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 "StringMatcher.h"
80 #include "llvm/ADT/OwningPtr.h"
81 #include "llvm/ADT/SmallVector.h"
82 #include "llvm/ADT/STLExtras.h"
83 #include "llvm/ADT/StringExtras.h"
84 #include "llvm/Support/CommandLine.h"
85 #include "llvm/Support/Debug.h"
91 static cl::opt<std::string>
92 MatchPrefix("match-prefix", cl::init(""),
93 cl::desc("Only match instructions with the given prefix"));
95 /// FlattenVariants - Flatten an .td file assembly string by selecting the
96 /// variant at index \arg N.
97 static std::string FlattenVariants(const std::string &AsmString,
99 StringRef Cur = AsmString;
100 std::string Res = "";
103 // Find the start of the next variant string.
104 size_t VariantsStart = 0;
105 for (size_t e = Cur.size(); VariantsStart != e; ++VariantsStart)
106 if (Cur[VariantsStart] == '{' &&
107 (VariantsStart == 0 || (Cur[VariantsStart-1] != '$' &&
108 Cur[VariantsStart-1] != '\\')))
111 // Add the prefix to the result.
112 Res += Cur.slice(0, VariantsStart);
113 if (VariantsStart == Cur.size())
116 ++VariantsStart; // Skip the '{'.
118 // Scan to the end of the variants string.
119 size_t VariantsEnd = VariantsStart;
120 unsigned NestedBraces = 1;
121 for (size_t e = Cur.size(); VariantsEnd != e; ++VariantsEnd) {
122 if (Cur[VariantsEnd] == '}' && Cur[VariantsEnd-1] != '\\') {
123 if (--NestedBraces == 0)
125 } else if (Cur[VariantsEnd] == '{')
129 // Select the Nth variant (or empty).
130 StringRef Selection = Cur.slice(VariantsStart, VariantsEnd);
131 for (unsigned i = 0; i != N; ++i)
132 Selection = Selection.split('|').second;
133 Res += Selection.split('|').first;
135 assert(VariantsEnd != Cur.size() &&
136 "Unterminated variants in assembly string!");
137 Cur = Cur.substr(VariantsEnd + 1);
143 /// TokenizeAsmString - Tokenize a simplified assembly string.
144 static void TokenizeAsmString(StringRef AsmString,
145 SmallVectorImpl<StringRef> &Tokens) {
148 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) {
149 switch (AsmString[i]) {
158 Tokens.push_back(AsmString.slice(Prev, i));
161 if (!isspace(AsmString[i]) && AsmString[i] != ',')
162 Tokens.push_back(AsmString.substr(i, 1));
168 Tokens.push_back(AsmString.slice(Prev, i));
172 assert(i != AsmString.size() && "Invalid quoted character");
173 Tokens.push_back(AsmString.substr(i, 1));
178 // If this isn't "${", treat like a normal token.
179 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') {
181 Tokens.push_back(AsmString.slice(Prev, i));
189 Tokens.push_back(AsmString.slice(Prev, i));
193 StringRef::iterator End =
194 std::find(AsmString.begin() + i, AsmString.end(), '}');
195 assert(End != AsmString.end() && "Missing brace in operand reference!");
196 size_t EndPos = End - AsmString.begin();
197 Tokens.push_back(AsmString.slice(i, EndPos+1));
205 Tokens.push_back(AsmString.slice(Prev, i));
215 if (InTok && Prev != AsmString.size())
216 Tokens.push_back(AsmString.substr(Prev));
219 static bool IsAssemblerInstruction(StringRef Name,
220 const CodeGenInstruction &CGI,
221 const SmallVectorImpl<StringRef> &Tokens) {
222 // Ignore "codegen only" instructions.
223 if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
226 // Ignore pseudo ops.
228 // FIXME: This is a hack; can we convert these instructions to set the
229 // "codegen only" bit instead?
230 if (const RecordVal *Form = CGI.TheDef->getValue("Form"))
231 if (Form->getValue()->getAsString() == "Pseudo")
234 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
236 // FIXME: This is a total hack.
237 if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int"))
240 // Ignore instructions with no .s string.
242 // FIXME: What are these?
243 if (CGI.AsmString.empty())
246 // FIXME: Hack; ignore any instructions with a newline in them.
247 if (std::find(CGI.AsmString.begin(),
248 CGI.AsmString.end(), '\n') != CGI.AsmString.end())
251 // Ignore instructions with attributes, these are always fake instructions for
252 // simplifying codegen.
254 // FIXME: Is this true?
256 // Also, check for instructions which reference the operand multiple times;
257 // this implies a constraint we would not honor.
258 std::set<std::string> OperandNames;
259 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
260 if (Tokens[i][0] == '$' &&
261 std::find(Tokens[i].begin(),
262 Tokens[i].end(), ':') != Tokens[i].end()) {
264 errs() << "warning: '" << Name << "': "
265 << "ignoring instruction; operand with attribute '"
266 << Tokens[i] << "'\n";
271 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
273 errs() << "warning: '" << Name << "': "
274 << "ignoring instruction with tied operand '"
275 << Tokens[i].str() << "'\n";
286 struct SubtargetFeatureInfo;
288 /// ClassInfo - Helper class for storing the information about a particular
289 /// class of operands which can be matched.
292 /// Invalid kind, for use as a sentinel value.
295 /// The class for a particular token.
298 /// The (first) register class, subsequent register classes are
299 /// RegisterClass0+1, and so on.
302 /// The (first) user defined class, subsequent user defined classes are
303 /// UserClass0+1, and so on.
307 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
308 /// N) for the Nth user defined class.
311 /// SuperClasses - The super classes of this class. Note that for simplicities
312 /// sake user operands only record their immediate super class, while register
313 /// operands include all superclasses.
314 std::vector<ClassInfo*> SuperClasses;
316 /// Name - The full class name, suitable for use in an enum.
319 /// ClassName - The unadorned generic name for this class (e.g., Token).
320 std::string ClassName;
322 /// ValueName - The name of the value this class represents; for a token this
323 /// is the literal token string, for an operand it is the TableGen class (or
324 /// empty if this is a derived class).
325 std::string ValueName;
327 /// PredicateMethod - The name of the operand method to test whether the
328 /// operand matches this class; this is not valid for Token or register kinds.
329 std::string PredicateMethod;
331 /// RenderMethod - The name of the operand method to add this operand to an
332 /// MCInst; this is not valid for Token or register kinds.
333 std::string RenderMethod;
335 /// For register classes, the records for all the registers in this class.
336 std::set<Record*> Registers;
339 /// isRegisterClass() - Check if this is a register class.
