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/SmallPtrSet.h"
82 #include "llvm/ADT/SmallVector.h"
83 #include "llvm/ADT/STLExtras.h"
84 #include "llvm/ADT/StringExtras.h"
85 #include "llvm/Support/CommandLine.h"
86 #include "llvm/Support/Debug.h"
92 static cl::opt<std::string>
93 MatchPrefix("match-prefix", cl::init(""),
94 cl::desc("Only match instructions with the given prefix"));
99 struct SubtargetFeatureInfo;
101 /// ClassInfo - Helper class for storing the information about a particular
102 /// class of operands which can be matched.
105 /// Invalid kind, for use as a sentinel value.
108 /// The class for a particular token.
111 /// The (first) register class, subsequent register classes are
112 /// RegisterClass0+1, and so on.
115 /// The (first) user defined class, subsequent user defined classes are
116 /// UserClass0+1, and so on.
120 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
121 /// N) for the Nth user defined class.
124 /// SuperClasses - The super classes of this class. Note that for simplicities
125 /// sake user operands only record their immediate super class, while register
126 /// operands include all superclasses.
127 std::vector<ClassInfo*> SuperClasses;
129 /// Name - The full class name, suitable for use in an enum.
132 /// ClassName - The unadorned generic name for this class (e.g., Token).
133 std::string ClassName;
135 /// ValueName - The name of the value this class represents; for a token this
136 /// is the literal token string, for an operand it is the TableGen class (or
137 /// empty if this is a derived class).
138 std::string ValueName;
140 /// PredicateMethod - The name of the operand method to test whether the
141 /// operand matches this class; this is not valid for Token or register kinds.
142 std::string PredicateMethod;
144 /// RenderMethod - The name of the operand method to add this operand to an
145 /// MCInst; this is not valid for Token or register kinds.
146 std::string RenderMethod;
148 /// For register classes, the records for all the registers in this class.
149 std::set<Record*> Registers;
152 /// isRegisterClass() - Check if this is a register class.
153 bool isRegisterClass() const {
154 return Kind >= RegisterClass0 && Kind < UserClass0;
157 /// isUserClass() - Check if this is a user defined class.
158 bool isUserClass() const {
159 return Kind >= UserClass0;
162 /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
163 /// are related if they are in the same class hierarchy.
164 bool isRelatedTo(const ClassInfo &RHS) const {
165 // Tokens are only related to tokens.
166 if (Kind == Token || RHS.Kind == Token)
167 return Kind == Token && RHS.Kind == Token;
169 // Registers classes are only related to registers classes, and only if
170 // their intersection is non-empty.
171 if (isRegisterClass() || RHS.isRegisterClass()) {
172 if (!isRegisterClass() || !RHS.isRegisterClass())
175 std::set<Record*> Tmp;
176 std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
177 std::set_intersection(Registers.begin(), Registers.end(),
178 RHS.Registers.begin(), RHS.Registers.end(),
184 // Otherwise we have two users operands; they are related if they are in the
185 // same class hierarchy.
187 // FIXME: This is an oversimplification, they should only be related if they
188 // intersect, however we don't have that information.
189 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
190 const ClassInfo *Root = this;
191 while (!Root->SuperClasses.empty())
192 Root = Root->SuperClasses.front();
194 const ClassInfo *RHSRoot = &RHS;
195 while (!RHSRoot->SuperClasses.empty())
196 RHSRoot = RHSRoot->SuperClasses.front();
198 return Root == RHSRoot;
201 /// isSubsetOf - Test whether this class is a subset of \arg RHS;
202 bool isSubsetOf(const ClassInfo &RHS) const {
203 // This is a subset of RHS if it is the same class...
207 // ... or if any of its super classes are a subset of RHS.
208 for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
209 ie = SuperClasses.end(); it != ie; ++it)
210 if ((*it)->isSubsetOf(RHS))
216 /// operator< - Compare two classes.
217 bool operator<(const ClassInfo &RHS) const {
221 // Unrelated classes can be ordered by kind.
222 if (!isRelatedTo(RHS))
223 return Kind < RHS.Kind;
227 assert(0 && "Invalid kind!");
229 // Tokens are comparable by value.
231 // FIXME: Compare by enum value.
232 return ValueName < RHS.ValueName;
235 // This class preceeds the RHS if it is a proper subset of the RHS.
238 if (RHS.isSubsetOf(*this))
241 // Otherwise, order by name to ensure we have a total ordering.
242 return ValueName < RHS.ValueName;
247 /// MatchableInfo - Helper class for storing the necessary information for an
248 /// instruction or alias which is capable of being matched.
249 struct MatchableInfo {
251 /// Token - This is the token that the operand came from.
254 /// The unique class instance this operand should match.
257 /// The original operand this corresponds to. This is unset for singleton
258 /// registers and tokens, because they don't have a list in the ins/outs
259 /// list. If an operand is tied ($a=$b), this refers to source operand: $b.
260 const CGIOperandList::OperandInfo *OperandInfo;
262 explicit Operand(StringRef T) : Token(T), Class(0), OperandInfo(0) {}
265 /// InstrName - The target name for this instruction.
266 std::string InstrName;
268 /// TheDef - This is the definition of the instruction or InstAlias that this
269 /// matchable came from.
270 Record *const TheDef;
272 /// OperandList - This is the operand list that came from the (ins) and (outs)
273 /// list of the alias or instruction.
274 const CGIOperandList &OperandList;
276 /// AsmString - The assembly string for this instruction (with variants
277 /// removed), e.g. "movsx $src, $dst".
278 std::string AsmString;
280 /// Mnemonic - This is the first token of the matched instruction, its
284 /// AsmOperands - The textual operands that this instruction matches,
285 /// annotated with a class and where in the OperandList they were defined.
286 /// This directly corresponds to the tokenized AsmString after the mnemonic is
288 SmallVector<Operand, 4> AsmOperands;
290 /// Predicates - The required subtarget features to match this instruction.
291 SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures;
293 /// ConversionFnKind - The enum value which is passed to the generated
294 /// ConvertToMCInst to convert parsed operands into an MCInst for this
296 std::string ConversionFnKind;
298 MatchableInfo(const CodeGenInstruction &CGI)
299 : TheDef(CGI.TheDef), OperandList(CGI.Operands), AsmString(CGI.AsmString) {
300 InstrName = TheDef->getName();
303 MatchableInfo(const CodeGenInstAlias *Alias)
304 : TheDef(Alias->TheDef), OperandList(Alias->Operands),
305 AsmString(Alias->AsmString) {
308 DefInit *DI = dynamic_cast<DefInit*>(Alias->Result->getOperator());
311 InstrName = DI->getDef()->getName();
314 void Initialize(const AsmMatcherInfo &Info,
315 SmallPtrSet<Record*, 16> &SingletonRegisters);
317 /// Validate - Return true if this matchable is a valid thing to match against
318 /// and perform a bunch of validity checking.
