AsmMatcherEmitter.cpp revision 2b54481a77696d47dc9220cd7a36155599750904
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 90namespace { 91static cl::opt<std::string> 92MatchPrefix("match-prefix", cl::init(""), 93 cl::desc("Only match instructions with the given prefix")); 94} 95 96/// FlattenVariants - Flatten an .td file assembly string by selecting the 97/// variant at index \arg N. 98static std::string FlattenVariants(const std::string &AsmString, 99 unsigned N) { 100 StringRef Cur = AsmString; 101 std::string Res = ""; 102 103 for (;;) { 104 // Find the start of the next variant string. 105 size_t VariantsStart = 0; 106 for (size_t e = Cur.size(); VariantsStart != e; ++VariantsStart) 107 if (Cur[VariantsStart] == '{' && 108 (VariantsStart == 0 || (Cur[VariantsStart-1] != '$' && 109 Cur[VariantsStart-1] != '\\'))) 110 break; 111 112 // Add the prefix to the result. 113 Res += Cur.slice(0, VariantsStart); 114 if (VariantsStart == Cur.size()) 115 break; 116 117 ++VariantsStart; // Skip the '{'. 118 119 // Scan to the end of the variants string. 120 size_t VariantsEnd = VariantsStart; 121 unsigned NestedBraces = 1; 122 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) { 123 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') { 124 if (--NestedBraces == 0) 125 break; 126 } else if (Cur[VariantsEnd] == '{') 127 ++NestedBraces; 128 } 129 130 // Select the Nth variant (or empty). 131 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd); 132 for (unsigned i = 0; i != N; ++i) 133 Selection = Selection.split('|').second; 134 Res += Selection.split('|').first; 135 136 assert(VariantsEnd != Cur.size() && 137 "Unterminated variants in assembly string!"); 138 Cur = Cur.substr(VariantsEnd + 1); 139 } 140 141 return Res; 142} 143 144/// TokenizeAsmString - Tokenize a simplified assembly string. 145static void TokenizeAsmString(const StringRef &AsmString, 146 SmallVectorImpl<StringRef> &Tokens) { 147 unsigned Prev = 0; 148 bool InTok = true; 149 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) { 150 switch (AsmString[i]) { 151 case '[': 152 case ']': 153 case '*': 154 case '!': 155 case ' ': 156 case '\t': 157 case ',': 158 if (InTok) { 159 Tokens.push_back(AsmString.slice(Prev, i)); 160 InTok = false; 161 } 162 if (!isspace(AsmString[i]) && AsmString[i] != ',') 163 Tokens.push_back(AsmString.substr(i, 1)); 164 Prev = i + 1; 165 break; 166 167 case '\\': 168 if (InTok) { 169 Tokens.push_back(AsmString.slice(Prev, i)); 170 InTok = false; 171 } 172 ++i; 173 assert(i != AsmString.size() && "Invalid quoted character"); 174 Tokens.push_back(AsmString.substr(i, 1)); 175 Prev = i + 1; 176 break; 177 178 case '$': { 179 // If this isn't "${", treat like a normal token. 180 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') { 181 if (InTok) { 182 Tokens.push_back(AsmString.slice(Prev, i)); 183 InTok = false; 184 } 185 Prev = i; 186 break; 187 } 188 189 if (InTok) { 190 Tokens.push_back(AsmString.slice(Prev, i)); 191 InTok = false; 192 } 193 194 StringRef::iterator End = 195 std::find(AsmString.begin() + i, AsmString.end(), '}'); 196 assert(End != AsmString.end() && "Missing brace in operand reference!"); 197 size_t EndPos = End - AsmString.begin(); 198 Tokens.push_back(AsmString.slice(i, EndPos+1)); 199 Prev = EndPos + 1; 200 i = EndPos; 201 break; 202 } 203 204 default: 205 InTok = true; 206 } 207 } 208 if (InTok && Prev != AsmString.size()) 209 Tokens.push_back(AsmString.substr(Prev)); 210} 211 212static bool IsAssemblerInstruction(const StringRef &Name, 213 const CodeGenInstruction &CGI, 214 const SmallVectorImpl<StringRef> &Tokens) { 215 // Ignore psuedo ops. 216 // 217 // FIXME: This is a hack. 218 if (const RecordVal *Form = CGI.TheDef->getValue("Form")) 219 if (Form->getValue()->getAsString() == "Pseudo") 220 return false; 221 222 // Ignore "PHI" node. 223 // 224 // FIXME: This is also a hack. 225 if (Name == "PHI") 226 return false; 227 228 // Ignore instructions with no .s string. 229 // 230 // FIXME: What are these? 231 if (CGI.AsmString.empty()) 232 return false; 233 234 // FIXME: Hack; ignore any instructions with a newline in them. 235 if (std::find(CGI.AsmString.begin(), 236 CGI.AsmString.end(), '\n') != CGI.AsmString.end()) 237 return false; 238 239 // Ignore instructions with attributes, these are always fake instructions for 240 // simplifying codegen. 