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"));
96 /// TokenizeAsmString - Tokenize a simplified assembly string.
97 static void TokenizeAsmString(StringRef AsmString,
98 SmallVectorImpl<StringRef> &Tokens) {
101 for (unsigned i = 0, e = AsmString.size(); i != e; ++i) {
102 switch (AsmString[i]) {
111 Tokens.push_back(AsmString.slice(Prev, i));
114 if (!isspace(AsmString[i]) && AsmString[i] != ',')
115 Tokens.push_back(AsmString.substr(i, 1));
121 Tokens.push_back(AsmString.slice(Prev, i));
125 assert(i != AsmString.size() && "Invalid quoted character");
126 Tokens.push_back(AsmString.substr(i, 1));
131 // If this isn't "${", treat like a normal token.
132 if (i + 1 == AsmString.size() || AsmString[i + 1] != '{') {
134 Tokens.push_back(AsmString.slice(Prev, i));
142 Tokens.push_back(AsmString.slice(Prev, i));
146 StringRef::iterator End =
147 std::find(AsmString.begin() + i, AsmString.end(), '}');
148 assert(End != AsmString.end() && "Missing brace in operand reference!");
149 size_t EndPos = End - AsmString.begin();
150 Tokens.push_back(AsmString.slice(i, EndPos+1));
158 Tokens.push_back(AsmString.slice(Prev, i));
168 if (InTok && Prev != AsmString.size())
169 Tokens.push_back(AsmString.substr(Prev));
172 static bool IsAssemblerInstruction(StringRef Name,
173 const CodeGenInstruction &CGI,
174 const SmallVectorImpl<StringRef> &Tokens) {
175 // Ignore "codegen only" instructions.
176 if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
179 // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
181 // FIXME: This is a total hack.
182 if (StringRef(Name).startswith("Int_") || StringRef(Name).endswith("_Int"))
185 // Reject instructions with no .s string.
186 if (CGI.AsmString.empty()) {
187 PrintError(CGI.TheDef->getLoc(),
188 "instruction with empty asm string");
189 throw std::string("ERROR: Invalid instruction for asm matcher");
192 // Reject any instructions with a newline in them, they should be marked
193 // isCodeGenOnly if they are pseudo instructions.
194 if (CGI.AsmString.find('\n') != std::string::npos) {
195 PrintError(CGI.TheDef->getLoc(),
196 "multiline instruction is not valid for the asmparser, "
197 "mark it isCodeGenOnly");
198 throw std::string("ERROR: Invalid instruction");
201 // Reject instructions with attributes, these aren't something we can handle,
202 // the target should be refactored to use operands instead of modifiers.
204 // Also, check for instructions which reference the operand multiple times;
205 // this implies a constraint we would not honor.
206 std::set<std::string> OperandNames;
207 for (unsigned i = 1, e = Tokens.size(); i < e; ++i) {
208 if (Tokens[i][0] == '$' &&
209 Tokens[i].find(':') != StringRef::npos) {
210 PrintError(CGI.TheDef->getLoc(),
211 "instruction with operand modifier '" + Tokens[i].str() +
212 "' not supported by asm matcher. Mark isCodeGenOnly!");
213 throw std::string("ERROR: Invalid instruction");
216 // FIXME: Should reject these. The ARM backend hits this with $lane in a
217 // bunch of instructions. It is unclear what the right answer is for this.
218 if (Tokens[i][0] == '$' && !OperandNames.insert(Tokens[i]).second) {
220 errs() << "warning: '" << Name << "': "
221 << "ignoring instruction with tied operand '"
222 << Tokens[i].str() << "'\n";
232 class AsmMatcherInfo;
233 struct SubtargetFeatureInfo;
235 /// ClassInfo - Helper class for storing the information about a particular
236 /// class of operands which can be matched.
239 /// Invalid kind, for use as a sentinel value.
242 /// The class for a particular token.
245 /// The (first) register class, subsequent register classes are
246 /// RegisterClass0+1, and so on.
249 /// The (first) user defined class, subsequent user defined classes are
250 /// UserClass0+1, and so on.
254 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
255 /// N) for the Nth user defined class.
258 /// SuperClasses - The super classes of this class. Note that for simplicities
259 /// sake user operands only record their immediate super class, while register
260 /// operands include all superclasses.
261 std::vector<ClassInfo*> SuperClasses;
263 /// Name - The full class name, suitable for use in an enum.
266 /// ClassName - The unadorned generic name for this class (e.g., Token).
267 std::string ClassName;
269 /// ValueName - The name of the value this class represents; for a token this
270 /// is the literal token string, for an operand it is the TableGen class (or
271 /// empty if this is a derived class).
272 std::string ValueName;
274 /// PredicateMethod - The name of the operand method to test whether the
275 /// operand matches this class; this is not valid for Token or register kinds.
276 std::string PredicateMethod;
278 /// RenderMethod - The name of the operand method to add this operand to an
279 /// MCInst; this is not valid for Token or register kinds.
280 std::string RenderMethod;
282 /// For register classes, the records for all the registers in this class.
283 std::set<Record*> Registers;
286 /// isRegisterClass() - Check if this is a register class.
287 bool isRegisterClass() const {
288 return Kind >= RegisterClass0 && Kind < UserClass0;
291 /// isUserClass() - Check if this is a user defined class.
292 bool isUserClass() const {
293 return Kind >= UserClass0;
296 /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
297 /// are related if they are in the same class hierarchy.
298 bool isRelatedTo(const ClassInfo &RHS) const {
299 // Tokens are only related to tokens.
300 if (Kind == Token || RHS.Kind == Token)
301 return Kind == Token && RHS.Kind == Token;
303 // Registers classes are only related to registers classes, and only if
304 // their intersection is non-empty.
305 if (isRegisterClass() || RHS.isRegisterClass()) {
306 if (!isRegisterClass() || !RHS.isRegisterClass())
309 std::set<Record*> Tmp;
310 std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
311 std::set_intersection(Registers.begin(), Registers.end(),
312 RHS.Registers.begin(), RHS.Registers.end(),
318 // Otherwise we have two users operands; they are related if they are in the
319 // same class hierarchy.
321 // FIXME: This is an oversimplification, they should only be related if they
322 // intersect, however we don't have that information.
323 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
324 const ClassInfo *Root = this;
325 while (!Root->SuperClasses.empty())
326 Root = Root->SuperClasses.front();
328 const ClassInfo *RHSRoot = &RHS;
329 while (!RHSRoot->SuperClasses.empty())
330 RHSRoot = RHSRoot->SuperClasses.front();
332 return Root == RHSRoot;
335 /// isSubsetOf - Test whether this class is a subset of \arg RHS;
336 bool isSubsetOf(const ClassInfo &RHS) const {
337 // This is a subset of RHS if it is the same class...
341 // ... or if any of its super classes are a subset of RHS.
342 for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
343 ie = SuperClasses.end(); it != ie; ++it)
344 if ((*it)->isSubsetOf(RHS))
350 /// operator< - Compare two classes.
