AsmMatcherEmitter.cpp revision 67cd20d0396a68e61858736b7943dd352f1b0d0b
1//===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This tablegen backend emits a target specifier matcher for converting parsed
11// assembly operands in the MCInst structures. It also emits a matcher for
12// custom operand parsing.
13//
14// Converting assembly operands into MCInst structures
15// ---------------------------------------------------
16//
17// The input to the target specific matcher is a list of literal tokens and
18// operands. The target specific parser should generally eliminate any syntax
19// which is not relevant for matching; for example, comma tokens should have
20// already been consumed and eliminated by the parser. Most instructions will
21// end up with a single literal token (the instruction name) and some number of
22// operands.
23//
24// Some example inputs, for X86:
25//   'addl' (immediate ...) (register ...)
26//   'add' (immediate ...) (memory ...)
27//   'call' '*' %epc
28//
29// The assembly matcher is responsible for converting this input into a precise
30// machine instruction (i.e., an instruction with a well defined encoding). This
31// mapping has several properties which complicate matching:
32//
33//  - It may be ambiguous; many architectures can legally encode particular
34//    variants of an instruction in different ways (for example, using a smaller
35//    encoding for small immediates). Such ambiguities should never be
36//    arbitrarily resolved by the assembler, the assembler is always responsible
37//    for choosing the "best" available instruction.
38//
39//  - It may depend on the subtarget or the assembler context. Instructions
40//    which are invalid for the current mode, but otherwise unambiguous (e.g.,
41//    an SSE instruction in a file being assembled for i486) should be accepted
42//    and rejected by the assembler front end. However, if the proper encoding
43//    for an instruction is dependent on the assembler context then the matcher
44//    is responsible for selecting the correct machine instruction for the
45//    current mode.
46//
47// The core matching algorithm attempts to exploit the regularity in most
48// instruction sets to quickly determine the set of possibly matching
49// instructions, and the simplify the generated code. Additionally, this helps
50// to ensure that the ambiguities are intentionally resolved by the user.
51//
52// The matching is divided into two distinct phases:
53//
54//   1. Classification: Each operand is mapped to the unique set which (a)
55//      contains it, and (b) is the largest such subset for which a single
56//      instruction could match all members.
57//
58//      For register classes, we can generate these subgroups automatically. For
59//      arbitrary operands, we expect the user to define the classes and their
60//      relations to one another (for example, 8-bit signed immediates as a
61//      subset of 32-bit immediates).
62//
63//      By partitioning the operands in this way, we guarantee that for any
64//      tuple of classes, any single instruction must match either all or none
65//      of the sets of operands which could classify to that tuple.
66//
67//      In addition, the subset relation amongst classes induces a partial order
68//      on such tuples, which we use to resolve ambiguities.
69//
70//   2. The input can now be treated as a tuple of classes (static tokens are
71//      simple singleton sets). Each such tuple should generally map to a single
72//      instruction (we currently ignore cases where this isn't true, whee!!!),
73//      which we can emit a simple matcher for.
74//
75// Custom Operand Parsing
76// ----------------------
77//
78//  Some targets need a custom way to parse operands, some specific instructions
79//  can contain arguments that can represent processor flags and other kinds of
80//  identifiers that need to be mapped to specific valeus in the final encoded
81//  instructions. The target specific custom operand parsing works in the
82//  following way:
83//
84//   1. A operand match table is built, each entry contains a mnemonic, an
85//      operand class, a mask for all operand positions for that same
86//      class/mnemonic and target features to be checked while trying to match.
87//
88//   2. The operand matcher will try every possible entry with the same
89//      mnemonic and will check if the target feature for this mnemonic also
90//      matches. After that, if the operand to be matched has its index
91//      present in the mask, a successful match occurs. Otherwise, fallback
92//      to the regular operand parsing.
93//
94//   3. For a match success, each operand class that has a 'ParserMethod'
95//      becomes part of a switch from where the custom method is called.
96//
97//===----------------------------------------------------------------------===//
98
99#include "AsmMatcherEmitter.h"
100#include "CodeGenTarget.h"
101#include "StringMatcher.h"
102#include "StringToOffsetTable.h"
103#include "llvm/ADT/OwningPtr.h"
104#include "llvm/ADT/PointerUnion.h"
105#include "llvm/ADT/SmallPtrSet.h"
106#include "llvm/ADT/SmallVector.h"
107#include "llvm/ADT/STLExtras.h"
108#include "llvm/ADT/StringExtras.h"
109#include "llvm/Support/CommandLine.h"
110#include "llvm/Support/Debug.h"
111#include "llvm/Support/ErrorHandling.h"
112#include "llvm/TableGen/Error.h"
113#include "llvm/TableGen/Record.h"
114#include <map>
115#include <set>
116using namespace llvm;
117
118static cl::opt<std::string>
119MatchPrefix("match-prefix", cl::init(""),
120            cl::desc("Only match instructions with the given prefix"));
121
122namespace {
123class AsmMatcherInfo;
124struct SubtargetFeatureInfo;
125
126/// ClassInfo - Helper class for storing the information about a particular
127/// class of operands which can be matched.
128struct ClassInfo {
129  enum ClassInfoKind {
130    /// Invalid kind, for use as a sentinel value.
131    Invalid = 0,
132
133    /// The class for a particular token.
134    Token,
135
136    /// The (first) register class, subsequent register classes are
137    /// RegisterClass0+1, and so on.
138    RegisterClass0,
139
140    /// The (first) user defined class, subsequent user defined classes are
141    /// UserClass0+1, and so on.
142    UserClass0 = 1<<16
143  };
144
145  /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
146  /// N) for the Nth user defined class.
147  unsigned Kind;
148
149  /// SuperClasses - The super classes of this class. Note that for simplicities
150  /// sake user operands only record their immediate super class, while register
151  /// operands include all superclasses.
152  std::vector<ClassInfo*> SuperClasses;
153
154  /// Name - The full class name, suitable for use in an enum.
155  std::string Name;
156
157  /// ClassName - The unadorned generic name for this class (e.g., Token).
158  std::string ClassName;
159
160  /// ValueName - The name of the value this class represents; for a token this
161  /// is the literal token string, for an operand it is the TableGen class (or
162  /// empty if this is a derived class).
163  std::string ValueName;
164
165  /// PredicateMethod - The name of the operand method to test whether the
166  /// operand matches this class; this is not valid for Token or register kinds.
167  std::string PredicateMethod;
168
169  /// RenderMethod - The name of the operand method to add this operand to an
170  /// MCInst; this is not valid for Token or register kinds.
171  std::string RenderMethod;
172
173  /// ParserMethod - The name of the operand method to do a target specific
174  /// parsing on the operand.
175  std::string ParserMethod;
176
177  /// For register classes, the records for all the registers in this class.
178  std::set<Record*> Registers;
179
180public:
181  /// isRegisterClass() - Check if this is a register class.
182  bool isRegisterClass() const {
183    return Kind >= RegisterClass0 && Kind < UserClass0;
184  }
185
186  /// isUserClass() - Check if this is a user defined class.
187  bool isUserClass() const {
188    return Kind >= UserClass0;
189  }
190
191  /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
192  /// are related if they are in the same class hierarchy.
193  bool isRelatedTo(const ClassInfo &RHS) const {
194    // Tokens are only related to tokens.
195    if (Kind == Token || RHS.Kind == Token)
196      return Kind == Token && RHS.Kind == Token;
197
198    // Registers classes are only related to registers classes, and only if
199    // their intersection is non-empty.
200    if (isRegisterClass() || RHS.isRegisterClass()) {
201      if (!isRegisterClass() || !RHS.isRegisterClass())
202        return false;
203
204      std::set<Record*> Tmp;
205      std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
206      std::set_intersection(Registers.begin(), Registers.end(),
207                            RHS.Registers.begin(), RHS.Registers.end(),
208                            II);
209
210      return !Tmp.empty();
211    }
212
213    // Otherwise we have two users operands; they are related if they are in the
214    // same class hierarchy.
215    //
216    // FIXME: This is an oversimplification, they should only be related if they
217    // intersect, however we don't have that information.
218    assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
219    const ClassInfo *Root = this;
220    while (!Root->SuperClasses.empty())
221      Root = Root->SuperClasses.front();
222
223    const ClassInfo *RHSRoot = &RHS;
224    while (!RHSRoot->SuperClasses.empty())
225      RHSRoot = RHSRoot->SuperClasses.front();
226
227    return Root == RHSRoot;
228  }
229
230  /// isSubsetOf - Test whether this class is a subset of \arg RHS;
231  bool isSubsetOf(const ClassInfo &RHS) const {
232    // This is a subset of RHS if it is the same class...
233    if (this == &RHS)
234      return true;
235
236    // ... or if any of its super classes are a subset of RHS.
237    for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
238           ie = SuperClasses.end(); it != ie; ++it)
239      if ((*it)->isSubsetOf(RHS))
240        return true;
241
242    return false;
243  }
244
245  /// operator< - Compare two classes.
246  bool operator<(const ClassInfo &RHS) const {
247    if (this == &RHS)
248      return false;
249
250    // Unrelated classes can be ordered by kind.
251    if (!isRelatedTo(RHS))
252      return Kind < RHS.Kind;
253
254    switch (Kind) {
255    case Invalid:
256      llvm_unreachable("Invalid kind!");
257
258    default:
259      // This class precedes the RHS if it is a proper subset of the RHS.
260      if (isSubsetOf(RHS))
261        return true;
262      if (RHS.isSubsetOf(*this))
263        return false;
264
265      // Otherwise, order by name to ensure we have a total ordering.
266      return ValueName < RHS.ValueName;
267    }
268  }
269};
270
271/// MatchableInfo - Helper class for storing the necessary information for an
272/// instruction or alias which is capable of being matched.
273struct MatchableInfo {
274  struct AsmOperand {
275    /// Token - This is the token that the operand came from.
276    StringRef Token;
277
278    /// The unique class instance this operand should match.
279    ClassInfo *Class;
280
281    /// The operand name this is, if anything.
282    StringRef SrcOpName;
283
284    /// The suboperand index within SrcOpName, or -1 for the entire operand.
285    int SubOpIdx;
286
287    /// Register record if this token is singleton register.
288    Record *SingletonReg;
289
290    explicit AsmOperand(StringRef T) : Token(T), Class(0), SubOpIdx(-1),
291                                       SingletonReg(0) {}
292  };
293
294  /// ResOperand - This represents a single operand in the result instruction
295  /// generated by the match.  In cases (like addressing modes) where a single
296  /// assembler operand expands to multiple MCOperands, this represents the
297  /// single assembler operand, not the MCOperand.
298  struct ResOperand {
299    enum {
300      /// RenderAsmOperand - This represents an operand result that is
301      /// generated by calling the render method on the assembly operand.  The
302      /// corresponding AsmOperand is specified by AsmOperandNum.
303      RenderAsmOperand,
304
305      /// TiedOperand - This represents a result operand that is a duplicate of
306      /// a previous result operand.
307      TiedOperand,
308
309      /// ImmOperand - This represents an immediate value that is dumped into
310      /// the operand.
311      ImmOperand,
312
313      /// RegOperand - This represents a fixed register that is dumped in.
314      RegOperand
315    } Kind;
316
317    union {
318      /// This is the operand # in the AsmOperands list that this should be
319      /// copied from.
320      unsigned AsmOperandNum;
321
322      /// TiedOperandNum - This is the (earlier) result operand that should be
323      /// copied from.
324      unsigned TiedOperandNum;
325
326      /// ImmVal - This is the immediate value added to the instruction.
327      int64_t ImmVal;
328
329      /// Register - This is the register record.
330      Record *Register;
331    };
332
333    /// MINumOperands - The number of MCInst operands populated by this
334    /// operand.