319 bool Validate(StringRef CommentDelimiter, bool Hack) const;
321 /// getSingletonRegisterForAsmOperand - If the specified token is a singleton
322 /// register, return the Record for it, otherwise return null.
323 Record *getSingletonRegisterForAsmOperand(unsigned i,
324 const AsmMatcherInfo &Info) const;
326 /// operator< - Compare two matchables.
327 bool operator<(const MatchableInfo &RHS) const {
328 // The primary comparator is the instruction mnemonic.
329 if (Mnemonic != RHS.Mnemonic)
330 return Mnemonic < RHS.Mnemonic;
332 if (AsmOperands.size() != RHS.AsmOperands.size())
333 return AsmOperands.size() < RHS.AsmOperands.size();
335 // Compare lexicographically by operand. The matcher validates that other
336 // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith().
337 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
338 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
340 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
347 /// CouldMatchAmiguouslyWith - Check whether this matchable could
348 /// ambiguously match the same set of operands as \arg RHS (without being a
349 /// strictly superior match).
350 bool CouldMatchAmiguouslyWith(const MatchableInfo &RHS) {
351 // The primary comparator is the instruction mnemonic.
352 if (Mnemonic != RHS.Mnemonic)
355 // The number of operands is unambiguous.
356 if (AsmOperands.size() != RHS.AsmOperands.size())
359 // Otherwise, make sure the ordering of the two instructions is unambiguous
360 // by checking that either (a) a token or operand kind discriminates them,
361 // or (b) the ordering among equivalent kinds is consistent.
363 // Tokens and operand kinds are unambiguous (assuming a correct target
365 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
366 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
367 AsmOperands[i].Class->Kind == ClassInfo::Token)
368 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
369 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
372 // Otherwise, this operand could commute if all operands are equivalent, or
373 // there is a pair of operands that compare less than and a pair that
374 // compare greater than.
375 bool HasLT = false, HasGT = false;
376 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
377 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
379 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
383 return !(HasLT ^ HasGT);
389 void TokenizeAsmString(const AsmMatcherInfo &Info);
392 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
393 /// feature which participates in instruction matching.
394 struct SubtargetFeatureInfo {
395 /// \brief The predicate record for this feature.
398 /// \brief An unique index assigned to represent this feature.
401 SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {}
403 /// \brief The name of the enumerated constant identifying this feature.
404 std::string getEnumName() const {
405 return "Feature_" + TheDef->getName();
409 class AsmMatcherInfo {
411 /// The tablegen AsmParser record.
414 /// Target - The target information.
415 CodeGenTarget &Target;
417 /// The AsmParser "RegisterPrefix" value.
418 std::string RegisterPrefix;
420 /// The classes which are needed for matching.
421 std::vector<ClassInfo*> Classes;
423 /// The information on the matchables to match.
424 std::vector<MatchableInfo*> Matchables;
426 /// Map of Register records to their class information.
427 std::map<Record*, ClassInfo*> RegisterClasses;
429 /// Map of Predicate records to their subtarget information.
430 std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures;
433 /// Map of token to class information which has already been constructed.
434 std::map<std::string, ClassInfo*> TokenClasses;
436 /// Map of RegisterClass records to their class information.
437 std::map<Record*, ClassInfo*> RegisterClassClasses;
439 /// Map of AsmOperandClass records to their class information.
440 std::map<Record*, ClassInfo*> AsmOperandClasses;
443 /// getTokenClass - Lookup or create the class for the given token.
444 ClassInfo *getTokenClass(StringRef Token);
446 /// getOperandClass - Lookup or create the class for the given operand.
447 ClassInfo *getOperandClass(StringRef Token,
448 const CGIOperandList::OperandInfo &OI);
450 /// BuildRegisterClasses - Build the ClassInfo* instances for register
452 void BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters);
454 /// BuildOperandClasses - Build the ClassInfo* instances for user defined
456 void BuildOperandClasses();
459 AsmMatcherInfo(Record *AsmParser, CodeGenTarget &Target);
461 /// BuildInfo - Construct the various tables used during matching.
464 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
466 SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
467 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
468 std::map<Record*, SubtargetFeatureInfo*>::const_iterator I =
469 SubtargetFeatures.find(Def);
470 return I == SubtargetFeatures.end() ? 0 : I->second;
476 void MatchableInfo::dump() {
477 errs() << InstrName << " -- " << "flattened:\"" << AsmString << "\"\n";
479 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
480 Operand &Op = AsmOperands[i];
481 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
482 if (Op.Class->Kind == ClassInfo::Token) {
483 errs() << '\"' << Op.Token << "\"\n";
487 if (!Op.OperandInfo) {
488 errs() << "(singleton register)\n";
492 const CGIOperandList::OperandInfo &OI = *Op.OperandInfo;
493 errs() << OI.Name << " " << OI.Rec->getName()
494 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
498 void MatchableInfo::Initialize(const AsmMatcherInfo &Info,
499 SmallPtrSet<Record*, 16> &SingletonRegisters) {
500 // TODO: Eventually support asmparser for Variant != 0.
501 AsmString = CodeGenInstruction::FlattenAsmStringVariants(AsmString, 0);
503 TokenizeAsmString(Info);
505 // Compute the require features.
506 std::vector<Record*> Predicates =TheDef->getValueAsListOfDefs("Predicates");
507 for (unsigned i = 0, e = Predicates.size(); i != e; ++i)
508 if (SubtargetFeatureInfo *Feature =
509 Info.getSubtargetFeature(Predicates[i]))
510 RequiredFeatures.push_back(Feature);
512 // Collect singleton registers, if used.
513 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
514 if (Record *Reg = getSingletonRegisterForAsmOperand(i, Info))
515 SingletonRegisters.insert(Reg);
519 /// TokenizeAsmString - Tokenize a simplified assembly string.
520 void MatchableInfo::TokenizeAsmString(const AsmMatcherInfo &Info) {
521 StringRef String = AsmString;
524 for (unsigned i = 0, e = String.size(); i != e; ++i) {
534 AsmOperands.push_back(Operand(String.slice(Prev, i)));
537 if (!isspace(String[i]) && String[i] != ',')
538 AsmOperands.push_back(Operand(String.substr(i, 1)));
544 AsmOperands.push_back(Operand(String.slice(Prev, i)));
548 assert(i != String.size() && "Invalid quoted character");
549 AsmOperands.push_back(Operand(String.substr(i, 1)));
554 // If this isn't "${", treat like a normal token.
