AsmMatcherEmitter.cpp revision 087642f8638b08bf236b2cf8f0feb4df34c568f4
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 "CodeGenTarget.h"
100#include "StringToOffsetTable.h"
101#include "llvm/ADT/OwningPtr.h"
102#include "llvm/ADT/PointerUnion.h"
103#include "llvm/ADT/SmallPtrSet.h"
104#include "llvm/ADT/SmallVector.h"
105#include "llvm/ADT/STLExtras.h"
106#include "llvm/ADT/StringExtras.h"
107#include "llvm/Support/CommandLine.h"
108#include "llvm/Support/Debug.h"
109#include "llvm/Support/ErrorHandling.h"
110#include "llvm/TableGen/Error.h"
111#include "llvm/TableGen/Record.h"
112#include "llvm/TableGen/StringMatcher.h"
113#include "llvm/TableGen/TableGenBackend.h"
114#include <cassert>
115#include <map>
116#include <set>
117using namespace llvm;
118
119static cl::opt<std::string>
120MatchPrefix("match-prefix", cl::init(""),
121            cl::desc("Only match instructions with the given prefix"));
122
123namespace {
124class AsmMatcherInfo;
125struct SubtargetFeatureInfo;
126
127class AsmMatcherEmitter {
128  RecordKeeper &Records;
129public:
130  AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
131
132  void run(raw_ostream &o);
133};
134
135/// ClassInfo - Helper class for storing the information about a particular
136/// class of operands which can be matched.
137struct ClassInfo {
138  enum ClassInfoKind {
139    /// Invalid kind, for use as a sentinel value.
140    Invalid = 0,
141
142    /// The class for a particular token.
143    Token,
144
145    /// The (first) register class, subsequent register classes are
146    /// RegisterClass0+1, and so on.
147    RegisterClass0,
148
149    /// The (first) user defined class, subsequent user defined classes are
150    /// UserClass0+1, and so on.
151    UserClass0 = 1<<16
152  };
153
154  /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
155  /// N) for the Nth user defined class.
156  unsigned Kind;
157
158  /// SuperClasses - The super classes of this class. Note that for simplicities
159  /// sake user operands only record their immediate super class, while register
160  /// operands include all superclasses.
161  std::vector<ClassInfo*> SuperClasses;
162
163  /// Name - The full class name, suitable for use in an enum.
164  std::string Name;
165
166  /// ClassName - The unadorned generic name for this class (e.g., Token).
167  std::string ClassName;
168
169  /// ValueName - The name of the value this class represents; for a token this
170  /// is the literal token string, for an operand it is the TableGen class (or
171  /// empty if this is a derived class).
172  std::string ValueName;
173
174  /// PredicateMethod - The name of the operand method to test whether the
175  /// operand matches this class; this is not valid for Token or register kinds.
176  std::string PredicateMethod;
177
178  /// RenderMethod - The name of the operand method to add this operand to an
179  /// MCInst; this is not valid for Token or register kinds.
180  std::string RenderMethod;
181
182  /// ParserMethod - The name of the operand method to do a target specific
183  /// parsing on the operand.
184  std::string ParserMethod;
185
186  /// For register classes, the records for all the registers in this class.
187  std::set<Record*> Registers;
188
189  /// For custom match classes, he diagnostic kind for when the predicate fails.
190  std::string DiagnosticType;
191public:
192  /// isRegisterClass() - Check if this is a register class.
193  bool isRegisterClass() const {
194    return Kind >= RegisterClass0 && Kind < UserClass0;
195  }
196
197  /// isUserClass() - Check if this is a user defined class.
198  bool isUserClass() const {
199    return Kind >= UserClass0;
200  }
201
202  /// isRelatedTo - Check whether this class is "related" to \arg RHS. Classes
203  /// are related if they are in the same class hierarchy.
204  bool isRelatedTo(const ClassInfo &RHS) const {
205    // Tokens are only related to tokens.
206    if (Kind == Token || RHS.Kind == Token)
207      return Kind == Token && RHS.Kind == Token;
208
209    // Registers classes are only related to registers classes, and only if
210    // their intersection is non-empty.
211    if (isRegisterClass() || RHS.isRegisterClass()) {
212      if (!isRegisterClass() || !RHS.isRegisterClass())
213        return false;
214
215      std::set<Record*> Tmp;
216      std::insert_iterator< std::set<Record*> > II(Tmp, Tmp.begin());
217      std::set_intersection(Registers.begin(), Registers.end(),
218                            RHS.Registers.begin(), RHS.Registers.end(),
219                            II);
220
221      return !Tmp.empty();
222    }
223
224    // Otherwise we have two users operands; they are related if they are in the
225    // same class hierarchy.
226    //
227    // FIXME: This is an oversimplification, they should only be related if they
228    // intersect, however we don't have that information.
229    assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
230    const ClassInfo *Root = this;
231    while (!Root->SuperClasses.empty())
232      Root = Root->SuperClasses.front();
233
234    const ClassInfo *RHSRoot = &RHS;
235    while (!RHSRoot->SuperClasses.empty())
236      RHSRoot = RHSRoot->SuperClasses.front();
237
238    return Root == RHSRoot;
239  }
240
241  /// isSubsetOf - Test whether this class is a subset of \arg RHS;
242  bool isSubsetOf(const ClassInfo &RHS) const {
243    // This is a subset of RHS if it is the same class...
244    if (this == &RHS)
245      return true;
246
247    // ... or if any of its super classes are a subset of RHS.
248    for (std::vector<ClassInfo*>::const_iterator it = SuperClasses.begin(),
249           ie = SuperClasses.end(); it != ie; ++it)
250      if ((*it)->isSubsetOf(RHS))
251        return true;
252
253    return false;
254  }
255
256  /// operator< - Compare two classes.
257  bool operator<(const ClassInfo &RHS) const {
258    if (this == &RHS)
259      return false;
260
261    // Unrelated classes can be ordered by kind.
262    if (!isRelatedTo(RHS))
263      return Kind < RHS.Kind;
264
265    switch (Kind) {
266    case Invalid:
267      llvm_unreachable("Invalid kind!");
268
269    default:
270      // This class precedes the RHS if it is a proper subset of the RHS.
271      if (isSubsetOf(RHS))
272        return true;
273      if (RHS.isSubsetOf(*this))
274        return false;
275
276      // Otherwise, order by name to ensure we have a total ordering.
277      return ValueName < RHS.ValueName;
278    }
279  }
280};
281
282/// MatchableInfo - Helper class for storing the necessary information for an
283/// instruction or alias which is capable of being matched.
284struct MatchableInfo {
285  struct AsmOperand {
286    /// Token - This is the token that the operand came from.
287    StringRef Token;
288
289    /// The unique class instance this operand should match.
290    ClassInfo *Class;
291
292    /// The operand name this is, if anything.
293    StringRef SrcOpName;
294
295    /// The suboperand index within SrcOpName, or -1 for the entire operand.
296    int SubOpIdx;
297
298    /// Register record if this token is singleton register.
299    Record *SingletonReg;
300
301    explicit AsmOperand(StringRef T) : Token(T), Class(0), SubOpIdx(-1),
302                                       SingletonReg(0) {}
303  };
304
305  /// ResOperand - This represents a single operand in the result instruction
306  /// generated by the match.  In cases (like addressing modes) where a single
307  /// assembler operand expands to multiple MCOperands, this represents the
308  /// single assembler operand, not the MCOperand.
309  struct ResOperand {
310    enum {
311      /// RenderAsmOperand - This represents an operand result that is
312      /// generated by calling the render method on the assembly operand.  The
313      /// corresponding AsmOperand is specified by AsmOperandNum.
314      RenderAsmOperand,
315
316      /// TiedOperand - This represents a result operand that is a duplicate of
317      /// a previous result operand.
318      TiedOperand,
319
320      /// ImmOperand - This represents an immediate value that is dumped into
321      /// the operand.
322      ImmOperand,
323
324      /// RegOperand - This represents a fixed register that is dumped in.
325      RegOperand
326    } Kind;
327
328    union {
329      /// This is the operand # in the AsmOperands list that this should be
330      /// copied from.
331      unsigned AsmOperandNum;
332
333      /// TiedOperandNum - This is the (earlier) result operand that should be
334      /// copied from.
335      unsigned TiedOperandNum;
336
337      /// ImmVal - This is the immediate value added to the instruction.
338      int64_t ImmVal;
339
340      /// Register - This is the register record.
341      Record *Register;
342    };
343
344    /// MINumOperands - The number of MCInst operands populated by this
345    /// operand.
346    unsigned MINumOperands;
347
348    static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
349      ResOperand X;
350      X.Kind = RenderAsmOperand;
351      X.AsmOperandNum = AsmOpNum;
352      X.MINumOperands = NumOperands;
353      return X;
354    }
355
356    static ResOperand getTiedOp(unsigned TiedOperandNum) {
357      ResOperand X;
358      X.Kind = TiedOperand;
359      X.TiedOperandNum = TiedOperandNum;
360      X.MINumOperands = 1;
361      return X;
362    }
363
364    static ResOperand getImmOp(int64_t Val) {
365      ResOperand X;
366      X.Kind = ImmOperand;
367      X.ImmVal = Val;
368      X.MINumOperands = 1;
369      return X;
370    }
371
372    static ResOperand getRegOp(Record *Reg) {
373      ResOperand X;
374      X.Kind = RegOperand;
375      X.Register = Reg;
376      X.MINumOperands = 1;
377      return X;
378    }
379  };
380
381  /// AsmVariantID - Target's assembly syntax variant no.
382  int AsmVariantID;
383
384  /// TheDef - This is the definition of the instruction or InstAlias that this
385  /// matchable came from.
386  Record *const TheDef;
387
388  /// DefRec - This is the definition that it came from.
389  PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
390
391  const CodeGenInstruction *getResultInst() const {
392    if (DefRec.is<const CodeGenInstruction*>())
393      return DefRec.get<const CodeGenInstruction*>();
394    return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
395  }
396
397  /// ResOperands - This is the operand list that should be built for the result
398  /// MCInst.
399  SmallVector<ResOperand, 8> ResOperands;
400
401  /// AsmString - The assembly string for this instruction (with variants
402  /// removed), e.g. "movsx $src, $dst".
403  std::string AsmString;
404
405  /// Mnemonic - This is the first token of the matched instruction, its
406  /// mnemonic.
407  StringRef Mnemonic;
408
409  /// AsmOperands - The textual operands that this instruction matches,
410  /// annotated with a class and where in the OperandList they were defined.
411  /// This directly corresponds to the tokenized AsmString after the mnemonic is
412  /// removed.
413  SmallVector<AsmOperand, 8> AsmOperands;
414
415  /// Predicates - The required subtarget features to match this instruction.
416  SmallVector<SubtargetFeatureInfo*, 4> RequiredFeatures;
417
418  /// ConversionFnKind - The enum value which is passed to the generated
419  /// ConvertToMCInst to convert parsed operands into an MCInst for this
420  /// function.
421  std::string ConversionFnKind;
422
423  MatchableInfo(const CodeGenInstruction &CGI)
424    : AsmVariantID(0), TheDef(CGI.TheDef), DefRec(&CGI),
425      AsmString(CGI.AsmString) {
426  }
427
428  MatchableInfo(const CodeGenInstAlias *Alias)
429    : AsmVariantID(0), TheDef(Alias->TheDef), DefRec(Alias),
430      AsmString(Alias->AsmString) {
431  }
432
433  // Two-operand aliases clone from the main matchable, but mark the second
434  // operand as a tied operand of the first for purposes of the assembler.
435  void formTwoOperandAlias(StringRef Constraint);
436
437  void initialize(const AsmMatcherInfo &Info,
438                  SmallPtrSet<Record*, 16> &SingletonRegisters,
439                  int AsmVariantNo, std::string &RegisterPrefix);
440
441  /// validate - Return true if this matchable is a valid thing to match against
442  /// and perform a bunch of validity checking.
443  bool validate(StringRef CommentDelimiter, bool Hack) const;
444
445  /// extractSingletonRegisterForAsmOperand - Extract singleton register,
446  /// if present, from specified token.
447  void
448  extractSingletonRegisterForAsmOperand(unsigned i, const AsmMatcherInfo &Info,
449                                        std::string &RegisterPrefix);
450
451  /// findAsmOperand - Find the AsmOperand with the specified name and
452  /// suboperand index.
453  int findAsmOperand(StringRef N, int SubOpIdx) const {
454    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
455      if (N == AsmOperands[i].SrcOpName &&
456          SubOpIdx == AsmOperands[i].SubOpIdx)
457        return i;
458    return -1;
459  }
460
461  /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
462  /// This does not check the suboperand index.