335    unsigned MINumOperands;
336
337    static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
338      ResOperand X;
339      X.Kind = RenderAsmOperand;
340      X.AsmOperandNum = AsmOpNum;
341      X.MINumOperands = NumOperands;
342      return X;
343    }
344
345    static ResOperand getTiedOp(unsigned TiedOperandNum) {
346      ResOperand X;
347      X.Kind = TiedOperand;
348      X.TiedOperandNum = TiedOperandNum;
349      X.MINumOperands = 1;
350      return X;
351    }
352
353    static ResOperand getImmOp(int64_t Val) {
354      ResOperand X;
355      X.Kind = ImmOperand;
356      X.ImmVal = Val;
357      X.MINumOperands = 1;
358      return X;
359    }
360
361    static ResOperand getRegOp(Record *Reg) {
362      ResOperand X;
363      X.Kind = RegOperand;
364      X.Register = Reg;
365      X.MINumOperands = 1;
366      return X;
367    }
368  };
369
370  /// AsmVariantID - Target's assembly syntax variant no.
371  int AsmVariantID;
372
373  /// TheDef - This is the definition of the instruction or InstAlias that this
374  /// matchable came from.
375  Record *const TheDef;
376
377  /// DefRec - This is the definition that it came from.
378  PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
379
380  const CodeGenInstruction *getResultInst() const {
381    if (DefRec.is<const CodeGenInstruction*>())
382      return DefRec.get<const CodeGenInstruction*>();
383    return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
384  }
385
386  /// ResOperands - This is the operand list that should be built for the result
387  /// MCInst.
388  std::vector<ResOperand> ResOperands;
389
390  /// AsmString - The assembly string for this instruction (with variants
391  /// removed), e.g. "movsx $src, $dst".
392  std::string AsmString;
393
394  /// Mnemonic - This is the first token of the matched instruction, its
395  /// mnemonic.
396  StringRef Mnemonic;
397
398  /// AsmOperands - The textual operands that this instruction matches,
399  /// annotated with a class and where in the OperandList they were defined.
400  /// This directly corresponds to the tokenized AsmString after the mnemonic is
401  /// removed.
402  SmallVector<AsmOperand, 4> AsmOperands;
403
404  /// Predicates - The required subtarget features to match this instruction.
405  SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures;
406
407  /// ConversionFnKind - The enum value which is passed to the generated
408  /// ConvertToMCInst to convert parsed operands into an MCInst for this
409  /// function.
410  std::string ConversionFnKind;
411
412  MatchableInfo(const CodeGenInstruction &CGI)
413    : AsmVariantID(0), TheDef(CGI.TheDef), DefRec(&CGI),
414      AsmString(CGI.AsmString) {
415  }
416
417  MatchableInfo(const CodeGenInstAlias *Alias)
418    : AsmVariantID(0), TheDef(Alias->TheDef), DefRec(Alias),
419      AsmString(Alias->AsmString) {
420  }
421
422  void Initialize(const AsmMatcherInfo &Info,
423                  SmallPtrSet<Record*, 16> &SingletonRegisters,
424                  int AsmVariantNo, std::string &RegisterPrefix);
425
426  /// Validate - Return true if this matchable is a valid thing to match against
427  /// and perform a bunch of validity checking.
428  bool Validate(StringRef CommentDelimiter, bool Hack) const;
429
430  /// extractSingletonRegisterForAsmOperand - Extract singleton register,
431  /// if present, from specified token.
432  void
433  extractSingletonRegisterForAsmOperand(unsigned i, const AsmMatcherInfo &Info,
434                                        std::string &RegisterPrefix);
435
436  /// FindAsmOperand - Find the AsmOperand with the specified name and
437  /// suboperand index.
438  int FindAsmOperand(StringRef N, int SubOpIdx) const {
439    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
440      if (N == AsmOperands[i].SrcOpName &&
441          SubOpIdx == AsmOperands[i].SubOpIdx)
442        return i;
443    return -1;
444  }
445
446  /// FindAsmOperandNamed - Find the first AsmOperand with the specified name.
447  /// This does not check the suboperand index.
448  int FindAsmOperandNamed(StringRef N) const {
449    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
450      if (N == AsmOperands[i].SrcOpName)
451        return i;
452    return -1;
453  }
454
455  void BuildInstructionResultOperands();
456  void BuildAliasResultOperands();
457
458  /// operator< - Compare two matchables.
459  bool operator<(const MatchableInfo &RHS) const {
460    // The primary comparator is the instruction mnemonic.
461    if (Mnemonic != RHS.Mnemonic)
462      return Mnemonic < RHS.Mnemonic;
463
464    if (AsmOperands.size() != RHS.AsmOperands.size())
465      return AsmOperands.size() < RHS.AsmOperands.size();
466
467    // Compare lexicographically by operand. The matcher validates that other
468    // orderings wouldn't be ambiguous using \see CouldMatchAmbiguouslyWith().
469    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
470      if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
471        return true;
472      if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
473        return false;
474    }
475
476    return false;
477  }
478
479  /// CouldMatchAmbiguouslyWith - Check whether this matchable could
480  /// ambiguously match the same set of operands as \arg RHS (without being a
481  /// strictly superior match).
482  bool CouldMatchAmbiguouslyWith(const MatchableInfo &RHS) {
483    // The primary comparator is the instruction mnemonic.
484    if (Mnemonic != RHS.Mnemonic)
485      return false;
486
487    // The number of operands is unambiguous.
488    if (AsmOperands.size() != RHS.AsmOperands.size())
489      return false;
490
491    // Otherwise, make sure the ordering of the two instructions is unambiguous
492    // by checking that either (a) a token or operand kind discriminates them,
493    // or (b) the ordering among equivalent kinds is consistent.
494
495    // Tokens and operand kinds are unambiguous (assuming a correct target
496    // specific parser).
497    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
498      if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
499          AsmOperands[i].Class->Kind == ClassInfo::Token)
500        if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
501            *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
502          return false;
503
504    // Otherwise, this operand could commute if all operands are equivalent, or
505    // there is a pair of operands that compare less than and a pair that
506    // compare greater than.
507    bool HasLT = false, HasGT = false;
508    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
509      if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
510        HasLT = true;
511      if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
512        HasGT = true;
513    }
514
515    return !(HasLT ^ HasGT);
516  }
517
518  void dump();
519
520private:
521  void TokenizeAsmString(const AsmMatcherInfo &Info);
522};
523
524/// SubtargetFeatureInfo - Helper class for storing information on a subtarget
525/// feature which participates in instruction matching.
526struct SubtargetFeatureInfo {
527  /// \brief The predicate record for this feature.
528  Record *TheDef;
529
530  /// \brief An unique index assigned to represent this feature.
531  unsigned Index;
532
533  SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {}
534
535  /// \brief The name of the enumerated constant identifying this feature.
536  std::string getEnumName() const {
537    return "Feature_" + TheDef->getName();
538  }
539};
540
541struct OperandMatchEntry {
542  unsigned OperandMask;
543  MatchableInfo* MI;
544  ClassInfo *CI;
545
546  static OperandMatchEntry Create(MatchableInfo* mi, ClassInfo *ci,
547                                  unsigned opMask) {
548    OperandMatchEntry X;
549    X.OperandMask = opMask;
550    X.CI = ci;
551    X.MI = mi;
552    return X;
553  }
554};
555
556
557class AsmMatcherInfo {
558public:
559  /// Tracked Records
560  RecordKeeper &Records;
561
562  /// The tablegen AsmParser record.
563  Record *AsmParser;
564
565  /// Target - The target information.
566  CodeGenTarget &Target;
567
568  /// The classes which are needed for matching.
569  std::vector<ClassInfo*> Classes;
570
571  /// The information on the matchables to match.
572  std::vector<MatchableInfo*> Matchables;
573
574  /// Info for custom matching operands by user defined methods.
575  std::vector<OperandMatchEntry> OperandMatchInfo;
576
577  /// Map of Register records to their class information.
578  std::map<Record*, ClassInfo*> RegisterClasses;
579
580  /// Map of Predicate records to their subtarget information.
581  std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures;
582
583private:
584  /// Map of token to class information which has already been constructed.
585  std::map<std::string, ClassInfo*> TokenClasses;
586
587  /// Map of RegisterClass records to their class information.
588  std::map<Record*, ClassInfo*> RegisterClassClasses;
589
590  /// Map of AsmOperandClass records to their class information.
591  std::map<Record*, ClassInfo*> AsmOperandClasses;
592
593private:
594  /// getTokenClass - Lookup or create the class for the given token.
595  ClassInfo *getTokenClass(StringRef Token);
596
597  /// getOperandClass - Lookup or create the class for the given operand.
598  ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
599                             int SubOpIdx);
600  ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
601
602  /// BuildRegisterClasses - Build the ClassInfo* instances for register
603  /// classes.
604  void BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters);
605
606  /// BuildOperandClasses - Build the ClassInfo* instances for user defined
607  /// operand classes.
608  void BuildOperandClasses();
609
610  void BuildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
611                                        unsigned AsmOpIdx);
612  void BuildAliasOperandReference(MatchableInfo *II, StringRef OpName,
613                                  MatchableInfo::AsmOperand &Op);
614
615public:
616  AsmMatcherInfo(Record *AsmParser,
617                 CodeGenTarget &Target,
618                 RecordKeeper &Records);
619
620  /// BuildInfo - Construct the various tables used during matching.
621  void BuildInfo();
622
623  /// BuildOperandMatchInfo - Build the necessary information to handle user
624  /// defined operand parsing methods.
625  void BuildOperandMatchInfo();
626
627  /// getSubtargetFeature - Lookup or create the subtarget feature info for the
628  /// given operand.
629  SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
630    assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
631    std::map<Record*, SubtargetFeatureInfo*>::const_iterator I =
632      SubtargetFeatures.find(Def);
633    return I == SubtargetFeatures.end() ? 0 : I->second;
634  }
635
636  RecordKeeper &getRecords() const {
637    return Records;
638  }
639};
640
641}
642
643void MatchableInfo::dump() {
644  errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
645
646  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
647    AsmOperand &Op = AsmOperands[i];
648    errs() << "  op[" << i << "] = " << Op.Class->ClassName << " - ";
649    errs() << '\"' << Op.Token << "\"\n";
650  }
651}
652
653void MatchableInfo::Initialize(const AsmMatcherInfo &Info,
654                               SmallPtrSet<Record*, 16> &SingletonRegisters,
655                               int AsmVariantNo, std::string &RegisterPrefix) {
656  AsmVariantID = AsmVariantNo;
657  AsmString =
658    CodeGenInstruction::FlattenAsmStringVariants(AsmString, AsmVariantNo);
659
660  TokenizeAsmString(Info);
661
662  // Compute the require features.
663  std::vector<Record*> Predicates =TheDef->getValueAsListOfDefs("Predicates");
664  for (unsigned i = 0, e = Predicates.size(); i != e; ++i)
665    if (SubtargetFeatureInfo *Feature =
666        Info.getSubtargetFeature(Predicates[i]))
667      RequiredFeatures.push_back(Feature);
668
669  // Collect singleton registers, if used.
670  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
671    extractSingletonRegisterForAsmOperand(i, Info, RegisterPrefix);
672    if (Record *Reg = AsmOperands[i].SingletonReg)
673      SingletonRegisters.insert(Reg);
674  }
675}
676
677/// TokenizeAsmString - Tokenize a simplified assembly string.
