1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tablegen backend emits a target specifier matcher for converting parsed
11 // assembly operands in the MCInst structures.
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
20 // Some example inputs, for X86:
21 // 'addl' (immediate ...) (register ...)
22 // 'add' (immediate ...) (memory ...)
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:
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.
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
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.
48 // The matching is divided into two distinct phases:
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.
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).
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.
63 // In addition, the subset relation amongst classes induces a partial order
64 // on such tuples, which we use to resolve ambiguities.
66 // FIXME: What do we do if a crazy case shows up where this is the wrong
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.
74 //===----------------------------------------------------------------------===//
76 #include "AsmMatcherEmitter.h"
77 #include "CodeGenTarget.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"
90 static cl::opt<std::string>
91 MatchPrefix("match-prefix", cl::init(""),
92 cl::desc("Only match instructions with the given prefix"));
94 /// FlattenVariants - Flatten an .td file assembly string by selecting the
95 /// variant at index \arg N.
96 static std::string FlattenVariants(const std::string &AsmString,
98 StringRef Cur = AsmString;
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] != '$' &&
107 Cur[VariantsStart-1] != '\\')))
110 // Add the prefix to the result.
111 Res += Cur.slice(0, VariantsStart);
112 if (VariantsStart == Cur.size())
115 ++VariantsStart; // Skip the '{'.
117 // Scan to the end of the variants string.
118 size_t VariantsEnd = VariantsStart;
119 unsigned NestedBraces = 1;
120 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) {
121 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') {
122 if (--NestedBraces == 0)
124 } else if (Cur[VariantsEnd] == '{')
128 // Select the Nth variant (or empty).
129 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd);
130 for (unsigned i = 0; i != N; ++i)
131 Selection = Selection.split('|').second;
132 Res += Selection.split('|').first;
134 assert(VariantsEnd != Cur.size() &&
135 "Unterminated variants in assembly string!");
136 Cur = Cur.substr(VariantsEnd + 1);
142 /// TokenizeAsmString - Tokenize a simplified assembly string.
143 static void TokenizeAsmString(StringRef AsmString,
144 SmallVectorImpl<StringRef> &Tokens) {
147 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) {
148 switch (AsmString[i]) {
157 Tokens.push_back(AsmString.slice(Prev, i));
160 if (!isspace(AsmString[i]) && AsmString[i] != ',')
161 Tokens.push_back(AsmString.substr(i, 1));
167 Tokens.push_back(AsmString.slice(Prev, i));
171 assert(i != AsmString.size() && "Invalid quoted character");
172 Tokens.push_back(AsmString.substr(i, 1));
177 // If this isn't "${", treat like a normal token.
178 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') {
180 Tokens.push_back(AsmString.slice(Prev, i));
188 Tokens.push_back(AsmString.slice(Prev, i));
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));
206 if (InTok && Prev != AsmString.size())
207 Tokens.push_back(AsmString.substr(Prev));
210 static bool IsAssemblerInstruction(StringRef Name,
211 const CodeGenInstruction &CGI,
212 const SmallVectorImpl<StringRef> &Tokens) {
213 // Ignore "codegen only" instructions.
214 if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
217 // Ignore pseudo ops.
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")
225 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
227 // FIXME: This is a total hack.
228 if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int"))
231 // Ignore instructions with no .s string.
233 // FIXME: What are these?
234 if (CGI.AsmString.empty())
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())
242 // Ignore instructions with attributes, these are always fake instructions for
243 // simplifying codegen.
245 // FIXME: Is this true?
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] == '$' &&
252 std::find(Tokens[i].begin(),
253 Tokens[i].end(), ':') != Tokens[i].end()) {
255 errs() << "warning: '" << Name << "': "
256 << "ignoring instruction; operand with attribute '"
257 << Tokens[i] << "'\n";
262 if (Tokens[i][0] == '$' && !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);
274 /// ClassInfo - Helper class for storing the information about a particular
275 /// class of operands which can be matched.
278 /// Invalid kind, for use as a sentinel value.
281 /// The class for a particular token.
284 /// The (first) register class, subsequent register classes are
285 /// RegisterClass0+1, and so on.
288 /// The (first) user defined class, subsequent user defined classes are
289 /// UserClass0+1, and so on.
293 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
294 /// N) for the Nth user defined class.
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;
302 /// Name - The full class name, suitable for use in an enum.
305 /// ClassName - The unadorned generic name for this class (e.g., Token).
306 std::string ClassName;
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;
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;
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;
321 /// For register classes, the records for all the registers in this class.
322 std::set<Record*> Registers;
325 /// isRegisterClass() - Check if this is a register class.
326 bool isRegisterClass() const {
327 return Kind >= RegisterClass0 && Kind < UserClass0;
330 /// isUserClass() - Check if this is a user defined class.
331 bool isUserClass() const {
332 return Kind >= UserClass0;
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;
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())
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(),
357 // Otherwise we have two users operands; they are related if they are in the
358 // same class hierarchy.
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();
367 const ClassInfo *RHSRoot = &RHS;
368 while (!RHSRoot->SuperClasses.empty())
369 RHSRoot = RHSRoot->SuperClasses.front();
371 return Root == RHSRoot;
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...
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))
389 /// operator< - Compare two classes.
390 bool operator<(const ClassInfo &RHS) const {
391 // Unrelated classes can be ordered by kind.
392 if (!isRelatedTo(RHS))
393 return Kind < RHS.Kind;
397 assert(0 && "Invalid kind!");
399 // Tokens are comparable by value.