463  int findAsmOperandNamed(StringRef N) const {
464    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
465      if (N == AsmOperands[i].SrcOpName)
466        return i;
467    return -1;
468  }
469
470  void buildInstructionResultOperands();
471  void buildAliasResultOperands();
472
473  /// operator< - Compare two matchables.
474  bool operator<(const MatchableInfo &RHS) const {
475    // The primary comparator is the instruction mnemonic.
476    if (Mnemonic != RHS.Mnemonic)
477      return Mnemonic < RHS.Mnemonic;
478
479    if (AsmOperands.size() != RHS.AsmOperands.size())
480      return AsmOperands.size() < RHS.AsmOperands.size();
481
482    // Compare lexicographically by operand. The matcher validates that other
483    // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
484    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
485      if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
486        return true;
487      if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
488        return false;
489    }
490
491    return false;
492  }
493
494  /// couldMatchAmbiguouslyWith - Check whether this matchable could
495  /// ambiguously match the same set of operands as \arg RHS (without being a
496  /// strictly superior match).
497  bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) {
498    // The primary comparator is the instruction mnemonic.
499    if (Mnemonic != RHS.Mnemonic)
500      return false;
501
502    // The number of operands is unambiguous.
503    if (AsmOperands.size() != RHS.AsmOperands.size())
504      return false;
505
506    // Otherwise, make sure the ordering of the two instructions is unambiguous
507    // by checking that either (a) a token or operand kind discriminates them,
508    // or (b) the ordering among equivalent kinds is consistent.
509
510    // Tokens and operand kinds are unambiguous (assuming a correct target
511    // specific parser).
512    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
513      if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
514          AsmOperands[i].Class->Kind == ClassInfo::Token)
515        if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
516            *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
517          return false;
518
519    // Otherwise, this operand could commute if all operands are equivalent, or
520    // there is a pair of operands that compare less than and a pair that
521    // compare greater than.
522    bool HasLT = false, HasGT = false;
523    for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
524      if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
525        HasLT = true;
526      if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
527        HasGT = true;
528    }
529
530    return !(HasLT ^ HasGT);
531  }
532
533  void dump();
534
535private:
536  void tokenizeAsmString(const AsmMatcherInfo &Info);
537};
538
539/// SubtargetFeatureInfo - Helper class for storing information on a subtarget
540/// feature which participates in instruction matching.
541struct SubtargetFeatureInfo {
542  /// \brief The predicate record for this feature.
543  Record *TheDef;
544
545  /// \brief An unique index assigned to represent this feature.
546  unsigned Index;
547
548  SubtargetFeatureInfo(Record *D, unsigned Idx) : TheDef(D), Index(Idx) {}
549
550  /// \brief The name of the enumerated constant identifying this feature.
551  std::string getEnumName() const {
552    return "Feature_" + TheDef->getName();
553  }
554};
555
556struct OperandMatchEntry {
557  unsigned OperandMask;
558  MatchableInfo* MI;
559  ClassInfo *CI;
560
561  static OperandMatchEntry create(MatchableInfo* mi, ClassInfo *ci,
562                                  unsigned opMask) {
563    OperandMatchEntry X;
564    X.OperandMask = opMask;
565    X.CI = ci;
566    X.MI = mi;
567    return X;
568  }
569};
570
571
572class AsmMatcherInfo {
573public:
574  /// Tracked Records
575  RecordKeeper &Records;
576
577  /// The tablegen AsmParser record.
578  Record *AsmParser;
579
580  /// Target - The target information.
581  CodeGenTarget &Target;
582
583  /// The classes which are needed for matching.
584  std::vector<ClassInfo*> Classes;
585
586  /// The information on the matchables to match.
587  std::vector<MatchableInfo*> Matchables;
588
589  /// Info for custom matching operands by user defined methods.
590  std::vector<OperandMatchEntry> OperandMatchInfo;
591
592  /// Map of Register records to their class information.
593  std::map<Record*, ClassInfo*> RegisterClasses;
594
595  /// Map of Predicate records to their subtarget information.
596  std::map<Record*, SubtargetFeatureInfo*> SubtargetFeatures;
597
598  /// Map of AsmOperandClass records to their class information.
599  std::map<Record*, ClassInfo*> AsmOperandClasses;
600
601private:
602  /// Map of token to class information which has already been constructed.
603  std::map<std::string, ClassInfo*> TokenClasses;
604
605  /// Map of RegisterClass records to their class information.
606  std::map<Record*, ClassInfo*> RegisterClassClasses;
607
608private:
609  /// getTokenClass - Lookup or create the class for the given token.
610  ClassInfo *getTokenClass(StringRef Token);
611
612  /// getOperandClass - Lookup or create the class for the given operand.
613  ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
614                             int SubOpIdx);
615  ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
616
617  /// buildRegisterClasses - Build the ClassInfo* instances for register
618  /// classes.
619  void buildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters);
620
621  /// buildOperandClasses - Build the ClassInfo* instances for user defined
622  /// operand classes.
623  void buildOperandClasses();
624
625  void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
626                                        unsigned AsmOpIdx);
627  void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
628                                  MatchableInfo::AsmOperand &Op);
629
630public:
631  AsmMatcherInfo(Record *AsmParser,
632                 CodeGenTarget &Target,
633                 RecordKeeper &Records);
634
635  /// buildInfo - Construct the various tables used during matching.
636  void buildInfo();
637
638  /// buildOperandMatchInfo - Build the necessary information to handle user
639  /// defined operand parsing methods.
640  void buildOperandMatchInfo();
641
642  /// getSubtargetFeature - Lookup or create the subtarget feature info for the
643  /// given operand.
644  SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
645    assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
646    std::map<Record*, SubtargetFeatureInfo*>::const_iterator I =
647      SubtargetFeatures.find(Def);
648    return I == SubtargetFeatures.end() ? 0 : I->second;
649  }
650
651  RecordKeeper &getRecords() const {
652    return Records;
653  }
654};
655
656} // End anonymous namespace
657
658void MatchableInfo::dump() {
659  errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
660
661  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
662    AsmOperand &Op = AsmOperands[i];
663    errs() << "  op[" << i << "] = " << Op.Class->ClassName << " - ";
664    errs() << '\"' << Op.Token << "\"\n";
665  }
666}
667
668static std::pair<StringRef, StringRef>
669parseTwoOperandConstraint(StringRef S, SMLoc Loc) {
670  // Split via the '='.
671  std::pair<StringRef, StringRef> Ops = S.split('=');
672  if (Ops.second == "")
673    throw TGError(Loc, "missing '=' in two-operand alias constraint");
674  // Trim whitespace and the leading '$' on the operand names.
675  size_t start = Ops.first.find_first_of('$');
676  if (start == std::string::npos)
677    throw TGError(Loc, "expected '$' prefix on asm operand name");
678  Ops.first = Ops.first.slice(start + 1, std::string::npos);
679  size_t end = Ops.first.find_last_of(" \t");
680  Ops.first = Ops.first.slice(0, end);
681  // Now the second operand.
682  start = Ops.second.find_first_of('$');
683  if (start == std::string::npos)
684    throw TGError(Loc, "expected '$' prefix on asm operand name");
685  Ops.second = Ops.second.slice(start + 1, std::string::npos);
686  end = Ops.second.find_last_of(" \t");
687  Ops.first = Ops.first.slice(0, end);
688  return Ops;
689}
690
691void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
692  // Figure out which operands are aliased and mark them as tied.
693  std::pair<StringRef, StringRef> Ops =
694    parseTwoOperandConstraint(Constraint, TheDef->getLoc());
695
696  // Find the AsmOperands that refer to the operands we're aliasing.
697  int SrcAsmOperand = findAsmOperandNamed(Ops.first);
698  int DstAsmOperand = findAsmOperandNamed(Ops.second);
699  if (SrcAsmOperand == -1)
700    throw TGError(TheDef->getLoc(),
701                  "unknown source two-operand alias operand '" +
702                  Ops.first.str() + "'.");
703  if (DstAsmOperand == -1)
704    throw TGError(TheDef->getLoc(),
705                  "unknown destination two-operand alias operand '" +
706                  Ops.second.str() + "'.");
707
708  // Find the ResOperand that refers to the operand we're aliasing away
709  // and update it to refer to the combined operand instead.
710  for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
711    ResOperand &Op = ResOperands[i];
712    if (Op.Kind == ResOperand::RenderAsmOperand &&
713        Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
714      Op.AsmOperandNum = DstAsmOperand;
715      break;
716    }
717  }
718  // Remove the AsmOperand for the alias operand.
719  AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
720  // Adjust the ResOperand references to any AsmOperands that followed
721  // the one we just deleted.
722  for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
723    ResOperand &Op = ResOperands[i];
724    switch(Op.Kind) {
725    default:
726      // Nothing to do for operands that don't reference AsmOperands.
727      break;
728    case ResOperand::RenderAsmOperand:
729      if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
730        --Op.AsmOperandNum;
731      break;
732    case ResOperand::TiedOperand:
733      if (Op.TiedOperandNum > (unsigned)SrcAsmOperand)
734        --Op.TiedOperandNum;
735      break;
736    }
737  }
738}
739
740void MatchableInfo::initialize(const AsmMatcherInfo &Info,
741                               SmallPtrSet<Record*, 16> &SingletonRegisters,
742                               int AsmVariantNo, std::string &RegisterPrefix) {
743  AsmVariantID = AsmVariantNo;
744  AsmString =
745    CodeGenInstruction::FlattenAsmStringVariants(AsmString, AsmVariantNo);
746
747  tokenizeAsmString(Info);
748
749  // Compute the require features.
750  std::vector<Record*> Predicates =TheDef->getValueAsListOfDefs("Predicates");
751  for (unsigned i = 0, e = Predicates.size(); i != e; ++i)
752    if (SubtargetFeatureInfo *Feature =
753        Info.getSubtargetFeature(Predicates[i]))
754      RequiredFeatures.push_back(Feature);
755
756  // Collect singleton registers, if used.
757  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
758    extractSingletonRegisterForAsmOperand(i, Info, RegisterPrefix);
759    if (Record *Reg = AsmOperands[i].SingletonReg)
760      SingletonRegisters.insert(Reg);
761  }
762}
763
764/// tokenizeAsmString - Tokenize a simplified assembly string.
765void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info) {
766  StringRef String = AsmString;
767  unsigned Prev = 0;
768  bool InTok = true;
769  for (unsigned i = 0, e = String.size(); i != e; ++i) {
770    switch (String[i]) {
771    case '[':
772    case ']':
773    case '*':
774    case '!':
775    case ' ':
776    case '\t':
777    case ',':
778      if (InTok) {
779        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
780        InTok = false;
781      }
782      if (!isspace(String[i]) && String[i] != ',')
783        AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
784      Prev = i + 1;
785      break;
786
787    case '\\':
788      if (InTok) {
789        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
790        InTok = false;
791      }
792      ++i;
793      assert(i != String.size() && "Invalid quoted character");
794      AsmOperands.push_back(AsmOperand(String.substr(i, 1)));
795      Prev = i + 1;
796      break;
797
798    case '$': {
799      if (InTok) {
800        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
801        InTok = false;
802      }
803
804      // If this isn't "${", treat like a normal token.
805      if (i + 1 == String.size() || String[i + 1] != '{') {
806        Prev = i;
807        break;
808      }
809
810      StringRef::iterator End = std::find(String.begin() + i, String.end(),'}');
811      assert(End != String.end() && "Missing brace in operand reference!");
812      size_t EndPos = End - String.begin();
813      AsmOperands.push_back(AsmOperand(String.slice(i, EndPos+1)));
814      Prev = EndPos + 1;
815      i = EndPos;
816      break;
817    }
818
819    case '.':
820      if (InTok)
821        AsmOperands.push_back(AsmOperand(String.slice(Prev, i)));
822      Prev = i;
823      InTok = true;
824      break;
825
826    default:
827      InTok = true;
828    }
829  }
830  if (InTok && Prev != String.size())
831    AsmOperands.push_back(AsmOperand(String.substr(Prev)));
832
833  // The first token of the instruction is the mnemonic, which must be a
834  // simple string, not a $foo variable or a singleton register.
835  if (AsmOperands.empty())
836    throw TGError(TheDef->getLoc(),
837                  "Instruction '" + TheDef->getName() + "' has no tokens");
838  Mnemonic = AsmOperands[0].Token;
839  if (Mnemonic.empty())
840    throw TGError(TheDef->getLoc(),
841                  "Missing instruction mnemonic");
842  // FIXME : Check and raise an error if it is a register.