241 // 242 // FIXME: Is this true? 243 // 244 // Also, we ignore instructions which reference the operand multiple times; 245 // this implies a constraint we would not currently honor. These are 246 // currently always fake instructions for simplifying codegen. 247 // 248 // FIXME: Encode this assumption in the .td, so we can error out here. 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 DEBUG({ 264 errs() << "warning: '" << Name << "': " 265 << "ignoring instruction; tied operand '" 266 << Tokens[i] << "'\n"; 267 }); 268 return false; 269 } 270 } 271 272 return true; 273} 274 275namespace { 276 277/// ClassInfo - Helper class for storing the information about a particular 278/// class of operands which can be matched. 279struct ClassInfo { 280 enum ClassInfoKind { 281 Invalid = 0, ///< Invalid kind, for use as a sentinel value. 282 Token, ///< The class for a particular token. 283 Register, ///< A register class. 284 UserClass0 ///< The (first) user defined class, subsequent user defined 285 /// classes are UserClass0+1, and so on. 286 }; 287 288 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + 289 /// N) for the Nth user defined class. 290 unsigned Kind; 291 292 /// Name - The full class name, suitable for use in an enum. 293 std::string Name; 294 295 /// ClassName - The unadorned generic name for this class (e.g., Token). 296 std::string ClassName; 297 298 /// ValueName - The name of the value this class represents; for a token this 299 /// is the literal token string, for an operand it is the TableGen class (or 300 /// empty if this is a derived class). 301 std::string ValueName; 302 303 /// PredicateMethod - The name of the operand method to test whether the 304 /// operand matches this class; this is not valid for Token kinds. 305 std::string PredicateMethod; 306 307 /// RenderMethod - The name of the operand method to add this operand to an 308 /// MCInst; this is not valid for Token kinds. 309 std::string RenderMethod; 310 311 /// operator< - Compare two classes. 312 bool operator<(const ClassInfo &RHS) const { 313 // Incompatible kinds are comparable. 314 if (Kind != RHS.Kind) 315 return Kind < RHS.Kind; 316 317 switch (Kind) { 318 case Invalid: 319 assert(0 && "Invalid kind!"); 320 case Token: 321 // Tokens are always comparable. 322 // 323 // FIXME: Compare by enum value. 324 return ValueName < RHS.ValueName; 325 326 case Register: 327 // FIXME: Compare by subset relation. 328 return false; 329 330 default: 331 // FIXME: Allow user defined relation. 332 return false; 333 } 334 } 335}; 336 337/// InstructionInfo - Helper class for storing the necessary information for an 338/// instruction which is capable of being matched. 339struct InstructionInfo { 340 struct Operand { 341 /// The unique class instance this operand should match. 342 ClassInfo *Class; 343 344 /// The original operand this corresponds to, if any. 345 const CodeGenInstruction::OperandInfo *OperandInfo; 346 }; 347 348 /// InstrName - The target name for this instruction. 349 std::string InstrName; 350 351 /// Instr - The instruction this matches. 352 const CodeGenInstruction *Instr; 353 354 /// AsmString - The assembly string for this instruction (with variants 355 /// removed). 356 std::string AsmString; 357 358 /// Tokens - The tokenized assembly pattern that this instruction matches. 359 SmallVector<StringRef, 4> Tokens; 360 361 /// Operands - The operands that this instruction matches. 362 SmallVector<Operand, 4> Operands; 363 364 /// ConversionFnKind - The enum value which is passed to the generated 365 /// ConvertToMCInst to convert parsed operands into an MCInst for this 366 /// function. 367 std::string ConversionFnKind; 368 369 /// operator< - Compare two instructions. 370 bool operator<(const InstructionInfo &RHS) const { 371 if (Operands.size() != RHS.Operands.size()) 372 return Operands.size() < RHS.Operands.size(); 373 374 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 375 if (*Operands[i].Class < *RHS.Operands[i].Class) 376 return true; 377 378 return false; 379 } 380 381 /// CouldMatchAmiguouslyWith - Check whether this instruction could 382 /// ambiguously match the same set of operands as \arg RHS (without being a 383 /// strictly superior match). 