401 // FIXME: Compare by enum value.
402 return ValueName < RHS.ValueName;
405 // This class preceeds the RHS if it is a proper subset of the RHS.
406 return this != &RHS && isSubsetOf(RHS);
411 /// InstructionInfo - Helper class for storing the necessary information for an
412 /// instruction which is capable of being matched.
413 struct InstructionInfo {
415 /// The unique class instance this operand should match.
418 /// The original operand this corresponds to, if any.
419 const CodeGenInstruction::OperandInfo *OperandInfo;
422 /// InstrName - The target name for this instruction.
423 std::string InstrName;
425 /// Instr - The instruction this matches.
426 const CodeGenInstruction *Instr;
428 /// AsmString - The assembly string for this instruction (with variants
430 std::string AsmString;
432 /// Tokens - The tokenized assembly pattern that this instruction matches.
433 SmallVector<StringRef, 4> Tokens;
435 /// Operands - The operands that this instruction matches.
436 SmallVector<Operand, 4> Operands;
438 /// ConversionFnKind - The enum value which is passed to the generated
439 /// ConvertToMCInst to convert parsed operands into an MCInst for this
441 std::string ConversionFnKind;
443 /// operator< - Compare two instructions.
444 bool operator<(const InstructionInfo &RHS) const {
445 if (Operands.size() != RHS.Operands.size())
446 return Operands.size() < RHS.Operands.size();
448 // Compare lexicographically by operand. The matcher validates that other
449 // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith().
450 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
451 if (*Operands[i].Class < *RHS.Operands[i].Class)
453 if (*RHS.Operands[i].Class < *Operands[i].Class)
460 /// CouldMatchAmiguouslyWith - Check whether this instruction could
461 /// ambiguously match the same set of operands as \arg RHS (without being a
462 /// strictly superior match).
463 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
464 // The number of operands is unambiguous.
465 if (Operands.size() != RHS.Operands.size())
468 // Otherwise, make sure the ordering of the two instructions is unambiguous
469 // by checking that either (a) a token or operand kind discriminates them,
470 // or (b) the ordering among equivalent kinds is consistent.
472 // Tokens and operand kinds are unambiguous (assuming a correct target
474 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
475 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
476 Operands[i].Class->Kind == ClassInfo::Token)
477 if (*Operands[i].Class < *RHS.Operands[i].Class ||
478 *RHS.Operands[i].Class < *Operands[i].Class)
481 // Otherwise, this operand could commute if all operands are equivalent, or
482 // there is a pair of operands that compare less than and a pair that
483 // compare greater than.
484 bool HasLT = false, HasGT = false;
485 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
486 if (*Operands[i].Class < *RHS.Operands[i].Class)
488 if (*RHS.Operands[i].Class < *Operands[i].Class)
492 return !(HasLT ^ HasGT);
499 class AsmMatcherInfo {
501 /// The tablegen AsmParser record.
504 /// The AsmParser "CommentDelimiter" value.
505 std::string CommentDelimiter;
507 /// The AsmParser "RegisterPrefix" value.
508 std::string RegisterPrefix;
510 /// The classes which are needed for matching.
511 std::vector<ClassInfo*> Classes;
513 /// The information on the instruction to match.
514 std::vector<InstructionInfo*> Instructions;
516 /// Map of Register records to their class information.
517 std::map<Record*, ClassInfo*> RegisterClasses;
520 /// Map of token to class information which has already been constructed.
521 std::map<std::string, ClassInfo*> TokenClasses;
523 /// Map of RegisterClass records to their class information.
524 std::map<Record*, ClassInfo*> RegisterClassClasses;
526 /// Map of AsmOperandClass records to their class information.
527 std::map<Record*, ClassInfo*> AsmOperandClasses;
530 /// getTokenClass - Lookup or create the class for the given token.
531 ClassInfo *getTokenClass(StringRef Token);
533 /// getOperandClass - Lookup or create the class for the given operand.
534 ClassInfo *getOperandClass(StringRef Token,
535 const CodeGenInstruction::OperandInfo &OI);
537 /// BuildRegisterClasses - Build the ClassInfo* instances for register
539 void BuildRegisterClasses(CodeGenTarget &Target,
540 std::set<std::string> &SingletonRegisterNames);
542 /// BuildOperandClasses - Build the ClassInfo* instances for user defined
544 void BuildOperandClasses(CodeGenTarget &Target);
547 AsmMatcherInfo(Record *_AsmParser);
549 /// BuildInfo - Construct the various tables used during matching.
550 void BuildInfo(CodeGenTarget &Target);
555 void InstructionInfo::dump() {
556 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
558 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
565 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
566 Operand &Op = Operands[i];
567 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
568 if (Op.Class->Kind == ClassInfo::Token) {
569 errs() << '\"' << Tokens[i] << "\"\n";
573 if (!Op.OperandInfo) {
574 errs() << "(singleton register)\n";
578 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
579 errs() << OI.Name << " " << OI.Rec->getName()
580 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
584 static std::string getEnumNameForToken(StringRef Str) {
587 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
589 case '*': Res += "_STAR_"; break;
590 case '%': Res += "_PCT_"; break;
591 case ':': Res += "_COLON_"; break;
597 Res += "_" + utostr((unsigned) *it) + "_";
605 /// getRegisterRecord - Get the register record for \arg name, or 0.