843  if (Mnemonic[0] == '$')
844    throw TGError(TheDef->getLoc(),
845                  "Invalid instruction mnemonic '" + Mnemonic.str() + "'!");
846
847  // Remove the first operand, it is tracked in the mnemonic field.
848  AsmOperands.erase(AsmOperands.begin());
849}
850
851bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const {
852  // Reject matchables with no .s string.
853  if (AsmString.empty())
854    throw TGError(TheDef->getLoc(), "instruction with empty asm string");
855
856  // Reject any matchables with a newline in them, they should be marked
857  // isCodeGenOnly if they are pseudo instructions.
858  if (AsmString.find('\n') != std::string::npos)
859    throw TGError(TheDef->getLoc(),
860                  "multiline instruction is not valid for the asmparser, "
861                  "mark it isCodeGenOnly");
862
863  // Remove comments from the asm string.  We know that the asmstring only
864  // has one line.
865  if (!CommentDelimiter.empty() &&
866      StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
867    throw TGError(TheDef->getLoc(),
868                  "asmstring for instruction has comment character in it, "
869                  "mark it isCodeGenOnly");
870
871  // Reject matchables with operand modifiers, these aren't something we can
872  // handle, the target should be refactored to use operands instead of
873  // modifiers.
874  //
875  // Also, check for instructions which reference the operand multiple times;
876  // this implies a constraint we would not honor.
877  std::set<std::string> OperandNames;
878  for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
879    StringRef Tok = AsmOperands[i].Token;
880    if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
881      throw TGError(TheDef->getLoc(),
882                    "matchable with operand modifier '" + Tok.str() +
883                    "' not supported by asm matcher.  Mark isCodeGenOnly!");
884
885    // Verify that any operand is only mentioned once.
886    // We reject aliases and ignore instructions for now.
887    if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
888      if (!Hack)
889        throw TGError(TheDef->getLoc(),
890                      "ERROR: matchable with tied operand '" + Tok.str() +
891                      "' can never be matched!");
892      // FIXME: Should reject these.  The ARM backend hits this with $lane in a
893      // bunch of instructions.  It is unclear what the right answer is.
894      DEBUG({
895        errs() << "warning: '" << TheDef->getName() << "': "
896               << "ignoring instruction with tied operand '"
897               << Tok.str() << "'\n";
898      });
899      return false;
900    }
901  }
902
903  return true;
904}
905
906/// extractSingletonRegisterForAsmOperand - Extract singleton register,
907/// if present, from specified token.
908void MatchableInfo::
909extractSingletonRegisterForAsmOperand(unsigned OperandNo,
910                                      const AsmMatcherInfo &Info,
911                                      std::string &RegisterPrefix) {
912  StringRef Tok = AsmOperands[OperandNo].Token;
913  if (RegisterPrefix.empty()) {
914    std::string LoweredTok = Tok.lower();
915    if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
916      AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
917    return;
918  }
919
920  if (!Tok.startswith(RegisterPrefix))
921    return;
922
923  StringRef RegName = Tok.substr(RegisterPrefix.size());
924  if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
925    AsmOperands[OperandNo].SingletonReg = Reg->TheDef;
926
927  // If there is no register prefix (i.e. "%" in "%eax"), then this may
928  // be some random non-register token, just ignore it.
929  return;
930}
931
932static std::string getEnumNameForToken(StringRef Str) {
933  std::string Res;
934
935  for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
936    switch (*it) {
937    case '*': Res += "_STAR_"; break;
938    case '%': Res += "_PCT_"; break;
939    case ':': Res += "_COLON_"; break;
940    case '!': Res += "_EXCLAIM_"; break;
941    case '.': Res += "_DOT_"; break;
942    default:
943      if (isalnum(*it))
944        Res += *it;
945      else
946        Res += "_" + utostr((unsigned) *it) + "_";
947    }
948  }
949
950  return Res;
951}
952
953ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
954  ClassInfo *&Entry = TokenClasses[Token];
955
956  if (!Entry) {
957    Entry = new ClassInfo();
958    Entry->Kind = ClassInfo::Token;
959    Entry->ClassName = "Token";
960    Entry->Name = "MCK_" + getEnumNameForToken(Token);
961    Entry->ValueName = Token;
962    Entry->PredicateMethod = "<invalid>";
963    Entry->RenderMethod = "<invalid>";
964    Entry->ParserMethod = "";
965    Entry->DiagnosticType = "";
966    Classes.push_back(Entry);
967  }
968
969  return Entry;
970}
971
972ClassInfo *
973AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
974                                int SubOpIdx) {
975  Record *Rec = OI.Rec;
976  if (SubOpIdx != -1)
977    Rec = dynamic_cast<DefInit*>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
978  return getOperandClass(Rec, SubOpIdx);
979}
980
981ClassInfo *
982AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
983  if (Rec->isSubClassOf("RegisterOperand")) {
984    // RegisterOperand may have an associated ParserMatchClass. If it does,
985    // use it, else just fall back to the underlying register class.
986    const RecordVal *R = Rec->getValue("ParserMatchClass");
987    if (R == 0 || R->getValue() == 0)
988      throw "Record `" + Rec->getName() +
989        "' does not have a ParserMatchClass!\n";
990
991    if (DefInit *DI= dynamic_cast<DefInit*>(R->getValue())) {
992      Record *MatchClass = DI->getDef();
993      if (ClassInfo *CI = AsmOperandClasses[MatchClass])
994        return CI;
995    }
996
997    // No custom match class. Just use the register class.
998    Record *ClassRec = Rec->getValueAsDef("RegClass");
999    if (!ClassRec)
1000      throw TGError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1001                    "' has no associated register class!\n");
1002    if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1003      return CI;
1004    throw TGError(Rec->getLoc(), "register class has no class info!");
1005  }
1006
1007
1008  if (Rec->isSubClassOf("RegisterClass")) {
1009    if (ClassInfo *CI = RegisterClassClasses[Rec])
1010      return CI;
1011    throw TGError(Rec->getLoc(), "register class has no class info!");
1012  }
1013
1014  assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1015  Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1016  if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1017    return CI;
1018
1019  throw TGError(Rec->getLoc(), "operand has no match class!");
1020}
1021
1022void AsmMatcherInfo::
1023buildRegisterClasses(SmallPtrSet<Record*, 16> &SingletonRegisters) {
1024  const std::vector<CodeGenRegister*> &Registers =
1025    Target.getRegBank().getRegisters();
1026  ArrayRef<CodeGenRegisterClass*> RegClassList =
1027    Target.getRegBank().getRegClasses();
1028
1029  // The register sets used for matching.
1030  std::set< std::set<Record*> > RegisterSets;
1031
1032  // Gather the defined sets.
1033  for (ArrayRef<CodeGenRegisterClass*>::const_iterator it =
1034       RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it)
1035    RegisterSets.insert(std::set<Record*>(
1036        (*it)->getOrder().begin(), (*it)->getOrder().end()));
1037
1038  // Add any required singleton sets.
1039  for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
1040       ie = SingletonRegisters.end(); it != ie; ++it) {
1041    Record *Rec = *it;
1042    RegisterSets.insert(std::set<Record*>(&Rec, &Rec + 1));
1043  }
1044
1045  // Introduce derived sets where necessary (when a register does not determine
1046  // a unique register set class), and build the mapping of registers to the set
1047  // they should classify to.
1048  std::map<Record*, std::set<Record*> > RegisterMap;
1049  for (std::vector<CodeGenRegister*>::const_iterator it = Registers.begin(),
1050         ie = Registers.end(); it != ie; ++it) {
1051    const CodeGenRegister &CGR = **it;
1052    // Compute the intersection of all sets containing this register.
1053    std::set<Record*> ContainingSet;
1054
1055    for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
1056           ie = RegisterSets.end(); it != ie; ++it) {
1057      if (!it->count(CGR.TheDef))
1058        continue;
1059
1060      if (ContainingSet.empty()) {
1061        ContainingSet = *it;
1062        continue;
1063      }
1064
1065      std::set<Record*> Tmp;
1066      std::swap(Tmp, ContainingSet);
1067      std::insert_iterator< std::set<Record*> > II(ContainingSet,
1068                                                   ContainingSet.begin());
1069      std::set_intersection(Tmp.begin(), Tmp.end(), it->begin(), it->end(), II);
1070    }
1071
1072    if (!ContainingSet.empty()) {
1073      RegisterSets.insert(ContainingSet);
1074      RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
1075    }
1076  }
1077
1078  // Construct the register classes.
1079  std::map<std::set<Record*>, ClassInfo*> RegisterSetClasses;
1080  unsigned Index = 0;
1081  for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
1082         ie = RegisterSets.end(); it != ie; ++it, ++Index) {
1083    ClassInfo *CI = new ClassInfo();
1084    CI->Kind = ClassInfo::RegisterClass0 + Index;
1085    CI->ClassName = "Reg" + utostr(Index);
1086    CI->Name = "MCK_Reg" + utostr(Index);
1087    CI->ValueName = "";
1088    CI->PredicateMethod = ""; // unused
1089    CI->RenderMethod = "addRegOperands";
1090    CI->Registers = *it;
1091    // FIXME: diagnostic type.
1092    CI->DiagnosticType = "";
1093    Classes.push_back(CI);
1094    RegisterSetClasses.insert(std::make_pair(*it, CI));
1095  }
1096
1097  // Find the superclasses; we could compute only the subgroup lattice edges,
1098  // but there isn't really a point.
1099  for (std::set< std::set<Record*> >::iterator it = RegisterSets.begin(),
1100         ie = RegisterSets.end(); it != ie; ++it) {
1101    ClassInfo *CI = RegisterSetClasses[*it];
1102    for (std::set< std::set<Record*> >::iterator it2 = RegisterSets.begin(),
1103           ie2 = RegisterSets.end(); it2 != ie2; ++it2)
1104      if (*it != *it2 &&
1105          std::includes(it2->begin(), it2->end(), it->begin(), it->end()))
1106        CI->SuperClasses.push_back(RegisterSetClasses[*it2]);
1107  }
1108
1109  // Name the register classes which correspond to a user defined RegisterClass.
1110  for (ArrayRef<CodeGenRegisterClass*>::const_iterator
1111       it = RegClassList.begin(), ie = RegClassList.end(); it != ie; ++it) {
1112    const CodeGenRegisterClass &RC = **it;
1113    // Def will be NULL for non-user defined register classes.
1114    Record *Def = RC.getDef();
1115    if (!Def)
1116      continue;
1117    ClassInfo *CI = RegisterSetClasses[std::set<Record*>(RC.getOrder().begin(),
1118                                                         RC.getOrder().end())];
1119    if (CI->ValueName.empty()) {
1120      CI->ClassName = RC.getName();
1121      CI->Name = "MCK_" + RC.getName();
1122      CI->ValueName = RC.getName();
1123    } else
1124      CI->ValueName = CI->ValueName + "," + RC.getName();
1125
1126    RegisterClassClasses.insert(std::make_pair(Def, CI));
1127  }
1128
1129  // Populate the map for individual registers.
1130  for (std::map<Record*, std::set<Record*> >::iterator it = RegisterMap.begin(),
1131         ie = RegisterMap.end(); it != ie; ++it)
1132    RegisterClasses[it->first] = RegisterSetClasses[it->second];
1133
1134  // Name the register classes which correspond to singleton registers.
1135  for (SmallPtrSet<Record*, 16>::iterator it = SingletonRegisters.begin(),
1136         ie = SingletonRegisters.end(); it != ie; ++it) {
1137    Record *Rec = *it;
1138    ClassInfo *CI = RegisterClasses[Rec];
1139    assert(CI && "Missing singleton register class info!");
1140
1141    if (CI->ValueName.empty()) {
1142      CI->ClassName = Rec->getName();
1143      CI->Name = "MCK_" + Rec->getName();
1144      CI->ValueName = Rec->getName();
1145    } else
1146      CI->ValueName = CI->ValueName + "," + Rec->getName();
1147  }
1148}
1149
1150void AsmMatcherInfo::buildOperandClasses() {
1151  std::vector<Record*> AsmOperands =
1152    Records.getAllDerivedDefinitions("AsmOperandClass");
1153
1154  // Pre-populate AsmOperandClasses map.
