AsmMatcherEmitter.cpp revision 3472766f9eb7d66f234c390ce1b3a8b76f0ee9ce
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. 12// 13// The input to the target specific matcher is a list of literal tokens and 14// operands. The target specific parser should generally eliminate any syntax 15// which is not relevant for matching; for example, comma tokens should have 16// already been consumed and eliminated by the parser. Most instructions will 17// end up with a single literal token (the instruction name) and some number of 18// operands. 19// 20// Some example inputs, for X86: 21// 'addl' (immediate ...) (register ...) 22// 'add' (immediate ...) (memory ...) 23// 'call' '*' %epc 24// 25// The assembly matcher is responsible for converting this input into a precise 26// machine instruction (i.e., an instruction with a well defined encoding). This 27// mapping has several properties which complicate matching: 28// 29// - It may be ambiguous; many architectures can legally encode particular 30// variants of an instruction in different ways (for example, using a smaller 31// encoding for small immediates). Such ambiguities should never be 32// arbitrarily resolved by the assembler, the assembler is always responsible 33// for choosing the "best" available instruction. 34// 35// - It may depend on the subtarget or the assembler context. Instructions 36// which are invalid for the current mode, but otherwise unambiguous (e.g., 37// an SSE instruction in a file being assembled for i486) should be accepted 38// and rejected by the assembler front end. However, if the proper encoding 39// for an instruction is dependent on the assembler context then the matcher 40// is responsible for selecting the correct machine instruction for the 41// current mode. 42// 43// The core matching algorithm attempts to exploit the regularity in most 44// instruction sets to quickly determine the set of possibly matching 45// instructions, and the simplify the generated code. Additionally, this helps 46// to ensure that the ambiguities are intentionally resolved by the user. 47// 48// The matching is divided into two distinct phases: 49// 50// 1. Classification: Each operand is mapped to the unique set which (a) 51// contains it, and (b) is the largest such subset for which a single 52// instruction could match all members. 53// 54// For register classes, we can generate these subgroups automatically. For 55// arbitrary operands, we expect the user to define the classes and their 56// relations to one another (for example, 8-bit signed immediates as a 57// subset of 32-bit immediates). 58// 59// By partitioning the operands in this way, we guarantee that for any 60// tuple of classes, any single instruction must match either all or none 61// of the sets of operands which could classify to that tuple. 62// 63// In addition, the subset relation amongst classes induces a partial order 64// on such tuples, which we use to resolve ambiguities. 65// 66// FIXME: What do we do if a crazy case shows up where this is the wrong 67// resolution? 68// 69// 2. The input can now be treated as a tuple of classes (static tokens are 70// simple singleton sets). Each such tuple should generally map to a single 71// instruction (we currently ignore cases where this isn't true, whee!!!), 72// which we can emit a simple matcher for. 73// 74//===----------------------------------------------------------------------===// 75 76#include "AsmMatcherEmitter.h" 77#include "CodeGenTarget.h" 78#include "Record.h" 79#include "llvm/ADT/OwningPtr.h" 80#include "llvm/ADT/SmallVector.h" 81#include "llvm/ADT/STLExtras.h" 82#include "llvm/ADT/StringExtras.h" 83#include "llvm/Support/CommandLine.h" 84#include "llvm/Support/Debug.h" 85#include <list> 86#include <map> 87#include <set> 88using namespace llvm; 89 90static cl::opt<std::string> 91MatchPrefix("match-prefix", cl::init(""), 92 cl::desc("Only match instructions with the given prefix")); 93 94/// FlattenVariants - Flatten an .td file assembly string by selecting the 95/// variant at index \arg N. 96static std::string FlattenVariants(const std::string &AsmString, 97 unsigned N) { 98 StringRef Cur = AsmString; 99 std::string Res = ""; 100 101 for (;;) { 102 // Find the start of the next variant string. 103 size_t VariantsStart = 0; 104 for (size_t e = Cur.size(); VariantsStart != e; ++VariantsStart) 105 if (Cur[VariantsStart] == '{' && 106 (VariantsStart == 0 || (Cur[VariantsStart-1] != '$' && 107 Cur[VariantsStart-1] != '\\'))) 108 break; 109 110 // Add the prefix to the result. 111 Res += Cur.slice(0, VariantsStart); 112 if (VariantsStart == Cur.size()) 113 break; 114 115 ++VariantsStart; // Skip the '{'. 116 117 // Scan to the end of the variants string. 118 size_t VariantsEnd = VariantsStart; 119 unsigned NestedBraces = 1; 120 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) { 121 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') { 122 if (--NestedBraces == 0) 123 break; 124 } else if (Cur[VariantsEnd] == '{') 125 ++NestedBraces; 126 } 127 128 // Select the Nth variant (or empty). 129 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd); 130 for (unsigned i = 0; i != N; ++i) 131 Selection = Selection.split('|').second; 132 Res += Selection.split('|').first; 133 134 assert(VariantsEnd != Cur.size() && 135 "Unterminated variants in assembly string!"); 136 Cur = Cur.substr(VariantsEnd + 1); 137 } 138 139 return Res; 140} 141 142/// TokenizeAsmString - Tokenize a simplified assembly string. 143static void TokenizeAsmString(StringRef AsmString, 144 SmallVectorImpl<StringRef> &Tokens) { 145 unsigned Prev = 0; 146 bool InTok = true; 147 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) { 148 switch (AsmString[i]) { 149 case '[': 150 case ']': 151 case '*': 152 case '!': 153 case ' ': 154 case '\t': 155 case ',': 156 if (InTok) { 157 Tokens.push_back(AsmString.slice(Prev, i)); 158 InTok = false; 159 } 160 if (!isspace(AsmString[i]) && AsmString[i] != ',') 161 Tokens.push_back(AsmString.substr(i, 1)); 162 Prev = i + 1; 163 break; 164 165 case '\\': 166 if (InTok) { 167 Tokens.push_back(AsmString.slice(Prev, i)); 168 InTok = false; 169 } 170 ++i; 171 assert(i != AsmString.size() && "Invalid quoted character"); 172 Tokens.push_back(AsmString.substr(i, 1)); 173 Prev = i + 1; 174 break; 175 176 case '$': { 177 // If this isn't "${", treat like a normal token. 178 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') { 179 if (InTok) { 180 Tokens.push_back(AsmString.slice(Prev, i)); 181 InTok = false; 182 } 183 Prev = i; 184 break; 185 } 186 187 if (InTok) { 188 Tokens.push_back(AsmString.slice(Prev, i)); 189 InTok = false; 190 } 191 192 StringRef::iterator End = 193 std::find(AsmString.begin() + i, AsmString.end(), '}'); 194 assert(End != AsmString.end() && "Missing brace in operand reference!"); 195 size_t EndPos = End - AsmString.begin(); 196 Tokens.push_back(AsmString.slice(i, EndPos+1)); 197 Prev = EndPos + 1; 198 i = EndPos; 199 break; 200 } 201 202 default: 203 InTok = true; 204 } 205 } 206 if (InTok && Prev != AsmString.size()) 207 Tokens.push_back(AsmString.substr(Prev)); 208} 209 210static bool IsAssemblerInstruction(StringRef Name, 211 const CodeGenInstruction &CGI, 212 const SmallVectorImpl<StringRef> &Tokens) { 213 // Ignore "codegen only" instructions. 