555 if (i + 1 == String.size() || String[i + 1] != '{') {
557 AsmOperands.push_back(Operand(String.slice(Prev, i)));
565 AsmOperands.push_back(Operand(String.slice(Prev, i)));
569 StringRef::iterator End = std::find(String.begin() + i, String.end(),'}');
570 assert(End != String.end() && "Missing brace in operand reference!");
571 size_t EndPos = End - String.begin();
572 AsmOperands.push_back(Operand(String.slice(i, EndPos+1)));
580 AsmOperands.push_back(Operand(String.slice(Prev, i)));
589 if (InTok && Prev != String.size())
590 AsmOperands.push_back(Operand(String.substr(Prev)));
592 // The first token of the instruction is the mnemonic, which must be a
593 // simple string, not a $foo variable or a singleton register.
594 assert(!AsmOperands.empty() && "Instruction has no tokens?");
595 Mnemonic = AsmOperands[0].Token;
596 if (Mnemonic[0] == '$' || getSingletonRegisterForAsmOperand(0, Info))
597 throw TGError(TheDef->getLoc(),
598 "Invalid instruction mnemonic '" + Mnemonic.str() + "'!");
600 // Remove the first operand, it is tracked in the mnemonic field.
601 AsmOperands.erase(AsmOperands.begin());
606 bool MatchableInfo::Validate(StringRef CommentDelimiter, bool Hack) const {
607 // Reject matchables with no .s string.
608 if (AsmString.empty())
609 throw TGError(TheDef->getLoc(), "instruction with empty asm string");
611 // Reject any matchables with a newline in them, they should be marked
612 // isCodeGenOnly if they are pseudo instructions.
613 if (AsmString.find('\n') != std::string::npos)
614 throw TGError(TheDef->getLoc(),
615 "multiline instruction is not valid for the asmparser, "
616 "mark it isCodeGenOnly");
618 // Remove comments from the asm string. We know that the asmstring only
620 if (!CommentDelimiter.empty() &&
621 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
622 throw TGError(TheDef->getLoc(),
623 "asmstring for instruction has comment character in it, "
624 "mark it isCodeGenOnly");
626 // Reject matchables with operand modifiers, these aren't something we can
627 /// handle, the target should be refactored to use operands instead of
630 // Also, check for instructions which reference the operand multiple times;
631 // this implies a constraint we would not honor.
632 std::set<std::string> OperandNames;
633 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
634 StringRef Tok = AsmOperands[i].Token;
635 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
636 throw TGError(TheDef->getLoc(),
637 "matchable with operand modifier '" + Tok.str() +
638 "' not supported by asm matcher. Mark isCodeGenOnly!");
640 // Verify that any operand is only mentioned once.
641 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
643 throw TGError(TheDef->getLoc(),
644 "ERROR: matchable with tied operand '" + Tok.str() +
645 "' can never be matched!");
646 // FIXME: Should reject these. The ARM backend hits this with $lane in a
647 // bunch of instructions. It is unclear what the right answer is.
649 errs() << "warning: '" << InstrName << "': "
650 << "ignoring instruction with tied operand '"
651 << Tok.str() << "'\n";
661 /// getSingletonRegisterForAsmOperand - If the specified token is a singleton
662 /// register, return the register name, otherwise return a null StringRef.
663 Record *MatchableInfo::
664 getSingletonRegisterForAsmOperand(unsigned i, const AsmMatcherInfo &Info) const{
665 StringRef Tok = AsmOperands[i].Token;
666 if (!Tok.startswith(Info.RegisterPrefix))
669 StringRef RegName = Tok.substr(Info.RegisterPrefix.size());
670 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
673 // If there is no register prefix (i.e. "%" in "%eax"), then this may
674 // be some random non-register token, just ignore it.
675 if (Info.RegisterPrefix.empty())
678 // Otherwise, we have something invalid prefixed with the register prefix,
680 std::string Err = "unable to find register for '" + RegName.str() +
681 "' (which matches register prefix)";
682 throw TGError(TheDef->getLoc(), Err);
686 static std::string getEnumNameForToken(StringRef Str) {
689 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
691 case '*': Res += "_STAR_"; break;
692 case '%': Res += "_PCT_"; break;
693 case ':': Res += "_COLON_"; break;
698 Res += "_" + utostr((unsigned) *it) + "_";
705 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
706 ClassInfo *&Entry = TokenClasses[Token];
709 Entry = new ClassInfo();
710 Entry->Kind = ClassInfo::Token;
711 Entry->ClassName = "Token";
712 Entry->Name = "MCK_" + getEnumNameForToken(Token);
713 Entry->ValueName = Token;
714 Entry->PredicateMethod = "<invalid>";
715 Entry->RenderMethod = "<invalid>";
716 Classes.push_back(Entry);
723 AsmMatcherInfo::getOperandClass(StringRef Token,
724 const CGIOperandList::OperandInfo &OI) {
725 if (OI.Rec->isSubClassOf("RegisterClass")) {
726 if (ClassInfo *CI = RegisterClassClasses[OI.Rec])
728 throw TGError(OI.Rec->getLoc(), "register class has no class info!");
731 assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!");
732 Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass");
733 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
736 throw TGError(OI.Rec->getLoc(), "operand has no match class!");
739 void AsmMatcherInfo::
740 BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters) {
741 const std::vector<CodeGenRegister> &Registers = Target.getRegisters();
742 const std::vector<CodeGenRegisterClass> &RegClassList =
743 Target.getRegisterClasses();
745 // The register sets used for matching.
746 std::set< std::set<Record*> > RegisterSets;
748 // Gather the defined sets.
749 for (std::vector<CodeGenRegisterClass>::const_iterator it =
750 RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it)
751 RegisterSets.insert(std::set<Record*>(it->Elements.begin(),
752 it->Elements.end()));
754 // Add any required singleton sets.
755 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
756 ie = SingletonRegisters.end(); it != ie; ++it) {
758 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
761 // Introduce derived sets where necessary (when a register does not determine
762 // a unique register set class), and build the mapping of registers to the set
763 // they should classify to.
764 std::map<Record*, std::set<Record*> > RegisterMap;
765 for (std::vector<CodeGenRegister>::const_iterator it = Registers.begin(),
766 ie = Registers.end(); it != ie; ++it) {
767 const CodeGenRegister &CGR = *it;
768 // Compute the intersection of all sets containing this register.