1155  for (std::vector<Record*>::iterator it = AsmOperands.begin(),
1156         ie = AsmOperands.end(); it != ie; ++it)
1157    AsmOperandClasses[*it] = new ClassInfo();
1158
1159  unsigned Index = 0;
1160  for (std::vector<Record*>::iterator it = AsmOperands.begin(),
1161         ie = AsmOperands.end(); it != ie; ++it, ++Index) {
1162    ClassInfo *CI = AsmOperandClasses[*it];
1163    CI->Kind = ClassInfo::UserClass0 + Index;
1164
1165    ListInit *Supers = (*it)->getValueAsListInit("SuperClasses");
1166    for (unsigned i = 0, e = Supers->getSize(); i != e; ++i) {
1167      DefInit *DI = dynamic_cast<DefInit*>(Supers->getElement(i));
1168      if (!DI) {
1169        PrintError((*it)->getLoc(), "Invalid super class reference!");
1170        continue;
1171      }
1172
1173      ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1174      if (!SC)
1175        PrintError((*it)->getLoc(), "Invalid super class reference!");
1176      else
1177        CI->SuperClasses.push_back(SC);
1178    }
1179    CI->ClassName = (*it)->getValueAsString("Name");
1180    CI->Name = "MCK_" + CI->ClassName;
1181    CI->ValueName = (*it)->getName();
1182
1183    // Get or construct the predicate method name.
1184    Init *PMName = (*it)->getValueInit("PredicateMethod");
1185    if (StringInit *SI = dynamic_cast<StringInit*>(PMName)) {
1186      CI->PredicateMethod = SI->getValue();
1187    } else {
1188      assert(dynamic_cast<UnsetInit*>(PMName) &&
1189             "Unexpected PredicateMethod field!");
1190      CI->PredicateMethod = "is" + CI->ClassName;
1191    }
1192
1193    // Get or construct the render method name.
1194    Init *RMName = (*it)->getValueInit("RenderMethod");
1195    if (StringInit *SI = dynamic_cast<StringInit*>(RMName)) {
1196      CI->RenderMethod = SI->getValue();
1197    } else {
1198      assert(dynamic_cast<UnsetInit*>(RMName) &&
1199             "Unexpected RenderMethod field!");
1200      CI->RenderMethod = "add" + CI->ClassName + "Operands";
1201    }
1202
1203    // Get the parse method name or leave it as empty.
1204    Init *PRMName = (*it)->getValueInit("ParserMethod");
1205    if (StringInit *SI = dynamic_cast<StringInit*>(PRMName))
1206      CI->ParserMethod = SI->getValue();
1207
1208    // Get the diagnostic type or leave it as empty.
1209    // Get the parse method name or leave it as empty.
1210    Init *DiagnosticType = (*it)->getValueInit("DiagnosticType");
1211    if (StringInit *SI = dynamic_cast<StringInit*>(DiagnosticType))
1212      CI->DiagnosticType = SI->getValue();
1213
1214    AsmOperandClasses[*it] = CI;
1215    Classes.push_back(CI);
1216  }
1217}
1218
1219AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1220                               CodeGenTarget &target,
1221                               RecordKeeper &records)
1222  : Records(records), AsmParser(asmParser), Target(target) {
1223}
1224
1225/// buildOperandMatchInfo - Build the necessary information to handle user
1226/// defined operand parsing methods.
1227void AsmMatcherInfo::buildOperandMatchInfo() {
1228
1229  /// Map containing a mask with all operands indices that can be found for
1230  /// that class inside a instruction.
1231  std::map<ClassInfo*, unsigned> OpClassMask;
1232
1233  for (std::vector<MatchableInfo*>::const_iterator it =
1234       Matchables.begin(), ie = Matchables.end();
1235       it != ie; ++it) {
1236    MatchableInfo &II = **it;
1237    OpClassMask.clear();
1238
1239    // Keep track of all operands of this instructions which belong to the
1240    // same class.
1241    for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
1242      MatchableInfo::AsmOperand &Op = II.AsmOperands[i];
1243      if (Op.Class->ParserMethod.empty())
1244        continue;
1245      unsigned &OperandMask = OpClassMask[Op.Class];
1246      OperandMask |= (1 << i);
1247    }
1248
1249    // Generate operand match info for each mnemonic/operand class pair.
1250    for (std::map<ClassInfo*, unsigned>::iterator iit = OpClassMask.begin(),
1251         iie = OpClassMask.end(); iit != iie; ++iit) {
1252      unsigned OpMask = iit->second;
1253      ClassInfo *CI = iit->first;
1254      OperandMatchInfo.push_back(OperandMatchEntry::create(&II, CI, OpMask));
1255    }
1256  }
1257}
1258
1259void AsmMatcherInfo::buildInfo() {
1260  // Build information about all of the AssemblerPredicates.
1261  std::vector<Record*> AllPredicates =
1262    Records.getAllDerivedDefinitions("Predicate");
1263  for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
1264    Record *Pred = AllPredicates[i];
1265    // Ignore predicates that are not intended for the assembler.
1266    if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1267      continue;
1268
1269    if (Pred->getName().empty())
1270      throw TGError(Pred->getLoc(), "Predicate has no name!");
1271
1272    unsigned FeatureNo = SubtargetFeatures.size();
1273    SubtargetFeatures[Pred] = new SubtargetFeatureInfo(Pred, FeatureNo);
1274    assert(FeatureNo < 32 && "Too many subtarget features!");
1275  }
1276
1277  // Parse the instructions; we need to do this first so that we can gather the
1278  // singleton register classes.
1279  SmallPtrSet<Record*, 16> SingletonRegisters;
1280  unsigned VariantCount = Target.getAsmParserVariantCount();
1281  for (unsigned VC = 0; VC != VariantCount; ++VC) {
1282    Record *AsmVariant = Target.getAsmParserVariant(VC);
1283    std::string CommentDelimiter =
1284      AsmVariant->getValueAsString("CommentDelimiter");
1285    std::string RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1286    int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1287
1288    for (CodeGenTarget::inst_iterator I = Target.inst_begin(),
1289           E = Target.inst_end(); I != E; ++I) {
1290      const CodeGenInstruction &CGI = **I;
1291
1292      // If the tblgen -match-prefix option is specified (for tblgen hackers),
1293      // filter the set of instructions we consider.
1294      if (!StringRef(CGI.TheDef->getName()).startswith(MatchPrefix))
1295        continue;
1296
1297      // Ignore "codegen only" instructions.
1298      if (CGI.TheDef->getValueAsBit("isCodeGenOnly"))
1299        continue;
1300
1301      // Validate the operand list to ensure we can handle this instruction.
1302      for (unsigned i = 0, e = CGI.Operands.size(); i != e; ++i) {
1303        const CGIOperandList::OperandInfo &OI = CGI.Operands[i];
1304
1305        // Validate tied operands.
1306        if (OI.getTiedRegister() != -1) {
1307          // If we have a tied operand that consists of multiple MCOperands,
1308          // reject it.  We reject aliases and ignore instructions for now.
1309          if (OI.MINumOperands != 1) {
1310            // FIXME: Should reject these.  The ARM backend hits this with $lane
1311            // in a bunch of instructions. The right answer is unclear.
1312            DEBUG({
1313                errs() << "warning: '" << CGI.TheDef->getName() << "': "
1314                     << "ignoring instruction with multi-operand tied operand '"
1315                     << OI.Name << "'\n";
1316              });
1317            continue;
1318          }
1319        }
1320      }
1321
1322      OwningPtr<MatchableInfo> II(new MatchableInfo(CGI));
1323
1324      II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1325
1326      // Ignore instructions which shouldn't be matched and diagnose invalid
1327      // instruction definitions with an error.
1328      if (!II->validate(CommentDelimiter, true))
1329        continue;
1330
1331      // Ignore "Int_*" and "*_Int" instructions, which are internal aliases.
1332      //
1333      // FIXME: This is a total hack.
1334      if (StringRef(II->TheDef->getName()).startswith("Int_") ||
1335          StringRef(II->TheDef->getName()).endswith("_Int"))
1336        continue;
1337
1338      Matchables.push_back(II.take());
1339    }
1340
1341    // Parse all of the InstAlias definitions and stick them in the list of
1342    // matchables.
1343    std::vector<Record*> AllInstAliases =
1344      Records.getAllDerivedDefinitions("InstAlias");
1345    for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1346      CodeGenInstAlias *Alias = new CodeGenInstAlias(AllInstAliases[i], Target);
1347
1348      // If the tblgen -match-prefix option is specified (for tblgen hackers),
1349      // filter the set of instruction aliases we consider, based on the target
1350      // instruction.
1351      if (!StringRef(Alias->ResultInst->TheDef->getName())
1352            .startswith( MatchPrefix))
1353        continue;
1354
1355      OwningPtr<MatchableInfo> II(new MatchableInfo(Alias));
1356
1357      II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix);
1358
1359      // Validate the alias definitions.
1360      II->validate(CommentDelimiter, false);
1361
1362      Matchables.push_back(II.take());
1363    }
1364  }
1365
1366  // Build info for the register classes.
1367  buildRegisterClasses(SingletonRegisters);
1368
1369  // Build info for the user defined assembly operand classes.
1370  buildOperandClasses();
1371
1372  // Build the information about matchables, now that we have fully formed
1373  // classes.
1374  std::vector<MatchableInfo*> NewMatchables;
1375  for (std::vector<MatchableInfo*>::iterator it = Matchables.begin(),
1376         ie = Matchables.end(); it != ie; ++it) {
1377    MatchableInfo *II = *it;
1378
1379    // Parse the tokens after the mnemonic.
1380    // Note: buildInstructionOperandReference may insert new AsmOperands, so
1381    // don't precompute the loop bound.
1382    for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1383      MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1384      StringRef Token = Op.Token;
1385
1386      // Check for singleton registers.
1387      if (Record *RegRecord = II->AsmOperands[i].SingletonReg) {
1388        Op.Class = RegisterClasses[RegRecord];
1389        assert(Op.Class && Op.Class->Registers.size() == 1 &&
1390               "Unexpected class for singleton register");
1391        continue;
1392      }
1393
1394      // Check for simple tokens.
1395      if (Token[0] != '$') {
1396        Op.Class = getTokenClass(Token);
1397        continue;
1398      }
1399
1400      if (Token.size() > 1 && isdigit(Token[1])) {
1401        Op.Class = getTokenClass(Token);
1402        continue;
1403      }
1404
1405      // Otherwise this is an operand reference.
1406      StringRef OperandName;
1407      if (Token[1] == '{')
1408        OperandName = Token.substr(2, Token.size() - 3);
1409      else
1410        OperandName = Token.substr(1);
1411
1412      if (II->DefRec.is<const CodeGenInstruction*>())
1413        buildInstructionOperandReference(II, OperandName, i);
1414      else
1415        buildAliasOperandReference(II, OperandName, Op);
1416    }
1417
1418    if (II->DefRec.is<const CodeGenInstruction*>()) {
1419      II->buildInstructionResultOperands();
1420      // If the instruction has a two-operand alias, build up the
1421      // matchable here. We'll add them in bulk at the end to avoid
1422      // confusing this loop.
1423      std::string Constraint =
1424        II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1425      if (Constraint != "") {
1426        // Start by making a copy of the original matchable.
1427        OwningPtr<MatchableInfo> AliasII(new MatchableInfo(*II));
1428
1429        // Adjust it to be a two-operand alias.
1430        AliasII->formTwoOperandAlias(Constraint);
1431
1432        // Add the alias to the matchables list.
1433        NewMatchables.push_back(AliasII.take());
1434      }
1435    } else
1436      II->buildAliasResultOperands();
1437  }
1438  if (!NewMatchables.empty())
1439    Matchables.insert(Matchables.end(), NewMatchables.begin(),
1440                      NewMatchables.end());
1441
1442  // Process token alias definitions and set up the associated superclass
1443  // information.
1444  std::vector<Record*> AllTokenAliases =
1445    Records.getAllDerivedDefinitions("TokenAlias");
1446  for (unsigned i = 0, e = AllTokenAliases.size(); i != e; ++i) {
1447    Record *Rec = AllTokenAliases[i];
1448    ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1449    ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1450    if (FromClass == ToClass)
1451      throw TGError(Rec->getLoc(),
1452                    "error: Destination value identical to source value.");
1453    FromClass->SuperClasses.push_back(ToClass);
1454  }
1455
1456  // Reorder classes so that classes precede super classes.
1457  std::sort(Classes.begin(), Classes.end(), less_ptr<ClassInfo>());
1458}
1459
1460/// buildInstructionOperandReference - The specified operand is a reference to a
1461/// named operand such as $src.  Resolve the Class and OperandInfo pointers.
1462void AsmMatcherInfo::
1463buildInstructionOperandReference(MatchableInfo *II,
1464                                 StringRef OperandName,
1465                                 unsigned AsmOpIdx) {
1466  const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1467  const CGIOperandList &Operands = CGI.Operands;
1468  MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1469
1470  // Map this token to an operand.
1471  unsigned Idx;
1472  if (!Operands.hasOperandNamed(OperandName, Idx))
1473    throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" +
1474                  OperandName.str() + "'");
1475
1476  // If the instruction operand has multiple suboperands, but the parser
1477  // match class for the asm operand is still the default "ImmAsmOperand",
1478  // then handle each suboperand separately.