214 if (CGI.TheDef->getValueAsBit("isCodeGenOnly")) 215 return false; 216 217 // Ignore pseudo ops. 218 // 219 // FIXME: This is a hack; can we convert these instructions to set the 220 // "codegen only" bit instead? 221 if (const RecordVal *Form = CGI.TheDef->getValue("Form")) 222 if (Form->getValue()->getAsString() == "Pseudo") 223 return false; 224 225 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases. 226 // 227 // FIXME: This is a total hack. 228 if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int")) 229 return false; 230 231 // Ignore instructions with no .s string. 232 // 233 // FIXME: What are these? 234 if (CGI.AsmString.empty()) 235 return false; 236 237 // FIXME: Hack; ignore any instructions with a newline in them. 238 if (std::find(CGI.AsmString.begin(), 239 CGI.AsmString.end(), '\n') != CGI.AsmString.end()) 240 return false; 241 242 // Ignore instructions with attributes, these are always fake instructions for 243 // simplifying codegen. 244 // 245 // FIXME: Is this true? 246 // 247 // Also, check for instructions which reference the operand multiple times; 248 // this implies a constraint we would not honor. 249 std::set<std::string> OperandNames; 250 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) { 251 if (Tokens[i][0] == '$' && 252 std::find(Tokens[i].begin(), 253 Tokens[i].end(), ':') != Tokens[i].end()) { 254 DEBUG({ 255 errs() << "warning: '" << Name << "': " 256 << "ignoring instruction; operand with attribute '" 257 << Tokens[i] << "'\n"; 258 }); 259 return false; 260 } 261 262 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) { 263 std::string Err = "'" + Name.str() + "': " + 264 "invalid assembler instruction; tied operand '" + Tokens[i].str() + "'"; 265 throw TGError(CGI.TheDef->getLoc(), Err); 266 } 267 } 268 269 return true; 270} 271 272namespace { 273 274/// ClassInfo - Helper class for storing the information about a particular 275/// class of operands which can be matched. 276struct ClassInfo { 277 enum ClassInfoKind { 278 /// Invalid kind, for use as a sentinel value. 279 Invalid = 0, 280 281 /// The class for a particular token. 282 Token, 283 284 /// The (first) register class, subsequent register classes are 285 /// RegisterClass0+1, and so on. 286 RegisterClass0, 287 288 /// The (first) user defined class, subsequent user defined classes are 289 /// UserClass0+1, and so on. 290 UserClass0 = 1<<16 291 }; 292 293 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + 294 /// N) for the Nth user defined class. 295 unsigned Kind; 296 297 /// SuperClasses - The super classes of this class. Note that for simplicities 298 /// sake user operands only record their immediate super class, while register 299 /// operands include all superclasses. 300 std::vector<ClassInfo*> SuperClasses; 301 302 /// Name - The full class name, suitable for use in an enum. 303 std::string Name; 304 305 /// ClassName - The unadorned generic name for this class (e.g., Token). 306 std::string ClassName; 307 308 /// ValueName - The name of the value this class represents; for a token this 309 /// is the literal token string, for an operand it is the TableGen class (or 310 /// empty if this is a derived class). 311 std::string ValueName; 312 313 /// PredicateMethod - The name of the operand method to test whether the 314 /// operand matches this class; this is not valid for Token or register kinds. 315 std::string PredicateMethod; 316 317 /// RenderMethod - The name of the operand method to add this operand to an 318 /// MCInst; this is not valid for Token or register kinds. 319 std::string RenderMethod; 320 321 /// For register classes, the records for all the registers in this class. 322 std::set<Record*> Registers; 323 324public: 325 /// isRegisterClass() - Check if this is a register class. 326 bool isRegisterClass() const { 327 return Kind >= RegisterClass0 && Kind < UserClass0; 328 } 329 330 /// isUserClass() - Check if this is a user defined class. 331 bool isUserClass() const { 332 return Kind >= UserClass0; 333 } 334 335 /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes 336 /// are related if they are in the same class hierarchy. 337 bool isRelatedTo(const ClassInfo &RHS) const { 338 // Tokens are only related to tokens. 339 if (Kind == Token || RHS.Kind == Token) 340 return Kind == Token && RHS.Kind == Token; 341 342 // Registers classes are only related to registers classes, and only if 343 // their intersection is non-empty. 344 if (isRegisterClass() || RHS.isRegisterClass()) { 345 if (!isRegisterClass() || !RHS.isRegisterClass()) 346 return false; 347 348 std::set<Record*> Tmp; 349 std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin()); 350 std::set_intersection(Registers.begin(), Registers.end(), 351 RHS.Registers.begin(), RHS.Registers.end(), 352 II); 353 354 return !Tmp.empty(); 355 } 356 357 // Otherwise we have two users operands; they are related if they are in the 358 // same class hierarchy. 359 // 360 // FIXME: This is an oversimplification, they should only be related if they 361 // intersect, however we don't have that information. 362 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!"); 363 const ClassInfo *Root = this; 364 while (!Root->SuperClasses.empty()) 365 Root = Root->SuperClasses.front(); 366 367 const ClassInfo *RHSRoot = &RHS; 368 while (!RHSRoot->SuperClasses.empty()) 369 RHSRoot = RHSRoot->SuperClasses.front(); 370 371 return Root == RHSRoot; 372 } 373 374 /// isSubsetOf - Test whether this class is a subset of \arg RHS; 375 bool isSubsetOf(const ClassInfo &RHS) const { 376 // This is a subset of RHS if it is the same class... 377 if (this == &RHS) 378 return true; 379 380 // ... or if any of its super classes are a subset of RHS. 381 for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(), 382 ie = SuperClasses.end(); it != ie; ++it) 383 if ((*it)->isSubsetOf(RHS)) 384 return true; 385 386 return false; 387 } 388 389 /// operator< - Compare two classes. 390 bool operator<(const ClassInfo &RHS) const { 391 if (this == &RHS) 392 return false; 393 394 // Unrelated classes can be ordered by kind. 395 if (!isRelatedTo(RHS)) 396 return Kind < RHS.Kind; 397 398 switch (Kind) { 399 case Invalid: 400 assert(0 && "Invalid kind!"); 401 case Token: 402 // Tokens are comparable by value. 403 // 404 // FIXME: Compare by enum value. 405 return ValueName < RHS.ValueName; 406 407 default: 408 // This class preceeds the RHS if it is a proper subset of the RHS. 409 if (isSubsetOf(RHS)) 410 return true; 411 if (RHS.isSubsetOf(*this)) 412 return false; 413 414 // Otherwise, order by name to ensure we have a total ordering. 415 return ValueName < RHS.ValueName; 416 } 417 } 418}; 419 420/// InstructionInfo - Helper class for storing the necessary information for an 421/// instruction which is capable of being matched. 422struct InstructionInfo { 423 struct Operand { 424 /// The unique class instance this operand should match. 425 ClassInfo *Class; 426 427 /// The original operand this corresponds to, if any. 428 const CodeGenInstruction::OperandInfo *OperandInfo; 429 }; 430 431 /// InstrName - The target name for this instruction. 