351 bool operator<(const ClassInfo &RHS) const {
355 // Unrelated classes can be ordered by kind.
356 if (!isRelatedTo(RHS))
357 return Kind < RHS.Kind;
361 assert(0 && "Invalid kind!");
363 // Tokens are comparable by value.
365 // FIXME: Compare by enum value.
366 return ValueName < RHS.ValueName;
369 // This class preceeds the RHS if it is a proper subset of the RHS.
372 if (RHS.isSubsetOf(*this))
375 // Otherwise, order by name to ensure we have a total ordering.
376 return ValueName < RHS.ValueName;
381 /// InstructionInfo - Helper class for storing the necessary information for an
382 /// instruction which is capable of being matched.
383 struct InstructionInfo {
385 /// The unique class instance this operand should match.
388 /// The original operand this corresponds to, if any.
389 const CGIOperandList::OperandInfo *OperandInfo;
392 /// InstrName - The target name for this instruction.
393 std::string InstrName;
395 /// Instr - The instruction this matches.
396 const CodeGenInstruction *Instr;
398 /// AsmString - The assembly string for this instruction (with variants
400 std::string AsmString;
402 /// Tokens - The tokenized assembly pattern that this instruction matches.
403 SmallVector<StringRef, 4> Tokens;
405 /// Operands - The operands that this instruction matches.
406 SmallVector<Operand, 4> Operands;
408 /// Predicates - The required subtarget features to match this instruction.
409 SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures;
411 /// ConversionFnKind - The enum value which is passed to the generated
412 /// ConvertToMCInst to convert parsed operands into an MCInst for this
414 std::string ConversionFnKind;
416 /// getSingletonRegisterForToken - If the specified token is a singleton
417 /// register, return the Record for it, otherwise return null.
418 Record *getSingletonRegisterForToken(unsigned i,
419 const AsmMatcherInfo &Info) const;
421 /// operator< - Compare two instructions.
422 bool operator<(const InstructionInfo &RHS) const {
423 // The primary comparator is the instruction mnemonic.
424 if (Tokens[0] != RHS.Tokens[0])
425 return Tokens[0] < RHS.Tokens[0];
427 if (Operands.size() != RHS.Operands.size())
428 return Operands.size() < RHS.Operands.size();
430 // Compare lexicographically by operand. The matcher validates that other
431 // orderings wouldn't be ambiguous using \see CouldMatchAmiguouslyWith().
432 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
433 if (*Operands[i].Class < *RHS.Operands[i].Class)
435 if (*RHS.Operands[i].Class < *Operands[i].Class)
442 /// CouldMatchAmiguouslyWith - Check whether this instruction could
443 /// ambiguously match the same set of operands as \arg RHS (without being a
444 /// strictly superior match).
445 bool CouldMatchAmiguouslyWith(const InstructionInfo &RHS) {
446 // The number of operands is unambiguous.
447 if (Operands.size() != RHS.Operands.size())
450 // Otherwise, make sure the ordering of the two instructions is unambiguous
451 // by checking that either (a) a token or operand kind discriminates them,
452 // or (b) the ordering among equivalent kinds is consistent.
454 // Tokens and operand kinds are unambiguous (assuming a correct target
456 for (unsigned i = 0, e = Operands.size(); i != e; ++i)
457 if (Operands[i].Class->Kind != RHS.Operands[i].Class->Kind ||
458 Operands[i].Class->Kind == ClassInfo::Token)
459 if (*Operands[i].Class < *RHS.Operands[i].Class ||
460 *RHS.Operands[i].Class < *Operands[i].Class)
463 // Otherwise, this operand could commute if all operands are equivalent, or
464 // there is a pair of operands that compare less than and a pair that
465 // compare greater than.
466 bool HasLT = false, HasGT = false;
467 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
468 if (*Operands[i].Class < *RHS.Operands[i].Class)
470 if (*RHS.Operands[i].Class < *Operands[i].Class)
474 return !(HasLT ^ HasGT);
480 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
481 /// feature which participates in instruction matching.
482 struct SubtargetFeatureInfo {
483 /// \brief The predicate record for this feature.
486 /// \brief An unique index assigned to represent this feature.
489 SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {}
491 /// \brief The name of the enumerated constant identifying this feature.
492 std::string getEnumName() const {
493 return "Feature_" + TheDef->getName();
497 class AsmMatcherInfo {
499 /// The tablegen AsmParser record.
502 /// Target - The target information.
503 CodeGenTarget &Target;
505 /// The AsmParser "CommentDelimiter" value.
506 std::string CommentDelimiter;
508 /// The AsmParser "RegisterPrefix" value.
509 std::string RegisterPrefix;
511 /// The classes which are needed for matching.
512 std::vector<ClassInfo*> Classes;
514 /// The information on the instruction to match.
515 std::vector<InstructionInfo*> Instructions;
517 /// Map of Register records to their class information.
518 std::map<Record*, ClassInfo*> RegisterClasses;
520 /// Map of Predicate records to their subtarget information.
521 std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures;
524 /// Map of token to class information which has already been constructed.
525 std::map<std::string, ClassInfo*> TokenClasses;
527 /// Map of RegisterClass records to their class information.
528 std::map<Record*, ClassInfo*> RegisterClassClasses;
530 /// Map of AsmOperandClass records to their class information.
531 std::map<Record*, ClassInfo*> AsmOperandClasses;
534 /// getTokenClass - Lookup or create the class for the given token.
535 ClassInfo *getTokenClass(StringRef Token);
537 /// getOperandClass - Lookup or create the class for the given operand.
538 ClassInfo *getOperandClass(StringRef Token,
539 const CGIOperandList::OperandInfo &OI);
541 /// BuildRegisterClasses - Build the ClassInfo* instances for register
543 void BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters);
545 /// BuildOperandClasses - Build the ClassInfo* instances for user defined
547 void BuildOperandClasses();
550 AsmMatcherInfo(Record *AsmParser, CodeGenTarget &Target);
552 /// BuildInfo - Construct the various tables used during matching.
555 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
557 SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
558 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
559 std::map<Record*, SubtargetFeatureInfo*>::const_iterator I =
560 SubtargetFeatures.find(Def);
561 return I == SubtargetFeatures.end() ? 0 : I->second;
567 void InstructionInfo::dump() {
568 errs() << InstrName << " -- " << "flattened:\"" << AsmString << '\"'
570 for (unsigned i = 0, e = Tokens.size(); i != e; ++i) {
577 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
578 Operand &Op = Operands[i];
579 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
580 if (Op.Class->Kind == ClassInfo::Token) {
581 errs() << '\"' << Tokens[i] << "\"\n";
585 if (!Op.OperandInfo) {
586 errs() << "(singleton register)\n";
590 const CGIOperandList::OperandInfo &OI = *Op.OperandInfo;
591 errs() << OI.Name << " " << OI.Rec->getName()
592 << " (" << OI.MIOperandNo << ", " << OI.MINumOperands << ")\n";
596 /// getRegisterRecord - Get the register record for \arg name, or 0.