384 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) { 385 // The number of operands is unambiguous. 386 if (Operands.size() != RHS.Operands.size()) 387 return false; 388 389 // Tokens and operand kinds are unambiguous (assuming a correct target 390 // specific parser). 391 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 392 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind || 393 Operands[i].Class->Kind == ClassInfo::Token) 394 if (*Operands[i].Class < *RHS.Operands[i].Class || 395 *RHS.Operands[i].Class < *Operands[i].Class) 396 return false; 397 398 // Otherwise, this operand could commute if all operands are equivalent, or 399 // there is a pair of operands that compare less than and a pair that 400 // compare greater than. 401 bool HasLT = false, HasGT = false; 402 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 403 if (*Operands[i].Class < *RHS.Operands[i].Class) 404 HasLT = true; 405 if (*RHS.Operands[i].Class < *Operands[i].Class) 406 HasGT = true; 407 } 408 409 return !(HasLT ^ HasGT); 410 } 411 412public: 413 void dump(); 414}; 415 416class AsmMatcherInfo { 417public: 418 /// The classes which are needed for matching. 419 std::vector<ClassInfo*> Classes; 420 421 /// The information on the instruction to match. 422 std::vector<InstructionInfo*> Instructions; 423 424private: 425 /// Map of token to class information which has already been constructed. 426 std::map<std::string, ClassInfo*> TokenClasses; 427 428 /// Map of operand name to class information which has already been 429 /// constructed. 430 std::map<std::string, ClassInfo*> OperandClasses; 431 432 /// Map of user class names to kind value. 433 std::map<std::string, unsigned> UserClasses; 434 435private: 436 /// getTokenClass - Lookup or create the class for the given token. 437 ClassInfo *getTokenClass(const StringRef &Token); 438 439 /// getUserClassKind - Lookup or create the kind value for the given class 440 /// name. 441 unsigned getUserClassKind(const StringRef &Name); 442 443 /// getOperandClass - Lookup or create the class for the given operand. 444 ClassInfo *getOperandClass(const StringRef &Token, 445 const CodeGenInstruction::OperandInfo &OI); 446 447public: 448 /// BuildInfo - Construct the various tables used during matching. 449 void BuildInfo(CodeGenTarget &Target); 450}; 451 452} 453 454void InstructionInfo::dump() { 455 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"' 456 << ", tokens:["; 457 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) { 458 errs() << Tokens[i]; 459 if (i + 1 != e) 460 errs() << ", "; 461 } 462 errs() << "]\n"; 463 464 for (unsigned i = 0, e = Operands.size(); i != e; ++i) { 465 Operand &Op = Operands[i]; 466 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - "; 467 if (Op.Class->Kind == ClassInfo::Token) { 468 errs() << '\"' << Tokens[i] << "\"\n"; 469 continue; 470 } 471 472 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo; 473 errs() << OI.Name << " " << OI.Rec->getName() 474 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n"; 475 } 476} 477 478static std::string getEnumNameForToken(const StringRef &Str) { 479 std::string Res; 480 481 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) { 482 switch (*it) { 483 case '*': Res += "_STAR_"; break; 484 case '%': Res += "_PCT_"; break; 485 case ':': Res += "_COLON_"; break; 486 487 default: 488 if (isalnum(*it)) { 489 Res += *it; 490 } else { 491 Res += "_" + utostr((unsigned) *it) + "_"; 492 } 493 } 494 } 495 496 return Res; 497} 498 499ClassInfo *AsmMatcherInfo::getTokenClass(const StringRef &Token) { 500 ClassInfo *&Entry = TokenClasses[Token]; 501 502 if (!Entry) { 503 Entry = new ClassInfo(); 504 Entry->Kind = ClassInfo::Token; 505 Entry->ClassName = "Token"; 506 Entry->Name = "MCK_" + getEnumNameForToken(Token); 507 Entry->ValueName = Token; 508 Entry->PredicateMethod = "<invalid>"; 509 Entry->RenderMethod = "<invalid>"; 510 Classes.push_back(Entry); 511 } 512 513 return Entry; 514} 515 516unsigned AsmMatcherInfo::getUserClassKind(const StringRef &Name) { 