432 std::string InstrName; 433 434 /// Instr - The instruction this matches. 435 const CodeGenInstruction *Instr; 436 437 /// AsmString - The assembly string for this instruction (with variants 438 /// removed). 439 std::string AsmString; 440 441 /// Tokens - The tokenized assembly pattern that this instruction matches. 442 SmallVector<StringRef, 4> Tokens; 443 444 /// Operands - The operands that this instruction matches. 445 SmallVector<Operand, 4> Operands; 446 447 /// ConversionFnKind - The enum value which is passed to the generated 448 /// ConvertToMCInst to convert parsed operands into an MCInst for this 449 /// function. 450 std::string ConversionFnKind; 451 452 /// operator< - Compare two instructions. 453 bool operator<(const InstructionInfo &RHS) const { 454 if (Operands.size() != RHS.Operands.size()) 455 return Operands.size() < RHS.Operands.size(); 456 457 // Compare lexicographically by operand. The matcher validates that other 458 // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith(). 459 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 460 if (*Operands[i].Class < *RHS.Operands[i].Class) 461 return true; 462 if (*RHS.Operands[i].Class < *Operands[i].Class) 463 return false; 464 } 465 466 return false; 467 } 468 469 /// CouldMatchAmiguouslyWith - Check whether this instruction could 470 /// ambiguously match the same set of operands as \arg RHS (without being a 471 /// strictly superior match). 472 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) { 473 // The number of operands is unambiguous. 474 if (Operands.size() != RHS.Operands.size()) 475 return false; 476 477 // Otherwise, make sure the ordering of the two instructions is unambiguous 478 // by checking that either (a) a token or operand kind discriminates them, 479 // or (b) the ordering among equivalent kinds is consistent. 480 481 // Tokens and operand kinds are unambiguous (assuming a correct target 482 // specific parser). 483 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 484 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind || 485 Operands[i].Class->Kind == ClassInfo::Token) 486 if (*Operands[i].Class < *RHS.Operands[i].Class || 487 *RHS.Operands[i].Class < *Operands[i].Class) 488 return false; 489 490 // Otherwise, this operand could commute if all operands are equivalent, or 491 // there is a pair of operands that compare less than and a pair that 492 // compare greater than. 493 bool HasLT = false, HasGT = false; 494 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 495 if (*Operands[i].Class < *RHS.Operands[i].Class) 496 HasLT = true; 497 if (*RHS.Operands[i].Class < *Operands[i].Class) 498 HasGT = true; 499 } 500 501 return !(HasLT ^ HasGT); 502 } 503 504public: 505 void dump(); 506}; 507 508class AsmMatcherInfo { 509public: 510 /// The tablegen AsmParser record. 511 Record *AsmParser; 512 513 /// The AsmParser "CommentDelimiter" value. 514 std::string CommentDelimiter; 515 516 /// The AsmParser "RegisterPrefix" value. 517 std::string RegisterPrefix; 518 519 /// The classes which are needed for matching. 520 std::vector<ClassInfo*> Classes; 521 522 /// The information on the instruction to match. 523 std::vector<InstructionInfo*> Instructions; 524 525 /// Map of Register records to their class information. 526 std::map<Record*, ClassInfo*> RegisterClasses; 527 528private: 529 /// Map of token to class information which has already been constructed. 530 std::map<std::string, ClassInfo*> TokenClasses; 531 532 /// Map of RegisterClass records to their class information. 533 std::map<Record*, ClassInfo*> RegisterClassClasses; 534 535 /// Map of AsmOperandClass records to their class information. 536 std::map<Record*, ClassInfo*> AsmOperandClasses; 537 538private: 539 /// getTokenClass - Lookup or create the class for the given token. 540 ClassInfo *getTokenClass(StringRef Token); 541 542 /// getOperandClass - Lookup or create the class for the given operand. 543 ClassInfo *getOperandClass(StringRef Token, 544 const CodeGenInstruction::OperandInfo &OI); 545 546 /// BuildRegisterClasses - Build the ClassInfo* instances for register 547 /// classes. 548 void BuildRegisterClasses(CodeGenTarget &Target, 549 std::set<std::string> &SingletonRegisterNames); 550 551 /// BuildOperandClasses - Build the ClassInfo* instances for user defined 552 /// operand classes. 553 void BuildOperandClasses(CodeGenTarget &Target); 554 555public: 556 AsmMatcherInfo(Record *_AsmParser); 557 558 /// BuildInfo - Construct the various tables used during matching. 559 void BuildInfo(CodeGenTarget &Target); 560}; 561 562} 563 564void InstructionInfo::dump() { 565 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"' 566 << ", tokens:["; 567 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) { 568 errs() << Tokens[i]; 569 if (i + 1 != e) 570 errs() << ", "; 571 } 572 errs() << "]\n"; 573 574 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 575 Operand &Op = Operands[i]; 576 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - "; 577 if (Op.Class->Kind == ClassInfo::Token) { 578 errs() << '\"' << Tokens[i] << "\"\n"; 579 continue; 580 } 581 582 if (!Op.OperandInfo) { 583 errs() << "(singleton register)\n"; 584 continue; 585 } 586 587 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo; 588 errs() << OI.Name << " " << OI.Rec->getName() 589 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n"; 590 } 591} 592 593static std::string getEnumNameForToken(StringRef Str) { 594 std::string Res; 595 596 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) { 597 switch (*it) { 598 case '*': Res += "_STAR_"; break; 599 case '%': Res += "_PCT_"; break; 600 case ':': Res += "_COLON_"; break; 601 602 default: 603 if (isalnum(*it)) { 604 Res += *it; 605 } else { 606 Res += "_" + utostr((unsigned) *it) + "_"; 607 } 608 } 609 } 610 611 return Res; 612} 613 614/// getRegisterRecord - Get the register record for \arg name, or 0. 615static Record *getRegisterRecord(CodeGenTarget &Target, StringRef Name) { 616 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) { 617 const CodeGenRegister &Reg = Target.getRegisters()[i]; 618 if (Name == Reg.TheDef->getValueAsString("AsmName")) 619 return Reg.TheDef; 620 } 621 622 return 0; 623} 624 625ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) { 626 ClassInfo *&Entry = TokenClasses[Token]; 627 628 if (!Entry) { 629 Entry = new ClassInfo(); 630 Entry->Kind = ClassInfo::Token; 631 Entry->ClassName = "Token"; 632 Entry->Name = "MCK_" + getEnumNameForToken(Token); 633 Entry->ValueName = Token; 634 Entry->PredicateMethod = "<invalid>"; 635 Entry->RenderMethod = "<invalid>"; 636 Classes.push_back(Entry); 637 } 638 639 return Entry; 640} 641 642ClassInfo * 643AsmMatcherInfo::getOperandClass(StringRef Token, 644 const CodeGenInstruction::OperandInfo &OI) { 645 if (OI.Rec->isSubClassOf("RegisterClass")) { 646 ClassInfo *CI = RegisterClassClasses[OI.Rec]; 647 648 if (!CI) { 649 PrintError(OI.Rec->getLoc(), "register class has no class info!"); 650 throw std::string("ERROR: Missing register class!"); 651 } 652 653 return CI; 654 } 655 656 assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!"); 657 Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass"); 658 ClassInfo *CI = AsmOperandClasses[MatchClass]; 659 660 if (!CI) { 661 PrintError(OI.Rec->getLoc(), "operand has no match class!"); 662 throw std::string("ERROR: Missing match class!"); 663 } 664 665 return CI; 666} 667 668void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target, 669 std::set<std::string> 670 &SingletonRegisterNames) { 671 std::vector<CodeGenRegisterClass> RegisterClasses; 672 std::vector<CodeGenRegister> Registers; 673 674 RegisterClasses = Target.getRegisterClasses(); 675 Registers = Target.getRegisters(); 676 677 // The register sets used for matching. 