597 static Record *getRegisterRecord(CodeGenTarget &Target, StringRef Name) {
598 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
599 const CodeGenRegister &Reg = Target.getRegisters()[i];
600 if (Name == Reg.TheDef->getValueAsString("AsmName"))
607 /// getSingletonRegisterForToken - If the specified token is a singleton
608 /// register, return the register name, otherwise return a null StringRef.
609 Record *InstructionInfo::
610 getSingletonRegisterForToken(unsigned i, const AsmMatcherInfo &Info) const {
611 StringRef Tok = Tokens[i];
612 if (!Tok.startswith(Info.RegisterPrefix))
615 StringRef RegName = Tok.substr(Info.RegisterPrefix.size());
616 if (Record *Rec = getRegisterRecord(Info.Target, RegName))
619 // If there is no register prefix (i.e. "%" in "%eax"), then this may
620 // be some random non-register token, just ignore it.
621 if (Info.RegisterPrefix.empty())
624 std::string Err = "unable to find register for '" + RegName.str() +
625 "' (which matches register prefix)";
626 throw TGError(Instr->TheDef->getLoc(), Err);
630 static std::string getEnumNameForToken(StringRef Str) {
633 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
635 case '*': Res += "_STAR_"; break;
636 case '%': Res += "_PCT_"; break;
637 case ':': Res += "_COLON_"; break;
642 Res += "_" + utostr((unsigned) *it) + "_";
649 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
650 ClassInfo *&Entry = TokenClasses[Token];
653 Entry = new ClassInfo();
654 Entry->Kind = ClassInfo::Token;
655 Entry->ClassName = "Token";
656 Entry->Name = "MCK_" + getEnumNameForToken(Token);
657 Entry->ValueName = Token;
658 Entry->PredicateMethod = "<invalid>";
659 Entry->RenderMethod = "<invalid>";
660 Classes.push_back(Entry);
667 AsmMatcherInfo::getOperandClass(StringRef Token,
668 const CGIOperandList::OperandInfo &OI) {
669 if (OI.Rec->isSubClassOf("RegisterClass")) {
670 ClassInfo *CI = RegisterClassClasses[OI.Rec];
673 PrintError(OI.Rec->getLoc(), "register class has no class info!");
674 throw std::string("ERROR: Missing register class!");
680 assert(OI.Rec->isSubClassOf("Operand") && "Unexpected operand!");
681 Record *MatchClass = OI.Rec->getValueAsDef("ParserMatchClass");
682 ClassInfo *CI = AsmOperandClasses[MatchClass];
685 PrintError(OI.Rec->getLoc(), "operand has no match class!");
686 throw std::string("ERROR: Missing match class!");
692 void AsmMatcherInfo::
693 BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters) {
694 std::vector<CodeGenRegisterClass> RegisterClasses;
695 std::vector<CodeGenRegister> Registers;
697 RegisterClasses = Target.getRegisterClasses();
698 Registers = Target.getRegisters();
700 // The register sets used for matching.
701 std::set< std::set<Record*> > RegisterSets;
703 // Gather the defined sets.
704 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
705 ie = RegisterClasses.end(); it != ie; ++it)
706 RegisterSets.insert(std::set<Record*>(it->Elements.begin(),
707 it->Elements.end()));
709 // Add any required singleton sets.
710 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
711 ie = SingletonRegisters.end(); it != ie; ++it) {
713 RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
716 // Introduce derived sets where necessary (when a register does not determine
717 // a unique register set class), and build the mapping of registers to the set
718 // they should classify to.
719 std::map<Record*, std::set<Record*> > RegisterMap;
720 for (std::vector<CodeGenRegister>::iterator it = Registers.begin(),
721 ie = Registers.end(); it != ie; ++it) {
722 CodeGenRegister &CGR = *it;
723 // Compute the intersection of all sets containing this register.
724 std::set<Record*> ContainingSet;
726 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
727 ie = RegisterSets.end(); it != ie; ++it) {
728 if (!it->count(CGR.TheDef))
731 if (ContainingSet.empty()) {
734 std::set<Record*> Tmp;
735 std::swap(Tmp, ContainingSet);
736 std::insert_iterator< std::set<Record*> > II(ContainingSet,
737 ContainingSet.begin());
738 std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(),
743 if (!ContainingSet.empty()) {
744 RegisterSets.insert(ContainingSet);
745 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
749 // Construct the register classes.
750 std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
752 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
753 ie = RegisterSets.end(); it != ie; ++it, ++Index) {
754 ClassInfo *CI = new ClassInfo();
755 CI->Kind = ClassInfo::RegisterClass0 + Index;
756 CI->ClassName = "Reg" + utostr(Index);
757 CI->Name = "MCK_Reg" + utostr(Index);
759 CI->PredicateMethod = ""; // unused
760 CI->RenderMethod = "addRegOperands";
762 Classes.push_back(CI);
763 RegisterSetClasses.insert(std::make_pair(*it, CI));
766 // Find the superclasses; we could compute only the subgroup lattice edges,
767 // but there isn't really a point.
768 for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
769 ie = RegisterSets.end(); it != ie; ++it) {
770 ClassInfo *CI = RegisterSetClasses[*it];
771 for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
772 ie2 = RegisterSets.end(); it2 != ie2; ++it2)
774 std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
775 CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
778 // Name the register classes which correspond to a user defined RegisterClass.
779 for (std::vector<CodeGenRegisterClass>::iterator it = RegisterClasses.begin(),
780 ie = RegisterClasses.end(); it != ie; ++it) {
781 ClassInfo *CI = RegisterSetClasses[std::set<Record*>(it->Elements.begin(),
782 it->Elements.end())];
783 if (CI->ValueName.empty()) {
784 CI->ClassName = it->getName();
785 CI->Name = "MCK_" + it->getName();
786 CI->ValueName = it->getName();
788 CI->ValueName = CI->ValueName + "," + it->getName();
790 RegisterClassClasses.insert(std::make_pair(it->TheDef, CI));
793 // Populate the map for individual registers.
794 for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
795 ie = RegisterMap.end(); it != ie; ++it)
796 this->RegisterClasses[it->first] = RegisterSetClasses[it->second];
798 // Name the register classes which correspond to singleton registers.
799 for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
800 ie = SingletonRegisters.end(); it != ie; ++it) {
802 ClassInfo *CI = this->RegisterClasses[Rec];
803 assert(CI && "Missing singleton register class info!");
805 if (CI->ValueName.empty()) {
806 CI->ClassName = Rec->getName();
807 CI->Name = "MCK_" + Rec->getName();
808 CI->ValueName = Rec->getName();
810 CI->ValueName = CI->ValueName + "," + Rec->getName();
814 void AsmMatcherInfo::BuildOperandClasses() {
815 std::vector<Record*> AsmOperands;
816 AsmOperands = Records.getAllDerivedDefinitions("AsmOperandClass");
818 // Pre-populate AsmOperandClasses map.