517 unsigned &Entry = UserClasses[Name]; 518 519 if (!Entry) 520 Entry = ClassInfo::UserClass0 + UserClasses.size() - 1; 521 522 return Entry; 523} 524 525ClassInfo * 526AsmMatcherInfo::getOperandClass(const StringRef &Token, 527 const CodeGenInstruction::OperandInfo &OI) { 528 std::string ClassName; 529 if (OI.Rec->isSubClassOf("RegisterClass")) { 530 ClassName = "Reg"; 531 } else { 532 try { 533 ClassName = OI.Rec->getValueAsString("ParserMatchClass"); 534 assert(ClassName != "Reg" && "'Reg' class name is reserved!"); 535 } catch(...) { 536 PrintError(OI.Rec->getLoc(), "operand has no match class!"); 537 ClassName = "Invalid"; 538 } 539 } 540 541 ClassInfo *&Entry = OperandClasses[ClassName]; 542 543 if (!Entry) { 544 Entry = new ClassInfo(); 545 // FIXME: Hack. 546 if (ClassName == "Reg") { 547 Entry->Kind = ClassInfo::Register; 548 } else { 549 Entry->Kind = getUserClassKind(ClassName); 550 } 551 Entry->ClassName = ClassName; 552 Entry->Name = "MCK_" + ClassName; 553 Entry->ValueName = OI.Rec->getName(); 554 Entry->PredicateMethod = "is" + ClassName; 555 Entry->RenderMethod = "add" + ClassName + "Operands"; 556 Classes.push_back(Entry); 557 } 558 559 return Entry; 560} 561 562void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) { 563 for (std::map<std::string, CodeGenInstruction>::const_iterator 564 it = Target.getInstructions().begin(), 565 ie = Target.getInstructions().end(); 566 it != ie; ++it) { 567 const CodeGenInstruction &CGI = it->second; 568 569 if (!StringRef(it->first).startswith(MatchPrefix)) 570 continue; 571 572 OwningPtr<InstructionInfo> II(new InstructionInfo); 573 574 II->InstrName = it->first; 575 II->Instr = &it->second; 576 II->AsmString = FlattenVariants(CGI.AsmString, 0); 577 578 TokenizeAsmString(II->AsmString, II->Tokens); 579 580 // Ignore instructions which shouldn't be matched. 581 if (!IsAssemblerInstruction(it->first, CGI, II->Tokens)) 582 continue; 583 584 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) { 585 StringRef Token = II->Tokens[i]; 586 587 // Check for simple tokens. 588 if (Token[0] != '$') { 589 InstructionInfo::Operand Op; 590 Op.Class = getTokenClass(Token); 591 Op.OperandInfo = 0; 592 II->Operands.push_back(Op); 593 continue; 594 } 595 596 // Otherwise this is an operand reference. 597 StringRef OperandName; 598 if (Token[1] == '{') 599 OperandName = Token.substr(2, Token.size() - 3); 600 else 601 OperandName = Token.substr(1); 602 603 // Map this token to an operand. FIXME: Move elsewhere. 604 unsigned Idx; 605 try { 606 Idx = CGI.getOperandNamed(OperandName); 607 } catch(...) { 608 errs() << "error: unable to find operand: '" << OperandName << "'!\n"; 609 break; 610 } 611 612 const CodeGenInstruction::OperandInfo &OI = CGI.OperandList[Idx]; 613 InstructionInfo::Operand Op; 614 Op.Class = getOperandClass(Token, OI); 615 Op.OperandInfo = &OI; 616 II->Operands.push_back(Op); 617 } 618 619 // If we broke out, ignore the instruction. 620 if (II->Operands.size() != II->Tokens.size()) 621 continue; 622 623 Instructions.push_back(II.take()); 624 } 625} 626 627static void EmitConvertToMCInst(CodeGenTarget &Target, 628 std::vector<InstructionInfo*> &Infos, 629 raw_ostream &OS) { 630 // Write the convert function to a separate stream, so we can drop it after 631 // the enum. 632 std::string ConvertFnBody; 633 raw_string_ostream CvtOS(ConvertFnBody); 634 635 // Function we have already generated. 636 std::set<std::string> GeneratedFns; 637 638 // Start the unified conversion function. 639 640 CvtOS << "static bool ConvertToMCInst(ConversionKind Kind, MCInst &Inst, " 641 << "unsigned Opcode,\n" 642 << " SmallVectorImpl<" 643 << Target.getName() << "Operand> &Operands) {\n"; 644 CvtOS << " Inst.setOpcode(Opcode);\n"; 645 CvtOS << " switch (Kind) {\n"; 646 CvtOS << " default:\n"; 647 648 // Start the enum, which we will generate inline. 649 650 OS << "// Unified function for converting operants to MCInst instances.\n\n"; 651 OS << "enum ConversionKind {\n"; 652 653 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(), 654 ie = Infos.end(); it != ie; ++it) { 655 InstructionInfo &II = **it; 656 657 // Order the (class) operands by the order to convert them into an MCInst. 