606 static Record *getRegisterRecord(CodeGenTarget &Target, StringRef Name) {
607 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
608 const CodeGenRegister &Reg = Target.getRegisters()[i];
609 if (Name == Reg.TheDef->getValueAsString("AsmName"))
616 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
617 ClassInfo *&Entry = TokenClasses[Token];
620 Entry = new ClassInfo();
621 Entry->Kind = ClassInfo::Token;
622 Entry->ClassName = "Token";
623 Entry->Name = "MCK_" + getEnumNameForToken(Token);
624 Entry->ValueName = Token;
625 Entry->PredicateMethod = "<invalid>";
626 Entry->RenderMethod = "<invalid>";
627 Classes.push_back(Entry);
634 AsmMatcherInfo::getOperandClass(StringRef Token,
635 const CodeGenInstruction::OperandInfo &OI) {
636 if (OI.Rec->isSubClassOf("RegisterClass")) {
637 ClassInfo *CI = RegisterClassClasses[OI.Rec];
640 PrintError(OI.Rec->getLoc(), "register class has no class info!");
641 throw std::string("ERROR: Missing register class!");
647 assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!");
648 Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass");
649 ClassInfo *CI = AsmOperandClasses[MatchClass];
652 PrintError(OI.Rec->getLoc(), "operand has no match class!");
653 throw std::string("ERROR: Missing match class!");
659 void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target,
660 std::set<std::string>
661 &SingletonRegisterNames) {
662 std::vector<CodeGenRegisterClass> RegisterClasses;
663 std::vector<CodeGenRegister> Registers;
665 RegisterClasses = Target.getRegisterClasses();
666 Registers = Target.getRegisters();
668 // The register sets used for matching.
669 std::set< std::set<Record*> > RegisterSets;
671 // Gather the defined sets.
672 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
673 ie = RegisterClasses.end(); it != ie; ++it)
674 RegisterSets.insert(std::set<Record*>(it->Elements.begin(),
675 it->Elements.end()));
677 // Add any required singleton sets.
678 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
679 ie = SingletonRegisterNames.end(); it != ie; ++it)
680 if (Record *Rec = getRegisterRecord(Target, *it))
681 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
683 // Introduce derived sets where necessary (when a register does not determine
684 // a unique register set class), and build the mapping of registers to the set
685 // they should classify to.
686 std::map<Record*, std::set<Record*> > RegisterMap;
687 for (std::vector<CodeGenRegister>::iterator it = Registers.begin(),
688 ie = Registers.end(); it != ie; ++it) {
689 CodeGenRegister &CGR = *it;
690 // Compute the intersection of all sets containing this register.
691 std::set<Record*> ContainingSet;
693 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
694 ie = RegisterSets.end(); it != ie; ++it) {
695 if (!it->count(CGR.TheDef))
698 if (ContainingSet.empty()) {
701 std::set<Record*> Tmp;
702 std::swap(Tmp, ContainingSet);
703 std::insert_iterator< std::set<Record*> > II(ContainingSet,
704 ContainingSet.begin());
705 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(),
710 if (!ContainingSet.empty()) {
711 RegisterSets.insert(ContainingSet);
712 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
716 // Construct the register classes.
717 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
719 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
720 ie = RegisterSets.end(); it != ie; ++it, ++Index) {
721 ClassInfo *CI = new ClassInfo();
722 CI->Kind = ClassInfo::RegisterClass0 + Index;
723 CI->ClassName = "Reg" + utostr(Index);
724 CI->Name = "MCK_Reg" + utostr(Index);
726 CI->PredicateMethod = ""; // unused
727 CI->RenderMethod = "addRegOperands";
729 Classes.push_back(CI);
730 RegisterSetClasses.insert(std::make_pair(*it, CI));
733 // Find the superclasses; we could compute only the subgroup lattice edges,
734 // but there isn't really a point.
735 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
736 ie = RegisterSets.end(); it != ie; ++it) {
737 ClassInfo *CI = RegisterSetClasses[*it];
738 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
739 ie2 = RegisterSets.end(); it2 != ie2; ++it2)
741 std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
742 CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
745 // Name the register classes which correspond to a user defined RegisterClass.
746 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
747 ie = RegisterClasses.end(); it != ie; ++it) {
748 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(),
749 it->Elements.end())];
750 if (CI->ValueName.empty()) {
751 CI->ClassName = it->getName();
752 CI->Name = "MCK_" + it->getName();
753 CI->ValueName = it->getName();
755 CI->ValueName = CI->ValueName + "," + it->getName();
757 RegisterClassClasses.insert(std::make_pair(it->TheDef, CI));
760 // Populate the map for individual registers.
761 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
762 ie = RegisterMap.end(); it != ie; ++it)
763 this->RegisterClasses[it->first] = RegisterSetClasses[it->second];
765 // Name the register classes which correspond to singleton registers.
766 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
767 ie = SingletonRegisterNames.end(); it != ie; ++it) {
768 if (Record *Rec = getRegisterRecord(Target, *it)) {
769 ClassInfo *CI = this->RegisterClasses[Rec];
770 assert(CI && "Missing singleton register class info!");
772 if (CI->ValueName.empty()) {
773 CI->ClassName = Rec->getName();
774 CI->Name = "MCK_" + Rec->getName();
775 CI->ValueName = Rec->getName();
777 CI->ValueName = CI->ValueName + "," + Rec->getName();
782 void AsmMatcherInfo::BuildOperandClasses(CodeGenTarget &Target) {
783 std::vector<Record*> AsmOperands;
784 AsmOperands = Records.getAllDerivedDefinitions("AsmOperandClass");
786 // Pre-populate AsmOperandClasses map.
