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 values 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 "llvm/ADT/PointerUnion.h"
101 #include "llvm/ADT/STLExtras.h"
102 #include "llvm/ADT/SmallPtrSet.h"
103 #include "llvm/ADT/SmallVector.h"
104 #include "llvm/ADT/StringExtras.h"
105 #include "llvm/Support/CommandLine.h"
106 #include "llvm/Support/Debug.h"
107 #include "llvm/Support/ErrorHandling.h"
108 #include "llvm/TableGen/Error.h"
109 #include "llvm/TableGen/Record.h"
110 #include "llvm/TableGen/StringMatcher.h"
111 #include "llvm/TableGen/StringToOffsetTable.h"
112 #include "llvm/TableGen/TableGenBackend.h"
113 #include <cassert>
114 #include <cctype>
115 #include <map>
116 #include <set>
117 #include <sstream>
118 #include <forward_list>
119 
120 using namespace llvm;
121 
122 #define DEBUG_TYPE "asm-matcher-emitter"
123 
124 static cl::opt<std::string>
125 MatchPrefix("match-prefix", cl::init(""),
126             cl::desc("Only match instructions with the given prefix"));
127 
128 namespace {
129 class AsmMatcherInfo;
130 struct SubtargetFeatureInfo;
131 
132 // Register sets are used as keys in some second-order sets TableGen creates
133 // when generating its data structures. This means that the order of two
134 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and
135 // can even affect compiler output (at least seen in diagnostics produced when
136 // all matches fail). So we use a type that sorts them consistently.
137 typedef std::set<Record*, LessRecordByID> RegisterSet;
138 
139 class AsmMatcherEmitter {
140   RecordKeeper &Records;
141 public:
142   AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
143 
144   void run(raw_ostream &o);
145 };
146 
147 /// ClassInfo - Helper class for storing the information about a particular
148 /// class of operands which can be matched.
149 struct ClassInfo {
150   enum ClassInfoKind {
151     /// Invalid kind, for use as a sentinel value.
152     Invalid = 0,
153 
154     /// The class for a particular token.
155     Token,
156 
157     /// The (first) register class, subsequent register classes are
158     /// RegisterClass0+1, and so on.
159     RegisterClass0,
160 
161     /// The (first) user defined class, subsequent user defined classes are
162     /// UserClass0+1, and so on.
163     UserClass0 = 1<<16
164   };
165 
166   /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
167   /// N) for the Nth user defined class.
168   unsigned Kind;
169 
170   /// SuperClasses - The super classes of this class. Note that for simplicities
171   /// sake user operands only record their immediate super class, while register
172   /// operands include all superclasses.
173   std::vector<ClassInfo*> SuperClasses;
174 
175   /// Name - The full class name, suitable for use in an enum.
176   std::string Name;
177 
178   /// ClassName - The unadorned generic name for this class (e.g., Token).
179   std::string ClassName;
180 
181   /// ValueName - The name of the value this class represents; for a token this
182   /// is the literal token string, for an operand it is the TableGen class (or
183   /// empty if this is a derived class).
184   std::string ValueName;
185 
186   /// PredicateMethod - The name of the operand method to test whether the
187   /// operand matches this class; this is not valid for Token or register kinds.
188   std::string PredicateMethod;
189 
190   /// RenderMethod - The name of the operand method to add this operand to an
191   /// MCInst; this is not valid for Token or register kinds.
192   std::string RenderMethod;
193 
194   /// ParserMethod - The name of the operand method to do a target specific
195   /// parsing on the operand.
196   std::string ParserMethod;
197 
198   /// For register classes: the records for all the registers in this class.
199   RegisterSet Registers;
200 
201   /// For custom match classes: the diagnostic kind for when the predicate fails.
202   std::string DiagnosticType;
203 
204   /// Is this operand optional and not always required.
205   bool IsOptional;
206 
207 public:
208   /// isRegisterClass() - Check if this is a register class.
209   bool isRegisterClass() const {
210     return Kind >= RegisterClass0 && Kind < UserClass0;
211   }
212 
213   /// isUserClass() - Check if this is a user defined class.
214   bool isUserClass() const {
215     return Kind >= UserClass0;
216   }
217 
218   /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
219   /// are related if they are in the same class hierarchy.
220   bool isRelatedTo(const ClassInfo &RHS) const {
221     // Tokens are only related to tokens.
222     if (Kind == Token || RHS.Kind == Token)
223       return Kind == Token && RHS.Kind == Token;
224 
225     // Registers classes are only related to registers classes, and only if
226     // their intersection is non-empty.