678void MatchableInfo::TokenizeAsmString(const AsmMatcherInfo &Info) {
679  StringRef String = AsmString;
680  unsigned Prev = 0;
681  bool InTok = true;
682  for (unsigned i = 0, e = String.size(); i != e; ++i) {
683    switch (String[i]) {
684    case '[':
685    case ']':
686    case '*':
687    case '!':
688    case ' ':
689    case '\t':
690    case ',':
691      if (InTok) {
692        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
693        InTok = false;
694      }
695      if (!isspace(String[i]) && String[i] != ',')
696        AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
697      Prev = i + 1;
698      break;
699
700    case '\\':
701      if (InTok) {
702        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
703        InTok = false;
704      }
705      ++i;
706      assert(i != String.size() && "Invalid quoted character");
707      AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
708      Prev = i + 1;
709      break;
710
711    case '$': {
712      if (InTok) {
713        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
714        InTok = false;
715      }
716
717      // If this isn't "${", treat like a normal token.
718      if (i + 1 == String.size() || String[i + 1] != '{') {
719        Prev = i;
720        break;
721      }
722
723      StringRef::iterator End = std::find(String.begin() + i, String.end(),'}');
724      assert(End != String.end() && "Missing brace in operand reference!");
725      size_t EndPos = End - String.begin();
726      AsmOperands.push_back(AsmOperand(String.slice(i, EndPos+1)));
727      Prev = EndPos + 1;
728      i = EndPos;
729      break;
730    }
731
732    case '.':
733      if (InTok)
734        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
735      Prev = i;
736      InTok = true;
737      break;
738
739    default:
740      InTok = true;
741    }
742  }
743  if (InTok && Prev != String.size())
744    AsmOperands.push_back(AsmOperand(String.substr(Prev)));
745
746  // The first token of the instruction is the mnemonic, which must be a
747  // simple string, not a $foo variable or a singleton register.
748  if (AsmOperands.empty())
749    throw TGError(TheDef->getLoc(),
750                  "Instruction '" + TheDef->getName() + "' has no tokens");
751  Mnemonic = AsmOperands[0].Token;
752  // FIXME : Check and raise an error if it is a register.
753  if (Mnemonic[0] == '$')
754    throw TGError(TheDef->getLoc(),
755                  "Invalid instruction mnemonic '" + Mnemonic.str() + "'!");
756
757  // Remove the first operand, it is tracked in the mnemonic field.
758  AsmOperands.erase(AsmOperands.begin());
759}
760
761bool MatchableInfo::Validate(StringRef CommentDelimiter, bool Hack) const {
762  // Reject matchables with no .s string.
763  if (AsmString.empty())
764    throw TGError(TheDef->getLoc(), "instruction with empty asm string");
765
766  // Reject any matchables with a newline in them, they should be marked
767  // isCodeGenOnly if they are pseudo instructions.
768  if (AsmString.find('\n') != std::string::npos)
769    throw TGError(TheDef->getLoc(),
770                  "multiline instruction is not valid for the asmparser, "
771                  "mark it isCodeGenOnly");
772
773  // Remove comments from the asm string.  We know that the asmstring only
774  // has one line.
775  if (!CommentDelimiter.empty() &&
776      StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
777    throw TGError(TheDef->getLoc(),
778                  "asmstring for instruction has comment character in it, "
779                  "mark it isCodeGenOnly");
780
781  // Reject matchables with operand modifiers, these aren't something we can
782  // handle, the target should be refactored to use operands instead of
783  // modifiers.
784  //
785  // Also, check for instructions which reference the operand multiple times;
786  // this implies a constraint we would not honor.
787  std::set<std::string> OperandNames;
788  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
789    StringRef Tok = AsmOperands[i].Token;
790    if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
791      throw TGError(TheDef->getLoc(),
792                    "matchable with operand modifier '" + Tok.str() +
793                    "' not supported by asm matcher.  Mark isCodeGenOnly!");
794
795    // Verify that any operand is only mentioned once.
796    // We reject aliases and ignore instructions for now.
797    if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
798      if (!Hack)
799        throw TGError(TheDef->getLoc(),
800                      "ERROR: matchable with tied operand '" + Tok.str() +
801                      "' can never be matched!");
802      // FIXME: Should reject these.  The ARM backend hits this with $lane in a
803      // bunch of instructions.  It is unclear what the right answer is.
804      DEBUG({
805        errs() << "warning: '" << TheDef->getName() << "': "
806               << "ignoring instruction with tied operand '"
807               << Tok.str() << "'\n";
808      });
809      return false;
810    }
811  }
812
813  return true;
814}
815
816/// extractSingletonRegisterForAsmOperand - Extract singleton register,
817/// if present, from specified token.
818void MatchableInfo::
819extractSingletonRegisterForAsmOperand(unsigned OperandNo,
820                                      const AsmMatcherInfo &Info,
821                                      std::string &RegisterPrefix) {
822  StringRef Tok = AsmOperands[OperandNo].Token;
823  if (RegisterPrefix.empty()) {
824    std::string LoweredTok = Tok.lower();
825    if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
826      AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
827    return;
828  }
829
830  if (!Tok.startswith(RegisterPrefix))
831    return;
832
833  StringRef RegName = Tok.substr(RegisterPrefix.size());
834  if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
835    AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
836
837  // If there is no register prefix (i.e. "%" in "%eax"), then this may
838  // be some random non-register token, just ignore it.
839  return;
840}
841
842static std::string getEnumNameForToken(StringRef Str) {
843  std::string Res;
844
845  for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
846    switch (*it) {
847    case '*': Res += "_STAR_"; break;
848    case '%': Res += "_PCT_"; break;
849    case ':': Res += "_COLON_"; break;
850    case '!': Res += "_EXCLAIM_"; break;
851    case '.': Res += "_DOT_"; break;
852    default:
853      if (isalnum(*it))
854        Res += *it;
855      else
856        Res += "_" + utostr((unsigned) *it) + "_";
857    }
858  }
859
860  return Res;
861}
862
863ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
864  ClassInfo *&Entry = TokenClasses[Token];
865
866  if (!Entry) {
867    Entry = new ClassInfo();
868    Entry->Kind = ClassInfo::Token;
869    Entry->ClassName = "Token";
870    Entry->Name = "MCK_" + getEnumNameForToken(Token);
871    Entry->ValueName = Token;
872    Entry->PredicateMethod = "<invalid>";
873    Entry->RenderMethod = "<invalid>";
874    Entry->ParserMethod = "";
875    Classes.push_back(Entry);
876  }
877
878  return Entry;
879}
880
881ClassInfo *
882AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
883                                int SubOpIdx) {
884  Record *Rec = OI.Rec;
885  if (SubOpIdx != -1)
886    Rec = dynamic_cast<DefInit*>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
887  return getOperandClass(Rec, SubOpIdx);
888}
889
890ClassInfo *
891AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
892  if (Rec->isSubClassOf("RegisterOperand")) {
893    // RegisterOperand may have an associated ParserMatchClass. If it does,
894    // use it, else just fall back to the underlying register class.
895    const RecordVal *R = Rec->getValue("ParserMatchClass");
896    if (R == 0 || R->getValue() == 0)
897      throw "Record `" + Rec->getName() +
898        "' does not have a ParserMatchClass!\n";
899
900    if (DefInit *DI= dynamic_cast<DefInit*>(R->getValue())) {
901      Record *MatchClass = DI->getDef();
902      if (ClassInfo *CI = AsmOperandClasses[MatchClass])
903        return CI;
904    }
905
906    // No custom match class. Just use the register class.
907    Record *ClassRec = Rec->getValueAsDef("RegClass");
908    if (!ClassRec)
909      throw TGError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
910                    "' has no associated register class!\n");
911    if (ClassInfo *CI = RegisterClassClasses[ClassRec])
912      return CI;
913    throw TGError(Rec->getLoc(), "register class has no class info!");
914  }
915
916
917  if (Rec->isSubClassOf("RegisterClass")) {
918    if (ClassInfo *CI = RegisterClassClasses[Rec])
919      return CI;
920    throw TGError(Rec->getLoc(), "register class has no class info!");
921  }
922
923  assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
924  Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
925  if (ClassInfo *CI = AsmOperandClasses[MatchClass])
926    return CI;
927
928  throw TGError(Rec->getLoc(), "operand has no match class!");
929}
930
931void AsmMatcherInfo::
932BuildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters) {
933  const std::vector<CodeGenRegister*> &Registers =
934    Target.getRegBank().getRegisters();
935  ArrayRef<CodeGenRegisterClass*> RegClassList =
936    Target.getRegBank().getRegClasses();
937
938  // The register sets used for matching.
939  std::set< std::set<Record*> > RegisterSets;
940
941  // Gather the defined sets.
942  for (ArrayRef<CodeGenRegisterClass*>::const_iterator it =
943       RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it)
944    RegisterSets.insert(std::set<Record*>(
945        (*it)->getOrder().begin(), (*it)->getOrder().end()));
946
947  // Add any required singleton sets.
948  for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
949       ie = SingletonRegisters.end(); it != ie; ++it) {
950    Record *Rec = *it;
951    RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
952  }
953
954  // Introduce derived sets where necessary (when a register does not determine
955  // a unique register set class), and build the mapping of registers to the set
956  // they should classify to.
957  std::map<Record*, std::set<Record*> > RegisterMap;
958  for (std::vector<CodeGenRegister*>::const_iterator it = Registers.begin(),
959         ie = Registers.end(); it != ie; ++it) {
960    const CodeGenRegister &CGR = **it;
961    // Compute the intersection of all sets containing this register.
962    std::set<Record*> ContainingSet;
963
964    for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
965           ie = RegisterSets.end(); it != ie; ++it) {
966      if (!it->count(CGR.TheDef))
967        continue;
968
969      if (ContainingSet.empty()) {
970        ContainingSet = *it;
971        continue;
972      }
973
974      std::set<Record*> Tmp;
975      std::swap(Tmp, ContainingSet);
976      std::insert_iterator< std::set<Record*> > II(ContainingSet,
977                                                   ContainingSet.begin());
978      std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(), II);
979    }
980
981    if (!ContainingSet.empty()) {
982      RegisterSets.insert(ContainingSet);
983      RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
984    }
985  }
986
987  // Construct the register classes.
988  std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
989  unsigned Index = 0;
990  for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
991         ie = RegisterSets.end(); it != ie; ++it, ++Index) {
992    ClassInfo *CI = new ClassInfo();
993    CI->Kind = ClassInfo::RegisterClass0 + Index;
994    CI->ClassName = "Reg" + utostr(Index);
995    CI->Name = "MCK_Reg" + utostr(Index);
996    CI->ValueName = "";
997    CI->PredicateMethod = ""; // unused
998    CI->RenderMethod = "addRegOperands";
999    CI->Registers = *it;
1000    Classes.push_back(CI);
1001    RegisterSetClasses.insert(std::make_pair(*it, CI));
1002  }
1003
1004  // Find the superclasses; we could compute only the subgroup lattice edges,
1005  // but there isn't really a point.
1006  for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
1007         ie = RegisterSets.end(); it != ie; ++it) {
1008    ClassInfo *CI = RegisterSetClasses[*it];
1009    for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
1010           ie2 = RegisterSets.end(); it2 != ie2; ++it2)
1011      if (*it != *it2 &&
1012          std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
1013        CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
1014  }
1015
1016  // Name the register classes which correspond to a user defined RegisterClass.
1017  for (ArrayRef<CodeGenRegisterClass*>::const_iterator
1018       it = RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it) {
1019    const CodeGenRegisterClass &RC = **it;
1020    // Def will be NULL for non-user defined register classes.