819 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
820 ie = AsmOperands.end(); it != ie; ++it)
821 AsmOperandClasses[*it] = new ClassInfo();
824 for (std::vector<Record*>::iterator it = AsmOperands.begin(),
825 ie = AsmOperands.end(); it != ie; ++it, ++Index) {
826 ClassInfo *CI = AsmOperandClasses[*it];
827 CI->Kind = ClassInfo::UserClass0 + Index;
829 ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
830 for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
831 DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
833 PrintError((*it)->getLoc(), "Invalid super class reference!");
837 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
839 PrintError((*it)->getLoc(), "Invalid super class reference!");
841 CI->SuperClasses.push_back(SC);
843 CI->ClassName = (*it)->getValueAsString("Name");
844 CI->Name = "MCK_" + CI->ClassName;
845 CI->ValueName = (*it)->getName();
847 // Get or construct the predicate method name.
848 Init *PMName = (*it)->getValueInit("PredicateMethod");
849 if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
850 CI->PredicateMethod = SI->getValue();
852 assert(dynamic_cast<UnsetInit*>(PMName) &&
853 "Unexpected PredicateMethod field!");
854 CI->PredicateMethod = "is" + CI->ClassName;
857 // Get or construct the render method name.
858 Init *RMName = (*it)->getValueInit("RenderMethod");
859 if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
860 CI->RenderMethod = SI->getValue();
862 assert(dynamic_cast<UnsetInit*>(RMName) &&
863 "Unexpected RenderMethod field!");
864 CI->RenderMethod = "add" + CI->ClassName + "Operands";
867 AsmOperandClasses[*it] = CI;
868 Classes.push_back(CI);
872 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, CodeGenTarget &target)
873 : AsmParser(asmParser), Target(target),
874 CommentDelimiter(AsmParser->getValueAsString("CommentDelimiter")),
875 RegisterPrefix(AsmParser->getValueAsString("RegisterPrefix"))
879 void AsmMatcherInfo::BuildInfo() {
880 // Build information about all of the AssemblerPredicates.
881 std::vector<Record*> AllPredicates =
882 Records.getAllDerivedDefinitions("Predicate");
883 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
884 Record *Pred = AllPredicates[i];
885 // Ignore predicates that are not intended for the assembler.
886 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
889 if (Pred->getName().empty()) {
890 PrintError(Pred->getLoc(), "Predicate has no name!");
891 throw std::string("ERROR: Predicate defs must be named");
894 unsigned FeatureNo = SubtargetFeatures.size();
895 SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo);
896 assert(FeatureNo < 32 && "Too many subtarget features!");
899 // Parse the instructions; we need to do this first so that we can gather the
900 // singleton register classes.
901 SmallPtrSet<Record*, 16> SingletonRegisters;
902 for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
903 E = Target.inst_end(); I != E; ++I) {
904 const CodeGenInstruction &CGI = **I;
906 // If the tblgen -match-prefix option is specified (for tblgen hackers),
907 // filter the set of instructions we consider.
908 if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
911 OwningPtr<InstructionInfo> II(new InstructionInfo());
913 II->InstrName = CGI.TheDef->getName();
915 // TODO: Eventually support asmparser for Variant != 0.
916 II->AsmString = CGI.FlattenAsmStringVariants(CGI.AsmString, 0);
918 // Remove comments from the asm string. We know that the asmstring only
920 if (!CommentDelimiter.empty()) {
921 size_t Idx = StringRef(II->AsmString).find(CommentDelimiter);
922 if (Idx != StringRef::npos)
923 II->AsmString = II->AsmString.substr(0, Idx);
926 TokenizeAsmString(II->AsmString, II->Tokens);
928 // Ignore instructions which shouldn't be matched and diagnose invalid
929 // instruction definitions with an error.
930 if (!IsAssemblerInstruction(CGI.TheDef->getName(), CGI, II->Tokens))
933 // Collect singleton registers, if used.
934 for (unsigned i = 0, e = II->Tokens.size(); i != e; ++i) {
935 if (Record *Reg = II->getSingletonRegisterForToken(i, *this))
936 SingletonRegisters.insert(Reg);
939 // Compute the require features.
940 std::vector<Record*> Predicates =
941 CGI.TheDef->getValueAsListOfDefs("Predicates");
942 for (unsigned i = 0, e = Predicates.size(); i != e; ++i)
943 if (SubtargetFeatureInfo *Feature = getSubtargetFeature(Predicates[i]))
944 II->RequiredFeatures.push_back(Feature);
946 Instructions.push_back(II.take());
950 // Build info for the register classes.
951 BuildRegisterClasses(SingletonRegisters);
953 // Build info for the user defined assembly operand classes.
954 BuildOperandClasses();
956 // Build the instruction information.
957 for (std::vector<InstructionInfo*>::iterator it = Instructions.begin(),
958 ie = Instructions.end(); it != ie; ++it) {
959 InstructionInfo *II = *it;
961 // The first token of the instruction is the mnemonic, which must be a
962 // simple string, not a $foo variable or a singleton register.
963 assert(!II->Tokens.empty() && "Instruction has no tokens?");
964 StringRef Mnemonic = II->Tokens[0];
965 if (Mnemonic[0] == '$' || II->getSingletonRegisterForToken(0, *this))
966 throw TGError(II->Instr->TheDef->getLoc(),
967 "Invalid instruction mnemonic '" + Mnemonic.str() + "'!");
969 // Parse the tokens after the mnemonic.
970 for (unsigned i = 1, e = II->Tokens.size(); i != e; ++i) {
971 StringRef Token = II->Tokens[i];
973 // Check for singleton registers.
974 if (Record *RegRecord = II->getSingletonRegisterForToken(i, *this)) {
975 InstructionInfo::Operand Op;
976 Op.Class = RegisterClasses[RegRecord];
978 assert(Op.Class && Op.Class->Registers.size() == 1 &&
979 "Unexpected class for singleton register");
980 II->Operands.push_back(Op);
984 // Check for simple tokens.
985 if (Token[0] != '$') {
986 InstructionInfo::Operand Op;
987 Op.Class = getTokenClass(Token);
989 II->Operands.push_back(Op);
993 // Otherwise this is an operand reference.
994 StringRef OperandName;
996 OperandName = Token.substr(2, Token.size() - 3);
998 OperandName = Token.substr(1);
1000 // Map this token to an operand. FIXME: Move elsewhere.
1002 if (!II->Instr->Operands.hasOperandNamed(OperandName, Idx))
1003 throw std::string("error: unable to find operand: '" +
1004 OperandName.str() + "'");
1006 // FIXME: This is annoying, the named operand may be tied (e.g.,
1007 // XCHG8rm). What we want is the untied operand, which we now have to
1008 // grovel for. Only worry about this for single entry operands, we have to
1009 // clean this up anyway.