227     if (isRegisterClass() || RHS.isRegisterClass()) {
228       if (!isRegisterClass() || !RHS.isRegisterClass())
229         return false;
230 
231       RegisterSet Tmp;
232       std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin());
233       std::set_intersection(Registers.begin(), Registers.end(),
234                             RHS.Registers.begin(), RHS.Registers.end(),
235                             II, LessRecordByID());
236 
237       return !Tmp.empty();
238     }
239 
240     // Otherwise we have two users operands; they are related if they are in the
241     // same class hierarchy.
242     //
243     // FIXME: This is an oversimplification, they should only be related if they
244     // intersect, however we don't have that information.
245     assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
246     const ClassInfo *Root = this;
247     while (!Root->SuperClasses.empty())
248       Root = Root->SuperClasses.front();
249 
250     const ClassInfo *RHSRoot = &RHS;
251     while (!RHSRoot->SuperClasses.empty())
252       RHSRoot = RHSRoot->SuperClasses.front();
253 
254     return Root == RHSRoot;
255   }
256 
257   /// isSubsetOf - Test whether this class is a subset of \p RHS.
258   bool isSubsetOf(const ClassInfo &RHS) const {
259     // This is a subset of RHS if it is the same class...
260     if (this == &RHS)
261       return true;
262 
263     // ... or if any of its super classes are a subset of RHS.
264     for (const ClassInfo *CI : SuperClasses)
265       if (CI->isSubsetOf(RHS))
266         return true;
267 
268     return false;
269   }
270 
271   int getTreeDepth() const {
272     int Depth = 0;
273     const ClassInfo *Root = this;
274     while (!Root->SuperClasses.empty()) {
275       Depth++;
276       Root = Root->SuperClasses.front();
277     }
278     return Depth;
279   }
280 
281   const ClassInfo *findRoot() const {
282     const ClassInfo *Root = this;
283     while (!Root->SuperClasses.empty())
284       Root = Root->SuperClasses.front();
285     return Root;
286   }
287 
288   /// Compare two classes. This does not produce a total ordering, but does
289   /// guarantee that subclasses are sorted before their parents, and that the
290   /// ordering is transitive.
291   bool operator<(const ClassInfo &RHS) const {
292     if (this == &RHS)
293       return false;
294 
295     // First, enforce the ordering between the three different types of class.
296     // Tokens sort before registers, which sort before user classes.
297     if (Kind == Token) {
298       if (RHS.Kind != Token)
299         return true;
300       assert(RHS.Kind == Token);
301     } else if (isRegisterClass()) {
302       if (RHS.Kind == Token)
303         return false;
304       else if (RHS.isUserClass())
305         return true;
306       assert(RHS.isRegisterClass());
307     } else if (isUserClass()) {
308       if (!RHS.isUserClass())
309         return false;
310       assert(RHS.isUserClass());
311     } else {
312       llvm_unreachable("Unknown ClassInfoKind");
313     }
314 
315     if (Kind == Token || isUserClass()) {
316       // Related tokens and user classes get sorted by depth in the inheritence
317       // tree (so that subclasses are before their parents).
318       if (isRelatedTo(RHS)) {
319         if (getTreeDepth() > RHS.getTreeDepth())
320           return true;
321         if (getTreeDepth() < RHS.getTreeDepth())
322           return false;
323       } else {
324         // Unrelated tokens and user classes are ordered by the name of their
325         // root nodes, so that there is a consistent ordering between
326         // unconnected trees.
327         return findRoot()->ValueName < RHS.findRoot()->ValueName;
328       }
329     } else if (isRegisterClass()) {
330       // For register sets, sort by number of registers. This guarantees that
331       // a set will always sort before all of it's strict supersets.
332       if (Registers.size() != RHS.Registers.size())
333         return Registers.size() < RHS.Registers.size();
334     } else {
335       llvm_unreachable("Unknown ClassInfoKind");
336     }
337 
338     // FIXME: We should be able to just return false here, as we only need a
339     // partial order (we use stable sorts, so this is deterministic) and the
340     // name of a class shouldn't be significant. However, some of the backends
341     // accidentally rely on this behaviour, so it will have to stay like this
342     // until they are fixed.