340 bool isRegisterClass() const {
341 return Kind >= RegisterClass0 && Kind < UserClass0;
344 /// isUserClass() - Check if this is a user defined class.
345 bool isUserClass() const {
346 return Kind >= UserClass0;
349 /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
350 /// are related if they are in the same class hierarchy.
351 bool isRelatedTo(const ClassInfo &RHS) const {
352 // Tokens are only related to tokens.
353 if (Kind == Token || RHS.Kind == Token)
354 return Kind == Token && RHS.Kind == Token;
356 // Registers classes are only related to registers classes, and only if
357 // their intersection is non-empty.
358 if (isRegisterClass() || RHS.isRegisterClass()) {
359 if (!isRegisterClass() || !RHS.isRegisterClass())
362 std::set<Record*> Tmp;
363 std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
364 std::set_intersection(Registers.begin(), Registers.end(),
365 RHS.Registers.begin(), RHS.Registers.end(),
371 // Otherwise we have two users operands; they are related if they are in the
372 // same class hierarchy.
374 // FIXME: This is an oversimplification, they should only be related if they
375 // intersect, however we don't have that information.
376 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
377 const ClassInfo *Root = this;
378 while (!Root->SuperClasses.empty())
379 Root = Root->SuperClasses.front();
381 const ClassInfo *RHSRoot = &RHS;
382 while (!RHSRoot->SuperClasses.empty())
383 RHSRoot = RHSRoot->SuperClasses.front();
385 return Root == RHSRoot;
388 /// isSubsetOf - Test whether this class is a subset of \arg RHS;
389 bool isSubsetOf(const ClassInfo &RHS) const {
390 // This is a subset of RHS if it is the same class...
394 // ... or if any of its super classes are a subset of RHS.
395 for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
396 ie = SuperClasses.end(); it != ie; ++it)
397 if ((*it)->isSubsetOf(RHS))
403 /// operator< - Compare two classes.
404 bool operator<(const ClassInfo &RHS) const {
408 // Unrelated classes can be ordered by kind.
409 if (!isRelatedTo(RHS))
410 return Kind < RHS.Kind;
414 assert(0 && "Invalid kind!");
416 // Tokens are comparable by value.
418 // FIXME: Compare by enum value.
419 return ValueName < RHS.ValueName;
422 // This class preceeds the RHS if it is a proper subset of the RHS.
425 if (RHS.isSubsetOf(*this))
428 // Otherwise, order by name to ensure we have a total ordering.
429 return ValueName < RHS.ValueName;
434 /// InstructionInfo - Helper class for storing the necessary information for an
435 /// instruction which is capable of being matched.
436 struct InstructionInfo {
438 /// The unique class instance this operand should match.
441 /// The original operand this corresponds to, if any.
442 const CodeGenInstruction::OperandInfo *OperandInfo;
445 /// InstrName - The target name for this instruction.
446 std::string InstrName;
448 /// Instr - The instruction this matches.
449 const CodeGenInstruction *Instr;
451 /// AsmString - The assembly string for this instruction (with variants
453 std::string AsmString;
455 /// Tokens - The tokenized assembly pattern that this instruction matches.
456 SmallVector<StringRef, 4> Tokens;
458 /// Operands - The operands that this instruction matches.
459 SmallVector<Operand, 4> Operands;
461 /// Predicates - The required subtarget features to match this instruction.
462 SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures;
464 /// ConversionFnKind - The enum value which is passed to the generated
465 /// ConvertToMCInst to convert parsed operands into an MCInst for this
467 std::string ConversionFnKind;
469 /// operator< - Compare two instructions.
470 bool operator<(const InstructionInfo &RHS) const {
471 // The primary comparator is the instruction mnemonic.
472 if (Tokens[0] != RHS.Tokens[0])
473 return Tokens[0] < RHS.Tokens[0];
475 if (Operands.size() != RHS.Operands.size())
476 return Operands.size() < RHS.Operands.size();
478 // Compare lexicographically by operand. The matcher validates that other
479 // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith().
480 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
481 if (*Operands[i].Class < *RHS.Operands[i].Class)
483 if (*RHS.Operands[i].Class < *Operands[i].Class)
490 /// CouldMatchAmiguouslyWith - Check whether this instruction could
491 /// ambiguously match the same set of operands as \arg RHS (without being a
492 /// strictly superior match).
493 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
494 // The number of operands is unambiguous.
495 if (Operands.size() != RHS.Operands.size())
498 // Otherwise, make sure the ordering of the two instructions is unambiguous
499 // by checking that either (a) a token or operand kind discriminates them,
500 // or (b) the ordering among equivalent kinds is consistent.
502 // Tokens and operand kinds are unambiguous (assuming a correct target
504 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
505 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
506 Operands[i].Class->Kind == ClassInfo::Token)
507 if (*Operands[i].Class < *RHS.Operands[i].Class ||
508 *RHS.Operands[i].Class < *Operands[i].Class)
511 // Otherwise, this operand could commute if all operands are equivalent, or
512 // there is a pair of operands that compare less than and a pair that
513 // compare greater than.
514 bool HasLT = false, HasGT = false;
515 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
516 if (*Operands[i].Class < *RHS.Operands[i].Class)
518 if (*RHS.Operands[i].Class < *Operands[i].Class)
522 return !(HasLT ^ HasGT);
529 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
530 /// feature which participates in instruction matching.
531 struct SubtargetFeatureInfo {
532 /// \brief The predicate record for this feature.
535 /// \brief An unique index assigned to represent this feature.
538 SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {}
540 /// \brief The name of the enumerated constant identifying this feature.
541 std::string getEnumName() const {
542 return "Feature_" + TheDef->getName();
546 class AsmMatcherInfo {
548 /// The tablegen AsmParser record.
551 /// The AsmParser "CommentDelimiter" value.
552 std::string CommentDelimiter;
554 /// The AsmParser "RegisterPrefix" value.
555 std::string RegisterPrefix;
557 /// The classes which are needed for matching.
558 std::vector<ClassInfo*> Classes;
560 /// The information on the instruction to match.
561 std::vector<InstructionInfo*> Instructions;
563 /// Map of Register records to their class information.
564 std::map<Record*, ClassInfo*> RegisterClasses;
566 /// Map of Predicate records to their subtarget information.
567 std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures;
570 /// Map of token to class information which has already been constructed.
571 std::map<std::string, ClassInfo*> TokenClasses;
573 /// Map of RegisterClass records to their class information.
574 std::map<Record*, ClassInfo*> RegisterClassClasses;
576 /// Map of AsmOperandClass records to their class information.
577 std::map<Record*, ClassInfo*> AsmOperandClasses;
580 /// getTokenClass - Lookup or create the class for the given token.
581 ClassInfo *getTokenClass(StringRef Token);
583 /// getOperandClass - Lookup or create the class for the given operand.