1479  if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1480    Record *Rec = Operands[Idx].Rec;
1481    assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1482    Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1483    if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1484      // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1485      StringRef Token = Op->Token; // save this in case Op gets moved
1486      for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1487        MatchableInfo::AsmOperand NewAsmOp(Token);
1488        NewAsmOp.SubOpIdx = SI;
1489        II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1490      }
1491      // Replace Op with first suboperand.
1492      Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1493      Op->SubOpIdx = 0;
1494    }
1495  }
1496
1497  // Set up the operand class.
1498  Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1499
1500  // If the named operand is tied, canonicalize it to the untied operand.
1501  // For example, something like:
1502  //   (outs GPR:$dst), (ins GPR:$src)
1503  // with an asmstring of
1504  //   "inc $src"
1505  // we want to canonicalize to:
1506  //   "inc $dst"
1507  // so that we know how to provide the $dst operand when filling in the result.
1508  int OITied = Operands[Idx].getTiedRegister();
1509  if (OITied != -1) {
1510    // The tied operand index is an MIOperand index, find the operand that
1511    // contains it.
1512    std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1513    OperandName = Operands[Idx.first].Name;
1514    Op->SubOpIdx = Idx.second;
1515  }
1516
1517  Op->SrcOpName = OperandName;
1518}
1519
1520/// buildAliasOperandReference - When parsing an operand reference out of the
1521/// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1522/// operand reference is by looking it up in the result pattern definition.
1523void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1524                                                StringRef OperandName,
1525                                                MatchableInfo::AsmOperand &Op) {
1526  const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1527
1528  // Set up the operand class.
1529  for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1530    if (CGA.ResultOperands[i].isRecord() &&
1531        CGA.ResultOperands[i].getName() == OperandName) {
1532      // It's safe to go with the first one we find, because CodeGenInstAlias
1533      // validates that all operands with the same name have the same record.
1534      Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1535      // Use the match class from the Alias definition, not the
1536      // destination instruction, as we may have an immediate that's
1537      // being munged by the match class.
1538      Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1539                                 Op.SubOpIdx);
1540      Op.SrcOpName = OperandName;
1541      return;
1542    }
1543
1544  throw TGError(II->TheDef->getLoc(), "error: unable to find operand: '" +
1545                OperandName.str() + "'");
1546}
1547
1548void MatchableInfo::buildInstructionResultOperands() {
1549  const CodeGenInstruction *ResultInst = getResultInst();
1550
1551  // Loop over all operands of the result instruction, determining how to
1552  // populate them.
1553  for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1554    const CGIOperandList::OperandInfo &OpInfo = ResultInst->Operands[i];
1555
1556    // If this is a tied operand, just copy from the previously handled operand.
1557    int TiedOp = OpInfo.getTiedRegister();
1558    if (TiedOp != -1) {
1559      ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1560      continue;
1561    }
1562
1563    // Find out what operand from the asmparser this MCInst operand comes from.
1564    int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1565    if (OpInfo.Name.empty() || SrcOperand == -1)
1566      throw TGError(TheDef->getLoc(), "Instruction '" +
1567                    TheDef->getName() + "' has operand '" + OpInfo.Name +
1568                    "' that doesn't appear in asm string!");
1569
1570    // Check if the one AsmOperand populates the entire operand.
1571    unsigned NumOperands = OpInfo.MINumOperands;
1572    if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1573      ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1574      continue;
1575    }
1576
1577    // Add a separate ResOperand for each suboperand.
1578    for (unsigned AI = 0; AI < NumOperands; ++AI) {
1579      assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1580             AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1581             "unexpected AsmOperands for suboperands");
1582      ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1583    }
1584  }
1585}
1586
1587void MatchableInfo::buildAliasResultOperands() {
1588  const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1589  const CodeGenInstruction *ResultInst = getResultInst();
1590
1591  // Loop over all operands of the result instruction, determining how to
1592  // populate them.
1593  unsigned AliasOpNo = 0;
1594  unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1595  for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1596    const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1597
1598    // If this is a tied operand, just copy from the previously handled operand.
1599    int TiedOp = OpInfo->getTiedRegister();
1600    if (TiedOp != -1) {
1601      ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1602      continue;
1603    }
1604
1605    // Handle all the suboperands for this operand.
1606    const std::string &OpName = OpInfo->Name;
1607    for ( ; AliasOpNo <  LastOpNo &&
1608            CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1609      int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1610
1611      // Find out what operand from the asmparser that this MCInst operand
1612      // comes from.
1613      switch (CGA.ResultOperands[AliasOpNo].Kind) {
1614      case CodeGenInstAlias::ResultOperand::K_Record: {
1615        StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1616        int SrcOperand = findAsmOperand(Name, SubIdx);
1617        if (SrcOperand == -1)
1618          throw TGError(TheDef->getLoc(), "Instruction '" +
1619                        TheDef->getName() + "' has operand '" + OpName +
1620                        "' that doesn't appear in asm string!");
1621        unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1622        ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1623                                                        NumOperands));
1624        break;
1625      }
1626      case CodeGenInstAlias::ResultOperand::K_Imm: {
1627        int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1628        ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1629        break;
1630      }
1631      case CodeGenInstAlias::ResultOperand::K_Reg: {
1632        Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1633        ResOperands.push_back(ResOperand::getRegOp(Reg));
1634        break;
1635      }
1636      }
1637    }
1638  }
1639}
1640
1641static void emitConvertToMCInst(CodeGenTarget &Target, StringRef ClassName,
1642                                std::vector<MatchableInfo*> &Infos,
1643                                raw_ostream &OS) {
1644  // Write the convert function to a separate stream, so we can drop it after
1645  // the enum.
1646  std::string ConvertFnBody;
1647  raw_string_ostream CvtOS(ConvertFnBody);
1648
1649  // Function we have already generated.
1650  std::set<std::string> GeneratedFns;
1651
1652  // Start the unified conversion function.
1653  CvtOS << "bool " << Target.getName() << ClassName << "::\n";
1654  CvtOS << "ConvertToMCInst(unsigned Kind, MCInst &Inst, "
1655        << "unsigned Opcode,\n"
1656        << "                      const SmallVectorImpl<MCParsedAsmOperand*"
1657        << "> &Operands) {\n";
1658  CvtOS << "  Inst.setOpcode(Opcode);\n";
1659  CvtOS << "  switch (Kind) {\n";
1660  CvtOS << "  default:\n";
1661
1662  // Start the enum, which we will generate inline.
1663
1664  OS << "// Unified function for converting operands to MCInst instances.\n\n";
1665  OS << "enum ConversionKind {\n";
1666
1667  // TargetOperandClass - This is the target's operand class, like X86Operand.
1668  std::string TargetOperandClass = Target.getName() + "Operand";
1669
1670  for (std::vector<MatchableInfo*>::const_iterator it = Infos.begin(),
1671         ie = Infos.end(); it != ie; ++it) {
1672    MatchableInfo &II = **it;
1673
1674    // Check if we have a custom match function.
1675    std::string AsmMatchConverter =
1676      II.getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1677    if (!AsmMatchConverter.empty()) {
1678      std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1679      II.ConversionFnKind = Signature;
1680
1681      // Check if we have already generated this signature.
1682      if (!GeneratedFns.insert(Signature).second)
1683        continue;
1684
1685      // If not, emit it now.  Add to the enum list.
1686      OS << "  " << Signature << ",\n";
1687
1688      CvtOS << "  case " << Signature << ":\n";
1689      CvtOS << "    return " << AsmMatchConverter
1690            << "(Inst, Opcode, Operands);\n";
1691      continue;
1692    }
1693
1694    // Build the conversion function signature.
1695    std::string Signature = "Convert";
1696    std::string CaseBody;
1697    raw_string_ostream CaseOS(CaseBody);
1698
1699    // Compute the convert enum and the case body.
1700    for (unsigned i = 0, e = II.ResOperands.size(); i != e; ++i) {
1701      const MatchableInfo::ResOperand &OpInfo = II.ResOperands[i];
1702
1703      // Generate code to populate each result operand.
1704      switch (OpInfo.Kind) {
1705      case MatchableInfo::ResOperand::RenderAsmOperand: {
1706        // This comes from something we parsed.
1707        MatchableInfo::AsmOperand &Op = II.AsmOperands[OpInfo.AsmOperandNum];
1708
1709        // Registers are always converted the same, don't duplicate the
1710        // conversion function based on them.
1711        Signature += "__";
1712        if (Op.Class->isRegisterClass())
1713          Signature += "Reg";
1714        else
1715          Signature += Op.Class->ClassName;
1716        Signature += utostr(OpInfo.MINumOperands);
1717        Signature += "_" + itostr(OpInfo.AsmOperandNum);
1718
1719        CaseOS << "    ((" << TargetOperandClass << "*)Operands["
1720               << (OpInfo.AsmOperandNum+1) << "])->" << Op.Class->RenderMethod
1721               << "(Inst, " << OpInfo.MINumOperands << ");\n";
1722        break;
1723      }
1724
1725      case MatchableInfo::ResOperand::TiedOperand: {
1726        // If this operand is tied to a previous one, just copy the MCInst
1727        // operand from the earlier one.We can only tie single MCOperand values.
1728        //assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
1729        unsigned TiedOp = OpInfo.TiedOperandNum;
1730        assert(i > TiedOp && "Tied operand precedes its target!");
1731        CaseOS << "    Inst.addOperand(Inst.getOperand(" << TiedOp << "));\n";
1732        Signature += "__Tie" + utostr(TiedOp);
1733        break;
1734      }
1735      case MatchableInfo::ResOperand::ImmOperand: {
1736        int64_t Val = OpInfo.ImmVal;
1737        CaseOS << "    Inst.addOperand(MCOperand::CreateImm(" << Val << "));\n";
1738        Signature += "__imm" + itostr(Val);
1739        break;
1740      }
1741      case MatchableInfo::ResOperand::RegOperand: {
1742        if (OpInfo.Register == 0) {
1743          CaseOS << "    Inst.addOperand(MCOperand::CreateReg(0));\n";
1744          Signature += "__reg0";
1745        } else {
1746          std::string N = getQualifiedName(OpInfo.Register);
1747          CaseOS << "    Inst.addOperand(MCOperand::CreateReg(" << N << "));\n";
1748          Signature += "__reg" + OpInfo.Register->getName();
1749        }
1750      }
1751      }
1752    }
1753
1754    II.ConversionFnKind = Signature;
1755
1756    // Check if we have already generated this signature.
1757    if (!GeneratedFns.insert(Signature).second)
1758      continue;
1759
1760    // If not, emit it now.  Add to the enum list.
1761    OS << "  " << Signature << ",\n";
1762
1763    CvtOS << "  case " << Signature << ":\n";
1764    CvtOS << CaseOS.str();
1765    CvtOS << "    return true;\n";
1766  }
1767
1768  // Finish the convert function.
1769
1770  CvtOS << "  }\n";
1771  CvtOS << "  return false;\n";
1772  CvtOS << "}\n\n";
1773
1774  // Finish the enum, and drop the convert function after it.
1775
1776  OS << "  NumConversionVariants\n";
1777  OS << "};\n\n";
1778
1779  OS << CvtOS.str();
1780}
1781
1782/// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
1783static void emitMatchClassEnumeration(CodeGenTarget &Target,
1784                                      std::vector<ClassInfo*> &Infos,
1785                                      raw_ostream &OS) {
1786  OS << "namespace {\n\n";
1787
1788  OS << "/// MatchClassKind - The kinds of classes which participate in\n"
1789     << "/// instruction matching.\n";
1790  OS << "enum MatchClassKind {\n";
1791  OS << "  InvalidMatchClass = 0,\n";
1792  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1793         ie = Infos.end(); it != ie; ++it) {
1794    ClassInfo &CI = **it;
1795    OS << "  " << CI.Name << ", // ";
1796    if (CI.Kind == ClassInfo::Token) {
1797      OS << "'" << CI.ValueName << "'\n";
1798    } else if (CI.isRegisterClass()) {
1799      if (!CI.ValueName.empty())
1800        OS << "register class '" << CI.ValueName << "'\n";
1801      else
1802        OS << "derived register class\n";
1803    } else {
1804      OS << "user defined class '" << CI.ValueName << "'\n";
1805    }
1806  }
1807  OS << "  NumMatchClassKinds\n";
1808  OS << "};\n\n";
1809
1810  OS << "}\n\n";
1811}
1812
1813/// emitValidateOperandClass - Emit the function to validate an operand class.