1021    Record *Def = RC.getDef();
1022    if (!Def)
1023      continue;
1024    ClassInfo *CI = RegisterSetClasses[std::set<Record*>(RC.getOrder().begin(),
1025                                                         RC.getOrder().end())];
1026    if (CI->ValueName.empty()) {
1027      CI->ClassName = RC.getName();
1028      CI->Name = "MCK_" + RC.getName();
1029      CI->ValueName = RC.getName();
1030    } else
1031      CI->ValueName = CI->ValueName + "," + RC.getName();
1032
1033    RegisterClassClasses.insert(std::make_pair(Def, CI));
1034  }
1035
1036  // Populate the map for individual registers.
1037  for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
1038         ie = RegisterMap.end(); it != ie; ++it)
1039    RegisterClasses[it->first] = RegisterSetClasses[it->second];
1040
1041  // Name the register classes which correspond to singleton registers.
1042  for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
1043         ie = SingletonRegisters.end(); it != ie; ++it) {
1044    Record *Rec = *it;
1045    ClassInfo *CI = RegisterClasses[Rec];
1046    assert(CI && "Missing singleton register class info!");
1047
1048    if (CI->ValueName.empty()) {
1049      CI->ClassName = Rec->getName();
1050      CI->Name = "MCK_" + Rec->getName();
1051      CI->ValueName = Rec->getName();
1052    } else
1053      CI->ValueName = CI->ValueName + "," + Rec->getName();
1054  }
1055}
1056
1057void AsmMatcherInfo::BuildOperandClasses() {
1058  std::vector<Record*> AsmOperands =
1059    Records.getAllDerivedDefinitions("AsmOperandClass");
1060
1061  // Pre-populate AsmOperandClasses map.
1062  for (std::vector<Record*>::iterator it = AsmOperands.begin(),
1063         ie = AsmOperands.end(); it != ie; ++it)
1064    AsmOperandClasses[*it] = new ClassInfo();
1065
1066  unsigned Index = 0;
1067  for (std::vector<Record*>::iterator it = AsmOperands.begin(),
1068         ie = AsmOperands.end(); it != ie; ++it, ++Index) {
1069    ClassInfo *CI = AsmOperandClasses[*it];
1070    CI->Kind = ClassInfo::UserClass0 + Index;
1071
1072    ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
1073    for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
1074      DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
1075      if (!DI) {
1076        PrintError((*it)->getLoc(), "Invalid super class reference!");
1077        continue;
1078      }
1079
1080      ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1081      if (!SC)
1082        PrintError((*it)->getLoc(), "Invalid super class reference!");
1083      else
1084        CI->SuperClasses.push_back(SC);
1085    }
1086    CI->ClassName = (*it)->getValueAsString("Name");
1087    CI->Name = "MCK_" + CI->ClassName;
1088    CI->ValueName = (*it)->getName();
1089
1090    // Get or construct the predicate method name.
1091    Init *PMName = (*it)->getValueInit("PredicateMethod");
1092    if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
1093      CI->PredicateMethod = SI->getValue();
1094    } else {
1095      assert(dynamic_cast<UnsetInit*>(PMName) &&
1096             "Unexpected PredicateMethod field!");
1097      CI->PredicateMethod = "is" + CI->ClassName;
1098    }
1099
1100    // Get or construct the render method name.
1101    Init *RMName = (*it)->getValueInit("RenderMethod");
1102    if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
1103      CI->RenderMethod = SI->getValue();
1104    } else {
1105      assert(dynamic_cast<UnsetInit*>(RMName) &&
1106             "Unexpected RenderMethod field!");
1107      CI->RenderMethod = "add" + CI->ClassName + "Operands";
1108    }
1109
1110    // Get the parse method name or leave it as empty.
1111    Init *PRMName = (*it)->getValueInit("ParserMethod");
1112    if (StringInit *SI = dynamic_cast<StringInit*>(PRMName))
1113      CI->ParserMethod = SI->getValue();
1114
1115    AsmOperandClasses[*it] = CI;
1116    Classes.push_back(CI);
1117  }
1118}
1119
1120AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1121                               CodeGenTarget &target,
1122                               RecordKeeper &records)
1123  : Records(records), AsmParser(asmParser), Target(target) {
1124}
1125
1126/// BuildOperandMatchInfo - Build the necessary information to handle user
1127/// defined operand parsing methods.
1128void AsmMatcherInfo::BuildOperandMatchInfo() {
1129
1130  /// Map containing a mask with all operands indicies that can be found for
1131  /// that class inside a instruction.
1132  std::map<ClassInfo*, unsigned> OpClassMask;
1133
1134  for (std::vector<MatchableInfo*>::const_iterator it =
1135       Matchables.begin(), ie = Matchables.end();
1136       it != ie; ++it) {
1137    MatchableInfo &II = **it;
1138    OpClassMask.clear();
1139
1140    // Keep track of all operands of this instructions which belong to the
1141    // same class.
1142    for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
1143      MatchableInfo::AsmOperand &Op = II.AsmOperands[i];
1144      if (Op.Class->ParserMethod.empty())
1145        continue;
1146      unsigned &OperandMask = OpClassMask[Op.Class];
1147      OperandMask |= (1 << i);
1148    }
1149
1150    // Generate operand match info for each mnemonic/operand class pair.
1151    for (std::map<ClassInfo*, unsigned>::iterator iit = OpClassMask.begin(),
1152         iie = OpClassMask.end(); iit != iie; ++iit) {
1153      unsigned OpMask = iit->second;
1154      ClassInfo *CI = iit->first;
1155      OperandMatchInfo.push_back(OperandMatchEntry::Create(&II, CI, OpMask));
1156    }
1157  }
1158}
1159
1160void AsmMatcherInfo::BuildInfo() {
1161  // Build information about all of the AssemblerPredicates.
1162  std::vector<Record*> AllPredicates =
1163    Records.getAllDerivedDefinitions("Predicate");
1164  for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
1165    Record *Pred = AllPredicates[i];
1166    // Ignore predicates that are not intended for the assembler.
1167    if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1168      continue;
1169
1170    if (Pred->getName().empty())
1171      throw TGError(Pred->getLoc(), "Predicate has no name!");
1172
1173    unsigned FeatureNo = SubtargetFeatures.size();
1174    SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo);
1175    assert(FeatureNo < 32 && "Too many subtarget features!");
1176  }
1177
1178  // Parse the instructions; we need to do this first so that we can gather the
1179  // singleton register classes.
1180  SmallPtrSet<Record*, 16> SingletonRegisters;
1181  unsigned VariantCount = Target.getAsmParserVariantCount();
1182  for (unsigned VC = 0; VC != VariantCount; ++VC) {
1183    Record *AsmVariant = Target.getAsmParserVariant(VC);
1184    std::string CommentDelimiter =
1185      AsmVariant->getValueAsString("CommentDelimiter");
1186    std::string RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1187    int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1188
1189    for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
1190           E = Target.inst_end(); I != E; ++I) {
1191      const CodeGenInstruction &CGI = **I;
1192
1193      // If the tblgen -match-prefix option is specified (for tblgen hackers),
1194      // filter the set of instructions we consider.
1195      if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
1196        continue;
1197
1198      // Ignore "codegen only" instructions.
1199      if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
1200        continue;
1201
1202      // Validate the operand list to ensure we can handle this instruction.
1203      for (unsigned i = 0, e = CGI.Operands.size(); i != e; ++i) {
1204        const CGIOperandList::OperandInfo &OI = CGI.Operands[i];
1205
1206        // Validate tied operands.
1207        if (OI.getTiedRegister() != -1) {
1208          // If we have a tied operand that consists of multiple MCOperands,
1209          // reject it.  We reject aliases and ignore instructions for now.
1210          if (OI.MINumOperands != 1) {
1211            // FIXME: Should reject these.  The ARM backend hits this with $lane
1212            // in a bunch of instructions. The right answer is unclear.
1213            DEBUG({
1214                errs() << "warning: '" << CGI.TheDef->getName() << "': "
1215                     << "ignoring instruction with multi-operand tied operand '"
1216                     << OI.Name << "'\n";
1217              });
1218            continue;
1219          }
1220        }
1221      }
1222
1223      OwningPtr<MatchableInfo> II(new MatchableInfo(CGI));
1224
1225      II->Initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1226
1227      // Ignore instructions which shouldn't be matched and diagnose invalid
1228      // instruction definitions with an error.
1229      if (!II->Validate(CommentDelimiter, true))
1230        continue;
1231
1232      // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
1233      //
1234      // FIXME: This is a total hack.
1235      if (StringRef(II->TheDef->getName()).startswith("Int_") ||
1236          StringRef(II->TheDef->getName()).endswith("_Int"))
1237        continue;
1238
1239      Matchables.push_back(II.take());
1240    }
1241
1242    // Parse all of the InstAlias definitions and stick them in the list of
1243    // matchables.
1244    std::vector<Record*> AllInstAliases =
1245      Records.getAllDerivedDefinitions("InstAlias");
1246    for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1247      CodeGenInstAlias *Alias = new CodeGenInstAlias(AllInstAliases[i], Target);
1248
1249      // If the tblgen -match-prefix option is specified (for tblgen hackers),
1250      // filter the set of instruction aliases we consider, based on the target
1251      // instruction.
1252      if (!StringRef(Alias->ResultInst->TheDef->getName())
1253            .startswith( MatchPrefix))
1254        continue;
1255
1256      OwningPtr<MatchableInfo> II(new MatchableInfo(Alias));
1257
1258      II->Initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1259
1260      // Validate the alias definitions.
1261      II->Validate(CommentDelimiter, false);
1262
1263      Matchables.push_back(II.take());
1264    }
1265  }
1266
1267  // Build info for the register classes.
1268  BuildRegisterClasses(SingletonRegisters);
1269
1270  // Build info for the user defined assembly operand classes.
1271  BuildOperandClasses();
1272
1273  // Build the information about matchables, now that we have fully formed
1274  // classes.
1275  for (std::vector<MatchableInfo*>::iterator it = Matchables.begin(),
1276         ie = Matchables.end(); it != ie; ++it) {
1277    MatchableInfo *II = *it;
1278
1279    // Parse the tokens after the mnemonic.
1280    // Note: BuildInstructionOperandReference may insert new AsmOperands, so
1281    // don't precompute the loop bound.
1282    for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1283      MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1284      StringRef Token = Op.Token;
1285
1286      // Check for singleton registers.
1287      if (Record *RegRecord = II->AsmOperands[i].SingletonReg) {
1288        Op.Class = RegisterClasses[RegRecord];
1289        assert(Op.Class && Op.Class->Registers.size() == 1 &&
1290               "Unexpected class for singleton register");
1291        continue;
1292      }
1293
1294      // Check for simple tokens.
1295      if (Token[0] != '$') {
1296        Op.Class = getTokenClass(Token);
1297        continue;
1298      }
1299
1300      if (Token.size() > 1 && isdigit(Token[1])) {
1301        Op.Class = getTokenClass(Token);
1302        continue;
1303      }
1304
1305      // Otherwise this is an operand reference.
1306      StringRef OperandName;
1307      if (Token[1] == '{')
1308        OperandName = Token.substr(2, Token.size() - 3);
1309      else
1310        OperandName = Token.substr(1);
1311
1312      if (II->DefRec.is<const CodeGenInstruction*>())
1313        BuildInstructionOperandReference(II, OperandName, i);
1314      else
1315        BuildAliasOperandReference(II, OperandName, Op);
1316    }
1317
1318    if (II->DefRec.is<const CodeGenInstruction*>())
1319      II->BuildInstructionResultOperands();
1320    else
1321      II->BuildAliasResultOperands();
1322  }
1323
1324  // Process token alias definitions and set up the associated superclass
1325  // information.
1326  std::vector<Record*> AllTokenAliases =
1327    Records.getAllDerivedDefinitions("TokenAlias");
1328  for (unsigned i = 0, e = AllTokenAliases.size(); i != e; ++i) {
1329    Record *Rec = AllTokenAliases[i];
1330    ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1331    ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1332    if (FromClass == ToClass)
1333      throw TGError(Rec->getLoc(),
1334                    "error: Destination value identical to source value.");
1335    FromClass->SuperClasses.push_back(ToClass);
1336  }
1337
1338  // Reorder classes so that classes precede super classes.