584 ClassInfo *getOperandClass(StringRef Token,
585 const CodeGenInstruction::OperandInfo &OI);
587 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
589 SubtargetFeatureInfo *getSubtargetFeature(Record *Def) {
590 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
591 return SubtargetFeatures[Def];
594 /// BuildRegisterClasses - Build the ClassInfo* instances for register
596 void BuildRegisterClasses(CodeGenTarget &Target,
597 std::set<std::string> &SingletonRegisterNames);
599 /// BuildOperandClasses - Build the ClassInfo* instances for user defined
601 void BuildOperandClasses(CodeGenTarget &Target);
604 AsmMatcherInfo(Record *_AsmParser);
606 /// BuildInfo - Construct the various tables used during matching.
607 void BuildInfo(CodeGenTarget &Target);
612 void InstructionInfo::dump() {
613 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
615 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
622 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
623 Operand &Op = Operands[i];
624 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
625 if (Op.Class->Kind == ClassInfo::Token) {
626 errs() << '\"' << Tokens[i] << "\"\n";
630 if (!Op.OperandInfo) {
631 errs() << "(singleton register)\n";
635 const CodeGenInstruction::OperandInfo &OI = *Op.OperandInfo;
636 errs() << OI.Name << " " << OI.Rec->getName()
637 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
641 static std::string getEnumNameForToken(StringRef Str) {
644 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
646 case '*': Res += "_STAR_"; break;
647 case '%': Res += "_PCT_"; break;
648 case ':': Res += "_COLON_"; break;
654 Res += "_" + utostr((unsigned) *it) + "_";
662 /// getRegisterRecord - Get the register record for \arg name, or 0.
663 static Record *getRegisterRecord(CodeGenTarget &Target, StringRef Name) {
664 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
665 const CodeGenRegister &Reg = Target.getRegisters()[i];
666 if (Name == Reg.TheDef->getValueAsString("AsmName"))
673 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
674 ClassInfo *&Entry = TokenClasses[Token];
677 Entry = new ClassInfo();
678 Entry->Kind = ClassInfo::Token;
679 Entry->ClassName = "Token";
680 Entry->Name = "MCK_" + getEnumNameForToken(Token);
681 Entry->ValueName = Token;
682 Entry->PredicateMethod = "<invalid>";
683 Entry->RenderMethod = "<invalid>";
684 Classes.push_back(Entry);
691 AsmMatcherInfo::getOperandClass(StringRef Token,
692 const CodeGenInstruction::OperandInfo &OI) {
693 if (OI.Rec->isSubClassOf("RegisterClass")) {
694 ClassInfo *CI = RegisterClassClasses[OI.Rec];
697 PrintError(OI.Rec->getLoc(), "register class has no class info!");
698 throw std::string("ERROR: Missing register class!");
704 assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!");
705 Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass");
706 ClassInfo *CI = AsmOperandClasses[MatchClass];
709 PrintError(OI.Rec->getLoc(), "operand has no match class!");
710 throw std::string("ERROR: Missing match class!");
716 void AsmMatcherInfo::BuildRegisterClasses(CodeGenTarget &Target,
717 std::set<std::string>
718 &SingletonRegisterNames) {
719 std::vector<CodeGenRegisterClass> RegisterClasses;
720 std::vector<CodeGenRegister> Registers;
722 RegisterClasses = Target.getRegisterClasses();
723 Registers = Target.getRegisters();
725 // The register sets used for matching.
726 std::set< std::set<Record*> > RegisterSets;
728 // Gather the defined sets.
729 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
730 ie = RegisterClasses.end(); it != ie; ++it)
731 RegisterSets.insert(std::set<Record*>(it->Elements.begin(),
732 it->Elements.end()));
734 // Add any required singleton sets.
735 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
736 ie = SingletonRegisterNames.end(); it != ie; ++it)
737 if (Record *Rec = getRegisterRecord(Target, *it))
738 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
740 // Introduce derived sets where necessary (when a register does not determine
741 // a unique register set class), and build the mapping of registers to the set
742 // they should classify to.
743 std::map<Record*, std::set<Record*> > RegisterMap;
744 for (std::vector<CodeGenRegister>::iterator it = Registers.begin(),
745 ie = Registers.end(); it != ie; ++it) {
746 CodeGenRegister &CGR = *it;
747 // Compute the intersection of all sets containing this register.
748 std::set<Record*> ContainingSet;
750 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
751 ie = RegisterSets.end(); it != ie; ++it) {
752 if (!it->count(CGR.TheDef))
755 if (ContainingSet.empty()) {
758 std::set<Record*> Tmp;
759 std::swap(Tmp, ContainingSet);
760 std::insert_iterator< std::set<Record*> > II(ContainingSet,
761 ContainingSet.begin());
762 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(),
767 if (!ContainingSet.empty()) {
768 RegisterSets.insert(ContainingSet);
769 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
773 // Construct the register classes.
774 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
776 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
777 ie = RegisterSets.end(); it != ie; ++it, ++Index) {
778 ClassInfo *CI = new ClassInfo();
779 CI->Kind = ClassInfo::RegisterClass0 + Index;
780 CI->ClassName = "Reg" + utostr(Index);
781 CI->Name = "MCK_Reg" + utostr(Index);
783 CI->PredicateMethod = ""; // unused
784 CI->RenderMethod = "addRegOperands";
786 Classes.push_back(CI);
787 RegisterSetClasses.insert(std::make_pair(*it, CI));
790 // Find the superclasses; we could compute only the subgroup lattice edges,
791 // but there isn't really a point.
792 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
793 ie = RegisterSets.end(); it != ie; ++it) {
794 ClassInfo *CI = RegisterSetClasses[*it];
795 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
796 ie2 = RegisterSets.end(); it2 != ie2; ++it2)
798 std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
799 CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
802 // Name the register classes which correspond to a user defined RegisterClass.
803 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
804 ie = RegisterClasses.end(); it != ie; ++it) {
805 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(),
806 it->Elements.end())];
807 if (CI->ValueName.empty()) {
808 CI->ClassName = it->getName();
809 CI->Name = "MCK_" + it->getName();
810 CI->ValueName = it->getName();
812 CI->ValueName = CI->ValueName + "," + it->getName();
814 RegisterClassClasses.insert(std::make_pair(it->TheDef, CI));
817 // Populate the map for individual registers.
818 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
819 ie = RegisterMap.end(); it != ie; ++it)
820 this->RegisterClasses[it->first] = RegisterSetClasses[it->second];
822 // Name the register classes which correspond to singleton registers.