678 std::set< std::set<Record*> > RegisterSets; 679 680 // Gather the defined sets. 681 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(), 682 ie = RegisterClasses.end(); it != ie; ++it) 683 RegisterSets.insert(std::set<Record*>(it->Elements.begin(), 684 it->Elements.end())); 685 686 // Add any required singleton sets. 687 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(), 688 ie = SingletonRegisterNames.end(); it != ie; ++it) 689 if (Record *Rec = getRegisterRecord(Target, *it)) 690 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1)); 691 692 // Introduce derived sets where necessary (when a register does not determine 693 // a unique register set class), and build the mapping of registers to the set 694 // they should classify to. 695 std::map<Record*, std::set<Record*> > RegisterMap; 696 for (std::vector<CodeGenRegister>::iterator it = Registers.begin(), 697 ie = Registers.end(); it != ie; ++it) { 698 CodeGenRegister &CGR = *it; 699 // Compute the intersection of all sets containing this register. 700 std::set<Record*> ContainingSet; 701 702 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 703 ie = RegisterSets.end(); it != ie; ++it) { 704 if (!it->count(CGR.TheDef)) 705 continue; 706 707 if (ContainingSet.empty()) { 708 ContainingSet = *it; 709 } else { 710 std::set<Record*> Tmp; 711 std::swap(Tmp, ContainingSet); 712 std::insert_iterator< std::set<Record*> > II(ContainingSet, 713 ContainingSet.begin()); 714 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(), 715 II); 716 } 717 } 718 719 if (!ContainingSet.empty()) { 720 RegisterSets.insert(ContainingSet); 721 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet)); 722 } 723 } 724 725 // Construct the register classes. 726 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses; 727 unsigned Index = 0; 728 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 729 ie = RegisterSets.end(); it != ie; ++it, ++Index) { 730 ClassInfo *CI = new ClassInfo(); 731 CI->Kind = ClassInfo::RegisterClass0 + Index; 732 CI->ClassName = "Reg" + utostr(Index); 733 CI->Name = "MCK_Reg" + utostr(Index); 734 CI->ValueName = ""; 735 CI->PredicateMethod = ""; // unused 736 CI->RenderMethod = "addRegOperands"; 737 CI->Registers = *it; 738 Classes.push_back(CI); 739 RegisterSetClasses.insert(std::make_pair(*it, CI)); 740 } 741 742 // Find the superclasses; we could compute only the subgroup lattice edges, 743 // but there isn't really a point. 744 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(), 745 ie = RegisterSets.end(); it != ie; ++it) { 746 ClassInfo *CI = RegisterSetClasses[*it]; 747 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(), 748 ie2 = RegisterSets.end(); it2 != ie2; ++it2) 749 if (*it != *it2 && 750 std::includes(it2->begin(), it2->end(), it->begin(), it->end())) 751 CI->SuperClasses.push_back(RegisterSetClasses[*it2]); 752 } 753 754 // Name the register classes which correspond to a user defined RegisterClass. 755 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(), 756 ie = RegisterClasses.end(); it != ie; ++it) { 757 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(), 758 it->Elements.end())]; 759 if (CI->ValueName.empty()) { 760 CI->ClassName = it->getName(); 761 CI->Name = "MCK_" + it->getName(); 762 CI->ValueName = it->getName(); 763 } else 764 CI->ValueName = CI->ValueName + "," + it->getName(); 765 766 RegisterClassClasses.insert(std::make_pair(it->TheDef, CI)); 767 } 768 769 // Populate the map for individual registers. 770 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(), 771 ie = RegisterMap.end(); it != ie; ++it) 772 this->RegisterClasses[it->first] = RegisterSetClasses[it->second]; 773 774 // Name the register classes which correspond to singleton registers. 775 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(), 776 ie = SingletonRegisterNames.end(); it != ie; ++it) { 777 if (Record *Rec = getRegisterRecord(Target, *it)) { 778 ClassInfo *CI = this->RegisterClasses[Rec]; 779 assert(CI && "Missing singleton register class info!"); 780 781 if (CI->ValueName.empty()) { 782 CI->ClassName = Rec->getName(); 783 CI->Name = "MCK_" + Rec->getName(); 784 CI->ValueName = Rec->getName(); 785 } else 786 CI->ValueName = CI->ValueName + "," + Rec->getName(); 787 } 788 } 789} 790 791void AsmMatcherInfo::BuildOperandClasses(CodeGenTarget &Target) { 792 std::vector<Record*> AsmOperands; 793 AsmOperands = Records.getAllDerivedDefinitions("AsmOperandClass"); 794 795 // Pre-populate AsmOperandClasses map. 796 for (std::vector<Record*>::iterator it = AsmOperands.begin(), 797 ie = AsmOperands.end(); it != ie; ++it) 798 AsmOperandClasses[*it] = new ClassInfo(); 799 800 unsigned Index = 0; 801 for (std::vector<Record*>::iterator it = AsmOperands.begin(), 802 ie = AsmOperands.end(); it != ie; ++it, ++Index) { 803 ClassInfo *CI = AsmOperandClasses[*it]; 804 CI->Kind = ClassInfo::UserClass0 + Index; 805 806 ListInit *Supers = (*it)->getValueAsListInit("SuperClasses"); 807 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) { 808 DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i)); 809 if (!DI) { 810 PrintError((*it)->getLoc(), "Invalid super class reference!"); 811 continue; 812 } 813 814 ClassInfo *SC = AsmOperandClasses[DI->getDef()]; 815 if (!SC) 816 PrintError((*it)->getLoc(), "Invalid super class reference!"); 817 else 818 CI->SuperClasses.push_back(SC); 819 } 820 CI->ClassName = (*it)->getValueAsString("Name"); 821 CI->Name = "MCK_" + CI->ClassName; 822 CI->ValueName = (*it)->getName(); 823 824 // Get or construct the predicate method name. 825 Init *PMName = (*it)->getValueInit("PredicateMethod"); 826 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) { 827 CI->PredicateMethod = SI->getValue(); 828 } else { 829 assert(dynamic_cast<UnsetInit*>(PMName) && 830 "Unexpected PredicateMethod field!"); 831 CI->PredicateMethod = "is" + CI->ClassName; 832 } 833 834 // Get or construct the render method name. 835 Init *RMName = (*it)->getValueInit("RenderMethod"); 836 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) { 837 CI->RenderMethod = SI->getValue(); 838 } else { 839 assert(dynamic_cast<UnsetInit*>(RMName) && 840 "Unexpected RenderMethod field!"); 841 CI->RenderMethod = "add" + CI->ClassName + "Operands"; 842 } 843 844 AsmOperandClasses[*it] = CI; 845 Classes.push_back(CI); 846 } 847} 848 849AsmMatcherInfo::AsmMatcherInfo(Record *_AsmParser) 850 : AsmParser(_AsmParser), 851 CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")), 852 RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix")) 853{ 854} 855 856void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) { 857 // Parse the instructions; we need to do this first so that we can gather the 858 // singleton register classes. 