1010 const CGIOperandList::OperandInfo *OI = &II->Instr->Operands[Idx];
1011 if (OI->Constraints[0].isTied()) {
1012 unsigned TiedOp = OI->Constraints[0].getTiedOperand();
1014 // The tied operand index is an MIOperand index, find the operand that
1016 for (unsigned i = 0, e = II->Instr->Operands.size(); i != e; ++i) {
1017 if (II->Instr->Operands[i].MIOperandNo == TiedOp) {
1018 OI = &II->Instr->Operands[i];
1023 assert(OI && "Unable to find tied operand target!");
1026 InstructionInfo::Operand Op;
1027 Op.Class = getOperandClass(Token, *OI);
1028 Op.OperandInfo = OI;
1029 II->Operands.push_back(Op);
1033 // Reorder classes so that classes preceed super classes.
1034 std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
1037 static std::pair<unsigned, unsigned> *
1038 GetTiedOperandAtIndex(SmallVectorImpl<std::pair<unsigned, unsigned> > &List,
1040 for (unsigned i = 0, e = List.size(); i != e; ++i)
1041 if (Index == List[i].first)
1047 static void EmitConvertToMCInst(CodeGenTarget &Target,
1048 std::vector<InstructionInfo*> &Infos,
1050 // Write the convert function to a separate stream, so we can drop it after
1052 std::string ConvertFnBody;
1053 raw_string_ostream CvtOS(ConvertFnBody);
1055 // Function we have already generated.
1056 std::set<std::string> GeneratedFns;
1058 // Start the unified conversion function.
1060 CvtOS << "static void ConvertToMCInst(ConversionKind Kind, MCInst &Inst, "
1061 << "unsigned Opcode,\n"
1062 << " const SmallVectorImpl<MCParsedAsmOperand*"
1063 << "> &Operands) {\n";
1064 CvtOS << " Inst.setOpcode(Opcode);\n";
1065 CvtOS << " switch (Kind) {\n";
1066 CvtOS << " default:\n";
1068 // Start the enum, which we will generate inline.
1070 OS << "// Unified function for converting operants to MCInst instances.\n\n";
1071 OS << "enum ConversionKind {\n";
1073 // TargetOperandClass - This is the target's operand class, like X86Operand.
1074 std::string TargetOperandClass = Target.getName() + "Operand";
1076 for (std::vector<InstructionInfo*>::const_iterator it = Infos.begin(),
1077 ie = Infos.end(); it != ie; ++it) {
1078 InstructionInfo &II = **it;
1080 // Order the (class) operands by the order to convert them into an MCInst.
1081 SmallVector<std::pair<unsigned, unsigned>, 4> MIOperandList;
1082 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1083 InstructionInfo::Operand &Op = II.Operands[i];
1085 MIOperandList.push_back(std::make_pair(Op.OperandInfo->MIOperandNo, i));
1088 // Find any tied operands.
1089 SmallVector<std::pair<unsigned, unsigned>, 4> TiedOperands;
1090 for (unsigned i = 0, e = II.Instr->Operands.size(); i != e; ++i) {
1091 const CGIOperandList::OperandInfo &OpInfo = II.Instr->Operands[i];
1092 for (unsigned j = 0, e = OpInfo.Constraints.size(); j != e; ++j) {
1093 const CGIOperandList::ConstraintInfo &CI = OpInfo.Constraints[j];
1095 TiedOperands.push_back(std::make_pair(OpInfo.MIOperandNo + j,
1096 CI.getTiedOperand()));
1100 std::sort(MIOperandList.begin(), MIOperandList.end());
1102 // Compute the total number of operands.
1103 unsigned NumMIOperands = 0;
1104 for (unsigned i = 0, e = II.Instr->Operands.size(); i != e; ++i) {
1105 const CGIOperandList::OperandInfo &OI = II.Instr->Operands[i];
1106 NumMIOperands = std::max(NumMIOperands,
1107 OI.MIOperandNo + OI.MINumOperands);
1110 // Build the conversion function signature.
1111 std::string Signature = "Convert";
1112 unsigned CurIndex = 0;
1113 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1114 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1115 assert(CurIndex <= Op.OperandInfo->MIOperandNo &&
1116 "Duplicate match for instruction operand!");
1118 // Skip operands which weren't matched by anything, this occurs when the
1119 // .td file encodes "implicit" operands as explicit ones.
1121 // FIXME: This should be removed from the MCInst structure.
1122 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1123 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1126 Signature += "__Imp";
1128 Signature += "__Tie" + utostr(Tie->second);
1133 // Registers are always converted the same, don't duplicate the conversion
1134 // function based on them.
1136 // FIXME: We could generalize this based on the render method, if it
1138 if (Op.Class->isRegisterClass())
1141 Signature += Op.Class->ClassName;
1142 Signature += utostr(Op.OperandInfo->MINumOperands);
1143 Signature += "_" + utostr(MIOperandList[i].second);
1145 CurIndex += Op.OperandInfo->MINumOperands;
1148 // Add any trailing implicit operands.
1149 for (; CurIndex != NumMIOperands; ++CurIndex) {
1150 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1153 Signature += "__Imp";
1155 Signature += "__Tie" + utostr(Tie->second);
1158 II.ConversionFnKind = Signature;
1160 // Check if we have already generated this signature.
1161 if (!GeneratedFns.insert(Signature).second)
1164 // If not, emit it now.
1166 // Add to the enum list.
1167 OS << " " << Signature << ",\n";
1169 // And to the convert function.
1170 CvtOS << " case " << Signature << ":\n";
1172 for (unsigned i = 0, e = MIOperandList.size(); i != e; ++i) {
1173 InstructionInfo::Operand &Op = II.Operands[MIOperandList[i].second];
1175 // Add the implicit operands.
1176 for (; CurIndex != Op.OperandInfo->MIOperandNo; ++CurIndex) {
1177 // See if this is a tied operand.
1178 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1182 // If not, this is some implicit operand. Just assume it is a register
1184 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1186 // Copy the tied operand.
1187 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1188 CvtOS << " Inst.addOperand(Inst.getOperand("
1189 << Tie->second << "));\n";
1193 CvtOS << " ((" << TargetOperandClass << "*)Operands["
1194 << MIOperandList[i].second
1195 << "+1])->" << Op.Class->RenderMethod
1196 << "(Inst, " << Op.OperandInfo->MINumOperands << ");\n";
1197 CurIndex += Op.OperandInfo->MINumOperands;
1200 // And add trailing implicit operands.
1201 for (; CurIndex != NumMIOperands; ++CurIndex) {
1202 std::pair<unsigned, unsigned> *Tie = GetTiedOperandAtIndex(TiedOperands,
1206 // If not, this is some implicit operand. Just assume it is a register
1208 CvtOS << " Inst.addOperand(MCOperand::CreateReg(0));\n";
1210 // Copy the tied operand.