769 std::set<Record*> ContainingSet;
771 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
772 ie = RegisterSets.end(); it != ie; ++it) {
773 if (!it->count(CGR.TheDef))
776 if (ContainingSet.empty()) {
781 std::set<Record*> Tmp;
782 std::swap(Tmp, ContainingSet);
783 std::insert_iterator< std::set<Record*> > II(ContainingSet,
784 ContainingSet.begin());
785 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(), II);
788 if (!ContainingSet.empty()) {
789 RegisterSets.insert(ContainingSet);
790 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
794 // Construct the register classes.
795 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
797 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
798 ie = RegisterSets.end(); it != ie; ++it, ++Index) {
799 ClassInfo *CI = new ClassInfo();
800 CI->Kind = ClassInfo::RegisterClass0 + Index;
801 CI->ClassName = "Reg" + utostr(Index);
802 CI->Name = "MCK_Reg" + utostr(Index);
804 CI->PredicateMethod = ""; // unused
805 CI->RenderMethod = "addRegOperands";
807 Classes.push_back(CI);
808 RegisterSetClasses.insert(std::make_pair(*it, CI));
811 // Find the superclasses; we could compute only the subgroup lattice edges,
812 // but there isn't really a point.
813 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
814 ie = RegisterSets.end(); it != ie; ++it) {
815 ClassInfo *CI = RegisterSetClasses[*it];
816 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
817 ie2 = RegisterSets.end(); it2 != ie2; ++it2)
819 std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
820 CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
823 // Name the register classes which correspond to a user defined RegisterClass.
824 for (std::vector<CodeGenRegisterClass>::const_iterator
825 it = RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it) {
826 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(),
827 it->Elements.end())];
828 if (CI->ValueName.empty()) {
829 CI->ClassName = it->getName();
830 CI->Name = "MCK_" + it->getName();
831 CI->ValueName = it->getName();
833 CI->ValueName = CI->ValueName + "," + it->getName();
835 RegisterClassClasses.insert(std::make_pair(it->TheDef, CI));
838 // Populate the map for individual registers.
839 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
840 ie = RegisterMap.end(); it != ie; ++it)
841 RegisterClasses[it->first] = RegisterSetClasses[it->second];
843 // Name the register classes which correspond to singleton registers.
844 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
845 ie = SingletonRegisters.end(); it != ie; ++it) {
847 ClassInfo *CI = RegisterClasses[Rec];
848 assert(CI && "Missing singleton register class info!");
850 if (CI->ValueName.empty()) {
851 CI->ClassName = Rec->getName();
852 CI->Name = "MCK_" + Rec->getName();
853 CI->ValueName = Rec->getName();
855 CI->ValueName = CI->ValueName + "," + Rec->getName();
859 void AsmMatcherInfo::BuildOperandClasses() {
860 std::vector<Record*> AsmOperands =
861 Records.getAllDerivedDefinitions("AsmOperandClass");
863 // Pre-populate AsmOperandClasses map.
864 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
865 ie = AsmOperands.end(); it != ie; ++it)
866 AsmOperandClasses[*it] = new ClassInfo();
869 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
870 ie = AsmOperands.end(); it != ie; ++it, ++Index) {
871 ClassInfo *CI = AsmOperandClasses[*it];
872 CI->Kind = ClassInfo::UserClass0 + Index;
874 ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
875 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
876 DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
878 PrintError((*it)->getLoc(), "Invalid super class reference!");
882 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
884 PrintError((*it)->getLoc(), "Invalid super class reference!");
886 CI->SuperClasses.push_back(SC);
888 CI->ClassName = (*it)->getValueAsString("Name");
889 CI->Name = "MCK_" + CI->ClassName;
890 CI->ValueName = (*it)->getName();
892 // Get or construct the predicate method name.
893 Init *PMName = (*it)->getValueInit("PredicateMethod");
894 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
895 CI->PredicateMethod = SI->getValue();
897 assert(dynamic_cast<UnsetInit*>(PMName) &&
898 "Unexpected PredicateMethod field!");
899 CI->PredicateMethod = "is" + CI->ClassName;
902 // Get or construct the render method name.
903 Init *RMName = (*it)->getValueInit("RenderMethod");
904 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
905 CI->RenderMethod = SI->getValue();
907 assert(dynamic_cast<UnsetInit*>(RMName) &&
908 "Unexpected RenderMethod field!");
909 CI->RenderMethod = "add" + CI->ClassName + "Operands";
912 AsmOperandClasses[*it] = CI;
913 Classes.push_back(CI);
917 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, CodeGenTarget &target)
918 : AsmParser(asmParser), Target(target),
919 RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix")) {
923 void AsmMatcherInfo::BuildInfo() {
924 // Build information about all of the AssemblerPredicates.
925 std::vector<Record*> AllPredicates =
926 Records.getAllDerivedDefinitions("Predicate");
927 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
928 Record *Pred = AllPredicates[i];
929 // Ignore predicates that are not intended for the assembler.
930 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
933 if (Pred->getName().empty())
934 throw TGError(Pred->getLoc(), "Predicate has no name!");
936 unsigned FeatureNo = SubtargetFeatures.size();
937 SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo);
938 assert(FeatureNo < 32 && "Too many subtarget features!");
941 StringRef CommentDelimiter = AsmParser->getValueAsString("CommentDelimiter");
943 // Parse the instructions; we need to do this first so that we can gather the
944 // singleton register classes.
945 SmallPtrSet<Record*, 16> SingletonRegisters;
946 for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
947 E = Target.inst_end(); I != E; ++I) {
948 const CodeGenInstruction &CGI = **I;
950 // If the tblgen -match-prefix option is specified (for tblgen hackers),
951 // filter the set of instructions we consider.
952 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
955 // Ignore "codegen only" instructions.
956 if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
959 OwningPtr<MatchableInfo> II(new MatchableInfo(CGI));
961 II->Initialize(*this, SingletonRegisters);
963 // Ignore instructions which shouldn't be matched and diagnose invalid
964 // instruction definitions with an error.
965 if (!II->Validate(CommentDelimiter, true))
968 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
970 // FIXME: This is a total hack.
971 if (StringRef(II->InstrName).startswith("Int_") ||
972 StringRef(II->InstrName).endswith("_Int"))
975 Matchables.push_back(II.take());
978 // Parse all of the InstAlias definitions and stick them in the list of
980 std::vector<Record*> AllInstAliases =
981 Records.getAllDerivedDefinitions("InstAlias");
982 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
983 CodeGenInstAlias *Alias = new CodeGenInstAlias(AllInstAliases[i]);
985 OwningPtr<MatchableInfo> II(new MatchableInfo(Alias));
987 II->Initialize(*this, SingletonRegisters);
989 // Validate the alias definitions.
990 II->Validate(CommentDelimiter, false);
992 Matchables.push_back(II.take());
995 // Build info for the register classes.
996 BuildRegisterClasses(SingletonRegisters);
998 // Build info for the user defined assembly operand classes.
