;\n"; 2584 OS << " it != ie; ++it) {\n"; 2585 2586 OS << " // equal_range guarantees that instruction mnemonic matches.\n"; 2587 OS << " assert(Mnemonic == it->getMnemonic());\n"; 2588 2589 // Emit check that the subclasses match. 2590 OS << " if (VariantID != it->AsmVariantID) continue;\n"; 2591 OS << " bool OperandsValid = true;\n"; 2592 OS << " for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n"; 2593 OS << " if (i + 1 >= Operands.size()) {\n"; 2594 OS << " OperandsValid = (it->Classes[i] == " <<"InvalidMatchClass);\n"; 2595 OS << " break;\n"; 2596 OS << " }\n"; 2597 OS << " if (validateOperandClass(Operands[i+1], " 2598 "(MatchClassKind)it->Classes[i]))\n"; 2599 OS << " continue;\n"; 2600 OS << " // If this operand is broken for all of the instances of this\n"; 2601 OS << " // mnemonic, keep track of it so we can report loc info.\n"; 2602 OS << " if (it == MnemonicRange.first || ErrorInfo <= i+1)\n"; 2603 OS << " ErrorInfo = i+1;\n"; 2604 OS << " // Otherwise, just reject this instance of the mnemonic.\n"; 2605 OS << " OperandsValid = false;\n"; 2606 OS << " break;\n"; 2607 OS << " }\n\n"; 2608 2609 OS << " if (!OperandsValid) continue;\n"; 2610 2611 // Emit check that the required features are available. 2612 OS << " if ((AvailableFeatures & it->RequiredFeatures) " 2613 << "!= it->RequiredFeatures) {\n"; 2614 OS << " HadMatchOtherThanFeatures = true;\n"; 2615 OS << " ErrorInfo = it->RequiredFeatures & ~AvailableFeatures;\n"; 2616 OS << " continue;\n"; 2617 OS << " }\n"; 2618 OS << "\n"; 2619 OS << " // We have selected a definite instruction, convert the parsed\n" 2620 << " // operands into the appropriate MCInst.\n"; 2621 OS << " if (!ConvertToMCInst(it->ConvertFn, Inst,\n" 2622 << " it->Opcode, Operands))\n"; 2623 OS << " return Match_ConversionFail;\n"; 2624 OS << "\n"; 2625 2626 // Verify the instruction with the target-specific match predicate function. 2627 OS << " // We have a potential match. Check the target predicate to\n" 2628 << " // handle any context sensitive constraints.\n" 2629 << " unsigned MatchResult;\n" 2630 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !=" 2631 << " Match_Success) {\n" 2632 << " Inst.clear();\n" 2633 << " RetCode = MatchResult;\n" 2634 << " HadMatchOtherThanPredicate = true;\n" 2635 << " continue;\n" 2636 << " }\n\n"; 2637 2638 // Call the post-processing function, if used. 2639 std::string InsnCleanupFn = 2640 AsmParser->getValueAsString("AsmParserInstCleanup"); 2641 if (!InsnCleanupFn.empty()) 2642 OS << " " << InsnCleanupFn << "(Inst);\n"; 2643 2644 OS << " return Match_Success;\n"; 2645 OS << " }\n\n"; 2646 2647 OS << " // Okay, we had no match. Try to return a useful error code.\n"; 2648 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)"; 2649 OS << " return RetCode;\n"; 2650 OS << " assert(ErrorInfo && \"missing feature(s) but what?!\");"; 2651 OS << " return Match_MissingFeature;\n"; 2652 OS << "}\n\n"; 2653 2654 if (Info.OperandMatchInfo.size()) 2655 emitCustomOperandParsing(OS, Target, Info, ClassName); 2656 2657 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n"; 2658 } 2659