1211 assert(Tie->first>Tie->second && "Tied operand preceeds its target!");
1212 CvtOS << " Inst.addOperand(Inst.getOperand("
1213 << Tie->second << "));\n";
1217 CvtOS << " return;\n";
1220 // Finish the convert function.
1225 // Finish the enum, and drop the convert function after it.
1227 OS << " NumConversionVariants\n";
1233 /// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
1234 static void EmitMatchClassEnumeration(CodeGenTarget &Target,
1235 std::vector<ClassInfo*> &Infos,
1237 OS << "namespace {\n\n";
1239 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1240 << "/// instruction matching.\n";
1241 OS << "enum MatchClassKind {\n";
1242 OS << " InvalidMatchClass = 0,\n";
1243 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1244 ie = Infos.end(); it != ie; ++it) {
1245 ClassInfo &CI = **it;
1246 OS << " " << CI.Name << ", // ";
1247 if (CI.Kind == ClassInfo::Token) {
1248 OS << "'" << CI.ValueName << "'\n";
1249 } else if (CI.isRegisterClass()) {
1250 if (!CI.ValueName.empty())
1251 OS << "register class '" << CI.ValueName << "'\n";
1253 OS << "derived register class\n";
1255 OS << "user defined class '" << CI.ValueName << "'\n";
1258 OS << " NumMatchClassKinds\n";
1264 /// EmitClassifyOperand - Emit the function to classify an operand.
1265 static void EmitClassifyOperand(AsmMatcherInfo &Info,
1267 OS << "static MatchClassKind ClassifyOperand(MCParsedAsmOperand *GOp) {\n"
1268 << " " << Info.Target.getName() << "Operand &Operand = *("
1269 << Info.Target.getName() << "Operand*)GOp;\n";
1272 OS << " if (Operand.isToken())\n";
1273 OS << " return MatchTokenString(Operand.getToken());\n\n";
1275 // Classify registers.
1277 // FIXME: Don't hardcode isReg, getReg.
1278 OS << " if (Operand.isReg()) {\n";
1279 OS << " switch (Operand.getReg()) {\n";
1280 OS << " default: return InvalidMatchClass;\n";
1281 for (std::map<Record*, ClassInfo*>::iterator
1282 it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1284 OS << " case " << Info.Target.getName() << "::"
1285 << it->first->getName() << ": return " << it->second->Name << ";\n";
1289 // Classify user defined operands.
1290 for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1291 ie = Info.Classes.end(); it != ie; ++it) {
1292 ClassInfo &CI = **it;
1294 if (!CI.isUserClass())
1297 OS << " // '" << CI.ClassName << "' class";
1298 if (!CI.SuperClasses.empty()) {
1299 OS << ", subclass of ";
1300 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i) {
1302 OS << "'" << CI.SuperClasses[i]->ClassName << "'";
1303 assert(CI < *CI.SuperClasses[i] && "Invalid class relation!");
1308 OS << " if (Operand." << CI.PredicateMethod << "()) {\n";
1310 // Validate subclass relationships.
1311 if (!CI.SuperClasses.empty()) {
1312 for (unsigned i = 0, e = CI.SuperClasses.size(); i != e; ++i)
1313 OS << " assert(Operand." << CI.SuperClasses[i]->PredicateMethod
1314 << "() && \"Invalid class relationship!\");\n";
1317 OS << " return " << CI.Name << ";\n";
1320 OS << " return InvalidMatchClass;\n";
1324 /// EmitIsSubclass - Emit the subclass predicate function.
1325 static void EmitIsSubclass(CodeGenTarget &Target,
1326 std::vector<ClassInfo*> &Infos,
1328 OS << "/// IsSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1329 OS << "static bool IsSubclass(MatchClassKind A, MatchClassKind B) {\n";
1330 OS << " if (A == B)\n";
1331 OS << " return true;\n\n";
1333 OS << " switch (A) {\n";
1334 OS << " default:\n";
1335 OS << " return false;\n";
1336 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1337 ie = Infos.end(); it != ie; ++it) {
1338 ClassInfo &A = **it;
1340 if (A.Kind != ClassInfo::Token) {
1341 std::vector<StringRef> SuperClasses;
1342 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1343 ie = Infos.end(); it != ie; ++it) {
1344 ClassInfo &B = **it;
1346 if (&A != &B && A.isSubsetOf(B))
1347 SuperClasses.push_back(B.Name);
1350 if (SuperClasses.empty())
1353 OS << "\n case " << A.Name << ":\n";
1355 if (SuperClasses.size() == 1) {
1356 OS << " return B == " << SuperClasses.back() << ";\n";
1360 OS << " switch (B) {\n";
1361 OS << " default: return false;\n";
1362 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1363 OS << " case " << SuperClasses[i] << ": return true;\n";
1373 /// EmitMatchTokenString - Emit the function to match a token string to the
1374 /// appropriate match class value.
1375 static void EmitMatchTokenString(CodeGenTarget &Target,
1376 std::vector<ClassInfo*> &Infos,
1378 // Construct the match list.
1379 std::vector<StringMatcher::StringPair> Matches;
1380 for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1381 ie = Infos.end(); it != ie; ++it) {
1382 ClassInfo &CI = **it;
1384 if (CI.Kind == ClassInfo::Token)
1385 Matches.push_back(StringMatcher::StringPair(CI.ValueName,
1386 "return " + CI.Name + ";"));
1389 OS << "static MatchClassKind MatchTokenString(StringRef Name) {\n";
1391 StringMatcher("Name", Matches, OS).Emit();
1393 OS << " return InvalidMatchClass;\n";
1397 /// EmitMatchRegisterName - Emit the function to match a string to the target
1398 /// specific register enum.
1399 static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1401 // Construct the match list.
1402 std::vector<StringMatcher::StringPair> Matches;
1403 for (unsigned i = 0, e = Target.getRegisters().size(); i != e; ++i) {
1404 const CodeGenRegister &Reg = Target.getRegisters()[i];
1405 if (Reg.TheDef->getValueAsString("AsmName").empty())
1408 Matches.push_back(StringMatcher::StringPair(
1409 Reg.TheDef->getValueAsString("AsmName"),
1410 "return " + utostr(i + 1) + ";"));
1413 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1415 StringMatcher("Name", Matches, OS).Emit();
1417 OS << " return 0;\n";
1421 /// EmitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
1423 static void EmitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
1425 OS << "// Flags for subtarget features that participate in "
1426 << "instruction matching.\n";
1427 OS << "enum SubtargetFeatureFlag {\n";
1428 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1429 it = Info.SubtargetFeatures.begin(),
1430 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1431 SubtargetFeatureInfo &SFI = *it->second;
1432 OS << " " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n";
1434 OS << " Feature_None = 0\n";
1438 /// EmitComputeAvailableFeatures - Emit the function to compute the list of
1439 /// available features given a subtarget.