658 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList; 659 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) { 660 InstructionInfo::Operand &Op = II.Operands[i]; 661 if (Op.OperandInfo) 662 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i)); 663 } 664 std::sort(MIOperandList.begin(), MIOperandList.end()); 665 666 // Compute the total number of operands. 667 unsigned NumMIOperands = 0; 668 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) { 669 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i]; 670 NumMIOperands = std::max(NumMIOperands, 671 OI.MIOperandNo + OI.MINumOperands); 672 } 673 674 // Build the conversion function signature. 675 std::string Signature = "Convert"; 676 unsigned CurIndex = 0; 677 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) { 678 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second]; 679 assert(CurIndex <= Op.OperandInfo->MIOperandNo && 680 "Duplicate match for instruction operand!"); 681 682 Signature += "_"; 683 684 // Skip operands which weren't matched by anything, this occurs when the 685 // .td file encodes "implicit" operands as explicit ones. 686 // 687 // FIXME: This should be removed from the MCInst structure. 688 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) 689 Signature += "Imp"; 690 691 Signature += Op.Class->ClassName; 692 Signature += utostr(Op.OperandInfo->MINumOperands); 693 Signature += "_" + utostr(MIOperandList[i].second); 694 695 CurIndex += Op.OperandInfo->MINumOperands; 696 } 697 698 // Add any trailing implicit operands. 699 for (; CurIndex != NumMIOperands; ++CurIndex) 700 Signature += "Imp"; 701 702 II.ConversionFnKind = Signature; 703 704 // Check if we have already generated this signature. 705 if (!GeneratedFns.insert(Signature).second) 706 continue; 707 708 // If not, emit it now. 709 710 // Add to the enum list. 711 OS << " " << Signature << ",\n"; 712 713 // And to the convert function. 714 CvtOS << " case " << Signature << ":\n"; 715 CurIndex = 0; 716 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) { 717 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second]; 718 719 // Add the implicit operands. 720 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) 721 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n"; 722 723 CvtOS << " Operands[" << MIOperandList[i].second 724 << "]." << Op.Class->RenderMethod 725 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n"; 726 CurIndex += Op.OperandInfo->MINumOperands; 727 } 728 729 // And add trailing implicit operands. 730 for (; CurIndex != NumMIOperands; ++CurIndex) 731 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n"; 732 CvtOS << " break;\n"; 733 } 734 735 // Finish the convert function. 736 737 CvtOS << " }\n"; 738 CvtOS << " return false;\n"; 739 CvtOS << "}\n\n"; 740 741 // Finish the enum, and drop the convert function after it. 742 743 OS << " NumConversionVariants\n"; 744 OS << "};\n\n"; 745 746 OS << CvtOS.str(); 747} 748 749/// EmitMatchClassEnumeration - Emit the enumeration for match class kinds. 750static void EmitMatchClassEnumeration(CodeGenTarget &Target, 751 std::vector<ClassInfo*> &Infos, 752 raw_ostream &OS) { 753 OS << "namespace {\n\n"; 754 755 OS << "/// MatchClassKind - The kinds of classes which participate in\n" 756 << "/// instruction matching.\n"; 757 OS << "enum MatchClassKind {\n"; 758 OS << " InvalidMatchClass = 0,\n"; 759 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 760 ie = Infos.end(); it != ie; ++it) { 761 ClassInfo &CI = **it; 762 OS << " " << CI.Name << ", // "; 763 if (CI.Kind == ClassInfo::Token) { 764 OS << "'" << CI.ValueName << "'\n"; 765 } else if (CI.Kind == ClassInfo::Register) { 766 if (!CI.ValueName.empty()) 767 OS << "register class '" << CI.ValueName << "'\n"; 768 else 769 OS << "derived register class\n"; 770 } else { 771 OS << "user defined class '" << CI.ValueName << "'\n"; 772 } 773 } 774 OS << " NumMatchClassKinds\n"; 775 OS << "};\n\n"; 776 777 OS << "}\n\n"; 778} 779 780/// EmitClassifyOperand - Emit the function to classify an operand. 