1814static void emitValidateOperandClass(AsmMatcherInfo &Info,
1815                                     raw_ostream &OS) {
1816  OS << "static unsigned validateOperandClass(MCParsedAsmOperand *GOp, "
1817     << "MatchClassKind Kind) {\n";
1818  OS << "  " << Info.Target.getName() << "Operand &Operand = *("
1819     << Info.Target.getName() << "Operand*)GOp;\n";
1820
1821  // The InvalidMatchClass is not to match any operand.
1822  OS << "  if (Kind == InvalidMatchClass)\n";
1823  OS << "    return MCTargetAsmParser::Match_InvalidOperand;\n\n";
1824
1825  // Check for Token operands first.
1826  // FIXME: Use a more specific diagnostic type.
1827  OS << "  if (Operand.isToken())\n";
1828  OS << "    return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
1829     << "             MCTargetAsmParser::Match_Success :\n"
1830     << "             MCTargetAsmParser::Match_InvalidOperand;\n\n";
1831
1832  // Check the user classes. We don't care what order since we're only
1833  // actually matching against one of them.
1834  for (std::vector<ClassInfo*>::iterator it = Info.Classes.begin(),
1835         ie = Info.Classes.end(); it != ie; ++it) {
1836    ClassInfo &CI = **it;
1837
1838    if (!CI.isUserClass())
1839      continue;
1840
1841    OS << "  // '" << CI.ClassName << "' class\n";
1842    OS << "  if (Kind == " << CI.Name << ") {\n";
1843    OS << "    if (Operand." << CI.PredicateMethod << "())\n";
1844    OS << "      return MCTargetAsmParser::Match_Success;\n";
1845    if (!CI.DiagnosticType.empty())
1846      OS << "    return " << Info.Target.getName() << "AsmParser::Match_"
1847         << CI.DiagnosticType << ";\n";
1848    OS << "  }\n\n";
1849  }
1850
1851  // Check for register operands, including sub-classes.
1852  OS << "  if (Operand.isReg()) {\n";
1853  OS << "    MatchClassKind OpKind;\n";
1854  OS << "    switch (Operand.getReg()) {\n";
1855  OS << "    default: OpKind = InvalidMatchClass; break;\n";
1856  for (std::map<Record*, ClassInfo*>::iterator
1857         it = Info.RegisterClasses.begin(), ie = Info.RegisterClasses.end();
1858       it != ie; ++it)
1859    OS << "    case " << Info.Target.getName() << "::"
1860       << it->first->getName() << ": OpKind = " << it->second->Name
1861       << "; break;\n";
1862  OS << "    }\n";
1863  OS << "    return isSubclass(OpKind, Kind) ? "
1864     << "MCTargetAsmParser::Match_Success :\n                             "
1865     << "         MCTargetAsmParser::Match_InvalidOperand;\n  }\n\n";
1866
1867  // Generic fallthrough match failure case for operands that don't have
1868  // specialized diagnostic types.
1869  OS << "  return MCTargetAsmParser::Match_InvalidOperand;\n";
1870  OS << "}\n\n";
1871}
1872
1873/// emitIsSubclass - Emit the subclass predicate function.
1874static void emitIsSubclass(CodeGenTarget &Target,
1875                           std::vector<ClassInfo*> &Infos,
1876                           raw_ostream &OS) {
1877  OS << "/// isSubclass - Compute whether \\arg A is a subclass of \\arg B.\n";
1878  OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
1879  OS << "  if (A == B)\n";
1880  OS << "    return true;\n\n";
1881
1882  OS << "  switch (A) {\n";
1883  OS << "  default:\n";
1884  OS << "    return false;\n";
1885  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1886         ie = Infos.end(); it != ie; ++it) {
1887    ClassInfo &A = **it;
1888
1889    std::vector<StringRef> SuperClasses;
1890    for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1891         ie = Infos.end(); it != ie; ++it) {
1892      ClassInfo &B = **it;
1893
1894      if (&A != &B && A.isSubsetOf(B))
1895        SuperClasses.push_back(B.Name);
1896    }
1897
1898    if (SuperClasses.empty())
1899      continue;
1900
1901    OS << "\n  case " << A.Name << ":\n";
1902
1903    if (SuperClasses.size() == 1) {
1904      OS << "    return B == " << SuperClasses.back() << ";\n";
1905      continue;
1906    }
1907
1908    OS << "    switch (B) {\n";
1909    OS << "    default: return false;\n";
1910    for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
1911      OS << "    case " << SuperClasses[i] << ": return true;\n";
1912    OS << "    }\n";
1913  }
1914  OS << "  }\n";
1915  OS << "}\n\n";
1916}
1917
1918/// emitMatchTokenString - Emit the function to match a token string to the
1919/// appropriate match class value.
1920static void emitMatchTokenString(CodeGenTarget &Target,
1921                                 std::vector<ClassInfo*> &Infos,
1922                                 raw_ostream &OS) {
1923  // Construct the match list.
1924  std::vector<StringMatcher::StringPair> Matches;
1925  for (std::vector<ClassInfo*>::iterator it = Infos.begin(),
1926         ie = Infos.end(); it != ie; ++it) {
1927    ClassInfo &CI = **it;
1928
1929    if (CI.Kind == ClassInfo::Token)
1930      Matches.push_back(StringMatcher::StringPair(CI.ValueName,
1931                                                  "return " + CI.Name + ";"));
1932  }
1933
1934  OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
1935
1936  StringMatcher("Name", Matches, OS).Emit();
1937
1938  OS << "  return InvalidMatchClass;\n";
1939  OS << "}\n\n";
1940}
1941
1942/// emitMatchRegisterName - Emit the function to match a string to the target
1943/// specific register enum.
1944static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
1945                                  raw_ostream &OS) {
1946  // Construct the match list.
1947  std::vector<StringMatcher::StringPair> Matches;
1948  const std::vector<CodeGenRegister*> &Regs =
1949    Target.getRegBank().getRegisters();
1950  for (unsigned i = 0, e = Regs.size(); i != e; ++i) {
1951    const CodeGenRegister *Reg = Regs[i];
1952    if (Reg->TheDef->getValueAsString("AsmName").empty())
1953      continue;
1954
1955    Matches.push_back(StringMatcher::StringPair(
1956                                     Reg->TheDef->getValueAsString("AsmName"),
1957                                     "return " + utostr(Reg->EnumValue) + ";"));
1958  }
1959
1960  OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
1961
1962  StringMatcher("Name", Matches, OS).Emit();
1963
1964  OS << "  return 0;\n";
1965  OS << "}\n\n";
1966}
1967
1968/// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
1969/// definitions.
1970static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
1971                                                raw_ostream &OS) {
1972  OS << "// Flags for subtarget features that participate in "
1973     << "instruction matching.\n";
1974  OS << "enum SubtargetFeatureFlag {\n";
1975  for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
1976         it = Info.SubtargetFeatures.begin(),
1977         ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
1978    SubtargetFeatureInfo &SFI = *it->second;
1979    OS << "  " << SFI.getEnumName() << " = (1 << " << SFI.Index << "),\n";
1980  }
1981  OS << "  Feature_None = 0\n";
1982  OS << "};\n\n";
1983}
1984
1985/// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
1986static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
1987  // Get the set of diagnostic types from all of the operand classes.
1988  std::set<StringRef> Types;
1989  for (std::map<Record*, ClassInfo*>::const_iterator
1990       I = Info.AsmOperandClasses.begin(),
1991       E = Info.AsmOperandClasses.end(); I != E; ++I) {
1992    if (!I->second->DiagnosticType.empty())
1993      Types.insert(I->second->DiagnosticType);
1994  }
1995
1996  if (Types.empty()) return;
1997
1998  // Now emit the enum entries.
1999  for (std::set<StringRef>::const_iterator I = Types.begin(), E = Types.end();
2000       I != E; ++I)
2001    OS << "  Match_" << *I << ",\n";
2002  OS << "  END_OPERAND_DIAGNOSTIC_TYPES\n";
2003}
2004
2005/// emitGetSubtargetFeatureName - Emit the helper function to get the
2006/// user-level name for a subtarget feature.
2007static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2008  OS << "// User-level names for subtarget features that participate in\n"
2009     << "// instruction matching.\n"
2010     << "static const char *getSubtargetFeatureName(unsigned Val) {\n"
2011     << "  switch(Val) {\n";
2012  for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
2013         it = Info.SubtargetFeatures.begin(),
2014         ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
2015    SubtargetFeatureInfo &SFI = *it->second;
2016    // FIXME: Totally just a placeholder name to get the algorithm working.
2017    OS << "  case " << SFI.getEnumName() << ": return \""
2018       << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2019  }
2020  OS << "  default: return \"(unknown)\";\n";
2021  OS << "  }\n}\n\n";
2022}
2023
2024/// emitComputeAvailableFeatures - Emit the function to compute the list of
2025/// available features given a subtarget.
2026static void emitComputeAvailableFeatures(AsmMatcherInfo &Info,
2027                                         raw_ostream &OS) {
2028  std::string ClassName =
2029    Info.AsmParser->getValueAsString("AsmParserClassName");
2030
2031  OS << "unsigned " << Info.Target.getName() << ClassName << "::\n"
2032     << "ComputeAvailableFeatures(uint64_t FB) const {\n";
2033  OS << "  unsigned Features = 0;\n";
2034  for (std::map<Record*, SubtargetFeatureInfo*>::const_iterator
2035         it = Info.SubtargetFeatures.begin(),
2036         ie = Info.SubtargetFeatures.end(); it != ie; ++it) {
2037    SubtargetFeatureInfo &SFI = *it->second;
2038
2039    OS << "  if (";
2040    std::string CondStorage =
2041      SFI.TheDef->getValueAsString("AssemblerCondString");
2042    StringRef Conds = CondStorage;
2043    std::pair<StringRef,StringRef> Comma = Conds.split(',');
2044    bool First = true;
2045    do {
2046      if (!First)
2047        OS << " && ";
2048
2049      bool Neg = false;
2050      StringRef Cond = Comma.first;
2051      if (Cond[0] == '!') {
2052        Neg = true;
2053        Cond = Cond.substr(1);
2054      }
2055
2056      OS << "((FB & " << Info.Target.getName() << "::" << Cond << ")";
2057      if (Neg)
2058        OS << " == 0";
2059      else
2060        OS << " != 0";
2061      OS << ")";
2062
2063      if (Comma.second.empty())
2064        break;
2065
2066      First = false;
2067      Comma = Comma.second.split(',');
2068    } while (true);
2069
2070    OS << ")\n";
2071    OS << "    Features |= " << SFI.getEnumName() << ";\n";
2072  }
2073  OS << "  return Features;\n";
2074  OS << "}\n\n";
2075}
2076
2077static std::string GetAliasRequiredFeatures(Record *R,
2078                                            const AsmMatcherInfo &Info) {
2079  std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2080  std::string Result;
2081  unsigned NumFeatures = 0;
2082  for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2083    SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2084
2085    if (F == 0)
2086      throw TGError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2087                    "' is not marked as an AssemblerPredicate!");
2088
2089    if (NumFeatures)
2090      Result += '|';
2091
2092    Result += F->getEnumName();
2093    ++NumFeatures;
2094  }
2095
2096  if (NumFeatures > 1)
2097    Result = '(' + Result + ')';
2098  return Result;
2099}
2100
2101/// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2102/// emit a function for them and return true, otherwise return false.
2103static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info) {
2104  // Ignore aliases when match-prefix is set.
2105  if (!MatchPrefix.empty())
2106    return false;
2107
2108  std::vector<Record*> Aliases =
2109    Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2110  if (Aliases.empty()) return false;
2111
2112  OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2113        "unsigned Features) {\n";
2114
2115  // Keep track of all the aliases from a mnemonic.  Use an std::map so that the
2116  // iteration order of the map is stable.
2117  std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2118
2119  for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
2120    Record *R = Aliases[i];
2121    AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2122  }
2123
2124  // Process each alias a "from" mnemonic at a time, building the code executed
2125  // by the string remapper.
2126  std::vector<StringMatcher::StringPair> Cases;
2127  for (std::map<std::string, std::vector<Record*> >::iterator
2128       I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
2129       I != E; ++I) {
2130    const std::vector<Record*> &ToVec = I->second;
2131
2132    // Loop through each alias and emit code that handles each case.  If there
2133    // are two instructions without predicates, emit an error.  If there is one,
2134    // emit it last.
2135    std::string MatchCode;
2136    int AliasWithNoPredicate = -1;
2137
2138    for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2139      Record *R = ToVec[i];
2140      std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2141
2142      // If this unconditionally matches, remember it for later and diagnose
2143      // duplicates.