823 for (std::set<std::string>::iterator it = SingletonRegisterNames.begin(),
824 ie = SingletonRegisterNames.end(); it != ie; ++it) {
825 if (Record *Rec = getRegisterRecord(Target, *it)) {
826 ClassInfo *CI = this->RegisterClasses[Rec];
827 assert(CI && "Missing singleton register class info!");
829 if (CI->ValueName.empty()) {
830 CI->ClassName = Rec->getName();
831 CI->Name = "MCK_" + Rec->getName();
832 CI->ValueName = Rec->getName();
834 CI->ValueName = CI->ValueName + "," + Rec->getName();
839 void AsmMatcherInfo::BuildOperandClasses(CodeGenTarget &Target) {
840 std::vector<Record*> AsmOperands;
841 AsmOperands = Records.getAllDerivedDefinitions("AsmOperandClass");
843 // Pre-populate AsmOperandClasses map.
844 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
845 ie = AsmOperands.end(); it != ie; ++it)
846 AsmOperandClasses[*it] = new ClassInfo();
849 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
850 ie = AsmOperands.end(); it != ie; ++it, ++Index) {
851 ClassInfo *CI = AsmOperandClasses[*it];
852 CI->Kind = ClassInfo::UserClass0 + Index;
854 ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
855 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
856 DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
858 PrintError((*it)->getLoc(), "Invalid super class reference!");
862 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
864 PrintError((*it)->getLoc(), "Invalid super class reference!");
866 CI->SuperClasses.push_back(SC);
868 CI->ClassName = (*it)->getValueAsString("Name");
869 CI->Name = "MCK_" + CI->ClassName;
870 CI->ValueName = (*it)->getName();
872 // Get or construct the predicate method name.
873 Init *PMName = (*it)->getValueInit("PredicateMethod");
874 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
875 CI->PredicateMethod = SI->getValue();
877 assert(dynamic_cast<UnsetInit*>(PMName) &&
878 "Unexpected PredicateMethod field!");
879 CI->PredicateMethod = "is" + CI->ClassName;
882 // Get or construct the render method name.
883 Init *RMName = (*it)->getValueInit("RenderMethod");
884 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
885 CI->RenderMethod = SI->getValue();
887 assert(dynamic_cast<UnsetInit*>(RMName) &&
888 "Unexpected RenderMethod field!");
889 CI->RenderMethod = "add" + CI->ClassName + "Operands";
892 AsmOperandClasses[*it] = CI;
893 Classes.push_back(CI);
897 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser)
898 : AsmParser(asmParser),
899 CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")),
900 RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix"))
904 void AsmMatcherInfo::BuildInfo(CodeGenTarget &Target) {
905 // Parse the instructions; we need to do this first so that we can gather the
906 // singleton register classes.
907 std::set<std::string> SingletonRegisterNames;
909 const std::vector<const CodeGenInstruction*> &InstrList =
910 Target.getInstructionsByEnumValue();
913 // Build information about all of the AssemblerPredicates.
914 std::vector<Record*> AllPredicates =
915 Records.getAllDerivedDefinitions("Predicate");
916 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
917 Record *Pred = AllPredicates[i];
918 // Ignore predicates that are not intended for the assembler.
919 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
922 if (Pred->getName().empty()) {
923 PrintError(Pred->getLoc(), "Predicate has no name!");
924 throw std::string("ERROR: Predicate defs must be named");
927 unsigned FeatureNo = SubtargetFeatures.size();
928 SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo);
929 assert(FeatureNo < 32 && "Too many subtarget features!");
932 for (unsigned i = 0, e = InstrList.size(); i != e; ++i) {
933 const CodeGenInstruction &CGI = *InstrList[i];
935 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
938 OwningPtr<InstructionInfo> II(new InstructionInfo());
940 II->InstrName = CGI.TheDef->getName();
942 II->AsmString = FlattenVariants(CGI.AsmString, 0);
944 // Remove comments from the asm string.
945 if (!CommentDelimiter.empty()) {
946 size_t Idx = StringRef(II->AsmString).find(CommentDelimiter);
947 if (Idx != StringRef::npos)
948 II->AsmString = II->AsmString.substr(0, Idx);
951 TokenizeAsmString(II->AsmString, II->Tokens);
953 // Ignore instructions which shouldn't be matched.
954 if (!IsAssemblerInstruction(CGI.TheDef->getName(), CGI, II->Tokens))
957 // Collect singleton registers, if used.
958 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
959 if (!II->Tokens[i].startswith(RegisterPrefix))
962 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
963 Record *Rec = getRegisterRecord(Target, RegName);
966 // If there is no register prefix (i.e. "%" in "%eax"), then this may
967 // be some random non-register token, just ignore it.
968 if (RegisterPrefix.empty())
971 std::string Err = "unable to find register for '" + RegName.str() +
972 "' (which matches register prefix)";
973 throw TGError(CGI.TheDef->getLoc(), Err);
976 SingletonRegisterNames.insert(RegName);
979 // Compute the require features.
980 std::vector<Record*> Predicates =
981 CGI.TheDef->getValueAsListOfDefs("Predicates");
982 for (unsigned i = 0, e = Predicates.size(); i != e; ++i) {
983 Record *Pred = Predicates[i];
984 // Ignore predicates that are not intended for the assembler.
985 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
988 II->RequiredFeatures.push_back(getSubtargetFeature(Pred));
991 Instructions.push_back(II.take());
994 // Build info for the register classes.
995 BuildRegisterClasses(Target, SingletonRegisterNames);
997 // Build info for the user defined assembly operand classes.
998 BuildOperandClasses(Target);
1000 // Build the instruction information.
1001 for (std::vector<InstructionInfo*>::iterator it = Instructions.begin(),
1002 ie = Instructions.end(); it != ie; ++it) {
1003 InstructionInfo *II = *it;
1005 // The first token of the instruction is the mnemonic, which must be a
1007 assert(!II->Tokens.empty() && "Instruction has no tokens?");
1008 StringRef Mnemonic = II->Tokens[0];
1009 assert(Mnemonic[0] != '$' &&
1010 (RegisterPrefix.empty() || !Mnemonic.startswith(RegisterPrefix)));
1012 // Parse the tokens after the mnemonic.
1013 for (unsigned i = 1, e = II->Tokens.size(); i != e; ++i) {
1014 StringRef Token = II->Tokens[i];
1016 // Check for singleton registers.
1017 if (Token.startswith(RegisterPrefix)) {
1018 StringRef RegName = II->Tokens[i].substr(RegisterPrefix.size());
1019 if (Record *RegRecord = getRegisterRecord(Target, RegName)) {
1020 InstructionInfo::Operand Op;
1021 Op.Class = RegisterClasses[RegRecord];
1023 assert(Op.Class && Op.Class->Registers.size() == 1 &&
1024 "Unexpected class for singleton register");
1025 II->Operands.push_back(Op);
1029 if (!RegisterPrefix.empty()) {
1030 std::string Err = "unable to find register for '" + RegName.str() +
1031 "' (which matches register prefix)";
1032 throw TGError(II->Instr->TheDef->getLoc(), Err);
1036 // Check for simple tokens.