859 std::set<std::string> SingletonRegisterNames; 860 861 const std::vector<const CodeGenInstruction*> &InstrList = 862 Target.getInstructionsByEnumValue(); 863 864 for (unsigned i = 0, e = InstrList.size(); i != e; ++i) { 865 const CodeGenInstruction &CGI = *InstrList[i]; 866 867 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix)) 868 continue; 869 870 OwningPtr<InstructionInfo> II(new InstructionInfo()); 871 872 II->InstrName = CGI.TheDef->getName(); 873 II->Instr = &CGI; 874 II->AsmString = FlattenVariants(CGI.AsmString, 0); 875 876 // Remove comments from the asm string. 877 if (!CommentDelimiter.empty()) { 878 size_t Idx = StringRef(II->AsmString).find(CommentDelimiter); 879 if (Idx != StringRef::npos) 880 II->AsmString = II->AsmString.substr(0, Idx); 881 } 882 883 TokenizeAsmString(II->AsmString, II->Tokens); 884 885 // Ignore instructions which shouldn't be matched. 886 if (!IsAssemblerInstruction(CGI.TheDef->getName(), CGI, II->Tokens)) 887 continue; 888 889 // Collect singleton registers, if used. 890 if (!RegisterPrefix.empty()) { 891 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) { 892 if (II->Tokens[i].startswith(RegisterPrefix)) { 893 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size()); 894 Record *Rec = getRegisterRecord(Target, RegName); 895 896 if (!Rec) { 897 std::string Err = "unable to find register for '" + RegName.str() + 898 "' (which matches register prefix)"; 899 throw TGError(CGI.TheDef->getLoc(), Err); 900 } 901 902 SingletonRegisterNames.insert(RegName); 903 } 904 } 905 } 906 907 Instructions.push_back(II.take()); 908 } 909 910 // Build info for the register classes. 911 BuildRegisterClasses(Target, SingletonRegisterNames); 912 913 // Build info for the user defined assembly operand classes. 914 BuildOperandClasses(Target); 915 916 // Build the instruction information. 917 for (std::vector<InstructionInfo*>::iterator it = Instructions.begin(), 918 ie = Instructions.end(); it != ie; ++it) { 919 InstructionInfo *II = *it; 920 921 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) { 922 StringRef Token = II->Tokens[i]; 923 924 // Check for singleton registers. 925 if (!RegisterPrefix.empty() && Token.startswith(RegisterPrefix)) { 926 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size()); 927 InstructionInfo::Operand Op; 928 Op.Class = RegisterClasses[getRegisterRecord(Target, RegName)]; 929 Op.OperandInfo = 0; 930 assert(Op.Class && Op.Class->Registers.size() == 1 && 931 "Unexpected class for singleton register"); 932 II->Operands.push_back(Op); 933 continue; 934 } 935 936 // Check for simple tokens. 937 if (Token[0] != '$') { 938 InstructionInfo::Operand Op; 939 Op.Class = getTokenClass(Token); 940 Op.OperandInfo = 0; 941 II->Operands.push_back(Op); 942 continue; 943 } 944 945 // Otherwise this is an operand reference. 946 StringRef OperandName; 947 if (Token[1] == '{') 948 OperandName = Token.substr(2, Token.size() - 3); 949 else 950 OperandName = Token.substr(1); 951 952 // Map this token to an operand. FIXME: Move elsewhere. 953 unsigned Idx; 954 try { 955 Idx = II->Instr->getOperandNamed(OperandName); 956 } catch(...) { 957 throw std::string("error: unable to find operand: '" + 958 OperandName.str() + "'"); 959 } 960 961 // FIXME: This is annoying, the named operand may be tied (e.g., 962 // XCHG8rm). What we want is the untied operand, which we now have to 963 // grovel for. Only worry about this for single entry operands, we have to 964 // clean this up anyway. 965 const CodeGenInstruction::OperandInfo *OI = &II->Instr->OperandList[Idx]; 966 if (OI->Constraints[0].isTied()) { 967 unsigned TiedOp = OI->Constraints[0].getTiedOperand(); 968 969 // The tied operand index is an MIOperand index, find the operand that 970 // contains it. 971 for (unsigned i = 0, e = II->Instr->OperandList.size(); i != e; ++i) { 972 if (II->Instr->OperandList[i].MIOperandNo == TiedOp) { 973 OI = &II->Instr->OperandList[i]; 974 break; 975 } 976 } 977 978 assert(OI && "Unable to find tied operand target!"); 979 } 980 981 InstructionInfo::Operand Op; 982 Op.Class = getOperandClass(Token, *OI); 983 Op.OperandInfo = OI; 984 II->Operands.push_back(Op); 985 } 986 } 987 988 // Reorder classes so that classes preceed super classes. 989 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>()); 990} 991 992static std::pair<unsigned, unsigned> * 993GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List, 994 unsigned Index) { 995 for (unsigned i = 0, e = List.size(); i != e; ++i) 996 if (Index == List[i].first) 997 return &List[i]; 998 999 return 0; 1000} 1001 1002static void EmitConvertToMCInst(CodeGenTarget &Target, 1003 std::vector<InstructionInfo*> &Infos, 1004 raw_ostream &OS) { 1005 // Write the convert function to a separate stream, so we can drop it after 1006 // the enum. 1007 std::string ConvertFnBody; 1008 raw_string_ostream CvtOS(ConvertFnBody); 1009 1010 // Function we have already generated. 1011 std::set<std::string> GeneratedFns; 1012 1013 // Start the unified conversion function. 1014 1015 CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, " 1016 << "unsigned Opcode,\n" 1017 << " const SmallVectorImpl<MCParsedAsmOperand*" 1018 << "> &Operands) {\n"; 1019 CvtOS << " Inst.setOpcode(Opcode);\n"; 1020 CvtOS << " switch (Kind) {\n"; 1021 CvtOS << " default:\n"; 1022 1023 // Start the enum, which we will generate inline. 1024 1025 OS << "// Unified function for converting operants to MCInst instances.\n\n"; 1026 OS << "enum ConversionKind {\n"; 1027 1028 // TargetOperandClass - This is the target's operand class, like X86Operand. 1029 std::string TargetOperandClass = Target.getName() + "Operand"; 1030 1031 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(), 1032 ie = Infos.end(); it != ie; ++it) { 1033 InstructionInfo &II = **it; 1034 1035 // Order the (class) operands by the order to convert them into an MCInst. 1036 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList; 1037 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) { 1038 InstructionInfo::Operand &Op = II.Operands[i]; 1039 if (Op.OperandInfo) 1040 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i)); 1041 } 1042 1043 // Find any tied operands. 1044 SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands; 1045 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) { 1046 const CodeGenInstruction::OperandInfo &OpInfo = II.Instr->OperandList[i]; 1047 for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) { 1048 const CodeGenInstruction::ConstraintInfo &CI = OpInfo.Constraints[j]; 1049 if (CI.isTied()) 1050 TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo + j, 1051 CI.getTiedOperand())); 1052 } 1053 } 1054 1055 std::sort(MIOperandList.begin(), MIOperandList.end()); 1056 1057 // Compute the total number of operands. 