2144      if (FeatureMask.empty()) {
2145        if (AliasWithNoPredicate != -1) {
2146          // We can't have two aliases from the same mnemonic with no predicate.
2147          PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2148                     "two MnemonicAliases with the same 'from' mnemonic!");
2149          throw TGError(R->getLoc(), "this is the other MnemonicAlias.");
2150        }
2151
2152        AliasWithNoPredicate = i;
2153        continue;
2154      }
2155      if (R->getValueAsString("ToMnemonic") == I->first)
2156        throw TGError(R->getLoc(), "MnemonicAlias to the same string");
2157
2158      if (!MatchCode.empty())
2159        MatchCode += "else ";
2160      MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
2161      MatchCode += "  Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
2162    }
2163
2164    if (AliasWithNoPredicate != -1) {
2165      Record *R = ToVec[AliasWithNoPredicate];
2166      if (!MatchCode.empty())
2167        MatchCode += "else\n  ";
2168      MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2169    }
2170
2171    MatchCode += "return;";
2172
2173    Cases.push_back(std::make_pair(I->first, MatchCode));
2174  }
2175
2176  StringMatcher("Mnemonic", Cases, OS).Emit();
2177  OS << "}\n\n";
2178
2179  return true;
2180}
2181
2182static const char *getMinimalTypeForRange(uint64_t Range) {
2183  assert(Range < 0xFFFFFFFFULL && "Enum too large");
2184  if (Range > 0xFFFF)
2185    return "uint32_t";
2186  if (Range > 0xFF)
2187    return "uint16_t";
2188  return "uint8_t";
2189}
2190
2191static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2192                              const AsmMatcherInfo &Info, StringRef ClassName) {
2193  // Emit the static custom operand parsing table;
2194  OS << "namespace {\n";
2195  OS << "  struct OperandMatchEntry {\n";
2196  OS << "    static const char *const MnemonicTable;\n";
2197  OS << "    uint32_t OperandMask;\n";
2198  OS << "    uint32_t Mnemonic;\n";
2199  OS << "    " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size())
2200               << " RequiredFeatures;\n";
2201  OS << "    " << getMinimalTypeForRange(Info.Classes.size())
2202               << " Class;\n\n";
2203  OS << "    StringRef getMnemonic() const {\n";
2204  OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2205  OS << "                       MnemonicTable[Mnemonic]);\n";
2206  OS << "    }\n";
2207  OS << "  };\n\n";
2208
2209  OS << "  // Predicate for searching for an opcode.\n";
2210  OS << "  struct LessOpcodeOperand {\n";
2211  OS << "    bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2212  OS << "      return LHS.getMnemonic()  < RHS;\n";
2213  OS << "    }\n";
2214  OS << "    bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2215  OS << "      return LHS < RHS.getMnemonic();\n";
2216  OS << "    }\n";
2217  OS << "    bool operator()(const OperandMatchEntry &LHS,";
2218  OS << " const OperandMatchEntry &RHS) {\n";
2219  OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2220  OS << "    }\n";
2221  OS << "  };\n";
2222
2223  OS << "} // end anonymous namespace.\n\n";
2224
2225  StringToOffsetTable StringTable;
2226
2227  OS << "static const OperandMatchEntry OperandMatchTable["
2228     << Info.OperandMatchInfo.size() << "] = {\n";
2229
2230  OS << "  /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2231  for (std::vector<OperandMatchEntry>::const_iterator it =
2232       Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2233       it != ie; ++it) {
2234    const OperandMatchEntry &OMI = *it;
2235    const MatchableInfo &II = *OMI.MI;
2236
2237    OS << "  { " << OMI.OperandMask;
2238
2239    OS << " /* ";
2240    bool printComma = false;
2241    for (int i = 0, e = 31; i !=e; ++i)
2242      if (OMI.OperandMask & (1 << i)) {
2243        if (printComma)
2244          OS << ", ";
2245        OS << i;
2246        printComma = true;
2247      }
2248    OS << " */";
2249
2250    // Store a pascal-style length byte in the mnemonic.
2251    std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2252    OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2253       << " /* " << II.Mnemonic << " */, ";
2254
2255    // Write the required features mask.
2256    if (!II.RequiredFeatures.empty()) {
2257      for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2258        if (i) OS << "|";
2259        OS << II.RequiredFeatures[i]->getEnumName();
2260      }
2261    } else
2262      OS << "0";
2263
2264    OS << ", " << OMI.CI->Name;
2265
2266    OS << " },\n";
2267  }
2268  OS << "};\n\n";
2269
2270  OS << "const char *const OperandMatchEntry::MnemonicTable =\n";
2271  StringTable.EmitString(OS);
2272  OS << ";\n\n";
2273
2274  // Emit the operand class switch to call the correct custom parser for
2275  // the found operand class.
2276  OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2277     << Target.getName() << ClassName << "::\n"
2278     << "tryCustomParseOperand(SmallVectorImpl<MCParsedAsmOperand*>"
2279     << " &Operands,\n                      unsigned MCK) {\n\n"
2280     << "  switch(MCK) {\n";
2281
2282  for (std::vector<ClassInfo*>::const_iterator it = Info.Classes.begin(),
2283       ie = Info.Classes.end(); it != ie; ++it) {
2284    ClassInfo *CI = *it;
2285    if (CI->ParserMethod.empty())
2286      continue;
2287    OS << "  case " << CI->Name << ":\n"
2288       << "    return " << CI->ParserMethod << "(Operands);\n";
2289  }
2290
2291  OS << "  default:\n";
2292  OS << "    return MatchOperand_NoMatch;\n";
2293  OS << "  }\n";
2294  OS << "  return MatchOperand_NoMatch;\n";
2295  OS << "}\n\n";
2296
2297  // Emit the static custom operand parser. This code is very similar with
2298  // the other matcher. Also use MatchResultTy here just in case we go for
2299  // a better error handling.
2300  OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2301     << Target.getName() << ClassName << "::\n"
2302     << "MatchOperandParserImpl(SmallVectorImpl<MCParsedAsmOperand*>"
2303     << " &Operands,\n                       StringRef Mnemonic) {\n";
2304
2305  // Emit code to get the available features.
2306  OS << "  // Get the current feature set.\n";
2307  OS << "  unsigned AvailableFeatures = getAvailableFeatures();\n\n";
2308
2309  OS << "  // Get the next operand index.\n";
2310  OS << "  unsigned NextOpNum = Operands.size()-1;\n";
2311
2312  // Emit code to search the table.
2313  OS << "  // Search the table.\n";
2314  OS << "  std::pair<const OperandMatchEntry*, const OperandMatchEntry*>";
2315  OS << " MnemonicRange =\n";
2316  OS << "    std::equal_range(OperandMatchTable, OperandMatchTable+"
2317     << Info.OperandMatchInfo.size() << ", Mnemonic,\n"
2318     << "                     LessOpcodeOperand());\n\n";
2319
2320  OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2321  OS << "    return MatchOperand_NoMatch;\n\n";
2322
2323  OS << "  for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2324     << "       *ie = MnemonicRange.second; it != ie; ++it) {\n";
2325
2326  OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2327  OS << "    assert(Mnemonic == it->getMnemonic());\n\n";
2328
2329  // Emit check that the required features are available.
2330  OS << "    // check if the available features match\n";
2331  OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2332     << "!= it->RequiredFeatures) {\n";
2333  OS << "      continue;\n";
2334  OS << "    }\n\n";
2335
2336  // Emit check to ensure the operand number matches.
2337  OS << "    // check if the operand in question has a custom parser.\n";
2338  OS << "    if (!(it->OperandMask & (1 << NextOpNum)))\n";
2339  OS << "      continue;\n\n";
2340
2341  // Emit call to the custom parser method
2342  OS << "    // call custom parse method to handle the operand\n";
2343  OS << "    OperandMatchResultTy Result = ";
2344  OS << "tryCustomParseOperand(Operands, it->Class);\n";
2345  OS << "    if (Result != MatchOperand_NoMatch)\n";
2346  OS << "      return Result;\n";
2347  OS << "  }\n\n";
2348
2349  OS << "  // Okay, we had no match.\n";
2350  OS << "  return MatchOperand_NoMatch;\n";
2351  OS << "}\n\n";
2352}
2353
2354void AsmMatcherEmitter::run(raw_ostream &OS) {
2355  CodeGenTarget Target(Records);
2356  Record *AsmParser = Target.getAsmParser();
2357  std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2358
2359  // Compute the information on the instructions to match.
2360  AsmMatcherInfo Info(AsmParser, Target, Records);
2361  Info.buildInfo();
2362
2363  // Sort the instruction table using the partial order on classes. We use
2364  // stable_sort to ensure that ambiguous instructions are still
2365  // deterministically ordered.
2366  std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
2367                   less_ptr<MatchableInfo>());
2368
2369  DEBUG_WITH_TYPE("instruction_info", {
2370      for (std::vector<MatchableInfo*>::iterator
2371             it = Info.Matchables.begin(), ie = Info.Matchables.end();
2372           it != ie; ++it)
2373        (*it)->dump();
2374    });
2375
2376  // Check for ambiguous matchables.
2377  DEBUG_WITH_TYPE("ambiguous_instrs", {
2378    unsigned NumAmbiguous = 0;
2379    for (unsigned i = 0, e = Info.Matchables.size(); i != e; ++i) {
2380      for (unsigned j = i + 1; j != e; ++j) {
2381        MatchableInfo &A = *Info.Matchables[i];
2382        MatchableInfo &B = *Info.Matchables[j];
2383
2384        if (A.couldMatchAmbiguouslyWith(B)) {
2385          errs() << "warning: ambiguous matchables:\n";
2386          A.dump();
2387          errs() << "\nis incomparable with:\n";
2388          B.dump();
2389          errs() << "\n\n";
2390          ++NumAmbiguous;
2391        }
2392      }
2393    }
2394    if (NumAmbiguous)
2395      errs() << "warning: " << NumAmbiguous
2396             << " ambiguous matchables!\n";
2397  });
2398
2399  // Compute the information on the custom operand parsing.
2400  Info.buildOperandMatchInfo();
2401
2402  // Write the output.
2403
2404  // Information for the class declaration.
2405  OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2406  OS << "#undef GET_ASSEMBLER_HEADER\n";
2407  OS << "  // This should be included into the middle of the declaration of\n";
2408  OS << "  // your subclasses implementation of MCTargetAsmParser.\n";
2409  OS << "  unsigned ComputeAvailableFeatures(uint64_t FeatureBits) const;\n";
2410  OS << "  bool ConvertToMCInst(unsigned Kind, MCInst &Inst, "
2411     << "unsigned Opcode,\n"
2412     << "                       const SmallVectorImpl<MCParsedAsmOperand*> "
2413     << "&Operands);\n";
2414  OS << "  bool MnemonicIsValid(StringRef Mnemonic);\n";
2415  OS << "  unsigned MatchInstructionImpl(\n";
2416  OS << "    const SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2417  OS << "    MCInst &Inst, unsigned &ErrorInfo, unsigned VariantID = 0);\n";
2418
2419  if (Info.OperandMatchInfo.size()) {
2420    OS << "\n  enum OperandMatchResultTy {\n";
2421    OS << "    MatchOperand_Success,    // operand matched successfully\n";
2422    OS << "    MatchOperand_NoMatch,    // operand did not match\n";
2423    OS << "    MatchOperand_ParseFail   // operand matched but had errors\n";
2424    OS << "  };\n";
2425    OS << "  OperandMatchResultTy MatchOperandParserImpl(\n";
2426    OS << "    SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2427    OS << "    StringRef Mnemonic);\n";
2428
2429    OS << "  OperandMatchResultTy tryCustomParseOperand(\n";
2430    OS << "    SmallVectorImpl<MCParsedAsmOperand*> &Operands,\n";
2431    OS << "    unsigned MCK);\n\n";
2432  }
2433
2434  OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2435
2436  // Emit the operand match diagnostic enum names.
2437  OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
2438  OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2439  emitOperandDiagnosticTypes(Info, OS);
2440  OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2441
2442
2443  OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2444  OS << "#undef GET_REGISTER_MATCHER\n\n";
2445
2446  // Emit the subtarget feature enumeration.
2447  emitSubtargetFeatureFlagEnumeration(Info, OS);
2448
2449  // Emit the function to match a register name to number.
2450  emitMatchRegisterName(Target, AsmParser, OS);
2451
2452  OS << "#endif // GET_REGISTER_MATCHER\n\n";
2453
2454  OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
2455  OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
2456
2457  // Generate the helper function to get the names for subtarget features.
2458  emitGetSubtargetFeatureName(Info, OS);
2459
2460  OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
2461
2462  OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2463  OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2464
2465  // Generate the function that remaps for mnemonic aliases.