1037 if (Token[0] != '$') {
1038 InstructionInfo::Operand Op;
1039 Op.Class = getTokenClass(Token);
1041 II->Operands.push_back(Op);
1045 // Otherwise this is an operand reference.
1046 StringRef OperandName;
1047 if (Token[1] == '{')
1048 OperandName = Token.substr(2, Token.size() - 3);
1050 OperandName = Token.substr(1);
1052 // Map this token to an operand. FIXME: Move elsewhere.
1055 Idx = II->Instr->getOperandNamed(OperandName);
1057 throw std::string("error: unable to find operand: '" +
1058 OperandName.str() + "'");
1061 // FIXME: This is annoying, the named operand may be tied (e.g.,
1062 // XCHG8rm). What we want is the untied operand, which we now have to
1063 // grovel for. Only worry about this for single entry operands, we have to
1064 // clean this up anyway.
1065 const CodeGenInstruction::OperandInfo *OI = &II->Instr->OperandList[Idx];
1066 if (OI->Constraints[0].isTied()) {
1067 unsigned TiedOp = OI->Constraints[0].getTiedOperand();
1069 // The tied operand index is an MIOperand index, find the operand that
1071 for (unsigned i = 0, e = II->Instr->OperandList.size(); i != e; ++i) {
1072 if (II->Instr->OperandList[i].MIOperandNo == TiedOp) {
1073 OI = &II->Instr->OperandList[i];
1078 assert(OI && "Unable to find tied operand target!");
1081 InstructionInfo::Operand Op;
1082 Op.Class = getOperandClass(Token, *OI);
1083 Op.OperandInfo = OI;
1084 II->Operands.push_back(Op);
1088 // Reorder classes so that classes preceed super classes.
1089 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
1092 static std::pair<unsigned, unsigned> *
1093 GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List,
1095 for (unsigned i = 0, e = List.size(); i != e; ++i)
1096 if (Index == List[i].first)
1102 static void EmitConvertToMCInst(CodeGenTarget &Target,
1103 std::vector<InstructionInfo*> &Infos,
1105 // Write the convert function to a separate stream, so we can drop it after
1107 std::string ConvertFnBody;
1108 raw_string_ostream CvtOS(ConvertFnBody);
1110 // Function we have already generated.
1111 std::set<std::string> GeneratedFns;
1113 // Start the unified conversion function.
1115 CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
1116 << "unsigned Opcode,\n"
1117 << " const SmallVectorImpl<MCParsedAsmOperand*"
1118 << "> &Operands) {\n";
1119 CvtOS << " Inst.setOpcode(Opcode);\n";
1120 CvtOS << " switch (Kind) {\n";
1121 CvtOS << " default:\n";
1123 // Start the enum, which we will generate inline.
1125 OS << "// Unified function for converting operants to MCInst instances.\n\n";
1126 OS << "enum ConversionKind {\n";
1128 // TargetOperandClass - This is the target's operand class, like X86Operand.
1129 std::string TargetOperandClass = Target.getName() + "Operand";
1131 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
1132 ie = Infos.end(); it != ie; ++it) {
1133 InstructionInfo &II = **it;
1135 // Order the (class) operands by the order to convert them into an MCInst.
1136 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
1137 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1138 InstructionInfo::Operand &Op = II.Operands[i];
1140 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
1143 // Find any tied operands.
1144 SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands;
1145 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
1146 const CodeGenInstruction::OperandInfo &OpInfo = II.Instr->OperandList[i];
1147 for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) {
1148 const CodeGenInstruction::ConstraintInfo &CI = OpInfo.Constraints[j];
1150 TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo + j,
1151 CI.getTiedOperand()));
1155 std::sort(MIOperandList.begin(), MIOperandList.end());
1157 // Compute the total number of operands.
1158 unsigned NumMIOperands = 0;
1159 for (unsigned i = 0, e = II.Instr->OperandList.size(); i != e; ++i) {
1160 const CodeGenInstruction::OperandInfo &OI = II.Instr->OperandList[i];
1161 NumMIOperands = std::max(NumMIOperands,
1162 OI.MIOperandNo + OI.MINumOperands);
1165 // Build the conversion function signature.
1166 std::string Signature = "Convert";
1167 unsigned CurIndex = 0;
1168 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1169 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1170 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
1171 "Duplicate match for instruction operand!");
1173 // Skip operands which weren't matched by anything, this occurs when the
1174 // .td file encodes "implicit" operands as explicit ones.
1176 // FIXME: This should be removed from the MCInst structure.
1177 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1178 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1181 Signature += "__Imp";
1183 Signature += "__Tie" + utostr(Tie->second);
1188 // Registers are always converted the same, don't duplicate the conversion
1189 // function based on them.
1191 // FIXME: We could generalize this based on the render method, if it
1193 if (Op.Class->isRegisterClass())
1196 Signature += Op.Class->ClassName;
1197 Signature += utostr(Op.OperandInfo->MINumOperands);
1198 Signature += "_" + utostr(MIOperandList[i].second);
1200 CurIndex += Op.OperandInfo->MINumOperands;
1203 // Add any trailing implicit operands.
1204 for (; CurIndex != NumMIOperands; ++CurIndex) {
1205 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1208 Signature += "__Imp";
1210 Signature += "__Tie" + utostr(Tie->second);
1213 II.ConversionFnKind = Signature;
1215 // Check if we have already generated this signature.
1216 if (!GeneratedFns.insert(Signature).second)
1219 // If not, emit it now.
1221 // Add to the enum list.
1222 OS << " " << Signature << ",\n";
1224 // And to the convert function.
1225 CvtOS << " case " << Signature << ":\n";
1227 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1228 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1230 // Add the implicit operands.
1231 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1232 // See if this is a tied operand.
1233 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1237 // If not, this is some implicit operand. Just assume it is a register
1239 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1241 // Copy the tied operand.
1242 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1243 CvtOS << " Inst.addOperand(Inst.getOperand("
1244 << Tie->second << "));\n";
1248 CvtOS << " ((" << TargetOperandClass << "*)Operands["
1249 << MIOperandList[i].second
1250 << "+1])->" << Op.Class->RenderMethod
1251 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
1252 CurIndex += Op.OperandInfo->MINumOperands;
1255 // And add trailing implicit operands.