1058 unsigned NumMIOperands = 0; 1059 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) { 1060 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i]; 1061 NumMIOperands = std::max(NumMIOperands, 1062 OI.MIOperandNo + OI.MINumOperands); 1063 } 1064 1065 // Build the conversion function signature. 1066 std::string Signature = "Convert"; 1067 unsigned CurIndex = 0; 1068 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) { 1069 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second]; 1070 assert(CurIndex <= Op.OperandInfo->MIOperandNo && 1071 "Duplicate match for instruction operand!"); 1072 1073 // Skip operands which weren't matched by anything, this occurs when the 1074 // .td file encodes "implicit" operands as explicit ones. 1075 // 1076 // FIXME: This should be removed from the MCInst structure. 1077 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) { 1078 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands, 1079 CurIndex); 1080 if (!Tie) 1081 Signature += "__Imp"; 1082 else 1083 Signature += "__Tie" + utostr(Tie->second); 1084 } 1085 1086 Signature += "__"; 1087 1088 // Registers are always converted the same, don't duplicate the conversion 1089 // function based on them. 1090 // 1091 // FIXME: We could generalize this based on the render method, if it 1092 // mattered. 1093 if (Op.Class->isRegisterClass()) 1094 Signature += "Reg"; 1095 else 1096 Signature += Op.Class->ClassName; 1097 Signature += utostr(Op.OperandInfo->MINumOperands); 1098 Signature += "_" + utostr(MIOperandList[i].second); 1099 1100 CurIndex += Op.OperandInfo->MINumOperands; 1101 } 1102 1103 // Add any trailing implicit operands. 1104 for (; CurIndex != NumMIOperands; ++CurIndex) { 1105 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands, 1106 CurIndex); 1107 if (!Tie) 1108 Signature += "__Imp"; 1109 else 1110 Signature += "__Tie" + utostr(Tie->second); 1111 } 1112 1113 II.ConversionFnKind = Signature; 1114 1115 // Check if we have already generated this signature. 1116 if (!GeneratedFns.insert(Signature).second) 1117 continue; 1118 1119 // If not, emit it now. 1120 1121 // Add to the enum list. 1122 OS << " " << Signature << ",\n"; 1123 1124 // And to the convert function. 1125 CvtOS << " case " << Signature << ":\n"; 1126 CurIndex = 0; 1127 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) { 1128 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second]; 1129 1130 // Add the implicit operands. 1131 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) { 1132 // See if this is a tied operand. 1133 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands, 1134 CurIndex); 1135 1136 if (!Tie) { 1137 // If not, this is some implicit operand. Just assume it is a register 1138 // for now. 1139 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n"; 1140 } else { 1141 // Copy the tied operand. 1142 assert(Tie->first>Tie->second && "Tied operand preceeds its target!"); 1143 CvtOS << " Inst.addOperand(Inst.getOperand(" 1144 << Tie->second << "));\n"; 1145 } 1146 } 1147 1148 CvtOS << " ((" << TargetOperandClass << "*)Operands[" 1149 << MIOperandList[i].second 1150 << "])->" << Op.Class->RenderMethod 1151 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n"; 1152 CurIndex += Op.OperandInfo->MINumOperands; 1153 } 1154 1155 // And add trailing implicit operands. 1156 for (; CurIndex != NumMIOperands; ++CurIndex) { 1157 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands, 1158 CurIndex); 1159 1160 if (!Tie) { 1161 // If not, this is some implicit operand. Just assume it is a register 1162 // for now. 1163 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n"; 1164 } else { 1165 // Copy the tied operand. 1166 assert(Tie->first>Tie->second && "Tied operand preceeds its target!"); 1167 CvtOS << " Inst.addOperand(Inst.getOperand(" 1168 << Tie->second << "));\n"; 1169 } 1170 } 1171 1172 CvtOS << " return;\n"; 1173 } 1174 1175 // Finish the convert function. 1176 1177 CvtOS << " }\n"; 1178 CvtOS << "}\n\n"; 1179 1180 // Finish the enum, and drop the convert function after it. 1181 1182 OS << " NumConversionVariants\n"; 1183 OS << "};\n\n"; 1184 1185 OS << CvtOS.str(); 1186} 1187 1188/// EmitMatchClassEnumeration - Emit the enumeration for match class kinds. 1189static void EmitMatchClassEnumeration(CodeGenTarget &Target, 1190 std::vector<ClassInfo*> &Infos, 1191 raw_ostream &OS) { 1192 OS << "namespace {\n\n"; 1193 1194 OS << "/// MatchClassKind - The kinds of classes which participate in\n" 1195 << "/// instruction matching.\n"; 1196 OS << "enum MatchClassKind {\n"; 1197 OS << " InvalidMatchClass = 0,\n"; 1198 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1199 ie = Infos.end(); it != ie; ++it) { 1200 ClassInfo &CI = **it; 1201 OS << " " << CI.Name << ", // "; 1202 if (CI.Kind == ClassInfo::Token) { 1203 OS << "'" << CI.ValueName << "'\n"; 1204 } else if (CI.isRegisterClass()) { 1205 if (!CI.ValueName.empty()) 1206 OS << "register class '" << CI.ValueName << "'\n"; 1207 else 1208 OS << "derived register class\n"; 1209 } else { 1210 OS << "user defined class '" << CI.ValueName << "'\n"; 1211 } 1212 } 1213 OS << " NumMatchClassKinds\n"; 1214 OS << "};\n\n"; 1215 1216 OS << "}\n\n"; 1217} 1218 1219/// EmitClassifyOperand - Emit the function to classify an operand. 1220static void EmitClassifyOperand(CodeGenTarget &Target, 1221 AsmMatcherInfo &Info, 1222 raw_ostream &OS) { 1223 OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n" 1224 << " " << Target.getName() << "Operand &Operand = *(" 1225 << Target.getName() << "Operand*)GOp;\n"; 1226 1227 // Classify tokens. 1228 OS << " if (Operand.isToken())\n"; 1229 OS << " return MatchTokenString(Operand.getToken());\n\n"; 1230 1231 // Classify registers. 1232 // 1233 // FIXME: Don't hardcode isReg, getReg. 1234 OS << " if (Operand.isReg()) {\n"; 1235 OS << " switch (Operand.getReg()) {\n"; 1236 OS << " default: return InvalidMatchClass;\n"; 1237 for (std::map<Record*, ClassInfo*>::iterator 1238 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end(); 1239 it != ie; ++it) 1240 OS << " case " << Target.getName() << "::" 1241 << it->first->getName() << ": return " << it->second->Name << ";\n"; 1242 OS << " }\n"; 1243 OS << " }\n\n"; 1244 1245 // Classify user defined operands. 1246 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(), 1247 ie = Info.Classes.end(); it != ie; ++it) { 1248 ClassInfo &CI = **it; 1249 1250 if (!CI.isUserClass()) 1251 continue; 1252 1253 OS << " // '" << CI.ClassName << "' class"; 1254 if (!CI.SuperClasses.empty()) { 1255 OS << ", subclass of "; 1256 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) { 1257 if (i) OS << ", "; 1258 OS << "'" << CI.SuperClasses[i]->ClassName << "'"; 1259 assert(CI < *CI.SuperClasses[i] && "Invalid class relation!"); 1260 } 1261 } 1262 OS << "\n"; 1263 1264 OS << " if (Operand." << CI.PredicateMethod << "()) {\n"; 1265 1266 // Validate subclass relationships. 1267 if (!CI.SuperClasses.empty()) { 1268 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) 1269 OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod 1270 << "() && \"Invalid class relationship!\");\n"; 1271 } 1272 1273 OS << " return " << CI.Name << ";\n"; 1274 OS << " }\n\n"; 1275 } 1276 OS << " return InvalidMatchClass;\n"; 1277 OS << "}\n\n"; 1278} 1279 1280/// EmitIsSubclass - Emit the subclass predicate function. 