781static void EmitClassifyOperand(CodeGenTarget &Target, 782 std::vector<ClassInfo*> &Infos, 783 raw_ostream &OS) { 784 OS << "static MatchClassKind ClassifyOperand(" 785 << Target.getName() << "Operand &Operand) {\n"; 786 OS << " if (Operand.isToken())\n"; 787 OS << " return MatchTokenString(Operand.getToken());\n\n"; 788 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 789 ie = Infos.end(); it != ie; ++it) { 790 ClassInfo &CI = **it; 791 792 if (CI.Kind != ClassInfo::Token) { 793 OS << " if (Operand." << CI.PredicateMethod << "())\n"; 794 OS << " return " << CI.Name << ";\n\n"; 795 } 796 } 797 OS << " return InvalidMatchClass;\n"; 798 OS << "}\n\n"; 799} 800 801typedef std::pair<std::string, std::string> StringPair; 802 803/// FindFirstNonCommonLetter - Find the first character in the keys of the 804/// string pairs that is not shared across the whole set of strings. All 805/// strings are assumed to have the same length. 806static unsigned 807FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) { 808 assert(!Matches.empty()); 809 for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) { 810 // Check to see if letter i is the same across the set. 811 char Letter = Matches[0]->first[i]; 812 813 for (unsigned str = 0, e = Matches.size(); str != e; ++str) 814 if (Matches[str]->first[i] != Letter) 815 return i; 816 } 817 818 return Matches[0]->first.size(); 819} 820 821/// EmitStringMatcherForChar - Given a set of strings that are known to be the 822/// same length and whose characters leading up to CharNo are the same, emit 823/// code to verify that CharNo and later are the same. 824/// 825/// \return - True if control can leave the emitted code fragment. 826static bool EmitStringMatcherForChar(const std::string &StrVariableName, 827 const std::vector<const StringPair*> &Matches, 828 unsigned CharNo, unsigned IndentCount, 829 raw_ostream &OS) { 830 assert(!Matches.empty() && "Must have at least one string to match!"); 831 std::string Indent(IndentCount*2+4, ' '); 832 833 // If we have verified that the entire string matches, we're done: output the 834 // matching code. 835 if (CharNo == Matches[0]->first.size()) { 836 assert(Matches.size() == 1 && "Had duplicate keys to match on"); 837 838 // FIXME: If Matches[0].first has embeded \n, this will be bad. 839 OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first 840 << "\"\n"; 841 return false; 842 } 843 844 // Bucket the matches by the character we are comparing. 845 std::map<char, std::vector<const StringPair*> > MatchesByLetter; 846 847 for (unsigned i = 0, e = Matches.size(); i != e; ++i) 848 MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]); 849 850 851 // If we have exactly one bucket to match, see how many characters are common 852 // across the whole set and match all of them at once. 853 if (MatchesByLetter.size() == 1) { 854 unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches); 855 unsigned NumChars = FirstNonCommonLetter-CharNo; 856 857 // Emit code to break out if the prefix doesn't match. 858 if (NumChars == 1) { 859 // Do the comparison with if (Str[1] != 'f') 860 // FIXME: Need to escape general characters. 861 OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '" 862 << Matches[0]->first[CharNo] << "')\n"; 863 OS << Indent << " break;\n"; 864 } else { 865 // Do the comparison with if (Str.substr(1,3) != "foo"). 866 // FIXME: Need to escape general strings. 867 OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << "," 868 << NumChars << ") != \""; 869 OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n"; 870 OS << Indent << " break;\n"; 871 } 872 873 return EmitStringMatcherForChar(StrVariableName, Matches, 874 FirstNonCommonLetter, IndentCount, OS); 875 } 876 877 // Otherwise, we have multiple possible things, emit a switch on the 878 // character. 879 OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n"; 880 OS << Indent << "default: break;\n"; 881 882 for (std::map<char, std::vector<const StringPair*> >::iterator LI = 883 MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) { 884 // TODO: escape hard stuff (like \n) if we ever care about it. 885 OS << Indent << "case '" << LI->first << "':\t // " 886 << LI->second.size() << " strings to match.