1339  std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
1340}
1341
1342/// BuildInstructionOperandReference - The specified operand is a reference to a
1343/// named operand such as $src.  Resolve the Class and OperandInfo pointers.
1344void AsmMatcherInfo::
1345BuildInstructionOperandReference(MatchableInfo *II,
1346                                 StringRef OperandName,
1347                                 unsigned AsmOpIdx) {
1348  const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1349  const CGIOperandList &Operands = CGI.Operands;
1350  MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1351
1352  // Map this token to an operand.
1353  unsigned Idx;
1354  if (!Operands.hasOperandNamed(OperandName, Idx))
1355    throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" +
1356                  OperandName.str() + "'");
1357
1358  // If the instruction operand has multiple suboperands, but the parser
1359  // match class for the asm operand is still the default "ImmAsmOperand",
1360  // then handle each suboperand separately.
1361  if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1362    Record *Rec = Operands[Idx].Rec;
1363    assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1364    Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1365    if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1366      // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1367      StringRef Token = Op->Token; // save this in case Op gets moved
1368      for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1369        MatchableInfo::AsmOperand NewAsmOp(Token);
1370        NewAsmOp.SubOpIdx = SI;
1371        II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1372      }
1373      // Replace Op with first suboperand.
1374      Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1375      Op->SubOpIdx = 0;
1376    }
1377  }
1378
1379  // Set up the operand class.
1380  Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1381
1382  // If the named operand is tied, canonicalize it to the untied operand.
1383  // For example, something like:
1384  //   (outs GPR:$dst), (ins GPR:$src)
1385  // with an asmstring of
1386  //   "inc $src"
1387  // we want to canonicalize to:
1388  //   "inc $dst"
1389  // so that we know how to provide the $dst operand when filling in the result.
1390  int OITied = Operands[Idx].getTiedRegister();
1391  if (OITied != -1) {
1392    // The tied operand index is an MIOperand index, find the operand that
1393    // contains it.
1394    std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1395    OperandName = Operands[Idx.first].Name;
1396    Op->SubOpIdx = Idx.second;
1397  }
1398
1399  Op->SrcOpName = OperandName;
1400}
1401
1402/// BuildAliasOperandReference - When parsing an operand reference out of the
1403/// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1404/// operand reference is by looking it up in the result pattern definition.
1405void AsmMatcherInfo::BuildAliasOperandReference(MatchableInfo *II,
1406                                                StringRef OperandName,
1407                                                MatchableInfo::AsmOperand &Op) {
1408  const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1409
1410  // Set up the operand class.
1411  for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1412    if (CGA.ResultOperands[i].isRecord() &&
1413        CGA.ResultOperands[i].getName() == OperandName) {
1414      // It's safe to go with the first one we find, because CodeGenInstAlias
1415      // validates that all operands with the same name have the same record.
1416      Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1417      // Use the match class from the Alias definition, not the
1418      // destination instruction, as we may have an immediate that's
1419      // being munged by the match class.
1420      Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1421                                 Op.SubOpIdx);
1422      Op.SrcOpName = OperandName;
1423      return;
1424    }
1425
1426  throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" +
1427                OperandName.str() + "'");
1428}
1429
1430void MatchableInfo::BuildInstructionResultOperands() {
1431  const CodeGenInstruction *ResultInst = getResultInst();
1432
1433  // Loop over all operands of the result instruction, determining how to
1434  // populate them.
1435  for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1436    const CGIOperandList::OperandInfo &OpInfo = ResultInst->Operands[i];
1437
1438    // If this is a tied operand, just copy from the previously handled operand.
1439    int TiedOp = OpInfo.getTiedRegister();
1440    if (TiedOp != -1) {
1441      ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1442      continue;
1443    }
1444
1445    // Find out what operand from the asmparser this MCInst operand comes from.
1446    int SrcOperand = FindAsmOperandNamed(OpInfo.Name);
1447    if (OpInfo.Name.empty() || SrcOperand == -1)
1448      throw TGError(TheDef->getLoc(), "Instruction '" +
1449                    TheDef->getName() + "' has operand '" + OpInfo.Name +
1450                    "' that doesn't appear in asm string!");
1451
1452    // Check if the one AsmOperand populates the entire operand.
1453    unsigned NumOperands = OpInfo.MINumOperands;
1454    if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1455      ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1456      continue;
1457    }
1458
1459    // Add a separate ResOperand for each suboperand.
1460    for (unsigned AI = 0; AI < NumOperands; ++AI) {
1461      assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1462             AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1463             "unexpected AsmOperands for suboperands");
1464      ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1465    }
1466  }
1467}
1468
1469void MatchableInfo::BuildAliasResultOperands() {
1470  const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1471  const CodeGenInstruction *ResultInst = getResultInst();
1472
1473  // Loop over all operands of the result instruction, determining how to
1474  // populate them.
1475  unsigned AliasOpNo = 0;
1476  unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1477  for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1478    const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1479
1480    // If this is a tied operand, just copy from the previously handled operand.
1481    int TiedOp = OpInfo->getTiedRegister();
1482    if (TiedOp != -1) {
1483      ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1484      continue;
1485    }
1486
1487    // Handle all the suboperands for this operand.
1488    const std::string &OpName = OpInfo->Name;
1489    for ( ; AliasOpNo <  LastOpNo &&
1490            CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1491      int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1492
1493      // Find out what operand from the asmparser that this MCInst operand
1494      // comes from.
1495      switch (CGA.ResultOperands[AliasOpNo].Kind) {
1496      case CodeGenInstAlias::ResultOperand::K_Record: {
1497        StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1498        int SrcOperand = FindAsmOperand(Name, SubIdx);
1499        if (SrcOperand == -1)
1500          throw TGError(TheDef->getLoc(), "Instruction '" +
1501                        TheDef->getName() + "' has operand '" + OpName +
1502                        "' that doesn't appear in asm string!");
1503        unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1504        ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1505                                                        NumOperands));
1506        break;
1507      }
1508      case CodeGenInstAlias::ResultOperand::K_Imm: {
1509        int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1510        ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1511        break;
1512      }
1513      case CodeGenInstAlias::ResultOperand::K_Reg: {
1514        Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1515        ResOperands.push_back(ResOperand::getRegOp(Reg));
1516        break;
1517      }
1518      }
1519    }
1520  }
1521}
1522
1523static void EmitConvertToMCInst(CodeGenTarget &Target, StringRef ClassName,
1524                                std::vector<MatchableInfo*> &Infos,
1525                                raw_ostream &OS) {
1526  // Write the convert function to a separate stream, so we can drop it after
1527  // the enum.
1528  std::string ConvertFnBody;
1529  raw_string_ostream CvtOS(ConvertFnBody);
1530
1531  // Function we have already generated.
1532  std::set<std::string> GeneratedFns;
1533
1534  // Start the unified conversion function.
1535  CvtOS << "bool " << Target.getName() << ClassName << "::\n";
1536  CvtOS << "ConvertToMCInst(unsigned Kind, MCInst &Inst, "
1537        << "unsigned Opcode,\n"
1538        << "                      const SmallVectorImpl<MCParsedAsmOperand*"
1539        << "> &Operands) {\n";
1540  CvtOS << "  Inst.setOpcode(Opcode);\n";
1541  CvtOS << "  switch (Kind) {\n";
1542  CvtOS << "  default:\n";
1543
1544  // Start the enum, which we will generate inline.
1545
1546  OS << "// Unified function for converting operands to MCInst instances.\n\n";
1547  OS << "enum ConversionKind {\n";
1548
1549  // TargetOperandClass - This is the target's operand class, like X86Operand.
1550  std::string TargetOperandClass = Target.getName() + "Operand";
1551
1552  for (std::vector<MatchableInfo*>::const_iterator it = Infos.begin(),
1553         ie = Infos.end(); it != ie; ++it) {
1554    MatchableInfo &II = **it;
1555
1556    // Check if we have a custom match function.
1557    std::string AsmMatchConverter =
1558      II.getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1559    if (!AsmMatchConverter.empty()) {
1560      std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1561      II.ConversionFnKind = Signature;
1562
1563      // Check if we have already generated this signature.
1564      if (!GeneratedFns.insert(Signature).second)
1565        continue;
1566
1567      // If not, emit it now.  Add to the enum list.
1568      OS << "  " << Signature << ",\n";
1569
1570      CvtOS << "  case " << Signature << ":\n";
1571      CvtOS << "    return " << AsmMatchConverter
1572            << "(Inst, Opcode, Operands);\n";
1573      continue;
1574    }
1575
1576    // Build the conversion function signature.
1577    std::string Signature = "Convert";
1578    std::string CaseBody;
1579    raw_string_ostream CaseOS(CaseBody);
1580
1581    // Compute the convert enum and the case body.
1582    for (unsigned i = 0, e = II.ResOperands.size(); i != e; ++i) {
1583      const MatchableInfo::ResOperand &OpInfo = II.ResOperands[i];
1584
1585      // Generate code to populate each result operand.
1586      switch (OpInfo.Kind) {
1587      case MatchableInfo::ResOperand::RenderAsmOperand: {
1588        // This comes from something we parsed.
1589        MatchableInfo::AsmOperand &Op = II.AsmOperands[OpInfo.AsmOperandNum];
1590
1591        // Registers are always converted the same, don't duplicate the
1592        // conversion function based on them.
1593        Signature += "__";
1594        if (Op.Class->isRegisterClass())
1595          Signature += "Reg";
1596        else
1597          Signature += Op.Class->ClassName;
1598        Signature += utostr(OpInfo.MINumOperands);
1599        Signature += "_" + itostr(OpInfo.AsmOperandNum);
1600
1601        CaseOS << "    ((" << TargetOperandClass << "*)Operands["
1602               << (OpInfo.AsmOperandNum+1) << "])->" << Op.Class->RenderMethod
1603               << "(Inst, " << OpInfo.MINumOperands << ");\n";
1604        break;
1605      }
1606
1607      case MatchableInfo::ResOperand::TiedOperand: {
1608        // If this operand is tied to a previous one, just copy the MCInst
1609        // operand from the earlier one.We can only tie single MCOperand values.
1610        //assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
1611        unsigned TiedOp = OpInfo.TiedOperandNum;
1612        assert(i > TiedOp && "Tied operand precedes its target!");
1613        CaseOS << "    Inst.addOperand(Inst.getOperand(" << TiedOp << "));\n";
1614        Signature += "__Tie" + utostr(TiedOp);
1615        break;
1616      }
1617      case MatchableInfo::ResOperand::ImmOperand: {
1618        int64_t Val = OpInfo.ImmVal;
1619        CaseOS << "    Inst.addOperand(MCOperand::CreateImm(" << Val << "));\n";
1620        Signature += "__imm" + itostr(Val);
1621        break;
1622      }
1623      case MatchableInfo::ResOperand::RegOperand: {
1624        if (OpInfo.Register == 0) {
1625          CaseOS << "    Inst.addOperand(MCOperand::CreateReg(0));\n";
1626          Signature += "__reg0";
1627        } else {
1628          std::string N = getQualifiedName(OpInfo.Register);
1629          CaseOS << "    Inst.addOperand(MCOperand::CreateReg(" << N << "));\n";
1630          Signature += "__reg" + OpInfo.Register->getName();
1631        }
1632      }
1633      }
1634    }
1635
1636    II.ConversionFnKind = Signature;
1637
1638    // Check if we have already generated this signature.