787 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
788 ie = AsmOperands.end(); it != ie; ++it)
789 AsmOperandClasses[*it] = new ClassInfo();
792 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
793 ie = AsmOperands.end(); it != ie; ++it, ++Index) {
794 ClassInfo *CI = AsmOperandClasses[*it];
795 CI->Kind = ClassInfo::UserClass0 + Index;
797 Init *Super = (*it)->getValueInit("SuperClass");
798 if (DefInit *DI = dynamic_cast<DefInit*>(Super)) {
799 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
801 PrintError((*it)->getLoc(), "Invalid super class reference!");
803 CI->SuperClasses.push_back(SC);
805 assert(dynamic_cast<UnsetInit*>(Super) && "Unexpected SuperClass field!");
807 CI->ClassName = (*it)->getValueAsString("Name");
808 CI->Name = "MCK_" + CI->ClassName;
809 CI->ValueName = (*it)->getName();
811 // Get or construct the predicate method name.
812 Init *PMName = (*it)->getValueInit("PredicateMethod");
813 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
814 CI->PredicateMethod = SI->getValue();
816 assert(dynamic_cast<UnsetInit*>(PMName) &&
817 "Unexpected PredicateMethod field!");
818 CI->PredicateMethod = "is" + CI->ClassName;
821 // Get or construct the render method name.
822 Init *RMName = (*it)->getValueInit("RenderMethod");
823 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
824 CI->RenderMethod = SI->getValue();
826 assert(dynamic_cast<UnsetInit*>(RMName) &&
827 "Unexpected RenderMethod field!");
828 CI->RenderMethod = "add" + CI->ClassName + "Operands";
831 AsmOperandClasses[*it] = CI;
832 Classes.push_back(CI);
836 AsmMatcherInfo::AsmMatcherInfo(Record *_AsmParser)
837 : AsmParser(_AsmParser),
838 CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")),
839 RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix"))
843 void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
844 // Parse the instructions; we need to do this first so that we can gather the
845 // singleton register classes.
846 std::set<std::string> SingletonRegisterNames;
848 const std::vector<const CodeGenInstruction*> &InstrList =
849 Target.getInstructionsByEnumValue();
851 for (unsigned i = 0, e = InstrList.size(); i != e; ++i) {
852 const CodeGenInstruction &CGI = *InstrList[i];
854 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
857 OwningPtr<InstructionInfo> II(new InstructionInfo());
859 II->InstrName = CGI.TheDef->getName();
861 II->AsmString = FlattenVariants(CGI.AsmString, 0);
863 // Remove comments from the asm string.
864 if (!CommentDelimiter.empty()) {
865 size_t Idx = StringRef(II->AsmString).find(CommentDelimiter);
866 if (Idx != StringRef::npos)
867 II->AsmString = II->AsmString.substr(0, Idx);
870 TokenizeAsmString(II->AsmString, II->Tokens);
872 // Ignore instructions which shouldn't be matched.
873 if (!IsAssemblerInstruction(CGI.TheDef->getName(), CGI, II->Tokens))
876 // Collect singleton registers, if used.
877 if (!RegisterPrefix.empty()) {
878 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
879 if (II->Tokens[i].startswith(RegisterPrefix)) {
880 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
881 Record *Rec = getRegisterRecord(Target, RegName);
884 std::string Err = "unable to find register for '" + RegName.str() +
885 "' (which matches register prefix)";
886 throw TGError(CGI.TheDef->getLoc(), Err);
889 SingletonRegisterNames.insert(RegName);
894 Instructions.push_back(II.take());
897 // Build info for the register classes.
898 BuildRegisterClasses(Target, SingletonRegisterNames);
900 // Build info for the user defined assembly operand classes.
901 BuildOperandClasses(Target);
903 // Build the instruction information.
904 for (std::vector<InstructionInfo*>::iterator it = Instructions.begin(),
905 ie = Instructions.end(); it != ie; ++it) {
906 InstructionInfo *II = *it;
908 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
909 StringRef Token = II->Tokens[i];
911 // Check for singleton registers.
912 if (!RegisterPrefix.empty() && Token.startswith(RegisterPrefix)) {
913 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
914 InstructionInfo::Operand Op;
915 Op.Class = RegisterClasses[getRegisterRecord(Target, RegName)];
917 assert(Op.Class && Op.Class->Registers.size() == 1 &&
918 "Unexpected class for singleton register");
919 II->Operands.push_back(Op);
923 // Check for simple tokens.
924 if (Token[0] != '$') {
925 InstructionInfo::Operand Op;
926 Op.Class = getTokenClass(Token);
928 II->Operands.push_back(Op);
932 // Otherwise this is an operand reference.
933 StringRef OperandName;
935 OperandName = Token.substr(2, Token.size() - 3);
937 OperandName = Token.substr(1);
939 // Map this token to an operand. FIXME: Move elsewhere.
942 Idx = II->Instr->getOperandNamed(OperandName);
944 throw std::string("error: unable to find operand: '" +
945 OperandName.str() + "'");
948 // FIXME: This is annoying, the named operand may be tied (e.g.,
949 // XCHG8rm). What we want is the untied operand, which we now have to
950 // grovel for. Only worry about this for single entry operands, we have to
951 // clean this up anyway.