999 BuildOperandClasses();
1001 // Build the information about matchables.
1002 for (std::vector<MatchableInfo*>::iterator it = Matchables.begin(),
1003 ie = Matchables.end(); it != ie; ++it) {
1004 MatchableInfo *II = *it;
1006 // Parse the tokens after the mnemonic.
1007 for (unsigned i = 0, e = II->AsmOperands.size(); i != e; ++i) {
1008 MatchableInfo::Operand &Op = II->AsmOperands[i];
1009 StringRef Token = Op.Token;
1011 // Check for singleton registers.
1012 if (Record *RegRecord = II->getSingletonRegisterForAsmOperand(i, *this)) {
1013 Op.Class = RegisterClasses[RegRecord];
1014 assert(Op.Class && Op.Class->Registers.size() == 1 &&
1015 "Unexpected class for singleton register");
1019 // Check for simple tokens.
1020 if (Token[0] != '$') {
1021 Op.Class = getTokenClass(Token);
1025 // Otherwise this is an operand reference.
1026 StringRef OperandName;
1027 if (Token[1] == '{')
1028 OperandName = Token.substr(2, Token.size() - 3);
1030 OperandName = Token.substr(1);
1032 // Map this token to an operand. FIXME: Move elsewhere.
1034 if (!II->OperandList.hasOperandNamed(OperandName, Idx))
1035 throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" +
1036 OperandName.str() + "'");
1038 // FIXME: This is annoying, the named operand may be tied (e.g.,
1039 // XCHG8rm). What we want is the untied operand, which we now have to
1040 // grovel for. Only worry about this for single entry operands, we have to
1041 // clean this up anyway.
1042 const CGIOperandList::OperandInfo *OI = &II->OperandList[Idx];
1043 if (OI->Constraints[0].isTied()) {
1044 unsigned TiedOp = OI->Constraints[0].getTiedOperand();
1046 // The tied operand index is an MIOperand index, find the operand that
1048 for (unsigned i = 0, e = II->OperandList.size(); i != e; ++i) {
1049 if (II->OperandList[i].MIOperandNo == TiedOp) {
1050 OI = &II->OperandList[i];
1055 assert(OI && "Unable to find tied operand target!");
1058 Op.Class = getOperandClass(Token, *OI);
1059 Op.OperandInfo = OI;
1063 // Reorder classes so that classes preceed super classes.
1064 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
1067 static std::pair<unsigned, unsigned> *
1068 GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List,
1070 for (unsigned i = 0, e = List.size(); i != e; ++i)
1071 if (Index == List[i].first)
1077 static void EmitConvertToMCInst(CodeGenTarget &Target,
1078 std::vector<MatchableInfo*> &Infos,
1080 // Write the convert function to a separate stream, so we can drop it after
1082 std::string ConvertFnBody;
1083 raw_string_ostream CvtOS(ConvertFnBody);
1085 // Function we have already generated.
1086 std::set<std::string> GeneratedFns;
1088 // Start the unified conversion function.
1090 CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
1091 << "unsigned Opcode,\n"
1092 << " const SmallVectorImpl<MCParsedAsmOperand*"
1093 << "> &Operands) {\n";
1094 CvtOS << " Inst.setOpcode(Opcode);\n";
1095 CvtOS << " switch (Kind) {\n";
1096 CvtOS << " default:\n";
1098 // Start the enum, which we will generate inline.
1100 OS << "// Unified function for converting operants to MCInst instances.\n\n";
1101 OS << "enum ConversionKind {\n";
1103 // TargetOperandClass - This is the target's operand class, like X86Operand.
1104 std::string TargetOperandClass = Target.getName() + "Operand";
1106 for (std::vector<MatchableInfo*>::const_iterator it = Infos.begin(),
1107 ie = Infos.end(); it != ie; ++it) {
1108 MatchableInfo &II = **it;
1110 // Order the (class) operands by the order to convert them into an MCInst.
1111 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
1112 for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
1113 MatchableInfo::Operand &Op = II.AsmOperands[i];
1115 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
1118 // Find any tied operands.
1119 SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands;
1120 for (unsigned i = 0, e = II.OperandList.size(); i != e; ++i) {
1121 const CGIOperandList::OperandInfo &OpInfo = II.OperandList[i];
1122 for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) {
1123 const CGIOperandList::ConstraintInfo &CI = OpInfo.Constraints[j];
1124 if (!CI.isTied()) continue;
1125 TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo,
1126 CI.getTiedOperand()));
1130 array_pod_sort(MIOperandList.begin(), MIOperandList.end());
1132 // Compute the total number of operands.
1133 unsigned NumMIOperands = 0;
1134 for (unsigned i = 0, e = II.OperandList.size(); i != e; ++i) {
1135 const CGIOperandList::OperandInfo &OI = II.OperandList[i];
1136 NumMIOperands = std::max(NumMIOperands, OI.MIOperandNo+OI.MINumOperands);
1139 // Build the conversion function signature.
1140 std::string Signature = "Convert";
1141 unsigned CurIndex = 0;
1143 std::string CaseBody;
1144 raw_string_ostream CaseOS(CaseBody);
1146 // Compute the convert enum and the case body.
1147 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1148 MatchableInfo::Operand &Op = II.AsmOperands[MIOperandList[i].second];
1149 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
1150 "Duplicate match for instruction operand!");
1152 // Add the implicit operands.
1153 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1154 // See if this is a tied operand.
1155 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1159 // If not, this is some implicit operand. Just assume it is a register
1161 CaseOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1162 Signature += "__Imp";
1164 // Copy the tied operand.
1165 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1166 CaseOS << " Inst.addOperand(Inst.getOperand("
1167 << Tie->second << "));\n";
1168 Signature += "__Tie" + utostr(Tie->second);
1172 // Registers are always converted the same, don't duplicate the conversion
1173 // function based on them.
1175 // FIXME: We could generalize this based on the render method, if it
1178 if (Op.Class->isRegisterClass())
1181 Signature += Op.Class->ClassName;
1182 Signature += utostr(Op.OperandInfo->MINumOperands);
1183 Signature += "_" + utostr(MIOperandList[i].second);
1186 CaseOS << " ((" << TargetOperandClass << "*)Operands["
1187 << MIOperandList[i].second << "+1])->" << Op.Class->RenderMethod
1188 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
1189 CurIndex += Op.OperandInfo->MINumOperands;
1192 // And add trailing implicit operands.
1193 for (; CurIndex != NumMIOperands; ++CurIndex) {
1194 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1198 // If not, this is some implicit operand. Just assume it is a register
1200 CaseOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1201 Signature += "__Imp";
1203 // Copy the tied operand.