1281static void EmitIsSubclass(CodeGenTarget &Target, 1282 std::vector<ClassInfo*> &Infos, 1283 raw_ostream &OS) { 1284 OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n"; 1285 OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n"; 1286 OS << " if (A == B)\n"; 1287 OS << " return true;\n\n"; 1288 1289 OS << " switch (A) {\n"; 1290 OS << " default:\n"; 1291 OS << " return false;\n"; 1292 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1293 ie = Infos.end(); it != ie; ++it) { 1294 ClassInfo &A = **it; 1295 1296 if (A.Kind != ClassInfo::Token) { 1297 std::vector<StringRef> SuperClasses; 1298 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1299 ie = Infos.end(); it != ie; ++it) { 1300 ClassInfo &B = **it; 1301 1302 if (&A != &B && A.isSubsetOf(B)) 1303 SuperClasses.push_back(B.Name); 1304 } 1305 1306 if (SuperClasses.empty()) 1307 continue; 1308 1309 OS << "\n case " << A.Name << ":\n"; 1310 1311 if (SuperClasses.size() == 1) { 1312 OS << " return B == " << SuperClasses.back() << ";\n"; 1313 continue; 1314 } 1315 1316 OS << " switch (B) {\n"; 1317 OS << " default: return false;\n"; 1318 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i) 1319 OS << " case " << SuperClasses[i] << ": return true;\n"; 1320 OS << " }\n"; 1321 } 1322 } 1323 OS << " }\n"; 1324 OS << "}\n\n"; 1325} 1326 1327typedef std::pair<std::string, std::string> StringPair; 1328 1329/// FindFirstNonCommonLetter - Find the first character in the keys of the 1330/// string pairs that is not shared across the whole set of strings. All 1331/// strings are assumed to have the same length. 1332static unsigned 1333FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) { 1334 assert(!Matches.empty()); 1335 for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) { 1336 // Check to see if letter i is the same across the set. 1337 char Letter = Matches[0]->first[i]; 1338 1339 for (unsigned str = 0, e = Matches.size(); str != e; ++str) 1340 if (Matches[str]->first[i] != Letter) 1341 return i; 1342 } 1343 1344 return Matches[0]->first.size(); 1345} 1346 1347/// EmitStringMatcherForChar - Given a set of strings that are known to be the 1348/// same length and whose characters leading up to CharNo are the same, emit 1349/// code to verify that CharNo and later are the same. 1350/// 1351/// \return - True if control can leave the emitted code fragment. 1352static bool EmitStringMatcherForChar(const std::string &StrVariableName, 1353 const std::vector<const StringPair*> &Matches, 1354 unsigned CharNo, unsigned IndentCount, 1355 raw_ostream &OS) { 1356 assert(!Matches.empty() && "Must have at least one string to match!"); 1357 std::string Indent(IndentCount*2+4, ' '); 1358 1359 // If we have verified that the entire string matches, we're done: output the 1360 // matching code. 1361 if (CharNo == Matches[0]->first.size()) { 1362 assert(Matches.size() == 1 && "Had duplicate keys to match on"); 1363 1364 // FIXME: If Matches[0].first has embeded \n, this will be bad. 1365 OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first 1366 << "\"\n"; 1367 return false; 1368 } 1369 1370 // Bucket the matches by the character we are comparing. 1371 std::map<char, std::vector<const StringPair*> > MatchesByLetter; 1372 1373 for (unsigned i = 0, e = Matches.size(); i != e; ++i) 1374 MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]); 1375 1376 1377 // If we have exactly one bucket to match, see how many characters are common 1378 // across the whole set and match all of them at once. 1379 if (MatchesByLetter.size() == 1) { 1380 unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches); 1381 unsigned NumChars = FirstNonCommonLetter-CharNo; 1382 1383 // Emit code to break out if the prefix doesn't match. 1384 if (NumChars == 1) { 1385 // Do the comparison with if (Str[1] != 'f') 1386 // FIXME: Need to escape general characters. 1387 OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '" 1388 << Matches[0]->first[CharNo] << "')\n"; 1389 OS << Indent << " break;\n"; 1390 } else { 1391 // Do the comparison with if (Str.substr(1,3) != "foo"). 1392 // FIXME: Need to escape general strings. 1393 OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << "," 1394 << NumChars << ") != \""; 1395 OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n"; 1396 OS << Indent << " break;\n"; 1397 } 1398 1399 return EmitStringMatcherForChar(StrVariableName, Matches, 1400 FirstNonCommonLetter, IndentCount, OS); 1401 } 1402 1403 // Otherwise, we have multiple possible things, emit a switch on the 1404 // character. 1405 OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n"; 1406 OS << Indent << "default: break;\n"; 1407 1408 for (std::map<char, std::vector<const StringPair*> >::iterator LI = 1409 MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) { 1410 // TODO: escape hard stuff (like \n) if we ever care about it. 1411 OS << Indent << "case '" << LI->first << "':\t // " 1412 << LI->second.size() << " strings to match.\n"; 1413 if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1, 1414 IndentCount+1, OS)) 1415 OS << Indent << " break;\n"; 1416 } 1417 1418 OS << Indent << "}\n"; 1419 return true; 1420} 1421 1422 1423/// EmitStringMatcher - Given a list of strings and code to execute when they 1424/// match, output a simple switch tree to classify the input string. 1425/// 1426/// If a match is found, the code in Vals[i].second is executed; control must 1427/// not exit this code fragment. If nothing matches, execution falls through. 1428/// 1429/// \param StrVariableName - The name of the variable to test. 1430static void EmitStringMatcher(const std::string &StrVariableName, 1431 const std::vector<StringPair> &Matches, 1432 raw_ostream &OS) { 1433 // First level categorization: group strings by length. 1434 std::map<unsigned, std::vector<const StringPair*> > MatchesByLength; 1435 1436 for (unsigned i = 0, e = Matches.size(); i != e; ++i) 1437 MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]); 1438 1439 // Output a switch statement on length and categorize the elements within each 1440 // bin. 1441 OS << " switch (" << StrVariableName << ".size()) {\n"; 1442 OS << " default: break;\n"; 1443 1444 for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI = 1445 MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) { 1446 OS << " case " << LI->first << ":\t // " << LI->second.size() 1447 << " strings to match.\n"; 1448 if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS)) 1449 OS << " break;\n"; 1450 } 1451 1452 OS << " }\n"; 1453} 1454 1455 1456/// EmitMatchTokenString - Emit the function to match a token string to the 1457/// appropriate match class value. 