\n"; 887 if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1, 888 IndentCount+1, OS)) 889 OS << Indent << " break;\n"; 890 } 891 892 OS << Indent << "}\n"; 893 return true; 894} 895 896 897/// EmitStringMatcher - Given a list of strings and code to execute when they 898/// match, output a simple switch tree to classify the input string. 899/// 900/// If a match is found, the code in Vals[i].second is executed; control must 901/// not exit this code fragment. If nothing matches, execution falls through. 902/// 903/// \param StrVariableName - The name of the variable to test. 904static void EmitStringMatcher(const std::string &StrVariableName, 905 const std::vector<StringPair> &Matches, 906 raw_ostream &OS) { 907 // First level categorization: group strings by length. 908 std::map<unsigned, std::vector<const StringPair*> > MatchesByLength; 909 910 for (unsigned i = 0, e = Matches.size(); i != e; ++i) 911 MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]); 912 913 // Output a switch statement on length and categorize the elements within each 914 // bin. 915 OS << " switch (" << StrVariableName << ".size()) {\n"; 916 OS << " default: break;\n"; 917 918 for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI = 919 MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) { 920 OS << " case " << LI->first << ":\t // " << LI->second.size() 921 << " strings to match.\n"; 922 if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS)) 923 OS << " break;\n"; 924 } 925 926 OS << " }\n"; 927} 928 929 930/// EmitMatchTokenString - Emit the function to match a token string to the 931/// appropriate match class value. 932static void EmitMatchTokenString(CodeGenTarget &Target, 933 std::vector<ClassInfo*> &Infos, 934 raw_ostream &OS) { 935 // Construct the match list. 936 std::vector<StringPair> Matches; 937 for (std::vector<ClassInfo*>::iterator it = Infos.begin(), 938 ie = Infos.end(); it != ie; ++it) { 939 ClassInfo &CI = **it; 940 941 if (CI.Kind == ClassInfo::Token) 942 Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";")); 943 } 944 945 OS << "static MatchClassKind MatchTokenString(const StringRef &Name) {\n"; 946 947 EmitStringMatcher("Name", Matches, OS); 948 949 OS << " return InvalidMatchClass;\n"; 950 OS << "}\n\n"; 951} 952 953/// EmitMatchRegisterName - Emit the function to match a string to the target 954/// specific register enum. 955static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser, 956 raw_ostream &OS) { 957 // Construct the match list. 958 std::vector<StringPair> Matches; 959 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) { 960 const CodeGenRegister &Reg = Target.getRegisters()[i]; 961 if (Reg.TheDef->getValueAsString("AsmName").empty()) 962 continue; 963 964 Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"), 965 "return " + utostr(i + 1) + ";")); 966 } 967 968 OS << "unsigned " << Target.getName() 969 << AsmParser->getValueAsString("AsmParserClassName") 970 << "::MatchRegisterName(const StringRef &Name) {\n"; 971 972 EmitStringMatcher("Name", Matches, OS); 973 974 OS << " return 0;\n"; 975 OS << "}\n\n"; 976} 977 978void AsmMatcherEmitter::run(raw_ostream &OS) { 979 CodeGenTarget Target; 980 Record *AsmParser = Target.getAsmParser(); 981 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName"); 982 983 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS); 984 985 // Emit the function to match a register name to number. 986 EmitMatchRegisterName(Target, AsmParser, OS); 987 988 // Compute the information on the instructions to match. 989 AsmMatcherInfo Info; 990 Info.BuildInfo(Target); 991 992 // Sort the instruction table using the partial order on classes. 993 std::sort(Info.Instructions.begin(), Info.Instructions.end(), 994 less_ptr<InstructionInfo>()); 995 996 DEBUG_WITH_TYPE("instruction_info", { 997 for (std::vector<InstructionInfo*>::iterator 998 it = Info.Instructions.begin(), ie = Info.Instructions.end(); 999 it != ie; ++it) 1000 (*it)->dump(); 1001 }); 1002 1003 // Check for ambiguous instructions. 