1204 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1205 CaseOS << " Inst.addOperand(Inst.getOperand("
1206 << Tie->second << "));\n";
1207 Signature += "__Tie" + utostr(Tie->second);
1211 II.ConversionFnKind = Signature;
1213 // Check if we have already generated this signature.
1214 if (!GeneratedFns.insert(Signature).second)
1217 // If not, emit it now. Add to the enum list.
1218 OS << " " << Signature << ",\n";
1220 CvtOS << " case " << Signature << ":\n";
1221 CvtOS << CaseOS.str();
1222 CvtOS << " return;\n";
1225 // Finish the convert function.
1230 // Finish the enum, and drop the convert function after it.
1232 OS << " NumConversionVariants\n";
1238 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
1239 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
1240 std::vector<ClassInfo*> &Infos,
1242 OS << "namespace {\n\n";
1244 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1245 << "/// instruction matching.\n";
1246 OS << "enum MatchClassKind {\n";
1247 OS << " InvalidMatchClass = 0,\n";
1248 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1249 ie = Infos.end(); it != ie; ++it) {
1250 ClassInfo &CI = **it;
1251 OS << " " << CI.Name << ", // ";
1252 if (CI.Kind == ClassInfo::Token) {
1253 OS << "'" << CI.ValueName << "'\n";
1254 } else if (CI.isRegisterClass()) {
1255 if (!CI.ValueName.empty())
1256 OS << "register class '" << CI.ValueName << "'\n";
1258 OS << "derived register class\n";
1260 OS << "user defined class '" << CI.ValueName << "'\n";
1263 OS << " NumMatchClassKinds\n";
1269 /// EmitClassifyOperand - Emit the function to classify an operand.
1270 static void EmitClassifyOperand(AsmMatcherInfo &Info,
1272 OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n"
1273 << " " << Info.Target.getName() << "Operand &Operand = *("
1274 << Info.Target.getName() << "Operand*)GOp;\n";
1277 OS << " if (Operand.isToken())\n";
1278 OS << " return MatchTokenString(Operand.getToken());\n\n";
1280 // Classify registers.
1282 // FIXME: Don't hardcode isReg, getReg.
1283 OS << " if (Operand.isReg()) {\n";
1284 OS << " switch (Operand.getReg()) {\n";
1285 OS << " default: return InvalidMatchClass;\n";
1286 for (std::map<Record*, ClassInfo*>::iterator
1287 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1289 OS << " case " << Info.Target.getName() << "::"
1290 << it->first->getName() << ": return " << it->second->Name << ";\n";
1294 // Classify user defined operands.
1295 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1296 ie = Info.Classes.end(); it != ie; ++it) {
1297 ClassInfo &CI = **it;
1299 if (!CI.isUserClass())
1302 OS << " // '" << CI.ClassName << "' class";
1303 if (!CI.SuperClasses.empty()) {
1304 OS << ", subclass of ";
1305 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) {
1307 OS << "'" << CI.SuperClasses[i]->ClassName << "'";
1308 assert(CI < *CI.SuperClasses[i] && "Invalid class relation!");
1313 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
1315 // Validate subclass relationships.
1316 if (!CI.SuperClasses.empty()) {
1317 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i)
1318 OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod
1319 << "() && \"Invalid class relationship!\");\n";
1322 OS << " return " << CI.Name << ";\n";
1325 OS << " return InvalidMatchClass;\n";
1329 /// EmitIsSubclass - Emit the subclass predicate function.
1330 static void EmitIsSubclass(CodeGenTarget &Target,
1331 std::vector<ClassInfo*> &Infos,
1333 OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1334 OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n";
1335 OS << " if (A == B)\n";
1336 OS << " return true;\n\n";
1338 OS << " switch (A) {\n";
1339 OS << " default:\n";
1340 OS << " return false;\n";
1341 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1342 ie = Infos.end(); it != ie; ++it) {
1343 ClassInfo &A = **it;
1345 if (A.Kind != ClassInfo::Token) {
1346 std::vector<StringRef> SuperClasses;
1347 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1348 ie = Infos.end(); it != ie; ++it) {
1349 ClassInfo &B = **it;
1351 if (&A != &B && A.isSubsetOf(B))
1352 SuperClasses.push_back(B.Name);
1355 if (SuperClasses.empty())
1358 OS << "\n case " << A.Name << ":\n";
1360 if (SuperClasses.size() == 1) {
1361 OS << " return B == " << SuperClasses.back() << ";\n";
1365 OS << " switch (B) {\n";
1366 OS << " default: return false;\n";
1367 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1368 OS << " case " << SuperClasses[i] << ": return true;\n";
1378 /// EmitMatchTokenString - Emit the function to match a token string to the
1379 /// appropriate match class value.
1380 static void EmitMatchTokenString(CodeGenTarget &Target,
1381 std::vector<ClassInfo*> &Infos,
1383 // Construct the match list.
1384 std::vector<StringMatcher::StringPair> Matches;
1385 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1386 ie = Infos.end(); it != ie; ++it) {
1387 ClassInfo &CI = **it;
1389 if (CI.Kind == ClassInfo::Token)
1390 Matches.push_back(StringMatcher::StringPair(CI.ValueName,
1391 "return " + CI.Name + ";"));
1394 OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
1396 StringMatcher("Name", Matches, OS).Emit();
1398 OS << " return InvalidMatchClass;\n";
1402 /// EmitMatchRegisterName - Emit the function to match a string to the target
1403 /// specific register enum.
1404 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1406 // Construct the match list.
1407 std::vector<StringMatcher::StringPair> Matches;
1408 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1409 const CodeGenRegister &Reg = Target.getRegisters()[i];
1410 if (Reg.TheDef->getValueAsString("AsmName").empty())
1413 Matches.push_back(StringMatcher::StringPair(
1414 Reg.TheDef->getValueAsString("AsmName"),
1415 "return " + utostr(i + 1) + ";"));
1418 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1420 StringMatcher("Name", Matches, OS).Emit();
1422 OS << " return 0;\n";
1426 /// EmitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
1428 static void EmitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
1430 OS << "// Flags for subtarget features that participate in "
1431 << "instruction matching.\n";
1432 OS << "enum SubtargetFeatureFlag {\n";
1433 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1434 it = Info.SubtargetFeatures.begin(),
1435 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1436 SubtargetFeatureInfo &SFI = *it->second;
1437 OS << " " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n";
1439 OS << " Feature_None = 0\n";
1443 /// EmitComputeAvailableFeatures - Emit the function to compute the list of
1444 /// available features given a subtarget.