1256 for (; CurIndex != NumMIOperands; ++CurIndex) {
1257 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1261 // If not, this is some implicit operand. Just assume it is a register
1263 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1265 // Copy the tied operand.
1266 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1267 CvtOS << " Inst.addOperand(Inst.getOperand("
1268 << Tie->second << "));\n";
1272 CvtOS << " return;\n";
1275 // Finish the convert function.
1280 // Finish the enum, and drop the convert function after it.
1282 OS << " NumConversionVariants\n";
1288 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
1289 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
1290 std::vector<ClassInfo*> &Infos,
1292 OS << "namespace {\n\n";
1294 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1295 << "/// instruction matching.\n";
1296 OS << "enum MatchClassKind {\n";
1297 OS << " InvalidMatchClass = 0,\n";
1298 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1299 ie = Infos.end(); it != ie; ++it) {
1300 ClassInfo &CI = **it;
1301 OS << " " << CI.Name << ", // ";
1302 if (CI.Kind == ClassInfo::Token) {
1303 OS << "'" << CI.ValueName << "'\n";
1304 } else if (CI.isRegisterClass()) {
1305 if (!CI.ValueName.empty())
1306 OS << "register class '" << CI.ValueName << "'\n";
1308 OS << "derived register class\n";
1310 OS << "user defined class '" << CI.ValueName << "'\n";
1313 OS << " NumMatchClassKinds\n";
1319 /// EmitClassifyOperand - Emit the function to classify an operand.
1320 static void EmitClassifyOperand(CodeGenTarget &Target,
1321 AsmMatcherInfo &Info,
1323 OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n"
1324 << " " << Target.getName() << "Operand &Operand = *("
1325 << Target.getName() << "Operand*)GOp;\n";
1328 OS << " if (Operand.isToken())\n";
1329 OS << " return MatchTokenString(Operand.getToken());\n\n";
1331 // Classify registers.
1333 // FIXME: Don't hardcode isReg, getReg.
1334 OS << " if (Operand.isReg()) {\n";
1335 OS << " switch (Operand.getReg()) {\n";
1336 OS << " default: return InvalidMatchClass;\n";
1337 for (std::map<Record*, ClassInfo*>::iterator
1338 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1340 OS << " case " << Target.getName() << "::"
1341 << it->first->getName() << ": return " << it->second->Name << ";\n";
1345 // Classify user defined operands.
1346 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1347 ie = Info.Classes.end(); it != ie; ++it) {
1348 ClassInfo &CI = **it;
1350 if (!CI.isUserClass())
1353 OS << " // '" << CI.ClassName << "' class";
1354 if (!CI.SuperClasses.empty()) {
1355 OS << ", subclass of ";
1356 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) {
1358 OS << "'" << CI.SuperClasses[i]->ClassName << "'";
1359 assert(CI < *CI.SuperClasses[i] && "Invalid class relation!");
1364 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
1366 // Validate subclass relationships.
1367 if (!CI.SuperClasses.empty()) {
1368 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i)
1369 OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod
1370 << "() && \"Invalid class relationship!\");\n";
1373 OS << " return " << CI.Name << ";\n";
1376 OS << " return InvalidMatchClass;\n";
1380 /// EmitIsSubclass - Emit the subclass predicate function.
1381 static void EmitIsSubclass(CodeGenTarget &Target,
1382 std::vector<ClassInfo*> &Infos,
1384 OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1385 OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n";
1386 OS << " if (A == B)\n";
1387 OS << " return true;\n\n";
1389 OS << " switch (A) {\n";
1390 OS << " default:\n";
1391 OS << " return false;\n";
1392 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1393 ie = Infos.end(); it != ie; ++it) {
1394 ClassInfo &A = **it;
1396 if (A.Kind != ClassInfo::Token) {
1397 std::vector<StringRef> SuperClasses;
1398 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1399 ie = Infos.end(); it != ie; ++it) {
1400 ClassInfo &B = **it;
1402 if (&A != &B && A.isSubsetOf(B))
1403 SuperClasses.push_back(B.Name);
1406 if (SuperClasses.empty())
1409 OS << "\n case " << A.Name << ":\n";
1411 if (SuperClasses.size() == 1) {
1412 OS << " return B == " << SuperClasses.back() << ";\n";
1416 OS << " switch (B) {\n";
1417 OS << " default: return false;\n";
1418 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1419 OS << " case " << SuperClasses[i] << ": return true;\n";
1429 /// EmitMatchTokenString - Emit the function to match a token string to the
1430 /// appropriate match class value.
1431 static void EmitMatchTokenString(CodeGenTarget &Target,
1432 std::vector<ClassInfo*> &Infos,
1434 // Construct the match list.
1435 std::vector<StringMatcher::StringPair> Matches;
1436 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1437 ie = Infos.end(); it != ie; ++it) {
1438 ClassInfo &CI = **it;
1440 if (CI.Kind == ClassInfo::Token)
1441 Matches.push_back(StringMatcher::StringPair(CI.ValueName,
1442 "return " + CI.Name + ";"));
1445 OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
1447 StringMatcher("Name", Matches, OS).Emit();
1449 OS << " return InvalidMatchClass;\n";
1453 /// EmitMatchRegisterName - Emit the function to match a string to the target
1454 /// specific register enum.
1455 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1457 // Construct the match list.
1458 std::vector<StringMatcher::StringPair> Matches;
1459 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1460 const CodeGenRegister &Reg = Target.getRegisters()[i];
1461 if (Reg.TheDef->getValueAsString("AsmName").empty())
1464 Matches.push_back(StringMatcher::StringPair(
1465 Reg.TheDef->getValueAsString("AsmName"),
1466 "return " + utostr(i + 1) + ";"));
1469 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1471 StringMatcher("Name", Matches, OS).Emit();
1473 OS << " return 0;\n";
1477 /// EmitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
1479 static void EmitSubtargetFeatureFlagEnumeration(CodeGenTarget &Target,
1480 AsmMatcherInfo &Info,
1482 OS << "// Flags for subtarget features that participate in "
1483 << "instruction matching.\n";
1484 OS << "enum SubtargetFeatureFlag {\n";
1485 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1486 it = Info.SubtargetFeatures.begin(),
1487 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1488 SubtargetFeatureInfo &SFI = *it->second;
1489 OS << " " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n";
1491 OS << " Feature_None = 0\n";
1495 /// EmitComputeAvailableFeatures - Emit the function to compute the list of
1496 /// available features given a subtarget.