1458static void EmitMatchTokenString(CodeGenTarget &Target, 1459 std::vector<ClassInfo*> &Infos, 1460 raw_ostream &OS) { 1461 // Construct the match list. 1462 std::vector<StringPair> Matches; 1463 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 1464 ie = Infos.end(); it != ie; ++it) { 1465 ClassInfo &CI = **it; 1466 1467 if (CI.Kind == ClassInfo::Token) 1468 Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";")); 1469 } 1470 1471 OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n"; 1472 1473 EmitStringMatcher("Name", Matches, OS); 1474 1475 OS << " return InvalidMatchClass;\n"; 1476 OS << "}\n\n"; 1477} 1478 1479/// EmitMatchRegisterName - Emit the function to match a string to the target 1480/// specific register enum. 1481static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser, 1482 raw_ostream &OS) { 1483 // Construct the match list. 1484 std::vector<StringPair> Matches; 1485 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) { 1486 const CodeGenRegister &Reg = Target.getRegisters()[i]; 1487 if (Reg.TheDef->getValueAsString("AsmName").empty()) 1488 continue; 1489 1490 Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"), 1491 "return " + utostr(i + 1) + ";")); 1492 } 1493 1494 OS << "static unsigned MatchRegisterName(StringRef Name) {\n"; 1495 1496 EmitStringMatcher("Name", Matches, OS); 1497 1498 OS << " return 0;\n"; 1499 OS << "}\n\n"; 1500} 1501 1502void AsmMatcherEmitter::run(raw_ostream &OS) { 1503 CodeGenTarget Target; 1504 Record *AsmParser = Target.getAsmParser(); 1505 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName"); 1506 1507 // Compute the information on the instructions to match. 1508 AsmMatcherInfo Info(AsmParser); 1509 Info.BuildInfo(Target); 1510 1511 // Sort the instruction table using the partial order on classes. We use 1512 // stable_sort to ensure that ambiguous instructions are still 1513 // deterministically ordered. 1514 std::stable_sort(Info.Instructions.begin(), Info.Instructions.end(), 1515 less_ptr<InstructionInfo>()); 1516 1517 DEBUG_WITH_TYPE("instruction_info", { 1518 for (std::vector<InstructionInfo*>::iterator 1519 it = Info.Instructions.begin(), ie = Info.Instructions.end(); 1520 it != ie; ++it) 1521 (*it)->dump(); 1522 }); 1523 1524 // Check for ambiguous instructions. 1525 unsigned NumAmbiguous = 0; 1526 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) { 1527 for (unsigned j = i + 1; j != e; ++j) { 1528 InstructionInfo &A = *Info.Instructions[i]; 1529 InstructionInfo &B = *Info.Instructions[j]; 1530 1531 if (A.CouldMatchAmiguouslyWith(B)) { 1532 DEBUG_WITH_TYPE("ambiguous_instrs", { 1533 errs() << "warning: ambiguous instruction match:\n"; 1534 A.dump(); 1535 errs() << "\nis incomparable with:\n"; 1536 B.dump(); 1537 errs() << "\n\n"; 1538 }); 1539 ++NumAmbiguous; 1540 } 1541 } 1542 } 1543 if (NumAmbiguous) 1544 DEBUG_WITH_TYPE("ambiguous_instrs", { 1545 errs() << "warning: " << NumAmbiguous 1546 << " ambiguous instructions!\n"; 1547 }); 1548 1549 // Write the output. 1550 1551 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS); 1552 1553 // Emit the function to match a register name to number. 1554 EmitMatchRegisterName(Target, AsmParser, OS); 1555 1556 OS << "#ifndef REGISTERS_ONLY\n\n"; 1557 1558 // Generate the unified function to convert operands into an MCInst. 1559 EmitConvertToMCInst(Target, Info.Instructions, OS); 1560 1561 // Emit the enumeration for classes which participate in matching. 1562 EmitMatchClassEnumeration(Target, Info.Classes, OS); 1563 1564 // Emit the routine to match token strings to their match class. 1565 EmitMatchTokenString(Target, Info.Classes, OS); 1566 1567 // Emit the routine to classify an operand. 1568 EmitClassifyOperand(Target, Info, OS); 1569 1570 // Emit the subclass predicate routine. 1571 EmitIsSubclass(Target, Info.Classes, OS); 1572 1573 // Finally, build the match function. 1574 1575 size_t MaxNumOperands = 0; 1576 for (std::vector<InstructionInfo*>::const_iterator it = 1577 Info.Instructions.begin(), ie = Info.Instructions.end(); 1578 it != ie; ++it) 1579 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size()); 1580 1581 const std::string &MatchName = 1582 AsmParser->getValueAsString("MatchInstructionName"); 1583 OS << "bool " << Target.getName() << ClassName << "::\n" 1584 << MatchName 1585 << "(const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n"; 1586 OS.indent(MatchName.size() + 1); 1587 OS << "MCInst &Inst) {\n"; 1588 1589 // Emit the static match table; unused classes get initalized to 0 which is 1590 // guaranteed to be InvalidMatchClass. 1591 // 1592 // FIXME: We can reduce the size of this table very easily. First, we change 1593 // it so that store the kinds in separate bit-fields for each index, which 1594 // only needs to be the max width used for classes at that index (we also need 1595 // to reject based on this during classification). If we then make sure to 1596 // order the match kinds appropriately (putting mnemonics last), then we 1597 // should only end up using a few bits for each class, especially the ones 1598 // following the mnemonic. 1599 OS << " static const struct MatchEntry {\n"; 1600 OS << " unsigned Opcode;\n"; 1601 OS << " ConversionKind ConvertFn;\n"; 1602 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n"; 1603 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n"; 1604 1605 for (std::vector<InstructionInfo*>::const_iterator it = 1606 Info.Instructions.begin(), ie = Info.Instructions.end(); 1607 it != ie; ++it) { 1608 InstructionInfo &II = **it; 1609 1610 OS << " { " << Target.getName() << "::" << II.InstrName 1611 << ", " << II.ConversionFnKind << ", { "; 1612 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) { 1613 InstructionInfo::Operand &Op = II.Operands[i]; 1614 1615 if (i) OS << ", "; 1616 OS << Op.Class->Name; 1617 } 1618 OS << " } },\n"; 1619 } 1620 1621 OS << " };\n\n"; 1622 1623 // Emit code to compute the class list for this operand vector. 1624 OS << " // Eliminate obvious mismatches.\n"; 1625 OS << " if (Operands.size() > " << MaxNumOperands << ")\n"; 1626 OS << " return true;\n\n"; 1627 1628 OS << " // Compute the class list for this operand vector.\n"; 1629 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n"; 1630 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n"; 1631 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n"; 1632 1633 OS << " // Check for invalid operands before matching.\n"; 1634 OS << " if (Classes[i] == InvalidMatchClass)\n"; 1635 OS << " return true;\n"; 1636 OS << " }\n\n"; 1637 1638 OS << " // Mark unused classes.\n"; 1639 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; " 1640 << "i != e; ++i)\n"; 1641 OS << " Classes[i] = InvalidMatchClass;\n\n"; 1642 1643 // Emit code to search the table. 1644 OS << " // Search the table.\n"; 1645 OS << " for (const MatchEntry *it = MatchTable, " 1646 << "*ie = MatchTable + " << Info.Instructions.size() 1647 << "; it != ie; ++it) {\n"; 1648 for (unsigned i = 0; i != MaxNumOperands; ++i) { 1649 OS << " if (!IsSubclass(Classes[" 1650 << i << "], it->Classes[" << i << "]))\n"; 1651 OS << " continue;\n"; 1652 } 1653 OS << "\n"; 1654 OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n"; 1655 1656 // Call the post-processing function, if used. 1657 std::string InsnCleanupFn = 1658 AsmParser->getValueAsString("AsmParserInstCleanup"); 1659 if (!InsnCleanupFn.empty()) 1660 OS << " " << InsnCleanupFn << "(Inst);\n"; 1661 1662 OS << " return false;\n"; 1663 OS << " }\n\n"; 1664 1665 OS << " return true;\n"; 1666 OS << "}\n\n"; 1667 1668 OS << "#endif // REGISTERS_ONLY\n"; 1669} 1670