1639    if (!GeneratedFns.insert(Signature).second)
1640      continue;
1641
1642    // If not, emit it now.  Add to the enum list.
1643    OS << "  " << Signature << ",\n";
1644
1645    CvtOS << "  case " << Signature << ":\n";
1646    CvtOS << CaseOS.str();
1647    CvtOS << "    return true;\n";
1648  }
1649
1650  // Finish the convert function.
1651
1652  CvtOS << "  }\n";
1653  CvtOS << "  return false;\n";
1654  CvtOS << "}\n\n";
1655
1656  // Finish the enum, and drop the convert function after it.
1657
1658  OS << "  NumConversionVariants\n";
1659  OS << "};\n\n";
1660
1661  OS << CvtOS.str();
1662}
1663
1664/// EmitMatchClassEnumeration - Emit the enumeration for match class kinds.
1665static void EmitMatchClassEnumeration(CodeGenTarget &Target,
1666                                      std::vector<ClassInfo*> &Infos,
1667                                      raw_ostream &OS) {
1668  OS << "namespace {\n\n";
1669
1670  OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1671     << "/// instruction matching.\n";
1672  OS << "enum MatchClassKind {\n";
1673  OS << "  InvalidMatchClass = 0,\n";
1674  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1675         ie = Infos.end(); it != ie; ++it) {
1676    ClassInfo &CI = **it;
1677    OS << "  " << CI.Name << ", // ";
1678    if (CI.Kind == ClassInfo::Token) {
1679      OS << "'" << CI.ValueName << "'\n";
1680    } else if (CI.isRegisterClass()) {
1681      if (!CI.ValueName.empty())
1682        OS << "register class '" << CI.ValueName << "'\n";
1683      else
1684        OS << "derived register class\n";
1685    } else {
1686      OS << "user defined class '" << CI.ValueName << "'\n";
1687    }
1688  }
1689  OS << "  NumMatchClassKinds\n";
1690  OS << "};\n\n";
1691
1692  OS << "}\n\n";
1693}
1694
1695/// EmitValidateOperandClass - Emit the function to validate an operand class.
1696static void EmitValidateOperandClass(AsmMatcherInfo &Info,
1697                                     raw_ostream &OS) {
1698  OS << "static bool validateOperandClass(MCParsedAsmOperand *GOp, "
1699     << "MatchClassKind Kind) {\n";
1700  OS << "  " << Info.Target.getName() << "Operand &Operand = *("
1701     << Info.Target.getName() << "Operand*)GOp;\n";
1702
1703  // The InvalidMatchClass is not to match any operand.
1704  OS << "  if (Kind == InvalidMatchClass)\n";
1705  OS << "    return false;\n\n";
1706
1707  // Check for Token operands first.
1708  OS << "  if (Operand.isToken())\n";
1709  OS << "    return isSubclass(matchTokenString(Operand.getToken()), Kind);"
1710     << "\n\n";
1711
1712  // Check for register operands, including sub-classes.
1713  OS << "  if (Operand.isReg()) {\n";
1714  OS << "    MatchClassKind OpKind;\n";
1715  OS << "    switch (Operand.getReg()) {\n";
1716  OS << "    default: OpKind = InvalidMatchClass; break;\n";
1717  for (std::map<Record*, ClassInfo*>::iterator
1718         it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1719       it != ie; ++it)
1720    OS << "    case " << Info.Target.getName() << "::"
1721       << it->first->getName() << ": OpKind = " << it->second->Name
1722       << "; break;\n";
1723  OS << "    }\n";
1724  OS << "    return isSubclass(OpKind, Kind);\n";
1725  OS << "  }\n\n";
1726
1727  // Check the user classes. We don't care what order since we're only
1728  // actually matching against one of them.
1729  for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1730         ie = Info.Classes.end(); it != ie; ++it) {
1731    ClassInfo &CI = **it;
1732
1733    if (!CI.isUserClass())
1734      continue;
1735
1736    OS << "  // '" << CI.ClassName << "' class\n";
1737    OS << "  if (Kind == " << CI.Name
1738       << " && Operand." << CI.PredicateMethod << "()) {\n";
1739    OS << "    return true;\n";
1740    OS << "  }\n\n";
1741  }
1742
1743  OS << "  return false;\n";
1744  OS << "}\n\n";
1745}
1746
1747/// EmitIsSubclass - Emit the subclass predicate function.
1748static void EmitIsSubclass(CodeGenTarget &Target,
1749                           std::vector<ClassInfo*> &Infos,
1750                           raw_ostream &OS) {
1751  OS << "/// isSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1752  OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
1753  OS << "  if (A == B)\n";
1754  OS << "    return true;\n\n";
1755
1756  OS << "  switch (A) {\n";
1757  OS << "  default:\n";
1758  OS << "    return false;\n";
1759  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1760         ie = Infos.end(); it != ie; ++it) {
1761    ClassInfo &A = **it;
1762
1763    std::vector<StringRef> SuperClasses;
1764    for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1765         ie = Infos.end(); it != ie; ++it) {
1766      ClassInfo &B = **it;
1767
1768      if (&A != &B && A.isSubsetOf(B))
1769        SuperClasses.push_back(B.Name);
1770    }
1771
1772    if (SuperClasses.empty())
1773      continue;
1774
1775    OS << "\n  case " << A.Name << ":\n";
1776
1777    if (SuperClasses.size() == 1) {
1778      OS << "    return B == " << SuperClasses.back() << ";\n";
1779      continue;
1780    }
1781
1782    OS << "    switch (B) {\n";
1783    OS << "    default: return false;\n";
1784    for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1785      OS << "    case " << SuperClasses[i] << ": return true;\n";
1786    OS << "    }\n";
1787  }
1788  OS << "  }\n";
1789  OS << "}\n\n";
1790}
1791
1792/// EmitMatchTokenString - Emit the function to match a token string to the
1793/// appropriate match class value.
1794static void EmitMatchTokenString(CodeGenTarget &Target,
1795                                 std::vector<ClassInfo*> &Infos,
1796                                 raw_ostream &OS) {
1797  // Construct the match list.
1798  std::vector<StringMatcher::StringPair> Matches;
1799  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1800         ie = Infos.end(); it != ie; ++it) {
1801    ClassInfo &CI = **it;
1802
1803    if (CI.Kind == ClassInfo::Token)
1804      Matches.push_back(StringMatcher::StringPair(CI.ValueName,
1805                                                  "return " + CI.Name + ";"));
1806  }
1807
1808  OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
1809
1810  StringMatcher("Name", Matches, OS).Emit();
1811
1812  OS << "  return InvalidMatchClass;\n";
1813  OS << "}\n\n";
1814}
1815
1816/// EmitMatchRegisterName - Emit the function to match a string to the target
1817/// specific register enum.
1818static void EmitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1819                                  raw_ostream &OS) {
1820  // Construct the match list.
1821  std::vector<StringMatcher::StringPair> Matches;
1822  const std::vector<CodeGenRegister*> &Regs =
1823    Target.getRegBank().getRegisters();
1824  for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
1825    const CodeGenRegister *Reg = Regs[i];
1826    if (Reg->TheDef->getValueAsString("AsmName").empty())
1827      continue;
1828
1829    Matches.push_back(StringMatcher::StringPair(
1830                                     Reg->TheDef->getValueAsString("AsmName"),
1831                                     "return " + utostr(Reg->EnumValue) + ";"));
1832  }
1833
1834  OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1835
1836  StringMatcher("Name", Matches, OS).Emit();
1837
1838  OS << "  return 0;\n";
1839  OS << "}\n\n";
1840}
1841
1842/// EmitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
1843/// definitions.
1844static void EmitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
1845                                                raw_ostream &OS) {
1846  OS << "// Flags for subtarget features that participate in "
1847     << "instruction matching.\n";
1848  OS << "enum SubtargetFeatureFlag {\n";
1849  for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1850         it = Info.SubtargetFeatures.begin(),
1851         ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1852    SubtargetFeatureInfo &SFI = *it->second;
1853    OS << "  " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n";
1854  }
1855  OS << "  Feature_None = 0\n";
1856  OS << "};\n\n";
1857}
1858
1859/// EmitComputeAvailableFeatures - Emit the function to compute the list of
1860/// available features given a subtarget.
1861static void EmitComputeAvailableFeatures(AsmMatcherInfo &Info,
1862                                         raw_ostream &OS) {
1863  std::string ClassName =
1864    Info.AsmParser->getValueAsString("AsmParserClassName");
1865
1866  OS << "unsigned " << Info.Target.getName() << ClassName << "::\n"
1867     << "ComputeAvailableFeatures(uint64_t FB) const {\n";
1868  OS << "  unsigned Features = 0;\n";
1869  for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1870         it = Info.SubtargetFeatures.begin(),
1871         ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1872    SubtargetFeatureInfo &SFI = *it->second;
1873
1874    OS << "  if (";
1875    std::string CondStorage =
1876      SFI.TheDef->getValueAsString("AssemblerCondString");
1877    StringRef Conds = CondStorage;
1878    std::pair<StringRef,StringRef> Comma = Conds.split(',');
1879    bool First = true;
1880    do {
1881      if (!First)
1882        OS << " && ";
1883
1884      bool Neg = false;
1885      StringRef Cond = Comma.first;
1886      if (Cond[0] == '!') {
1887        Neg = true;
1888        Cond = Cond.substr(1);
1889      }
1890
1891      OS << "((FB & " << Info.Target.getName() << "::" << Cond << ")";
1892      if (Neg)
1893        OS << " == 0";
1894      else
1895        OS << " != 0";
1896      OS << ")";
1897
1898      if (Comma.second.empty())
1899        break;
1900
1901      First = false;
1902      Comma = Comma.second.split(',');
1903    } while (true);
1904
1905    OS << ")\n";
1906    OS << "    Features |= " << SFI.getEnumName() << ";\n";
1907  }
1908  OS << "  return Features;\n";
1909  OS << "}\n\n";
1910}
1911
1912static std::string GetAliasRequiredFeatures(Record *R,
1913                                            const AsmMatcherInfo &Info) {
1914  std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
1915  std::string Result;
1916  unsigned NumFeatures = 0;
1917  for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
1918    SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
1919
1920    if (F == 0)
1921      throw TGError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
1922                    "' is not marked as an AssemblerPredicate!");
1923
1924    if (NumFeatures)
1925      Result += '|';
1926
1927    Result += F->getEnumName();
1928    ++NumFeatures;
1929  }
1930
1931  if (NumFeatures > 1)
1932    Result = '(' + Result + ')';
1933  return Result;
1934}
1935
1936/// EmitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
1937/// emit a function for them and return true, otherwise return false.
1938static bool EmitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) {
1939  // Ignore aliases when match-prefix is set.
1940  if (!MatchPrefix.empty())
1941    return false;
1942
1943  std::vector<Record*> Aliases =
1944    Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
1945  if (Aliases.empty()) return false;
1946
1947  OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
1948        "unsigned Features) {\n";
1949
1950  // Keep track of all the aliases from a mnemonic.  Use an std::map so that the
1951  // iteration order of the map is stable.
1952  std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
1953
1954  for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
1955    Record *R = Aliases[i];
1956    AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
1957  }
1958
1959  // Process each alias a "from" mnemonic at a time, building the code executed
1960  // by the string remapper.
1961  std::vector<StringMatcher::StringPair> Cases;
1962  for (std::map<std::string, std::vector<Record*> >::iterator
1963       I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
1964       I != E; ++I) {
1965    const std::vector<Record*> &ToVec = I->second;
1966
1967    // Loop through each alias and emit code that handles each case.  If there
1968    // are two instructions without predicates, emit an error.  If there is one,
1969    // emit it last.