1004 unsigned NumAmbiguous = 0; 1005 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) { 1006 for (unsigned j = i + 1; j != e; ++j) { 1007 InstructionInfo &A = *Info.Instructions[i]; 1008 InstructionInfo &B = *Info.Instructions[j]; 1009 1010 if (A.CouldMatchAmiguouslyWith(B)) { 1011 DEBUG_WITH_TYPE("ambiguous_instrs", { 1012 errs() << "warning: ambiguous instruction match:\n"; 1013 A.dump(); 1014 errs() << "\nis incomparable with:\n"; 1015 B.dump(); 1016 errs() << "\n\n"; 1017 }); 1018 ++NumAmbiguous; 1019 } 1020 } 1021 } 1022 if (NumAmbiguous) 1023 DEBUG_WITH_TYPE("ambiguous_instrs", { 1024 errs() << "warning: " << NumAmbiguous 1025 << " ambiguous instructions!\n"; 1026 }); 1027 1028 // Generate the unified function to convert operands into an MCInst. 1029 EmitConvertToMCInst(Target, Info.Instructions, OS); 1030 1031 // Emit the enumeration for classes which participate in matching. 1032 EmitMatchClassEnumeration(Target, Info.Classes, OS); 1033 1034 // Emit the routine to match token strings to their match class. 1035 EmitMatchTokenString(Target, Info.Classes, OS); 1036 1037 // Emit the routine to classify an operand. 1038 EmitClassifyOperand(Target, Info.Classes, OS); 1039 1040 // Finally, build the match function. 1041 1042 size_t MaxNumOperands = 0; 1043 for (std::vector<InstructionInfo*>::const_iterator it = 1044 Info.Instructions.begin(), ie = Info.Instructions.end(); 1045 it != ie; ++it) 1046 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size()); 1047 1048 OS << "bool " << Target.getName() << ClassName 1049 << "::MatchInstruction(" 1050 << "SmallVectorImpl<" << Target.getName() << "Operand> &Operands, " 1051 << "MCInst &Inst) {\n"; 1052 1053 // Emit the static match table; unused classes get initalized to 0 which is 1054 // guaranteed to be InvalidMatchClass. 1055 // 1056 // FIXME: We can reduce the size of this table very easily. First, we change 1057 // it so that store the kinds in separate bit-fields for each index, which 1058 // only needs to be the max width used for classes at that index (we also need 1059 // to reject based on this during classification). If we then make sure to 1060 // order the match kinds appropriately (putting mnemonics last), then we 1061 // should only end up using a few bits for each class, especially the ones 1062 // following the mnemonic. 1063 OS << " static const struct MatchEntry {\n"; 1064 OS << " unsigned Opcode;\n"; 1065 OS << " ConversionKind ConvertFn;\n"; 1066 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n"; 1067 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n"; 1068 1069 for (std::vector<InstructionInfo*>::const_iterator it = 1070 Info.Instructions.begin(), ie = Info.Instructions.end(); 1071 it != ie; ++it) { 1072 InstructionInfo &II = **it; 1073 1074 OS << " { " << Target.getName() << "::" << II.InstrName 1075 << ", " << II.ConversionFnKind << ", { "; 1076 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) { 1077 InstructionInfo::Operand &Op = II.Operands[i]; 1078 1079 if (i) OS << ", "; 1080 OS << Op.Class->Name; 1081 } 1082 OS << " } },\n"; 1083 } 1084 1085 OS << " };\n\n"; 1086 1087 // Emit code to compute the class list for this operand vector. 1088 OS << " // Eliminate obvious mismatches.\n"; 1089 OS << " if (Operands.size() > " << MaxNumOperands << ")\n"; 1090 OS << " return true;\n\n"; 1091 1092 OS << " // Compute the class list for this operand vector.\n"; 1093 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n"; 1094 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n"; 1095 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n"; 1096 1097 OS << " // Check for invalid operands before matching.\n"; 1098 OS << " if (Classes[i] == InvalidMatchClass)\n"; 1099 OS << " return true;\n"; 1100 OS << " }\n\n"; 1101 1102 OS << " // Mark unused classes.\n"; 1103 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; " 1104 << "i != e; ++i)\n"; 1105 OS << " Classes[i] = InvalidMatchClass;\n\n"; 1106 1107 // Emit code to search the table. 1108 OS << " // Search the table.\n"; 1109 OS << " for (const MatchEntry *it = MatchTable, " 1110 << "*ie = MatchTable + " << Info.Instructions.size() 1111 << "; it != ie; ++it) {\n"; 1112 for (unsigned i = 0; i != MaxNumOperands; ++i) { 1113 OS << " if (Classes[" << i << "] != it->Classes[" << i << "])\n"; 1114 OS << " continue;\n"; 1115 } 1116 OS << "\n"; 1117 OS << " return ConvertToMCInst(it->ConvertFn, Inst, " 1118 << "it->Opcode, Operands);\n"; 1119 OS << " }\n\n"; 1120 1121 OS << " return true;\n"; 1122 OS << "}\n\n"; 1123} 1124