1440 static void EmitComputeAvailableFeatures(AsmMatcherInfo &Info,
1442 std::string ClassName =
1443 Info.AsmParser->getValueAsString("AsmParserClassName");
1445 OS << "unsigned " << Info.Target.getName() << ClassName << "::\n"
1446 << "ComputeAvailableFeatures(const " << Info.Target.getName()
1447 << "Subtarget *Subtarget) const {\n";
1448 OS << " unsigned Features = 0;\n";
1449 for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1450 it = Info.SubtargetFeatures.begin(),
1451 ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1452 SubtargetFeatureInfo &SFI = *it->second;
1453 OS << " if (" << SFI.TheDef->getValueAsString("CondString")
1455 OS << " Features |= " << SFI.getEnumName() << ";\n";
1457 OS << " return Features;\n";
1461 static std::string GetAliasRequiredFeatures(Record *R,
1462 const AsmMatcherInfo &Info) {
1463 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
1465 unsigned NumFeatures = 0;
1466 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
1467 SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
1470 throw TGError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
1471 "' is not marked as an AssemblerPredicate!");
1476 Result += F->getEnumName();
1480 if (NumFeatures > 1)
1481 Result = '(' + Result + ')';
1485 /// EmitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
1486 /// emit a function for them and return true, otherwise return false.
1487 static bool EmitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) {
1488 std::vector<Record*> Aliases =
1489 Records.getAllDerivedDefinitions("MnemonicAlias");
1490 if (Aliases.empty()) return false;
1492 OS << "static void ApplyMnemonicAliases(StringRef &Mnemonic, "
1493 "unsigned Features) {\n";
1495 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
1496 // iteration order of the map is stable.
1497 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
1499 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
1500 Record *R = Aliases[i];
1501 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
1504 // Process each alias a "from" mnemonic at a time, building the code executed
1505 // by the string remapper.
1506 std::vector<StringMatcher::StringPair> Cases;
1507 for (std::map<std::string, std::vector<Record*> >::iterator
1508 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
1510 const std::vector<Record*> &ToVec = I->second;
1512 // Loop through each alias and emit code that handles each case. If there
1513 // are two instructions without predicates, emit an error. If there is one,
1515 std::string MatchCode;
1516 int AliasWithNoPredicate = -1;
1518 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
1519 Record *R = ToVec[i];
1520 std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
1522 // If this unconditionally matches, remember it for later and diagnose
1524 if (FeatureMask.empty()) {
1525 if (AliasWithNoPredicate != -1) {
1526 // We can't have two aliases from the same mnemonic with no predicate.
1527 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
1528 "two MnemonicAliases with the same 'from' mnemonic!");
1529 PrintError(R->getLoc(), "this is the other MnemonicAlias.");
1530 throw std::string("ERROR: Invalid MnemonicAlias definitions!");
1533 AliasWithNoPredicate = i;
1537 if (!MatchCode.empty())
1538 MatchCode += "else ";
1539 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
1540 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
1543 if (AliasWithNoPredicate != -1) {
1544 Record *R = ToVec[AliasWithNoPredicate];
1545 if (!MatchCode.empty())
1546 MatchCode += "else\n ";
1547 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
1550 MatchCode += "return;";
1552 Cases.push_back(std::make_pair(I->first, MatchCode));
1556 StringMatcher("Mnemonic", Cases, OS).Emit();
1562 void AsmMatcherEmitter::run(raw_ostream &OS) {
1563 CodeGenTarget Target;
1564 Record *AsmParser = Target.getAsmParser();
1565 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
1567 // Compute the information on the instructions to match.
1568 AsmMatcherInfo Info(AsmParser, Target);
1571 // Sort the instruction table using the partial order on classes. We use
1572 // stable_sort to ensure that ambiguous instructions are still
1573 // deterministically ordered.
1574 std::stable_sort(Info.Instructions.begin(), Info.Instructions.end(),
1575 less_ptr<InstructionInfo>());
1577 DEBUG_WITH_TYPE("instruction_info", {
1578 for (std::vector<InstructionInfo*>::iterator
1579 it = Info.Instructions.begin(), ie = Info.Instructions.end();
1584 // Check for ambiguous instructions.
1585 DEBUG_WITH_TYPE("ambiguous_instrs", {
1586 unsigned NumAmbiguous = 0;
1587 for (unsigned i = 0, e = Info.Instructions.size(); i != e; ++i) {
1588 for (unsigned j = i + 1; j != e; ++j) {
1589 InstructionInfo &A = *Info.Instructions[i];
1590 InstructionInfo &B = *Info.Instructions[j];
1592 if (A.CouldMatchAmiguouslyWith(B)) {
1593 errs() << "warning: ambiguous instruction match:\n";
1595 errs() << "\nis incomparable with:\n";
1603 errs() << "warning: " << NumAmbiguous
1604 << " ambiguous instructions!\n";
1607 // Write the output.
1609 EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
1611 // Information for the class declaration.
1612 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
1613 OS << "#undef GET_ASSEMBLER_HEADER\n";
1614 OS << " // This should be included into the middle of the declaration of \n";
1615 OS << " // your subclasses implementation of TargetAsmParser.\n";
1616 OS << " unsigned ComputeAvailableFeatures(const " <<
1617 Target.getName() << "Subtarget *Subtarget) const;\n";
1618 OS << " enum MatchResultTy {\n";
1619 OS << " Match_Success, Match_MnemonicFail, Match_InvalidOperand,\n";
1620 OS << " Match_MissingFeature\n";
1622 OS << " MatchResultTy MatchInstructionImpl(const "
1623 << "SmallVectorImpl<MCParsedAsmOperand*>"
1624 << " &Operands, MCInst &Inst, unsigned &ErrorInfo);\n\n";
1625 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
1630 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
1631 OS << "#undef GET_REGISTER_MATCHER\n\n";
1633 // Emit the subtarget feature enumeration.
1634 EmitSubtargetFeatureFlagEnumeration(Info, OS);
1636 // Emit the function to match a register name to number.
1637 EmitMatchRegisterName(Target, AsmParser, OS);
1639 OS << "#endif // GET_REGISTER_MATCHER\n\n";
1642 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
1643 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
1645 // Generate the function that remaps for mnemonic aliases.
1646 bool HasMnemonicAliases = EmitMnemonicAliases(OS, Info);
1648 // Generate the unified function to convert operands into an MCInst.
1649 EmitConvertToMCInst(Target, Info.Instructions, OS);
1651 // Emit the enumeration for classes which participate in matching.
1652 EmitMatchClassEnumeration(Target, Info.Classes, OS);
1654 // Emit the routine to match token strings to their match class.
1655 EmitMatchTokenString(Target, Info.Classes, OS);
1657 // Emit the routine to classify an operand.
1658 EmitClassifyOperand(Info, OS);
1660 // Emit the subclass predicate routine.
1661 EmitIsSubclass(Target, Info.Classes, OS);
1663 // Emit the available features compute function.