952 const CodeGenInstruction::OperandInfo *OI = &II->Instr->OperandList[Idx];
953 if (OI->Constraints[0].isTied()) {
954 unsigned TiedOp = OI->Constraints[0].getTiedOperand();
956 // The tied operand index is an MIOperand index, find the operand that
958 for (unsigned i = 0, e = II->Instr->OperandList.size(); i != e; ++i) {
959 if (II->Instr->OperandList[i].MIOperandNo == TiedOp) {
960 OI = &II->Instr->OperandList[i];
965 assert(OI && "Unable to find tied operand target!");
968 InstructionInfo::Operand Op;
969 Op.Class = getOperandClass(Token, *OI);
971 II->Operands.push_back(Op);
975 // Reorder classes so that classes preceed super classes.
976 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
979 static std::pair<unsigned, unsigned> *
980 GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List,
982 for (unsigned i = 0, e = List.size(); i != e; ++i)
983 if (Index == List[i].first)
989 static void EmitConvertToMCInst(CodeGenTarget &Target,
990 std::vector<InstructionInfo*> &Infos,
992 // Write the convert function to a separate stream, so we can drop it after
994 std::string ConvertFnBody;
995 raw_string_ostream CvtOS(ConvertFnBody);
997 // Function we have already generated.
998 std::set<std::string> GeneratedFns;
1000 // Start the unified conversion function.
1002 CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
1003 << "unsigned Opcode,\n"
1004 << " const SmallVectorImpl<MCParsedAsmOperand*"
1005 << "> &Operands) {\n";
1006 CvtOS << " Inst.setOpcode(Opcode);\n";
1007 CvtOS << " switch (Kind) {\n";
1008 CvtOS << " default:\n";
1010 // Start the enum, which we will generate inline.
1012 OS << "// Unified function for converting operants to MCInst instances.\n\n";
1013 OS << "enum ConversionKind {\n";
1015 // TargetOperandClass - This is the target's operand class, like X86Operand.
1016 std::string TargetOperandClass = Target.getName() + "Operand";
1018 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
1019 ie = Infos.end(); it != ie; ++it) {
1020 InstructionInfo &II = **it;
1022 // Order the (class) operands by the order to convert them into an MCInst.
1023 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
1024 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1025 InstructionInfo::Operand &Op = II.Operands[i];
1027 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
1030 // Find any tied operands.
1031 SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands;
1032 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
1033 const CodeGenInstruction::OperandInfo &OpInfo = II.Instr->OperandList[i];
1034 for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) {
1035 const CodeGenInstruction::ConstraintInfo &CI = OpInfo.Constraints[j];
1037 TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo + j,
1038 CI.getTiedOperand()));
1042 std::sort(MIOperandList.begin(), MIOperandList.end());
1044 // Compute the total number of operands.
1045 unsigned NumMIOperands = 0;
1046 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
1047 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i];
1048 NumMIOperands = std::max(NumMIOperands,
1049 OI.MIOperandNo + OI.MINumOperands);
1052 // Build the conversion function signature.
1053 std::string Signature = "Convert";
1054 unsigned CurIndex = 0;
1055 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1056 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1057 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
1058 "Duplicate match for instruction operand!");
1060 // Skip operands which weren't matched by anything, this occurs when the
1061 // .td file encodes "implicit" operands as explicit ones.
1063 // FIXME: This should be removed from the MCInst structure.
1064 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1065 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1068 Signature += "__Imp";
1070 Signature += "__Tie" + utostr(Tie->second);
1075 // Registers are always converted the same, don't duplicate the conversion
1076 // function based on them.
1078 // FIXME: We could generalize this based on the render method, if it
1080 if (Op.Class->isRegisterClass())
1083 Signature += Op.Class->ClassName;
1084 Signature += utostr(Op.OperandInfo->MINumOperands);
1085 Signature += "_" + utostr(MIOperandList[i].second);
1087 CurIndex += Op.OperandInfo->MINumOperands;
1090 // Add any trailing implicit operands.
1091 for (; CurIndex != NumMIOperands; ++CurIndex) {
1092 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1095 Signature += "__Imp";
1097 Signature += "__Tie" + utostr(Tie->second);
1100 II.ConversionFnKind = Signature;
1102 // Check if we have already generated this signature.
1103 if (!GeneratedFns.insert(Signature).second)
1106 // If not, emit it now.
1108 // Add to the enum list.
1109 OS << " " << Signature << ",\n";
1111 // And to the convert function.
1112 CvtOS << " case " << Signature << ":\n";
1114 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1115 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1117 // Add the implicit operands.
1118 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1119 // See if this is a tied operand.
1120 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1124 // If not, this is some implicit operand. Just assume it is a register
1126 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1128 // Copy the tied operand.
1129 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1130 CvtOS << " Inst.addOperand(Inst.getOperand("
1131 << Tie->second << "));\n";
1135 CvtOS << " ((" << TargetOperandClass << "*)Operands["
1136 << MIOperandList[i].second
1137 << "])->" << Op.Class->RenderMethod
1138 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
1139 CurIndex += Op.OperandInfo->MINumOperands;
1142 // And add trailing implicit operands.
1143 for (; CurIndex != NumMIOperands; ++CurIndex) {
1144 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1148 // If not, this is some implicit operand. Just assume it is a register
1150 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1152 // Copy the tied operand.
1153 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1154 CvtOS << " Inst.addOperand(Inst.getOperand("
1155 << Tie->second << "));\n";
1159 CvtOS << " return;\n";
1162 // Finish the convert function.
1167 // Finish the enum, and drop the convert function after it.