1970    std::string MatchCode;
1971    int AliasWithNoPredicate = -1;
1972
1973    for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
1974      Record *R = ToVec[i];
1975      std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
1976
1977      // If this unconditionally matches, remember it for later and diagnose
1978      // duplicates.
1979      if (FeatureMask.empty()) {
1980        if (AliasWithNoPredicate != -1) {
1981          // We can't have two aliases from the same mnemonic with no predicate.
1982          PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
1983                     "two MnemonicAliases with the same 'from' mnemonic!");
1984          throw TGError(R->getLoc(), "this is the other MnemonicAlias.");
1985        }
1986
1987        AliasWithNoPredicate = i;
1988        continue;
1989      }
1990      if (R->getValueAsString("ToMnemonic") == I->first)
1991        throw TGError(R->getLoc(), "MnemonicAlias to the same string");
1992
1993      if (!MatchCode.empty())
1994        MatchCode += "else ";
1995      MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
1996      MatchCode += "  Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
1997    }
1998
1999    if (AliasWithNoPredicate != -1) {
2000      Record *R = ToVec[AliasWithNoPredicate];
2001      if (!MatchCode.empty())
2002        MatchCode += "else\n  ";
2003      MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2004    }
2005
2006    MatchCode += "return;";
2007
2008    Cases.push_back(std::make_pair(I->first, MatchCode));
2009  }
2010
2011  StringMatcher("Mnemonic", Cases, OS).Emit();
2012  OS << "}\n\n";
2013
2014  return true;
2015}
2016
2017static const char *getMinimalTypeForRange(uint64_t Range) {
2018  assert(Range < 0xFFFFFFFFULL && "Enum too large");
2019  if (Range > 0xFFFF)
2020    return "uint32_t";
2021  if (Range > 0xFF)
2022    return "uint16_t";
2023  return "uint8_t";
2024}
2025
2026static void EmitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2027                              const AsmMatcherInfo &Info, StringRef ClassName) {
2028  // Emit the static custom operand parsing table;
2029  OS << "namespace {\n";
2030  OS << "  struct OperandMatchEntry {\n";
2031  OS << "    static const char *const MnemonicTable;\n";
2032  OS << "    uint32_t OperandMask;\n";
2033  OS << "    uint32_t Mnemonic;\n";
2034  OS << "    " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size())
2035               << " RequiredFeatures;\n";
2036  OS << "    " << getMinimalTypeForRange(Info.Classes.size())
2037               << " Class;\n\n";
2038  OS << "    StringRef getMnemonic() const {\n";
2039  OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2040  OS << "                       MnemonicTable[Mnemonic]);\n";
2041  OS << "    }\n";
2042  OS << "  };\n\n";
2043
2044  OS << "  // Predicate for searching for an opcode.\n";
2045  OS << "  struct LessOpcodeOperand {\n";
2046  OS << "    bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2047  OS << "      return LHS.getMnemonic()  < RHS;\n";
2048  OS << "    }\n";
2049  OS << "    bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2050  OS << "      return LHS < RHS.getMnemonic();\n";
2051  OS << "    }\n";
2052  OS << "    bool operator()(const OperandMatchEntry &LHS,";
2053  OS << " const OperandMatchEntry &RHS) {\n";
2054  OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2055  OS << "    }\n";
2056  OS << "  };\n";
2057
2058  OS << "} // end anonymous namespace.\n\n";
2059
2060  StringToOffsetTable StringTable;
2061
2062  OS << "static const OperandMatchEntry OperandMatchTable["
2063     << Info.OperandMatchInfo.size() << "] = {\n";
2064
2065  OS << "  /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2066  for (std::vector<OperandMatchEntry>::const_iterator it =
2067       Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2068       it != ie; ++it) {
2069    const OperandMatchEntry &OMI = *it;
2070    const MatchableInfo &II = *OMI.MI;
2071
2072    OS << "  { " << OMI.OperandMask;
2073
2074    OS << " /* ";
2075    bool printComma = false;
2076    for (int i = 0, e = 31; i !=e; ++i)
2077      if (OMI.OperandMask & (1 << i)) {
2078        if (printComma)
2079          OS << ", ";
2080        OS << i;
2081        printComma = true;
2082      }
2083    OS << " */";
2084
2085    // Store a pascal-style length byte in the mnemonic.
2086    std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2087    OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2088       << " /* " << II.Mnemonic << " */, ";
2089
2090    // Write the required features mask.
2091    if (!II.RequiredFeatures.empty()) {
2092      for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2093        if (i) OS << "|";
2094        OS << II.RequiredFeatures[i]->getEnumName();
2095      }
2096    } else
2097      OS << "0";
2098
2099    OS << ", " << OMI.CI->Name;
2100
2101    OS << " },\n";
2102  }
2103  OS << "};\n\n";
2104
2105  OS << "const char *const OperandMatchEntry::MnemonicTable =\n";
2106  StringTable.EmitString(OS);
2107  OS << ";\n\n";
2108
2109  // Emit the operand class switch to call the correct custom parser for
2110  // the found operand class.
2111  OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2112     << Target.getName() << ClassName << "::\n"
2113     << "tryCustomParseOperand(SmallVectorImpl<MCParsedAsmOperand*>"
2114     << " &Operands,\n                      unsigned MCK) {\n\n"
2115     << "  switch(MCK) {\n";
2116
2117  for (std::vector<ClassInfo*>::const_iterator it = Info.Classes.begin(),
2118       ie = Info.Classes.end(); it != ie; ++it) {
2119    ClassInfo *CI = *it;
2120    if (CI->ParserMethod.empty())
2121      continue;
2122    OS << "  case " << CI->Name << ":\n"
2123       << "    return " << CI->ParserMethod << "(Operands);\n";
2124  }
2125
2126  OS << "  default:\n";
2127  OS << "    return MatchOperand_NoMatch;\n";
2128  OS << "  }\n";
2129  OS << "  return MatchOperand_NoMatch;\n";
2130  OS << "}\n\n";
2131
2132  // Emit the static custom operand parser. This code is very similar with
2133  // the other matcher. Also use MatchResultTy here just in case we go for
2134  // a better error handling.
2135  OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2136     << Target.getName() << ClassName << "::\n"
2137     << "MatchOperandParserImpl(SmallVectorImpl<MCParsedAsmOperand*>"
2138     << " &Operands,\n                       StringRef Mnemonic) {\n";
2139
2140  // Emit code to get the available features.
2141  OS << "  // Get the current feature set.\n";
2142  OS << "  unsigned AvailableFeatures = getAvailableFeatures();\n\n";
2143
2144  OS << "  // Get the next operand index.\n";
2145  OS << "  unsigned NextOpNum = Operands.size()-1;\n";
2146
2147  // Emit code to search the table.
2148  OS << "  // Search the table.\n";
2149  OS << "  std::pair<const OperandMatchEntry*, const OperandMatchEntry*>";
2150  OS << " MnemonicRange =\n";
2151  OS << "    std::equal_range(OperandMatchTable, OperandMatchTable+"
2152     << Info.OperandMatchInfo.size() << ", Mnemonic,\n"
2153     << "                     LessOpcodeOperand());\n\n";
2154
2155  OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2156  OS << "    return MatchOperand_NoMatch;\n\n";
2157
2158  OS << "  for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2159     << "       *ie = MnemonicRange.second; it != ie; ++it) {\n";
2160
2161  OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2162  OS << "    assert(Mnemonic == it->getMnemonic());\n\n";
2163
2164  // Emit check that the required features are available.
2165  OS << "    // check if the available features match\n";
2166  OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2167     << "!= it->RequiredFeatures) {\n";
2168  OS << "      continue;\n";
2169  OS << "    }\n\n";
2170
2171  // Emit check to ensure the operand number matches.
2172  OS << "    // check if the operand in question has a custom parser.\n";
2173  OS << "    if (!(it->OperandMask & (1 << NextOpNum)))\n";
2174  OS << "      continue;\n\n";
2175
2176  // Emit call to the custom parser method
2177  OS << "    // call custom parse method to handle the operand\n";
2178  OS << "    OperandMatchResultTy Result = ";
2179  OS << "tryCustomParseOperand(Operands, it->Class);\n";
2180  OS << "    if (Result != MatchOperand_NoMatch)\n";
2181  OS << "      return Result;\n";
2182  OS << "  }\n\n";
2183
2184  OS << "  // Okay, we had no match.\n";
2185  OS << "  return MatchOperand_NoMatch;\n";
2186  OS << "}\n\n";
2187}
2188
2189void AsmMatcherEmitter::run(raw_ostream &OS) {
2190  CodeGenTarget Target(Records);
2191  Record *AsmParser = Target.getAsmParser();
2192  std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2193
2194  // Compute the information on the instructions to match.
2195  AsmMatcherInfo Info(AsmParser, Target, Records);
2196  Info.BuildInfo();
2197
2198  // Sort the instruction table using the partial order on classes. We use
2199  // stable_sort to ensure that ambiguous instructions are still
2200  // deterministically ordered.
2201  std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
2202                   less_ptr<MatchableInfo>());
2203
2204  DEBUG_WITH_TYPE("instruction_info", {
2205      for (std::vector<MatchableInfo*>::iterator
2206             it = Info.Matchables.begin(), ie = Info.Matchables.end();
2207           it != ie; ++it)
2208        (*it)->dump();
2209    });
2210
2211  // Check for ambiguous matchables.
2212  DEBUG_WITH_TYPE("ambiguous_instrs", {
2213    unsigned NumAmbiguous = 0;
2214    for (unsigned i = 0, e = Info.Matchables.size(); i != e; ++i) {
2215      for (unsigned j = i + 1; j != e; ++j) {
2216        MatchableInfo &A = *Info.Matchables[i];
2217        MatchableInfo &B = *Info.Matchables[j];
2218
2219        if (A.CouldMatchAmbiguouslyWith(B)) {
2220          errs() << "warning: ambiguous matchables:\n";
2221          A.dump();
2222          errs() << "\nis incomparable with:\n";
2223          B.dump();
2224          errs() << "\n\n";
2225          ++NumAmbiguous;
2226        }
2227      }
2228    }
2229    if (NumAmbiguous)
2230      errs() << "warning: " << NumAmbiguous
2231             << " ambiguous matchables!\n";
2232  });
2233
2234  // Compute the information on the custom operand parsing.
2235  Info.BuildOperandMatchInfo();
2236
2237  // Write the output.
2238
2239  EmitSourceFileHeader("Assembly Matcher Source Fragment", OS);
2240
2241  // Information for the class declaration.
2242  OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2243  OS << "#undef GET_ASSEMBLER_HEADER\n";
2244  OS << "  // This should be included into the middle of the declaration of\n";
2245  OS << "  // your subclasses implementation of MCTargetAsmParser.\n";
2246  OS << "  unsigned ComputeAvailableFeatures(uint64_t FeatureBits) const;\n";
2247  OS << "  bool ConvertToMCInst(unsigned Kind, MCInst &Inst, "
2248     << "unsigned Opcode,\n"
2249     << "                       const SmallVectorImpl<MCParsedAsmOperand*> "
2250     << "&Operands);\n";
2251  OS << "  bool MnemonicIsValid(StringRef Mnemonic);\n";
2252  OS << "  unsigned MatchInstructionImpl(\n";
2253  OS << "    const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2254  OS << "    MCInst &Inst, unsigned &ErrorInfo, unsigned VariantID = 0);\n";
2255
2256  if (Info.OperandMatchInfo.size()) {
2257    OS << "\n  enum OperandMatchResultTy {\n";
2258    OS << "    MatchOperand_Success,    // operand matched successfully\n";
2259    OS << "    MatchOperand_NoMatch,    // operand did not match\n";
2260    OS << "    MatchOperand_ParseFail   // operand matched but had errors\n";
2261    OS << "  };\n";
2262    OS << "  OperandMatchResultTy MatchOperandParserImpl(\n";
2263    OS << "    SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2264    OS << "    StringRef Mnemonic);\n";
2265
2266    OS << "  OperandMatchResultTy tryCustomParseOperand(\n";
2267    OS << "    SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2268    OS << "    unsigned MCK);\n\n";
2269  }
2270
2271  OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2272
2273  OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2274  OS << "#undef GET_REGISTER_MATCHER\n\n";
2275
2276  // Emit the subtarget feature enumeration.