1664 EmitComputeAvailableFeatures(Info, OS);
1667 size_t MaxNumOperands = 0;
1668 for (std::vector<InstructionInfo*>::const_iterator it =
1669 Info.Instructions.begin(), ie = Info.Instructions.end();
1671 MaxNumOperands = std::max(MaxNumOperands, (*it)->Operands.size());
1674 // Emit the static match table; unused classes get initalized to 0 which is
1675 // guaranteed to be InvalidMatchClass.
1677 // FIXME: We can reduce the size of this table very easily. First, we change
1678 // it so that store the kinds in separate bit-fields for each index, which
1679 // only needs to be the max width used for classes at that index (we also need
1680 // to reject based on this during classification). If we then make sure to
1681 // order the match kinds appropriately (putting mnemonics last), then we
1682 // should only end up using a few bits for each class, especially the ones
1683 // following the mnemonic.
1684 OS << "namespace {\n";
1685 OS << " struct MatchEntry {\n";
1686 OS << " unsigned Opcode;\n";
1687 OS << " const char *Mnemonic;\n";
1688 OS << " ConversionKind ConvertFn;\n";
1689 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1690 OS << " unsigned RequiredFeatures;\n";
1693 OS << "// Predicate for searching for an opcode.\n";
1694 OS << " struct LessOpcode {\n";
1695 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
1696 OS << " return StringRef(LHS.Mnemonic) < RHS;\n";
1698 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
1699 OS << " return LHS < StringRef(RHS.Mnemonic);\n";
1701 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
1702 OS << " return StringRef(LHS.Mnemonic) < StringRef(RHS.Mnemonic);\n";
1706 OS << "} // end anonymous namespace.\n\n";
1708 OS << "static const MatchEntry MatchTable["
1709 << Info.Instructions.size() << "] = {\n";
1711 for (std::vector<InstructionInfo*>::const_iterator it =
1712 Info.Instructions.begin(), ie = Info.Instructions.end();
1714 InstructionInfo &II = **it;
1716 OS << " { " << Target.getName() << "::" << II.InstrName
1717 << ", \"" << II.Tokens[0] << "\""
1718 << ", " << II.ConversionFnKind << ", { ";
1719 for (unsigned i = 0, e = II.Operands.size(); i != e; ++i) {
1720 InstructionInfo::Operand &Op = II.Operands[i];
1723 OS << Op.Class->Name;
1727 // Write the required features mask.
1728 if (!II.RequiredFeatures.empty()) {
1729 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
1731 OS << II.RequiredFeatures[i]->getEnumName();
1741 // Finally, build the match function.
1742 OS << Target.getName() << ClassName << "::MatchResultTy "
1743 << Target.getName() << ClassName << "::\n"
1744 << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>"
1746 OS << " MCInst &Inst, unsigned &ErrorInfo) {\n";
1748 // Emit code to get the available features.
1749 OS << " // Get the current feature set.\n";
1750 OS << " unsigned AvailableFeatures = getAvailableFeatures();\n\n";
1752 OS << " // Get the instruction mnemonic, which is the first token.\n";
1753 OS << " StringRef Mnemonic = ((" << Target.getName()
1754 << "Operand*)Operands[0])->getToken();\n\n";
1756 if (HasMnemonicAliases) {
1757 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
1758 OS << " ApplyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n";
1761 // Emit code to compute the class list for this operand vector.
1762 OS << " // Eliminate obvious mismatches.\n";
1763 OS << " if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
1764 OS << " ErrorInfo = " << (MaxNumOperands+1) << ";\n";
1765 OS << " return Match_InvalidOperand;\n";
1768 OS << " // Compute the class list for this operand vector.\n";
1769 OS << " MatchClassKind Classes[" << MaxNumOperands << "];\n";
1770 OS << " for (unsigned i = 1, e = Operands.size(); i != e; ++i) {\n";
1771 OS << " Classes[i-1] = ClassifyOperand(Operands[i]);\n\n";
1773 OS << " // Check for invalid operands before matching.\n";
1774 OS << " if (Classes[i-1] == InvalidMatchClass) {\n";
1775 OS << " ErrorInfo = i;\n";
1776 OS << " return Match_InvalidOperand;\n";
1780 OS << " // Mark unused classes.\n";
1781 OS << " for (unsigned i = Operands.size()-1, e = " << MaxNumOperands << "; "
1782 << "i != e; ++i)\n";
1783 OS << " Classes[i] = InvalidMatchClass;\n\n";
1785 OS << " // Some state to try to produce better error messages.\n";
1786 OS << " bool HadMatchOtherThanFeatures = false;\n\n";
1787 OS << " // Set ErrorInfo to the operand that mismatches if it is \n";
1788 OS << " // wrong for all instances of the instruction.\n";
1789 OS << " ErrorInfo = ~0U;\n";
1791 // Emit code to search the table.
1792 OS << " // Search the table.\n";
1793 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
1794 OS << " std::equal_range(MatchTable, MatchTable+"
1795 << Info.Instructions.size() << ", Mnemonic, LessOpcode());\n\n";
1797 OS << " // Return a more specific error code if no mnemonics match.\n";
1798 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
1799 OS << " return Match_MnemonicFail;\n\n";
1801 OS << " for (const MatchEntry *it = MnemonicRange.first, "
1802 << "*ie = MnemonicRange.second;\n";
1803 OS << " it != ie; ++it) {\n";
1805 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
1806 OS << " assert(Mnemonic == it->Mnemonic);\n";
1808 // Emit check that the subclasses match.
1809 OS << " bool OperandsValid = true;\n";
1810 OS << " for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
1811 OS << " if (IsSubclass(Classes[i], it->Classes[i]))\n";
1812 OS << " continue;\n";
1813 OS << " // If this operand is broken for all of the instances of this\n";
1814 OS << " // mnemonic, keep track of it so we can report loc info.\n";
1815 OS << " if (it == MnemonicRange.first || ErrorInfo == i+1)\n";
1816 OS << " ErrorInfo = i+1;\n";
1818 OS << " ErrorInfo = ~0U;";
1819 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
1820 OS << " OperandsValid = false;\n";
1824 OS << " if (!OperandsValid) continue;\n";
1826 // Emit check that the required features are available.
1827 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
1828 << "!= it->RequiredFeatures) {\n";
1829 OS << " HadMatchOtherThanFeatures = true;\n";
1830 OS << " continue;\n";
1834 OS << " ConvertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
1836 // Call the post-processing function, if used.
1837 std::string InsnCleanupFn =
1838 AsmParser->getValueAsString("AsmParserInstCleanup");
1839 if (!InsnCleanupFn.empty())
1840 OS << " " << InsnCleanupFn << "(Inst);\n";
1842 OS << " return Match_Success;\n";
1845 OS << " // Okay, we had no match. Try to return a useful error code.\n";
1846 OS << " if (HadMatchOtherThanFeatures) return Match_MissingFeature;\n";
1847 OS << " return Match_InvalidOperand;\n";
1850 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";