343     return ValueName < RHS.ValueName;
344   }
345 };
346 
347 class AsmVariantInfo {
348 public:
349   std::string RegisterPrefix;
350   std::string TokenizingCharacters;
351   std::string SeparatorCharacters;
352   std::string BreakCharacters;
353   int AsmVariantNo;
354 };
355 
356 /// MatchableInfo - Helper class for storing the necessary information for an
357 /// instruction or alias which is capable of being matched.
358 struct MatchableInfo {
359   struct AsmOperand {
360     /// Token - This is the token that the operand came from.
361     StringRef Token;
362 
363     /// The unique class instance this operand should match.
364     ClassInfo *Class;
365 
366     /// The operand name this is, if anything.
367     StringRef SrcOpName;
368 
369     /// The suboperand index within SrcOpName, or -1 for the entire operand.
370     int SubOpIdx;
371 
372     /// Whether the token is "isolated", i.e., it is preceded and followed
373     /// by separators.
374     bool IsIsolatedToken;
375 
376     /// Register record if this token is singleton register.
377     Record *SingletonReg;
378 
379     explicit AsmOperand(bool IsIsolatedToken, StringRef T)
380         : Token(T), Class(nullptr), SubOpIdx(-1),
381           IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {}
382   };
383 
384   /// ResOperand - This represents a single operand in the result instruction
385   /// generated by the match.  In cases (like addressing modes) where a single
386   /// assembler operand expands to multiple MCOperands, this represents the
387   /// single assembler operand, not the MCOperand.
388   struct ResOperand {
389     enum {
390       /// RenderAsmOperand - This represents an operand result that is
391       /// generated by calling the render method on the assembly operand.  The
392       /// corresponding AsmOperand is specified by AsmOperandNum.
393       RenderAsmOperand,
394 
395       /// TiedOperand - This represents a result operand that is a duplicate of
396       /// a previous result operand.
397       TiedOperand,
398 
399       /// ImmOperand - This represents an immediate value that is dumped into
400       /// the operand.
401       ImmOperand,
402 
403       /// RegOperand - This represents a fixed register that is dumped in.
404       RegOperand
405     } Kind;
406 
407     union {
408       /// This is the operand # in the AsmOperands list that this should be
409       /// copied from.
410       unsigned AsmOperandNum;
411 
412       /// TiedOperandNum - This is the (earlier) result operand that should be
413       /// copied from.
414       unsigned TiedOperandNum;
415 
416       /// ImmVal - This is the immediate value added to the instruction.
417       int64_t ImmVal;
418 
419       /// Register - This is the register record.
420       Record *Register;
421     };
422 
423     /// MINumOperands - The number of MCInst operands populated by this
424     /// operand.
425     unsigned MINumOperands;
426 
427     static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
428       ResOperand X;
429       X.Kind = RenderAsmOperand;
430       X.AsmOperandNum = AsmOpNum;
431       X.MINumOperands = NumOperands;
432       return X;
433     }
434 
435     static ResOperand getTiedOp(unsigned TiedOperandNum) {
436       ResOperand X;
437       X.Kind = TiedOperand;
438       X.TiedOperandNum = TiedOperandNum;
439       X.MINumOperands = 1;
440       return X;
441     }
442 
443     static ResOperand getImmOp(int64_t Val) {
444       ResOperand X;
445       X.Kind = ImmOperand;
446       X.ImmVal = Val;
447       X.MINumOperands = 1;
448       return X;
449     }
450 
451     static ResOperand getRegOp(Record *Reg) {
452       ResOperand X;
453       X.Kind = RegOperand;
454       X.Register = Reg;
455       X.MINumOperands = 1;
456       return X;
457     }
458   };
459 
460   /// AsmVariantID - Target's assembly syntax variant no.
461   int AsmVariantID;
462 
463   /// AsmString - The assembly string for this instruction (with variants
464   /// removed), e.g. "movsx $src, $dst".
465   std::string AsmString;
466 
467   /// TheDef - This is the definition of the instruction or InstAlias that this
468   /// matchable came from.
469   Record *const TheDef;
470 
471   /// DefRec - This is the definition that it came from.
472   PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
473 
474   const CodeGenInstruction *getResultInst() const {
475     if (DefRec.is<const CodeGenInstruction*>())
476       return DefRec.get<const CodeGenInstruction*>();
477     return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
478   }
479 
480   /// ResOperands - This is the operand list that should be built for the result
481   /// MCInst.
482   SmallVector<ResOperand, 8> ResOperands;
483 
484   /// Mnemonic - This is the first token of the matched instruction, its
485   /// mnemonic.