1169 OS << " NumConversionVariants\n";
1175 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
1176 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
1177 std::vector<ClassInfo*> &Infos,
1179 OS << "namespace {\n\n";
1181 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1182 << "/// instruction matching.\n";
1183 OS << "enum MatchClassKind {\n";
1184 OS << " InvalidMatchClass = 0,\n";
1185 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1186 ie = Infos.end(); it != ie; ++it) {
1187 ClassInfo &CI = **it;
1188 OS << " " << CI.Name << ", // ";
1189 if (CI.Kind == ClassInfo::Token) {
1190 OS << "'" << CI.ValueName << "'\n";
1191 } else if (CI.isRegisterClass()) {
1192 if (!CI.ValueName.empty())
1193 OS << "register class '" << CI.ValueName << "'\n";
1195 OS << "derived register class\n";
1197 OS << "user defined class '" << CI.ValueName << "'\n";
1200 OS << " NumMatchClassKinds\n";
1206 /// EmitClassifyOperand - Emit the function to classify an operand.
1207 static void EmitClassifyOperand(CodeGenTarget &Target,
1208 AsmMatcherInfo &Info,
1210 OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n"
1211 << " " << Target.getName() << "Operand &Operand = *("
1212 << Target.getName() << "Operand*)GOp;\n";
1215 OS << " if (Operand.isToken())\n";
1216 OS << " return MatchTokenString(Operand.getToken());\n\n";
1218 // Classify registers.
1220 // FIXME: Don't hardcode isReg, getReg.
1221 OS << " if (Operand.isReg()) {\n";
1222 OS << " switch (Operand.getReg()) {\n";
1223 OS << " default: return InvalidMatchClass;\n";
1224 for (std::map<Record*, ClassInfo*>::iterator
1225 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1227 OS << " case " << Target.getName() << "::"
1228 << it->first->getName() << ": return " << it->second->Name << ";\n";
1232 // Classify user defined operands.
1233 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1234 ie = Info.Classes.end(); it != ie; ++it) {
1235 ClassInfo &CI = **it;
1237 if (!CI.isUserClass())
1240 OS << " // '" << CI.ClassName << "' class";
1241 if (!CI.SuperClasses.empty()) {
1242 OS << ", subclass of ";
1243 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) {
1245 OS << "'" << CI.SuperClasses[i]->ClassName << "'";
1246 assert(CI < *CI.SuperClasses[i] && "Invalid class relation!");
1251 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
1253 // Validate subclass relationships.
1254 if (!CI.SuperClasses.empty()) {
1255 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i)
1256 OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod
1257 << "() && \"Invalid class relationship!\");\n";
1260 OS << " return " << CI.Name << ";\n";
1263 OS << " return InvalidMatchClass;\n";
1267 /// EmitIsSubclass - Emit the subclass predicate function.
1268 static void EmitIsSubclass(CodeGenTarget &Target,
1269 std::vector<ClassInfo*> &Infos,
1271 OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1272 OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n";
1273 OS << " if (A == B)\n";
1274 OS << " return true;\n\n";
1276 OS << " switch (A) {\n";
1277 OS << " default:\n";
1278 OS << " return false;\n";
1279 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1280 ie = Infos.end(); it != ie; ++it) {
1281 ClassInfo &A = **it;
1283 if (A.Kind != ClassInfo::Token) {
1284 std::vector<StringRef> SuperClasses;
1285 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1286 ie = Infos.end(); it != ie; ++it) {
1287 ClassInfo &B = **it;
1289 if (&A != &B && A.isSubsetOf(B))
1290 SuperClasses.push_back(B.Name);
1293 if (SuperClasses.empty())
1296 OS << "\n case " << A.Name << ":\n";
1298 if (SuperClasses.size() == 1) {
1299 OS << " return B == " << SuperClasses.back() << ";\n";
1303 OS << " switch (B) {\n";
1304 OS << " default: return false;\n";
1305 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1306 OS << " case " << SuperClasses[i] << ": return true;\n";
1314 typedef std::pair<std::string, std::string> StringPair;
1316 /// FindFirstNonCommonLetter - Find the first character in the keys of the
1317 /// string pairs that is not shared across the whole set of strings. All
1318 /// strings are assumed to have the same length.
1320 FindFirstNonCommonLetter(const std::vector<const StringPair*> &Matches) {
1321 assert(!Matches.empty());
1322 for (unsigned i = 0, e = Matches[0]->first.size(); i != e; ++i) {
1323 // Check to see if letter i is the same across the set.
1324 char Letter = Matches[0]->first[i];
1326 for (unsigned str = 0, e = Matches.size(); str != e; ++str)
1327 if (Matches[str]->first[i] != Letter)
1331 return Matches[0]->first.size();
1334 /// EmitStringMatcherForChar - Given a set of strings that are known to be the
1335 /// same length and whose characters leading up to CharNo are the same, emit
1336 /// code to verify that CharNo and later are the same.
1338 /// \return - True if control can leave the emitted code fragment.
1339 static bool EmitStringMatcherForChar(const std::string &StrVariableName,
1340 const std::vector<const StringPair*> &Matches,
1341 unsigned CharNo, unsigned IndentCount,
1343 assert(!Matches.empty() && "Must have at least one string to match!");
1344 std::string Indent(IndentCount*2+4, ' ');
1346 // If we have verified that the entire string matches, we're done: output the
1348 if (CharNo == Matches[0]->first.size()) {
1349 assert(Matches.size() == 1 && "Had duplicate keys to match on");
1351 // FIXME: If Matches[0].first has embeded \n, this will be bad.