2277  EmitSubtargetFeatureFlagEnumeration(Info, OS);
2278
2279  // Emit the function to match a register name to number.
2280  EmitMatchRegisterName(Target, AsmParser, OS);
2281
2282  OS << "#endif // GET_REGISTER_MATCHER\n\n";
2283
2284
2285  OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2286  OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2287
2288  // Generate the function that remaps for mnemonic aliases.
2289  bool HasMnemonicAliases = EmitMnemonicAliases(OS, Info);
2290
2291  // Generate the unified function to convert operands into an MCInst.
2292  EmitConvertToMCInst(Target, ClassName, Info.Matchables, OS);
2293
2294  // Emit the enumeration for classes which participate in matching.
2295  EmitMatchClassEnumeration(Target, Info.Classes, OS);
2296
2297  // Emit the routine to match token strings to their match class.
2298  EmitMatchTokenString(Target, Info.Classes, OS);
2299
2300  // Emit the subclass predicate routine.
2301  EmitIsSubclass(Target, Info.Classes, OS);
2302
2303  // Emit the routine to validate an operand against a match class.
2304  EmitValidateOperandClass(Info, OS);
2305
2306  // Emit the available features compute function.
2307  EmitComputeAvailableFeatures(Info, OS);
2308
2309
2310  size_t MaxNumOperands = 0;
2311  for (std::vector<MatchableInfo*>::const_iterator it =
2312         Info.Matchables.begin(), ie = Info.Matchables.end();
2313       it != ie; ++it)
2314    MaxNumOperands = std::max(MaxNumOperands, (*it)->AsmOperands.size());
2315
2316  // Emit the static match table; unused classes get initalized to 0 which is
2317  // guaranteed to be InvalidMatchClass.
2318  //
2319  // FIXME: We can reduce the size of this table very easily. First, we change
2320  // it so that store the kinds in separate bit-fields for each index, which
2321  // only needs to be the max width used for classes at that index (we also need
2322  // to reject based on this during classification). If we then make sure to
2323  // order the match kinds appropriately (putting mnemonics last), then we
2324  // should only end up using a few bits for each class, especially the ones
2325  // following the mnemonic.
2326  OS << "namespace {\n";
2327  OS << "  struct MatchEntry {\n";
2328  OS << "    static const char *const MnemonicTable;\n";
2329  OS << "    uint32_t Mnemonic;\n";
2330  OS << "    uint16_t Opcode;\n";
2331  OS << "    " << getMinimalTypeForRange(Info.Matchables.size())
2332               << " ConvertFn;\n";
2333  OS << "    " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size())
2334               << " RequiredFeatures;\n";
2335  OS << "    " << getMinimalTypeForRange(Info.Classes.size())
2336               << " Classes[" << MaxNumOperands << "];\n";
2337  OS << "    uint8_t AsmVariantID;\n\n";
2338  OS << "    StringRef getMnemonic() const {\n";
2339  OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2340  OS << "                       MnemonicTable[Mnemonic]);\n";
2341  OS << "    }\n";
2342  OS << "  };\n\n";
2343
2344  OS << "  // Predicate for searching for an opcode.\n";
2345  OS << "  struct LessOpcode {\n";
2346  OS << "    bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
2347  OS << "      return LHS.getMnemonic() < RHS;\n";
2348  OS << "    }\n";
2349  OS << "    bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
2350  OS << "      return LHS < RHS.getMnemonic();\n";
2351  OS << "    }\n";
2352  OS << "    bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
2353  OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2354  OS << "    }\n";
2355  OS << "  };\n";
2356
2357  OS << "} // end anonymous namespace.\n\n";
2358
2359  StringToOffsetTable StringTable;
2360
2361  OS << "static const MatchEntry MatchTable["
2362     << Info.Matchables.size() << "] = {\n";
2363
2364  for (std::vector<MatchableInfo*>::const_iterator it =
2365       Info.Matchables.begin(), ie = Info.Matchables.end();
2366       it != ie; ++it) {
2367    MatchableInfo &II = **it;
2368
2369    // Store a pascal-style length byte in the mnemonic.
2370    std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2371    OS << "  { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2372       << " /* " << II.Mnemonic << " */, "
2373       << Target.getName() << "::"
2374       << II.getResultInst()->TheDef->getName() << ", "
2375       << II.ConversionFnKind << ", ";
2376
2377    // Write the required features mask.
2378    if (!II.RequiredFeatures.empty()) {
2379      for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2380        if (i) OS << "|";
2381        OS << II.RequiredFeatures[i]->getEnumName();
2382      }
2383    } else
2384      OS << "0";
2385
2386    OS << ", { ";
2387    for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
2388      MatchableInfo::AsmOperand &Op = II.AsmOperands[i];
2389
2390      if (i) OS << ", ";
2391      OS << Op.Class->Name;
2392    }
2393    OS << " }, " << II.AsmVariantID;
2394    OS << "},\n";
2395  }
2396
2397  OS << "};\n\n";
2398
2399  OS << "const char *const MatchEntry::MnemonicTable =\n";
2400  StringTable.EmitString(OS);
2401  OS << ";\n\n";
2402
2403  // A method to determine if a mnemonic is in the list.
2404  OS << "bool " << Target.getName() << ClassName << "::\n"
2405     << "MnemonicIsValid(StringRef Mnemonic) {\n";
2406  OS << "  // Search the table.\n";
2407  OS << "  std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2408  OS << "    std::equal_range(MatchTable, MatchTable+"
2409     << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n";
2410  OS << "  return MnemonicRange.first != MnemonicRange.second;\n";
2411  OS << "}\n\n";
2412
2413  // Finally, build the match function.
2414  OS << "unsigned "
2415     << Target.getName() << ClassName << "::\n"
2416     << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>"
2417     << " &Operands,\n";
2418  OS << "                     MCInst &Inst, unsigned &ErrorInfo,\n";
2419  OS << "                     unsigned VariantID) {\n";
2420
2421  // Emit code to get the available features.
2422  OS << "  // Get the current feature set.\n";
2423  OS << "  unsigned AvailableFeatures = getAvailableFeatures();\n\n";
2424
2425  OS << "  // Get the instruction mnemonic, which is the first token.\n";
2426  OS << "  StringRef Mnemonic = ((" << Target.getName()
2427     << "Operand*)Operands[0])->getToken();\n\n";
2428
2429  if (HasMnemonicAliases) {
2430    OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
2431    OS << "  // FIXME : Add an entry in AsmParserVariant to check this.\n";
2432    OS << "  if (!VariantID)\n";
2433    OS << "    applyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n";
2434  }
2435
2436  // Emit code to compute the class list for this operand vector.
2437  OS << "  // Eliminate obvious mismatches.\n";
2438  OS << "  if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
2439  OS << "    ErrorInfo = " << (MaxNumOperands+1) << ";\n";
2440  OS << "    return Match_InvalidOperand;\n";
2441  OS << "  }\n\n";
2442
2443  OS << "  // Some state to try to produce better error messages.\n";
2444  OS << "  bool HadMatchOtherThanFeatures = false;\n";
2445  OS << "  bool HadMatchOtherThanPredicate = false;\n";
2446  OS << "  unsigned RetCode = Match_InvalidOperand;\n";
2447  OS << "  // Set ErrorInfo to the operand that mismatches if it is\n";
2448  OS << "  // wrong for all instances of the instruction.\n";
2449  OS << "  ErrorInfo = ~0U;\n";
2450
2451  // Emit code to search the table.
2452  OS << "  // Search the table.\n";
2453  OS << "  std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2454  OS << "    std::equal_range(MatchTable, MatchTable+"
2455     << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n\n";
2456
2457  OS << "  // Return a more specific error code if no mnemonics match.\n";
2458  OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2459  OS << "    return Match_MnemonicFail;\n\n";
2460
2461  OS << "  for (const MatchEntry *it = MnemonicRange.first, "
2462     << "*ie = MnemonicRange.second;\n";
2463  OS << "       it != ie; ++it) {\n";
2464
2465  OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2466  OS << "    assert(Mnemonic == it->getMnemonic());\n";
2467
2468  // Emit check that the subclasses match.
2469  OS << "    if (VariantID != it->AsmVariantID) continue;\n";
2470  OS << "    bool OperandsValid = true;\n";
2471  OS << "    for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
2472  OS << "      if (i + 1 >= Operands.size()) {\n";
2473  OS << "        OperandsValid = (it->Classes[i] == " <<"InvalidMatchClass);\n";
2474  OS << "        break;\n";
2475  OS << "      }\n";
2476  OS << "      if (validateOperandClass(Operands[i+1], "
2477                                       "(MatchClassKind)it->Classes[i]))\n";
2478  OS << "        continue;\n";
2479  OS << "      // If this operand is broken for all of the instances of this\n";
2480  OS << "      // mnemonic, keep track of it so we can report loc info.\n";
2481  OS << "      if (it == MnemonicRange.first || ErrorInfo <= i+1)\n";
2482  OS << "        ErrorInfo = i+1;\n";
2483  OS << "      // Otherwise, just reject this instance of the mnemonic.\n";
2484  OS << "      OperandsValid = false;\n";
2485  OS << "      break;\n";
2486  OS << "    }\n\n";
2487
2488  OS << "    if (!OperandsValid) continue;\n";
2489
2490  // Emit check that the required features are available.
2491  OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2492     << "!= it->RequiredFeatures) {\n";
2493  OS << "      HadMatchOtherThanFeatures = true;\n";
2494  OS << "      continue;\n";
2495  OS << "    }\n";
2496  OS << "\n";
2497  OS << "    // We have selected a definite instruction, convert the parsed\n"
2498     << "    // operands into the appropriate MCInst.\n";
2499  OS << "    if (!ConvertToMCInst(it->ConvertFn, Inst,\n"
2500     << "                         it->Opcode, Operands))\n";
2501  OS << "      return Match_ConversionFail;\n";
2502  OS << "\n";
2503
2504  // Verify the instruction with the target-specific match predicate function.
2505  OS << "    // We have a potential match. Check the target predicate to\n"
2506     << "    // handle any context sensitive constraints.\n"
2507     << "    unsigned MatchResult;\n"
2508     << "    if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
2509     << " Match_Success) {\n"
2510     << "      Inst.clear();\n"
2511     << "      RetCode = MatchResult;\n"
2512     << "      HadMatchOtherThanPredicate = true;\n"
2513     << "      continue;\n"
2514     << "    }\n\n";
2515
2516  // Call the post-processing function, if used.
2517  std::string InsnCleanupFn =
2518    AsmParser->getValueAsString("AsmParserInstCleanup");
2519  if (!InsnCleanupFn.empty())
2520    OS << "    " << InsnCleanupFn << "(Inst);\n";
2521
2522  OS << "    return Match_Success;\n";
2523  OS << "  }\n\n";
2524
2525  OS << "  // Okay, we had no match.  Try to return a useful error code.\n";
2526  OS << "  if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)";
2527  OS << " return RetCode;\n";
2528  OS << "  return Match_MissingFeature;\n";
2529  OS << "}\n\n";
2530
2531  if (Info.OperandMatchInfo.size())
2532    EmitCustomOperandParsing(OS, Target, Info, ClassName);
2533
2534  OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
2535}
2536