1497 static void EmitComputeAvailableFeatures(CodeGenTarget &Target,
1498 AsmMatcherInfo &Info,
1500 std::string ClassName =
1501 Info.AsmParser->getValueAsString("AsmParserClassName");
1503 OS << "unsigned " << Target.getName() << ClassName << "::\n"
1504 << "ComputeAvailableFeatures(const " << Target.getName()
1505 << "Subtarget *Subtarget) const {\n";
1506 OS << " unsigned Features = 0;\n";
1507 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1508 it = Info.SubtargetFeatures.begin(),
1509 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1510 SubtargetFeatureInfo &SFI = *it->second;
1511 OS << " if (" << SFI.TheDef->getValueAsString("CondString")
1513 OS << " Features |= " << SFI.getEnumName() << ";\n";
1515 OS << " return Features;\n";
1519 static std::string GetAliasRequiredFeatures(Record *R) {
1520 // FIXME: This is a total hack.
1521 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
1524 unsigned NumFeatures = 0;
1525 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
1526 Record *Pred = ReqFeatures[i];
1528 // Ignore predicates that are not intended for the assembler.
1529 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1535 Result += "Feature_" + Pred->getName();
1539 if (NumFeatures > 1)
1540 Result = '(' + Result + ')';
1544 /// EmitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
1545 /// emit a function for them and return true, otherwise return false.
1546 static bool EmitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) {
1547 std::vector<Record*> Aliases =
1548 Records.getAllDerivedDefinitions("MnemonicAlias");
1549 if (Aliases.empty()) return false;
1551 OS << "static void ApplyMnemonicAliases(StringRef &Mnemonic, "
1552 "unsigned Features) {\n";
1554 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
1555 // iteration order of the map is stable.
1556 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
1558 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
1559 Record *R = Aliases[i];
1560 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
1563 // Process each alias a "from" mnemonic at a time, building the code executed
1564 // by the string remapper.
1565 std::vector<StringMatcher::StringPair> Cases;
1566 for (std::map<std::string, std::vector<Record*> >::iterator
1567 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
1569 const std::vector<Record*> &ToVec = I->second;
1571 // Loop through each alias and emit code that handles each case. If there
1572 // are two instructions without predicates, emit an error. If there is one,
1574 std::string MatchCode;
1575 int AliasWithNoPredicate = -1;
1577 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
1578 Record *R = ToVec[i];
1579 std::string FeatureMask = GetAliasRequiredFeatures(R);
1581 // If this unconditionally matches, remember it for later and diagnose
1583 if (FeatureMask.empty()) {
1584 if (AliasWithNoPredicate != -1) {
1585 // We can't have two aliases from the same mnemonic with no predicate.
1586 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
1587 "two MnemonicAliases with the same 'from' mnemonic!");
1588 PrintError(R->getLoc(), "this is the other MnemonicAlias.");
1589 throw std::string("ERROR: Invalid MnemonicAlias definitions!");
1592 AliasWithNoPredicate = i;
1596 if (!MatchCode.empty())
1597 MatchCode += "else ";
1598 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
1599 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
1602 if (AliasWithNoPredicate != -1) {
1603 Record *R = ToVec[AliasWithNoPredicate];
1604 if (!MatchCode.empty())
1605 MatchCode += "else\n ";
1606 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
1609 MatchCode += "return;";
1611 Cases.push_back(std::make_pair(I->first, MatchCode));
1615 StringMatcher("Mnemonic", Cases, OS).Emit();
1621 void AsmMatcherEmitter::run(raw_ostream &OS) {
1622 CodeGenTarget Target;
1623 Record *AsmParser = Target.getAsmParser();
1624 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1626 // Compute the information on the instructions to match.
1627 AsmMatcherInfo Info(AsmParser);
1628 Info.BuildInfo(Target);
1630 // Sort the instruction table using the partial order on classes. We use
1631 // stable_sort to ensure that ambiguous instructions are still
1632 // deterministically ordered.
1633 std::stable_sort(Info.Instructions.begin(), Info.Instructions.end(),
1634 less_ptr<InstructionInfo>());
1636 DEBUG_WITH_TYPE("instruction_info", {
1637 for (std::vector<InstructionInfo*>::iterator
1638 it = Info.Instructions.begin(), ie = Info.Instructions.end();
1643 // Check for ambiguous instructions.
1644 DEBUG_WITH_TYPE("ambiguous_instrs", {
1645 unsigned NumAmbiguous = 0;
1646 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
1647 for (unsigned j = i + 1; j != e; ++j) {
1648 InstructionInfo &A = *Info.Instructions[i];
1649 InstructionInfo &B = *Info.Instructions[j];
1651 if (A.CouldMatchAmiguouslyWith(B)) {
1652 errs() << "warning: ambiguous instruction match:\n";
1654 errs() << "\nis incomparable with:\n";
1662 errs() << "warning: " << NumAmbiguous
1663 << " ambiguous instructions!\n";
1666 // Write the output.
1668 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1670 // Information for the class declaration.
1671 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
1672 OS << "#undef GET_ASSEMBLER_HEADER\n";
1673 OS << " // This should be included into the middle of the declaration of \n";
1674 OS << " // your subclasses implementation of TargetAsmParser.\n";
1675 OS << " unsigned ComputeAvailableFeatures(const " <<
1676 Target.getName() << "Subtarget *Subtarget) const;\n";
1677 OS << " enum MatchResultTy {\n";
1678 OS << " Match_Success, Match_MnemonicFail, Match_InvalidOperand,\n";
1679 OS << " Match_MissingFeature\n";
1681 OS << " MatchResultTy MatchInstructionImpl(const "
1682 << "SmallVectorImpl<MCParsedAsmOperand*>"
1683 << " &Operands, MCInst &Inst, unsigned &ErrorInfo);\n\n";
1684 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
1689 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
1690 OS << "#undef GET_REGISTER_MATCHER\n\n";
1692 // Emit the subtarget feature enumeration.
1693 EmitSubtargetFeatureFlagEnumeration(Target, Info, OS);
1695 // Emit the function to match a register name to number.
1696 EmitMatchRegisterName(Target, AsmParser, OS);
1698 OS << "#endif // GET_REGISTER_MATCHER\n\n";
1701 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
1702 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
1704 // Generate the function that remaps for mnemonic aliases.
1705 bool HasMnemonicAliases = EmitMnemonicAliases(OS, Info);
1707 // Generate the unified function to convert operands into an MCInst.
1708 EmitConvertToMCInst(Target, Info.Instructions, OS);
1710 // Emit the enumeration for classes which participate in matching.
1711 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1713 // Emit the routine to match token strings to their match class.
1714 EmitMatchTokenString(Target, Info.Classes, OS);
1716 // Emit the routine to classify an operand.
1717 EmitClassifyOperand(Target, Info, OS);
1719 // Emit the subclass predicate routine.
1720 EmitIsSubclass(Target, Info.Classes, OS);
1722 // Emit the available features compute function.