2466  bool HasMnemonicAliases = emitMnemonicAliases(OS, Info);
2467
2468  // Generate the unified function to convert operands into an MCInst.
2469  emitConvertToMCInst(Target, ClassName, Info.Matchables, OS);
2470
2471  // Emit the enumeration for classes which participate in matching.
2472  emitMatchClassEnumeration(Target, Info.Classes, OS);
2473
2474  // Emit the routine to match token strings to their match class.
2475  emitMatchTokenString(Target, Info.Classes, OS);
2476
2477  // Emit the subclass predicate routine.
2478  emitIsSubclass(Target, Info.Classes, OS);
2479
2480  // Emit the routine to validate an operand against a match class.
2481  emitValidateOperandClass(Info, OS);
2482
2483  // Emit the available features compute function.
2484  emitComputeAvailableFeatures(Info, OS);
2485
2486
2487  size_t MaxNumOperands = 0;
2488  for (std::vector<MatchableInfo*>::const_iterator it =
2489         Info.Matchables.begin(), ie = Info.Matchables.end();
2490       it != ie; ++it)
2491    MaxNumOperands = std::max(MaxNumOperands, (*it)->AsmOperands.size());
2492
2493  // Emit the static match table; unused classes get initalized to 0 which is
2494  // guaranteed to be InvalidMatchClass.
2495  //
2496  // FIXME: We can reduce the size of this table very easily. First, we change
2497  // it so that store the kinds in separate bit-fields for each index, which
2498  // only needs to be the max width used for classes at that index (we also need
2499  // to reject based on this during classification). If we then make sure to
2500  // order the match kinds appropriately (putting mnemonics last), then we
2501  // should only end up using a few bits for each class, especially the ones
2502  // following the mnemonic.
2503  OS << "namespace {\n";
2504  OS << "  struct MatchEntry {\n";
2505  OS << "    static const char *const MnemonicTable;\n";
2506  OS << "    uint32_t Mnemonic;\n";
2507  OS << "    uint16_t Opcode;\n";
2508  OS << "    " << getMinimalTypeForRange(Info.Matchables.size())
2509               << " ConvertFn;\n";
2510  OS << "    " << getMinimalTypeForRange(1ULL << Info.SubtargetFeatures.size())
2511               << " RequiredFeatures;\n";
2512  OS << "    " << getMinimalTypeForRange(Info.Classes.size())
2513               << " Classes[" << MaxNumOperands << "];\n";
2514  OS << "    uint8_t AsmVariantID;\n\n";
2515  OS << "    StringRef getMnemonic() const {\n";
2516  OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2517  OS << "                       MnemonicTable[Mnemonic]);\n";
2518  OS << "    }\n";
2519  OS << "  };\n\n";
2520
2521  OS << "  // Predicate for searching for an opcode.\n";
2522  OS << "  struct LessOpcode {\n";
2523  OS << "    bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
2524  OS << "      return LHS.getMnemonic() < RHS;\n";
2525  OS << "    }\n";
2526  OS << "    bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
2527  OS << "      return LHS < RHS.getMnemonic();\n";
2528  OS << "    }\n";
2529  OS << "    bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
2530  OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2531  OS << "    }\n";
2532  OS << "  };\n";
2533
2534  OS << "} // end anonymous namespace.\n\n";
2535
2536  StringToOffsetTable StringTable;
2537
2538  OS << "static const MatchEntry MatchTable["
2539     << Info.Matchables.size() << "] = {\n";
2540
2541  for (std::vector<MatchableInfo*>::const_iterator it =
2542       Info.Matchables.begin(), ie = Info.Matchables.end();
2543       it != ie; ++it) {
2544    MatchableInfo &II = **it;
2545
2546    // Store a pascal-style length byte in the mnemonic.
2547    std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2548    OS << "  { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2549       << " /* " << II.Mnemonic << " */, "
2550       << Target.getName() << "::"
2551       << II.getResultInst()->TheDef->getName() << ", "
2552       << II.ConversionFnKind << ", ";
2553
2554    // Write the required features mask.
2555    if (!II.RequiredFeatures.empty()) {
2556      for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2557        if (i) OS << "|";
2558        OS << II.RequiredFeatures[i]->getEnumName();
2559      }
2560    } else
2561      OS << "0";
2562
2563    OS << ", { ";
2564    for (unsigned i = 0, e = II.AsmOperands.size(); i != e; ++i) {
2565      MatchableInfo::AsmOperand &Op = II.AsmOperands[i];
2566
2567      if (i) OS << ", ";
2568      OS << Op.Class->Name;
2569    }
2570    OS << " }, " << II.AsmVariantID;
2571    OS << "},\n";
2572  }
2573
2574  OS << "};\n\n";
2575
2576  OS << "const char *const MatchEntry::MnemonicTable =\n";
2577  StringTable.EmitString(OS);
2578  OS << ";\n\n";
2579
2580  // A method to determine if a mnemonic is in the list.
2581  OS << "bool " << Target.getName() << ClassName << "::\n"
2582     << "MnemonicIsValid(StringRef Mnemonic) {\n";
2583  OS << "  // Search the table.\n";
2584  OS << "  std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2585  OS << "    std::equal_range(MatchTable, MatchTable+"
2586     << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n";
2587  OS << "  return MnemonicRange.first != MnemonicRange.second;\n";
2588  OS << "}\n\n";
2589
2590  // Finally, build the match function.
2591  OS << "unsigned "
2592     << Target.getName() << ClassName << "::\n"
2593     << "MatchInstructionImpl(const SmallVectorImpl<MCParsedAsmOperand*>"
2594     << " &Operands,\n";
2595  OS << "                     MCInst &Inst, unsigned &ErrorInfo, ";
2596  OS << "unsigned VariantID) {\n";
2597
2598  // Emit code to get the available features.
2599  OS << "  // Get the current feature set.\n";
2600  OS << "  unsigned AvailableFeatures = getAvailableFeatures();\n\n";
2601
2602  OS << "  // Get the instruction mnemonic, which is the first token.\n";
2603  OS << "  StringRef Mnemonic = ((" << Target.getName()
2604     << "Operand*)Operands[0])->getToken();\n\n";
2605
2606  if (HasMnemonicAliases) {
2607    OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
2608    OS << "  // FIXME : Add an entry in AsmParserVariant to check this.\n";
2609    OS << "  if (!VariantID)\n";
2610    OS << "    applyMnemonicAliases(Mnemonic, AvailableFeatures);\n\n";
2611  }
2612
2613  // Emit code to compute the class list for this operand vector.
2614  OS << "  // Eliminate obvious mismatches.\n";
2615  OS << "  if (Operands.size() > " << (MaxNumOperands+1) << ") {\n";
2616  OS << "    ErrorInfo = " << (MaxNumOperands+1) << ";\n";
2617  OS << "    return Match_InvalidOperand;\n";
2618  OS << "  }\n\n";
2619
2620  OS << "  // Some state to try to produce better error messages.\n";
2621  OS << "  bool HadMatchOtherThanFeatures = false;\n";
2622  OS << "  bool HadMatchOtherThanPredicate = false;\n";
2623  OS << "  unsigned RetCode = Match_InvalidOperand;\n";
2624  OS << "  unsigned MissingFeatures = ~0U;\n";
2625  OS << "  // Set ErrorInfo to the operand that mismatches if it is\n";
2626  OS << "  // wrong for all instances of the instruction.\n";
2627  OS << "  ErrorInfo = ~0U;\n";
2628
2629  // Emit code to search the table.
2630  OS << "  // Search the table.\n";
2631  OS << "  std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2632  OS << "    std::equal_range(MatchTable, MatchTable+"
2633     << Info.Matchables.size() << ", Mnemonic, LessOpcode());\n\n";
2634
2635  OS << "  // Return a more specific error code if no mnemonics match.\n";
2636  OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2637  OS << "    return Match_MnemonicFail;\n\n";
2638
2639  OS << "  for (const MatchEntry *it = MnemonicRange.first, "
2640     << "*ie = MnemonicRange.second;\n";
2641  OS << "       it != ie; ++it) {\n";
2642
2643  OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2644  OS << "    assert(Mnemonic == it->getMnemonic());\n";
2645
2646  // Emit check that the subclasses match.
2647  OS << "    if (VariantID != it->AsmVariantID) continue;\n";
2648  OS << "    bool OperandsValid = true;\n";
2649  OS << "    for (unsigned i = 0; i != " << MaxNumOperands << "; ++i) {\n";
2650  OS << "      if (i + 1 >= Operands.size()) {\n";
2651  OS << "        OperandsValid = (it->Classes[i] == " <<"InvalidMatchClass);\n";
2652  OS << "        if (!OperandsValid) ErrorInfo = i + 1;\n";
2653  OS << "        break;\n";
2654  OS << "      }\n";
2655  OS << "      unsigned Diag = validateOperandClass(Operands[i+1],\n";
2656  OS.indent(43);
2657  OS << "(MatchClassKind)it->Classes[i]);\n";
2658  OS << "      if (Diag == Match_Success)\n";
2659  OS << "        continue;\n";
2660  OS << "      // If this operand is broken for all of the instances of this\n";
2661  OS << "      // mnemonic, keep track of it so we can report loc info.\n";
2662  OS << "      // If we already had a match that only failed due to a\n";
2663  OS << "      // target predicate, that diagnostic is preferred.\n";
2664  OS << "      if (!HadMatchOtherThanPredicate &&\n";
2665  OS << "          (it == MnemonicRange.first || ErrorInfo <= i+1)) {\n";
2666  OS << "        ErrorInfo = i+1;\n";
2667  OS << "        // InvalidOperand is the default. Prefer specificity.\n";
2668  OS << "        if (Diag != Match_InvalidOperand)\n";
2669  OS << "          RetCode = Diag;\n";
2670  OS << "      }\n";
2671  OS << "      // Otherwise, just reject this instance of the mnemonic.\n";
2672  OS << "      OperandsValid = false;\n";
2673  OS << "      break;\n";
2674  OS << "    }\n\n";
2675
2676  OS << "    if (!OperandsValid) continue;\n";
2677
2678  // Emit check that the required features are available.
2679  OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2680     << "!= it->RequiredFeatures) {\n";
2681  OS << "      HadMatchOtherThanFeatures = true;\n";
2682  OS << "      unsigned NewMissingFeatures = it->RequiredFeatures & "
2683        "~AvailableFeatures;\n";
2684  OS << "      if (CountPopulation_32(NewMissingFeatures) <= "
2685        "CountPopulation_32(MissingFeatures))\n";
2686  OS << "        MissingFeatures = NewMissingFeatures;\n";
2687  OS << "      continue;\n";
2688  OS << "    }\n";
2689  OS << "\n";
2690  OS << "    // We have selected a definite instruction, convert the parsed\n"
2691     << "    // operands into the appropriate MCInst.\n";
2692  OS << "    if (!ConvertToMCInst(it->ConvertFn, Inst,\n"
2693     << "                         it->Opcode, Operands))\n";
2694  OS << "      return Match_ConversionFail;\n";
2695  OS << "\n";
2696
2697  // Verify the instruction with the target-specific match predicate function.
2698  OS << "    // We have a potential match. Check the target predicate to\n"
2699     << "    // handle any context sensitive constraints.\n"
2700     << "    unsigned MatchResult;\n"
2701     << "    if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
2702     << " Match_Success) {\n"
2703     << "      Inst.clear();\n"
2704     << "      RetCode = MatchResult;\n"
2705     << "      HadMatchOtherThanPredicate = true;\n"
2706     << "      continue;\n"
2707     << "    }\n\n";
2708
2709  // Call the post-processing function, if used.
2710  std::string InsnCleanupFn =
2711    AsmParser->getValueAsString("AsmParserInstCleanup");
2712  if (!InsnCleanupFn.empty())
2713    OS << "    " << InsnCleanupFn << "(Inst);\n";
2714
2715  OS << "    return Match_Success;\n";
2716  OS << "  }\n\n";
2717
2718  OS << "  // Okay, we had no match.  Try to return a useful error code.\n";
2719  OS << "  if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)";
2720  OS << "  return RetCode;\n";
2721  OS << "  // Missing feature matches return which features were missing\n";
2722  OS << "  ErrorInfo = MissingFeatures;\n";
2723  OS << "  return Match_MissingFeature;\n";
2724  OS << "}\n\n";
2725
2726  if (Info.OperandMatchInfo.size())
2727    emitCustomOperandParsing(OS, Target, Info, ClassName);
2728
2729  OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
2730}
2731
2732namespace llvm {
2733
2734void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
2735  emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
2736  AsmMatcherEmitter(RK).run(OS);
2737}
2738
2739} // End llvm namespace
2740