1445 static void EmitComputeAvailableFeatures(AsmMatcherInfo &Info,
1447 std::string ClassName =
1448 Info.AsmParser->getValueAsString("AsmParserClassName");
1450 OS << "unsigned " << Info.Target.getName() << ClassName << "::\n"
1451 << "ComputeAvailableFeatures(const " << Info.Target.getName()
1452 << "Subtarget *Subtarget) const {\n";
1453 OS << " unsigned Features = 0;\n";
1454 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1455 it = Info.SubtargetFeatures.begin(),
1456 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1457 SubtargetFeatureInfo &SFI = *it->second;
1458 OS << " if (" << SFI.TheDef->getValueAsString("CondString")
1460 OS << " Features |= " << SFI.getEnumName() << ";\n";
1462 OS << " return Features;\n";
1466 static std::string GetAliasRequiredFeatures(Record *R,
1467 const AsmMatcherInfo &Info) {
1468 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
1470 unsigned NumFeatures = 0;
1471 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
1472 SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
1475 throw TGError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
1476 "' is not marked as an AssemblerPredicate!");
1481 Result += F->getEnumName();
1485 if (NumFeatures > 1)
1486 Result = '(' + Result + ')';
1490 /// EmitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
1491 /// emit a function for them and return true, otherwise return false.
1492 static bool EmitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) {
1493 std::vector<Record*> Aliases =
1494 Records.getAllDerivedDefinitions("MnemonicAlias");
1495 if (Aliases.empty()) return false;
1497 OS << "static void ApplyMnemonicAliases(StringRef &Mnemonic, "
1498 "unsigned Features) {\n";
1500 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
1501 // iteration order of the map is stable.
1502 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
1504 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
1505 Record *R = Aliases[i];
1506 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
1509 // Process each alias a "from" mnemonic at a time, building the code executed
1510 // by the string remapper.
1511 std::vector<StringMatcher::StringPair> Cases;
1512 for (std::map<std::string, std::vector<Record*> >::iterator
1513 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
1515 const std::vector<Record*> &ToVec = I->second;
1517 // Loop through each alias and emit code that handles each case. If there
1518 // are two instructions without predicates, emit an error. If there is one,
1520 std::string MatchCode;
1521 int AliasWithNoPredicate = -1;
1523 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
1524 Record *R = ToVec[i];
1525 std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
1527 // If this unconditionally matches, remember it for later and diagnose
1529 if (FeatureMask.empty()) {
1530 if (AliasWithNoPredicate != -1) {
1531 // We can't have two aliases from the same mnemonic with no predicate.
1532 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
1533 "two MnemonicAliases with the same 'from' mnemonic!");
1534 throw TGError(R->getLoc(), "this is the other MnemonicAlias.");
1537 AliasWithNoPredicate = i;
1541 if (!MatchCode.empty())
1542 MatchCode += "else ";
1543 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
1544 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
1547 if (AliasWithNoPredicate != -1) {
1548 Record *R = ToVec[AliasWithNoPredicate];
1549 if (!MatchCode.empty())
1550 MatchCode += "else\n ";
1551 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
1554 MatchCode += "return;";
1556 Cases.push_back(std::make_pair(I->first, MatchCode));
1560 StringMatcher("Mnemonic", Cases, OS).Emit();
1566 void AsmMatcherEmitter::run(raw_ostream &OS) {
1567 CodeGenTarget Target;
1568 Record *AsmParser = Target.getAsmParser();
1569 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1571 // Compute the information on the instructions to match.
1572 AsmMatcherInfo Info(AsmParser, Target);
1575 // Sort the instruction table using the partial order on classes. We use
1576 // stable_sort to ensure that ambiguous instructions are still
1577 // deterministically ordered.
1578 std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
1579 less_ptr<MatchableInfo>());
1581 DEBUG_WITH_TYPE("instruction_info", {
1582 for (std::vector<MatchableInfo*>::iterator
1583 it = Info.Matchables.begin(), ie = Info.Matchables.end();
1588 // Check for ambiguous matchables.
1589 DEBUG_WITH_TYPE("ambiguous_instrs", {
1590 unsigned NumAmbiguous = 0;
1591 for (unsigned i = 0, e = Info.Matchables.size(); i != e; ++i) {
1592 for (unsigned j = i + 1; j != e; ++j) {
1593 MatchableInfo &A = *Info.Matchables[i];
1594 MatchableInfo &B = *Info.Matchables[j];
1596 if (A.CouldMatchAmiguouslyWith(B)) {
1597 errs() << "warning: ambiguous matchables:\n";
1599 errs() << "\nis incomparable with:\n";
1607 errs() << "warning: " << NumAmbiguous
1608 << " ambiguous matchables!\n";
1611 // Write the output.
1613 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1615 // Information for the class declaration.
1616 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
1617 OS << "#undef GET_ASSEMBLER_HEADER\n";
1618 OS << " // This should be included into the middle of the declaration of \n";
1619 OS << " // your subclasses implementation of TargetAsmParser.\n";
1620 OS << " unsigned ComputeAvailableFeatures(const " <<
1621 Target.getName() << "Subtarget *Subtarget) const;\n";
1622 OS << " enum MatchResultTy {\n";
1623 OS << " Match_Success, Match_MnemonicFail, Match_InvalidOperand,\n";
1624 OS << " Match_MissingFeature\n";
1626 OS << " MatchResultTy MatchInstructionImpl(const "
1627 << "SmallVectorImpl<MCParsedAsmOperand*>"
1628 << " &Operands, MCInst &Inst, unsigned &ErrorInfo);\n\n";
1629 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
1634 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
1635 OS << "#undef GET_REGISTER_MATCHER\n\n";
1637 // Emit the subtarget feature enumeration.
1638 EmitSubtargetFeatureFlagEnumeration(Info, OS);
1640 // Emit the function to match a register name to number.
1641 EmitMatchRegisterName(Target, AsmParser, OS);
1643 OS << "#endif // GET_REGISTER_MATCHER\n\n";
1646 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
1647 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
1649 // Generate the function that remaps for mnemonic aliases.
1650 bool HasMnemonicAliases = EmitMnemonicAliases(OS, Info);
1652 // Generate the unified function to convert operands into an MCInst.
1653 EmitConvertToMCInst(Target, Info.Matchables, OS);
1655 // Emit the enumeration for classes which participate in matching.
1656 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1658 // Emit the routine to match token strings to their match class.
1659 EmitMatchTokenString(Target, Info.Classes, OS);
1661 // Emit the routine to classify an operand.
1662 EmitClassifyOperand(Info, OS);
1664 // Emit the subclass predicate routine.
1665 EmitIsSubclass(Target, Info.Classes, OS);
1667 // Emit the available features compute function.