1352 OS << Indent << Matches[0]->second << "\t // \"" << Matches[0]->first
1357 // Bucket the matches by the character we are comparing.
1358 std::map<char, std::vector<const StringPair*> > MatchesByLetter;
1360 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
1361 MatchesByLetter[Matches[i]->first[CharNo]].push_back(Matches[i]);
1364 // If we have exactly one bucket to match, see how many characters are common
1365 // across the whole set and match all of them at once.
1366 if (MatchesByLetter.size() == 1) {
1367 unsigned FirstNonCommonLetter = FindFirstNonCommonLetter(Matches);
1368 unsigned NumChars = FirstNonCommonLetter-CharNo;
1370 // Emit code to break out if the prefix doesn't match.
1371 if (NumChars == 1) {
1372 // Do the comparison with if (Str[1] != 'f')
1373 // FIXME: Need to escape general characters.
1374 OS << Indent << "if (" << StrVariableName << "[" << CharNo << "] != '"
1375 << Matches[0]->first[CharNo] << "')\n";
1376 OS << Indent << " break;\n";
1378 // Do the comparison with if (Str.substr(1,3) != "foo").
1379 // FIXME: Need to escape general strings.
1380 OS << Indent << "if (" << StrVariableName << ".substr(" << CharNo << ","
1381 << NumChars << ") != \"";
1382 OS << Matches[0]->first.substr(CharNo, NumChars) << "\")\n";
1383 OS << Indent << " break;\n";
1386 return EmitStringMatcherForChar(StrVariableName, Matches,
1387 FirstNonCommonLetter, IndentCount, OS);
1390 // Otherwise, we have multiple possible things, emit a switch on the
1392 OS << Indent << "switch (" << StrVariableName << "[" << CharNo << "]) {\n";
1393 OS << Indent << "default: break;\n";
1395 for (std::map<char, std::vector<const StringPair*> >::iterator LI =
1396 MatchesByLetter.begin(), E = MatchesByLetter.end(); LI != E; ++LI) {
1397 // TODO: escape hard stuff (like \n) if we ever care about it.
1398 OS << Indent << "case '" << LI->first << "':\t // "
1399 << LI->second.size() << " strings to match.\n";
1400 if (EmitStringMatcherForChar(StrVariableName, LI->second, CharNo+1,
1402 OS << Indent << " break;\n";
1405 OS << Indent << "}\n";
1410 /// EmitStringMatcher - Given a list of strings and code to execute when they
1411 /// match, output a simple switch tree to classify the input string.
1413 /// If a match is found, the code in Vals[i].second is executed; control must
1414 /// not exit this code fragment. If nothing matches, execution falls through.
1416 /// \param StrVariableName - The name of the variable to test.
1417 static void EmitStringMatcher(const std::string &StrVariableName,
1418 const std::vector<StringPair> &Matches,
1420 // First level categorization: group strings by length.
1421 std::map<unsigned, std::vector<const StringPair*> > MatchesByLength;
1423 for (unsigned i = 0, e = Matches.size(); i != e; ++i)
1424 MatchesByLength[Matches[i].first.size()].push_back(&Matches[i]);
1426 // Output a switch statement on length and categorize the elements within each
1428 OS << " switch (" << StrVariableName << ".size()) {\n";
1429 OS << " default: break;\n";
1431 for (std::map<unsigned, std::vector<const StringPair*> >::iterator LI =
1432 MatchesByLength.begin(), E = MatchesByLength.end(); LI != E; ++LI) {
1433 OS << " case " << LI->first << ":\t // " << LI->second.size()
1434 << " strings to match.\n";
1435 if (EmitStringMatcherForChar(StrVariableName, LI->second, 0, 0, OS))
1443 /// EmitMatchTokenString - Emit the function to match a token string to the
1444 /// appropriate match class value.
1445 static void EmitMatchTokenString(CodeGenTarget &Target,
1446 std::vector<ClassInfo*> &Infos,
1448 // Construct the match list.
1449 std::vector<StringPair> Matches;
1450 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1451 ie = Infos.end(); it != ie; ++it) {
1452 ClassInfo &CI = **it;
1454 if (CI.Kind == ClassInfo::Token)
1455 Matches.push_back(StringPair(CI.ValueName, "return " + CI.Name + ";"));
1458 OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
1460 EmitStringMatcher("Name", Matches, OS);
1462 OS << " return InvalidMatchClass;\n";
1466 /// EmitMatchRegisterName - Emit the function to match a string to the target
1467 /// specific register enum.
1468 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1470 // Construct the match list.
1471 std::vector<StringPair> Matches;
1472 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1473 const CodeGenRegister &Reg = Target.getRegisters()[i];
1474 if (Reg.TheDef->getValueAsString("AsmName").empty())
1477 Matches.push_back(StringPair(Reg.TheDef->getValueAsString("AsmName"),
1478 "return " + utostr(i + 1) + ";"));
1481 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1483 EmitStringMatcher("Name", Matches, OS);
1485 OS << " return 0;\n";
1489 void AsmMatcherEmitter::run(raw_ostream &OS) {
1490 CodeGenTarget Target;
1491 Record *AsmParser = Target.getAsmParser();
1492 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1494 // Compute the information on the instructions to match.
1495 AsmMatcherInfo Info(AsmParser);
1496 Info.BuildInfo(Target);
1498 // Sort the instruction table using the partial order on classes. We use
1499 // stable_sort to ensure that ambiguous instructions are still
1500 // deterministically ordered.