1723 EmitComputeAvailableFeatures(Target, Info, OS);
1726 size_t MaxNumOperands = 0;
1727 for (std::vector<InstructionInfo*>::const_iterator it =
1728 Info.Instructions.begin(), ie = Info.Instructions.end();
1730 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
1733 // Emit the static match table; unused classes get initalized to 0 which is
1734 // guaranteed to be InvalidMatchClass.
1736 // FIXME: We can reduce the size of this table very easily. First, we change
1737 // it so that store the kinds in separate bit-fields for each index, which
1738 // only needs to be the max width used for classes at that index (we also need
1739 // to reject based on this during classification). If we then make sure to
1740 // order the match kinds appropriately (putting mnemonics last), then we
1741 // should only end up using a few bits for each class, especially the ones
1742 // following the mnemonic.
1743 OS << "namespace {\n";
1744 OS << " struct MatchEntry {\n";
1745 OS << " unsigned Opcode;\n";
1746 OS << " const char *Mnemonic;\n";
1747 OS << " ConversionKind ConvertFn;\n";
1748 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1749 OS << " unsigned RequiredFeatures;\n";
1752 OS << "// Predicate for searching for an opcode.\n";
1753 OS << " struct LessOpcode {\n";
1754 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
1755 OS << " return StringRef(LHS.Mnemonic) < RHS;\n";
1757 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
1758 OS << " return LHS < StringRef(RHS.Mnemonic);\n";
1760 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
1761 OS << " return StringRef(LHS.Mnemonic) < StringRef(RHS.Mnemonic);\n";
1765 OS << "} // end anonymous namespace.\n\n";
1767 OS << "static const MatchEntry MatchTable["
1768 << Info.Instructions.size() << "] = {\n";
1770 for (std::vector<InstructionInfo*>::const_iterator it =
1771 Info.Instructions.begin(), ie = Info.Instructions.end();
1773 InstructionInfo &II = **it;
1775 OS << " { " << Target.getName() << "::" << II.InstrName
1776 << ", \"" << II.Tokens[0] << "\""
1777 << ", " << II.ConversionFnKind << ", { ";
1778 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1779 InstructionInfo::Operand &Op = II.Operands[i];
1782 OS << Op.Class->Name;
1786 // Write the required features mask.
1787 if (!II.RequiredFeatures.empty()) {
1788 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
1790 OS << II.RequiredFeatures[i]->getEnumName();
1800 // Finally, build the match function.
1801 OS << Target.getName() << ClassName << "::MatchResultTy "
1802 << Target.getName() << ClassName << "::\n"
1803 << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>"
1805 OS << " MCInst &Inst, unsigned &ErrorInfo) {\n";
1807 // Emit code to get the available features.
1808 OS << " // Get the current feature set.\n";
1809 OS << " unsigned AvailableFeatures = getAvailableFeatures();\n\n";
1811 OS << " // Get the instruction mnemonic, which is the first token.\n";
1812 OS << " StringRef Mnemonic = ((" << Target.getName()
1813 << "Operand*)Operands[0])->getToken();\n\n";
1815 if (HasMnemonicAliases) {
1816 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
1817 OS << " ApplyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n";
1820 // Emit code to compute the class list for this operand vector.
1821 OS << " // Eliminate obvious mismatches.\n";
1822 OS << " if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
1823 OS << " ErrorInfo = " << (MaxNumOperands+1) << ";\n";
1824 OS << " return Match_InvalidOperand;\n";
1827 OS << " // Compute the class list for this operand vector.\n";
1828 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1829 OS << " for (unsigned i = 1, e = Operands.size(); i != e; ++i) {\n";
1830 OS << " Classes[i-1] = ClassifyOperand(Operands[i]);\n\n";
1832 OS << " // Check for invalid operands before matching.\n";
1833 OS << " if (Classes[i-1] == InvalidMatchClass) {\n";
1834 OS << " ErrorInfo = i;\n";
1835 OS << " return Match_InvalidOperand;\n";
1839 OS << " // Mark unused classes.\n";
1840 OS << " for (unsigned i = Operands.size()-1, e = " << MaxNumOperands << "; "
1841 << "i != e; ++i)\n";
1842 OS << " Classes[i] = InvalidMatchClass;\n\n";
1844 OS << " // Some state to try to produce better error messages.\n";
1845 OS << " bool HadMatchOtherThanFeatures = false;\n\n";
1846 OS << " // Set ErrorInfo to the operand that mismatches if it is \n";
1847 OS << " // wrong for all instances of the instruction.\n";
1848 OS << " ErrorInfo = ~0U;\n";
1850 // Emit code to search the table.
1851 OS << " // Search the table.\n";
1852 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
1853 OS << " std::equal_range(MatchTable, MatchTable+"
1854 << Info.Instructions.size() << ", Mnemonic, LessOpcode());\n\n";
1856 OS << " // Return a more specific error code if no mnemonics match.\n";
1857 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
1858 OS << " return Match_MnemonicFail;\n\n";
1860 OS << " for (const MatchEntry *it = MnemonicRange.first, "
1861 << "*ie = MnemonicRange.second;\n";
1862 OS << " it != ie; ++it) {\n";
1864 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
1865 OS << " assert(Mnemonic == it->Mnemonic);\n";
1867 // Emit check that the subclasses match.
1868 OS << " bool OperandsValid = true;\n";
1869 OS << " for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
1870 OS << " if (IsSubclass(Classes[i], it->Classes[i]))\n";
1871 OS << " continue;\n";
1872 OS << " // If this operand is broken for all of the instances of this\n";
1873 OS << " // mnemonic, keep track of it so we can report loc info.\n";
1874 OS << " if (it == MnemonicRange.first || ErrorInfo == i+1)\n";
1875 OS << " ErrorInfo = i+1;\n";
1877 OS << " ErrorInfo = ~0U;";
1878 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
1879 OS << " OperandsValid = false;\n";
1883 OS << " if (!OperandsValid) continue;\n";
1885 // Emit check that the required features are available.
1886 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
1887 << "!= it->RequiredFeatures) {\n";
1888 OS << " HadMatchOtherThanFeatures = true;\n";
1889 OS << " continue;\n";
1893 OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
1895 // Call the post-processing function, if used.
1896 std::string InsnCleanupFn =
1897 AsmParser->getValueAsString("AsmParserInstCleanup");
1898 if (!InsnCleanupFn.empty())
1899 OS << " " << InsnCleanupFn << "(Inst);\n";
1901 OS << " return Match_Success;\n";
1904 OS << " // Okay, we had no match. Try to return a useful error code.\n";
1905 OS << " if (HadMatchOtherThanFeatures) return Match_MissingFeature;\n";
1906 OS << " return Match_InvalidOperand;\n";
1909 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";