486   StringRef Mnemonic;
487 
488   /// AsmOperands - The textual operands that this instruction matches,
489   /// annotated with a class and where in the OperandList they were defined.
490   /// This directly corresponds to the tokenized AsmString after the mnemonic is
491   /// removed.
492   SmallVector<AsmOperand, 8> AsmOperands;
493 
494   /// Predicates - The required subtarget features to match this instruction.
495   SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
496 
497   /// ConversionFnKind - The enum value which is passed to the generated
498   /// convertToMCInst to convert parsed operands into an MCInst for this
499   /// function.
500   std::string ConversionFnKind;
501 
502   /// If this instruction is deprecated in some form.
503   bool HasDeprecation;
504 
505   /// If this is an alias, this is use to determine whether or not to using
506   /// the conversion function defined by the instruction's AsmMatchConverter
507   /// or to use the function generated by the alias.
508   bool UseInstAsmMatchConverter;
509 
510   MatchableInfo(const CodeGenInstruction &CGI)
511     : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI),
512       UseInstAsmMatchConverter(true) {
513   }
514 
515   MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias)
516     : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef),
517       DefRec(Alias.release()),
518       UseInstAsmMatchConverter(
519         TheDef->getValueAsBit("UseInstAsmMatchConverter")) {
520   }
521 
522   // Could remove this and the dtor if PointerUnion supported unique_ptr
523   // elements with a dynamic failure/assertion (like the one below) in the case
524   // where it was copied while being in an owning state.
525   MatchableInfo(const MatchableInfo &RHS)
526       : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString),
527         TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands),
528         Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands),
529         RequiredFeatures(RHS.RequiredFeatures),
530         ConversionFnKind(RHS.ConversionFnKind),
531         HasDeprecation(RHS.HasDeprecation),
532         UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) {
533     assert(!DefRec.is<const CodeGenInstAlias *>());
534   }
535 
536   ~MatchableInfo() {
537     delete DefRec.dyn_cast<const CodeGenInstAlias*>();
538   }
539 
540   // Two-operand aliases clone from the main matchable, but mark the second
541   // operand as a tied operand of the first for purposes of the assembler.
542   void formTwoOperandAlias(StringRef Constraint);
543 
544   void initialize(const AsmMatcherInfo &Info,
545                   SmallPtrSetImpl<Record*> &SingletonRegisters,
546                   AsmVariantInfo const &Variant,
547                   bool HasMnemonicFirst);
548 
549   /// validate - Return true if this matchable is a valid thing to match against
550   /// and perform a bunch of validity checking.
551   bool validate(StringRef CommentDelimiter, bool Hack) const;
552 
553   /// findAsmOperand - Find the AsmOperand with the specified name and
554   /// suboperand index.
555   int findAsmOperand(StringRef N, int SubOpIdx) const {
556     auto I = std::find_if(AsmOperands.begin(), AsmOperands.end(),
557                           [&](const AsmOperand &Op) {
558                             return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx;
559                           });
560     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
561   }
562 
563   /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
564   /// This does not check the suboperand index.
565   int findAsmOperandNamed(StringRef N) const {
566     auto I = std::find_if(AsmOperands.begin(), AsmOperands.end(),
567                           [&](const AsmOperand &Op) {
568                             return Op.SrcOpName == N;
569                           });
570     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
571   }
572 
573   void buildInstructionResultOperands();
574   void buildAliasResultOperands();
575 
576   /// operator< - Compare two matchables.
577   bool operator<(const MatchableInfo &RHS) const {
578     // The primary comparator is the instruction mnemonic.
579     if (Mnemonic != RHS.Mnemonic)
580       return Mnemonic < RHS.Mnemonic;
581 
582     if (AsmOperands.size() != RHS.AsmOperands.size())
583       return AsmOperands.size() < RHS.AsmOperands.size();
584 
585     // Compare lexicographically by operand. The matcher validates that other
586     // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
587     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
588       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
589         return true;
590       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
591         return false;
592     }
593 
594     // Give matches that require more features higher precedence. This is useful
595     // because we cannot define AssemblerPredicates with the negation of
596     // processor features. For example, ARM v6 "nop" may be either a HINT or
597     // MOV. With v6, we want to match HINT. The assembler has no way to
598     // predicate MOV under "NoV6", but HINT will always match first because it
599     // requires V6 while MOV does not.
600     if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
601       return RequiredFeatures.size() > RHS.RequiredFeatures.size();
602 
603     return false;
604   }
605 
606   /// couldMatchAmbiguouslyWith - Check whether this matchable could
607   /// ambiguously match the same set of operands as \p RHS (without being a
608   /// strictly superior match).