1668 EmitComputeAvailableFeatures(Info, OS);
1671 size_t MaxNumOperands = 0;
1672 for (std::vector<MatchableInfo*>::const_iterator it =
1673 Info.Matchables.begin(), ie = Info.Matchables.end();
1675 MaxNumOperands = std::max(MaxNumOperands, (*it)->AsmOperands.size());
1678 // Emit the static match table; unused classes get initalized to 0 which is
1679 // guaranteed to be InvalidMatchClass.
1681 // FIXME: We can reduce the size of this table very easily. First, we change
1682 // it so that store the kinds in separate bit-fields for each index, which
1683 // only needs to be the max width used for classes at that index (we also need
1684 // to reject based on this during classification). If we then make sure to
1685 // order the match kinds appropriately (putting mnemonics last), then we
1686 // should only end up using a few bits for each class, especially the ones
1687 // following the mnemonic.
1688 OS << "namespace {\n";
1689 OS << " struct MatchEntry {\n";
1690 OS << " unsigned Opcode;\n";
1691 OS << " const char *Mnemonic;\n";
1692 OS << " ConversionKind ConvertFn;\n";
1693 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1694 OS << " unsigned RequiredFeatures;\n";
1697 OS << "// Predicate for searching for an opcode.\n";
1698 OS << " struct LessOpcode {\n";
1699 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
1700 OS << " return StringRef(LHS.Mnemonic) < RHS;\n";
1702 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
1703 OS << " return LHS < StringRef(RHS.Mnemonic);\n";
1705 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
1706 OS << " return StringRef(LHS.Mnemonic) < StringRef(RHS.Mnemonic);\n";
1710 OS << "} // end anonymous namespace.\n\n";
1712 OS << "static const MatchEntry MatchTable["
1713 << Info.Matchables.size() << "] = {\n";
1715 for (std::vector<MatchableInfo*>::const_iterator it =
1716 Info.Matchables.begin(), ie = Info.Matchables.end();
1718 MatchableInfo &II = **it;
1720 OS << " { " << Target.getName() << "::" << II.InstrName
1721 << ", \"" << II.Mnemonic << "\""
1722 << ", " << II.ConversionFnKind << ", { ";
1723 for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
1724 MatchableInfo::Operand &Op = II.AsmOperands[i];
1727 OS << Op.Class->Name;
1731 // Write the required features mask.
1732 if (!II.RequiredFeatures.empty()) {
1733 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
1735 OS << II.RequiredFeatures[i]->getEnumName();
1745 // Finally, build the match function.
1746 OS << Target.getName() << ClassName << "::MatchResultTy "
1747 << Target.getName() << ClassName << "::\n"
1748 << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>"
1750 OS << " MCInst &Inst, unsigned &ErrorInfo) {\n";
1752 // Emit code to get the available features.
1753 OS << " // Get the current feature set.\n";
1754 OS << " unsigned AvailableFeatures = getAvailableFeatures();\n\n";
1756 OS << " // Get the instruction mnemonic, which is the first token.\n";
1757 OS << " StringRef Mnemonic = ((" << Target.getName()
1758 << "Operand*)Operands[0])->getToken();\n\n";
1760 if (HasMnemonicAliases) {
1761 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
1762 OS << " ApplyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n";
1765 // Emit code to compute the class list for this operand vector.
1766 OS << " // Eliminate obvious mismatches.\n";
1767 OS << " if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
1768 OS << " ErrorInfo = " << (MaxNumOperands+1) << ";\n";
1769 OS << " return Match_InvalidOperand;\n";
1772 OS << " // Compute the class list for this operand vector.\n";
1773 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1774 OS << " for (unsigned i = 1, e = Operands.size(); i != e; ++i) {\n";
1775 OS << " Classes[i-1] = ClassifyOperand(Operands[i]);\n\n";
1777 OS << " // Check for invalid operands before matching.\n";
1778 OS << " if (Classes[i-1] == InvalidMatchClass) {\n";
1779 OS << " ErrorInfo = i;\n";
1780 OS << " return Match_InvalidOperand;\n";
1784 OS << " // Mark unused classes.\n";
1785 OS << " for (unsigned i = Operands.size()-1, e = " << MaxNumOperands << "; "
1786 << "i != e; ++i)\n";
1787 OS << " Classes[i] = InvalidMatchClass;\n\n";
1789 OS << " // Some state to try to produce better error messages.\n";
1790 OS << " bool HadMatchOtherThanFeatures = false;\n\n";
1791 OS << " // Set ErrorInfo to the operand that mismatches if it is \n";
1792 OS << " // wrong for all instances of the instruction.\n";
1793 OS << " ErrorInfo = ~0U;\n";
1795 // Emit code to search the table.
1796 OS << " // Search the table.\n";
1797 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
1798 OS << " std::equal_range(MatchTable, MatchTable+"
1799 << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n\n";
1801 OS << " // Return a more specific error code if no mnemonics match.\n";
1802 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
1803 OS << " return Match_MnemonicFail;\n\n";
1805 OS << " for (const MatchEntry *it = MnemonicRange.first, "
1806 << "*ie = MnemonicRange.second;\n";
1807 OS << " it != ie; ++it) {\n";
1809 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
1810 OS << " assert(Mnemonic == it->Mnemonic);\n";
1812 // Emit check that the subclasses match.
1813 OS << " bool OperandsValid = true;\n";
1814 OS << " for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
1815 OS << " if (IsSubclass(Classes[i], it->Classes[i]))\n";
1816 OS << " continue;\n";
1817 OS << " // If this operand is broken for all of the instances of this\n";
1818 OS << " // mnemonic, keep track of it so we can report loc info.\n";
1819 OS << " if (it == MnemonicRange.first || ErrorInfo == i+1)\n";
1820 OS << " ErrorInfo = i+1;\n";
1822 OS << " ErrorInfo = ~0U;";
1823 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
1824 OS << " OperandsValid = false;\n";
1828 OS << " if (!OperandsValid) continue;\n";
1830 // Emit check that the required features are available.
1831 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
1832 << "!= it->RequiredFeatures) {\n";
1833 OS << " HadMatchOtherThanFeatures = true;\n";
1834 OS << " continue;\n";
1838 OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
1840 // Call the post-processing function, if used.
1841 std::string InsnCleanupFn =
1842 AsmParser->getValueAsString("AsmParserInstCleanup");
1843 if (!InsnCleanupFn.empty())
1844 OS << " " << InsnCleanupFn << "(Inst);\n";
1846 OS << " return Match_Success;\n";
1849 OS << " // Okay, we had no match. Try to return a useful error code.\n";
1850 OS << " if (HadMatchOtherThanFeatures) return Match_MissingFeature;\n";
1851 OS << " return Match_InvalidOperand;\n";
1854 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";