1501 std::stable_sort(Info.Instructions.begin(), Info.Instructions.end(),
1502 less_ptr<InstructionInfo>());
1504 DEBUG_WITH_TYPE("instruction_info", {
1505 for (std::vector<InstructionInfo*>::iterator
1506 it = Info.Instructions.begin(), ie = Info.Instructions.end();
1511 // Check for ambiguous instructions.
1512 unsigned NumAmbiguous = 0;
1513 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
1514 for (unsigned j = i + 1; j != e; ++j) {
1515 InstructionInfo &A = *Info.Instructions[i];
1516 InstructionInfo &B = *Info.Instructions[j];
1518 if (A.CouldMatchAmiguouslyWith(B)) {
1519 DEBUG_WITH_TYPE("ambiguous_instrs", {
1520 errs() << "warning: ambiguous instruction match:\n";
1522 errs() << "\nis incomparable with:\n";
1531 DEBUG_WITH_TYPE("ambiguous_instrs", {
1532 errs() << "warning: " << NumAmbiguous
1533 << " ambiguous instructions!\n";
1536 // Write the output.
1538 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1540 // Emit the function to match a register name to number.
1541 EmitMatchRegisterName(Target, AsmParser, OS);
1543 OS << "#ifndef REGISTERS_ONLY\n\n";
1545 // Generate the unified function to convert operands into an MCInst.
1546 EmitConvertToMCInst(Target, Info.Instructions, OS);
1548 // Emit the enumeration for classes which participate in matching.
1549 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1551 // Emit the routine to match token strings to their match class.
1552 EmitMatchTokenString(Target, Info.Classes, OS);
1554 // Emit the routine to classify an operand.
1555 EmitClassifyOperand(Target, Info, OS);
1557 // Emit the subclass predicate routine.
1558 EmitIsSubclass(Target, Info.Classes, OS);
1560 // Finally, build the match function.
1562 size_t MaxNumOperands = 0;
1563 for (std::vector<InstructionInfo*>::const_iterator it =
1564 Info.Instructions.begin(), ie = Info.Instructions.end();
1566 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
1568 const std::string &MatchName =
1569 AsmParser->getValueAsString("MatchInstructionName");
1570 OS << "bool " << Target.getName() << ClassName << "::\n"
1572 << "(const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
1573 OS.indent(MatchName.size() + 1);
1574 OS << "MCInst &Inst) {\n";
1576 // Emit the static match table; unused classes get initalized to 0 which is
1577 // guaranteed to be InvalidMatchClass.
1579 // FIXME: We can reduce the size of this table very easily. First, we change
1580 // it so that store the kinds in separate bit-fields for each index, which
1581 // only needs to be the max width used for classes at that index (we also need
1582 // to reject based on this during classification). If we then make sure to
1583 // order the match kinds appropriately (putting mnemonics last), then we
1584 // should only end up using a few bits for each class, especially the ones
1585 // following the mnemonic.
1586 OS << " static const struct MatchEntry {\n";
1587 OS << " unsigned Opcode;\n";
1588 OS << " ConversionKind ConvertFn;\n";
1589 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1590 OS << " } MatchTable[" << Info.Instructions.size() << "] = {\n";
1592 for (std::vector<InstructionInfo*>::const_iterator it =
1593 Info.Instructions.begin(), ie = Info.Instructions.end();
1595 InstructionInfo &II = **it;
1597 OS << " { " << Target.getName() << "::" << II.InstrName
1598 << ", " << II.ConversionFnKind << ", { ";
1599 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1600 InstructionInfo::Operand &Op = II.Operands[i];
1603 OS << Op.Class->Name;
1610 // Emit code to compute the class list for this operand vector.
1611 OS << " // Eliminate obvious mismatches.\n";
1612 OS << " if (Operands.size() > " << MaxNumOperands << ")\n";
1613 OS << " return true;\n\n";
1615 OS << " // Compute the class list for this operand vector.\n";
1616 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1617 OS << " for (unsigned i = 0, e = Operands.size(); i != e; ++i) {\n";
1618 OS << " Classes[i] = ClassifyOperand(Operands[i]);\n\n";
1620 OS << " // Check for invalid operands before matching.\n";
1621 OS << " if (Classes[i] == InvalidMatchClass)\n";
1622 OS << " return true;\n";
1625 OS << " // Mark unused classes.\n";
1626 OS << " for (unsigned i = Operands.size(), e = " << MaxNumOperands << "; "
1627 << "i != e; ++i)\n";
1628 OS << " Classes[i] = InvalidMatchClass;\n\n";
1630 // Emit code to search the table.
1631 OS << " // Search the table.\n";
1632 OS << " for (const MatchEntry *it = MatchTable, "
1633 << "*ie = MatchTable + " << Info.Instructions.size()
1634 << "; it != ie; ++it) {\n";
1635 for (unsigned i = 0; i != MaxNumOperands; ++i) {
1636 OS << " if (!IsSubclass(Classes["
1637 << i << "], it->Classes[" << i << "]))\n";
1638 OS << " continue;\n";
1641 OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
1643 // Call the post-processing function, if used.
1644 std::string InsnCleanupFn =
1645 AsmParser->getValueAsString("AsmParserInstCleanup");
1646 if (!InsnCleanupFn.empty())
1647 OS << " " << InsnCleanupFn << "(Inst);\n";
1649 OS << " return false;\n";
1652 OS << " return true;\n";
1655 OS << "#endif // REGISTERS_ONLY\n";