609   bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const {
610     // The primary comparator is the instruction mnemonic.
611     if (Mnemonic != RHS.Mnemonic)
612       return false;
613 
614     // The number of operands is unambiguous.
615     if (AsmOperands.size() != RHS.AsmOperands.size())
616       return false;
617 
618     // Otherwise, make sure the ordering of the two instructions is unambiguous
619     // by checking that either (a) a token or operand kind discriminates them,
620     // or (b) the ordering among equivalent kinds is consistent.
621 
622     // Tokens and operand kinds are unambiguous (assuming a correct target
623     // specific parser).
624     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
625       if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
626           AsmOperands[i].Class->Kind == ClassInfo::Token)
627         if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
628             *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
629           return false;
630 
631     // Otherwise, this operand could commute if all operands are equivalent, or
632     // there is a pair of operands that compare less than and a pair that
633     // compare greater than.
634     bool HasLT = false, HasGT = false;
635     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
636       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
637         HasLT = true;
638       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
639         HasGT = true;
640     }
641 
642     return HasLT == HasGT;
643   }
644 
645   void dump() const;
646 
647 private:
648   void tokenizeAsmString(AsmMatcherInfo const &Info,
649                          AsmVariantInfo const &Variant);
650   void addAsmOperand(StringRef Token, bool IsIsolatedToken = false);
651 };
652 
653 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
654 /// feature which participates in instruction matching.
655 struct SubtargetFeatureInfo {
656   /// \brief The predicate record for this feature.
657   Record *TheDef;
658 
659   /// \brief An unique index assigned to represent this feature.
660   uint64_t Index;
661 
662   SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {}
663 
664   /// \brief The name of the enumerated constant identifying this feature.
665   std::string getEnumName() const {
666     return "Feature_" + TheDef->getName();
667   }
668 
669   void dump() const {
670     errs() << getEnumName() << " " << Index << "\n";
671     TheDef->dump();
672   }
673 };
674 
675 struct OperandMatchEntry {
676   unsigned OperandMask;
677   const MatchableInfo* MI;
678   ClassInfo *CI;
679 
680   static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
681                                   unsigned opMask) {
682     OperandMatchEntry X;
683     X.OperandMask = opMask;
684     X.CI = ci;
685     X.MI = mi;
686     return X;
687   }
688 };
689 
690 class AsmMatcherInfo {
691 public:
692   /// Tracked Records
693   RecordKeeper &Records;
694 
695   /// The tablegen AsmParser record.
696   Record *AsmParser;
697 
698   /// Target - The target information.
699   CodeGenTarget &Target;
700 
701   /// The classes which are needed for matching.
702   std::forward_list<ClassInfo> Classes;
703 
704   /// The information on the matchables to match.
705   std::vector<std::unique_ptr<MatchableInfo>> Matchables;
706 
707   /// Info for custom matching operands by user defined methods.
708   std::vector<OperandMatchEntry> OperandMatchInfo;
709 
710   /// Map of Register records to their class information.
711   typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy;
712   RegisterClassesTy RegisterClasses;
713 
714   /// Map of Predicate records to their subtarget information.
715   std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
716 
717   /// Map of AsmOperandClass records to their class information.
718   std::map<Record*, ClassInfo*> AsmOperandClasses;
719 
720 private:
721   /// Map of token to class information which has already been constructed.
722   std::map<std::string, ClassInfo*> TokenClasses;
723 
724   /// Map of RegisterClass records to their class information.
725   std::map<Record*, ClassInfo*> RegisterClassClasses;
726 
727 private:
728   /// getTokenClass - Lookup or create the class for the given token.
729   ClassInfo *getTokenClass(StringRef Token);
730 
731   /// getOperandClass - Lookup or create the class for the given operand.
732   ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
733                              int SubOpIdx);
734   ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
735 
736   /// buildRegisterClasses - Build the ClassInfo* instances for register
737   /// classes.
738   void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters);
739 
740   /// buildOperandClasses - Build the ClassInfo* instances for user defined
741   /// operand classes.
742   void buildOperandClasses();
743 
744   void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
745                                         unsigned AsmOpIdx);
746   void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
747                                   MatchableInfo::AsmOperand &Op);
748 
749 public:
750   AsmMatcherInfo(Record *AsmParser,
751                  CodeGenTarget &Target,
752                  RecordKeeper &Records);
753 
754   /// buildInfo - Construct the various tables used during matching.
755   void buildInfo();
756 
757   /// buildOperandMatchInfo - Build the necessary information to handle user
758   /// defined operand parsing methods.
759   void buildOperandMatchInfo();
760 
761   /// getSubtargetFeature - Lookup or create the subtarget feature info for the
762   /// given operand.
763   const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
764     assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
765     const auto &I = SubtargetFeatures.find(Def);
766     return I == SubtargetFeatures.end() ? nullptr : &I->second;
767   }
768 
769   RecordKeeper &getRecords() const {
770     return Records;
771   }
772 };
773 
774 } // end anonymous namespace
775 
776 void MatchableInfo::dump() const {
777   errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
778 
779   for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
780     const AsmOperand &Op = AsmOperands[i];
781     errs() << "  op[" << i << "] = " << Op.Class->ClassName << " - ";
782     errs() << '\"' << Op.Token << "\"\n";
783   }
784 }
785 
786 static std::pair<StringRef, StringRef>
787 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
788   // Split via the '='.
789   std::pair<StringRef, StringRef> Ops = S.split('=');
790   if (Ops.second == "")
791     PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
792   // Trim whitespace and the leading '$' on the operand names.
793   size_t start = Ops.first.find_first_of('$');
794   if (start == std::string::npos)
795     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
796   Ops.first = Ops.first.slice(start + 1, std::string::npos);
797   size_t end = Ops.first.find_last_of(" \t");
798   Ops.first = Ops.first.slice(0, end);
799   // Now the second operand.
800   start = Ops.second.find_first_of('$');
801   if (start == std::string::npos)
802     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
803   Ops.second = Ops.second.slice(start + 1, std::string::npos);
804   end = Ops.second.find_last_of(" \t");
805   Ops.first = Ops.first.slice(0, end);
806   return Ops;
807 }
808 
809 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
810   // Figure out which operands are aliased and mark them as tied.
811   std::pair<StringRef, StringRef> Ops =
812     parseTwoOperandConstraint(Constraint, TheDef->getLoc());
813 
814   // Find the AsmOperands that refer to the operands we're aliasing.
815   int SrcAsmOperand = findAsmOperandNamed(Ops.first);
816   int DstAsmOperand = findAsmOperandNamed(Ops.second);
817   if (SrcAsmOperand == -1)
818     PrintFatalError(TheDef->getLoc(),
819                     "unknown source two-operand alias operand '" + Ops.first +
820                     "'.");
821   if (DstAsmOperand == -1)
822     PrintFatalError(TheDef->getLoc(),
823                     "unknown destination two-operand alias operand '" +
824                     Ops.second + "'.");
825 
826   // Find the ResOperand that refers to the operand we're aliasing away
827   // and update it to refer to the combined operand instead.
828   for (ResOperand &Op : ResOperands) {
829     if (Op.Kind == ResOperand::RenderAsmOperand &&
830         Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
831       Op.AsmOperandNum = DstAsmOperand;
832       break;
833     }
834   }
835   // Remove the AsmOperand for the alias operand.
836   AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
837   // Adjust the ResOperand references to any AsmOperands that followed
838   // the one we just deleted.
839   for (ResOperand &Op : ResOperands) {
840     switch(Op.Kind) {
841     default:
842       // Nothing to do for operands that don't reference AsmOperands.
843       break;
844     case ResOperand::RenderAsmOperand:
845       if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
846         --Op.AsmOperandNum;
847       break;
848     case ResOperand::TiedOperand:
849       if (Op.TiedOperandNum > (unsigned)SrcAsmOperand)
850         --Op.TiedOperandNum;
851       break;
852     }
853   }
854 }
855 
856 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
857 /// if present, from specified token.
858 static void
859 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op,
860                                       const AsmMatcherInfo &Info,
861                                       StringRef RegisterPrefix) {
862   StringRef Tok = Op.Token;
863 
864   // If this token is not an isolated token, i.e., it isn't separated from
865   // other tokens (e.g. with whitespace), don't interpret it as a register name.
866   if (!Op.IsIsolatedToken)
867     return;
868 
869   if (RegisterPrefix.empty()) {
870     std::string LoweredTok = Tok.lower();
871     if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
872       Op.SingletonReg = Reg->TheDef;
873     return;
874   }
875 
876   if (!Tok.startswith(RegisterPrefix))
877     return;
878 
879   StringRef RegName = Tok.substr(RegisterPrefix.size());
880   if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
881     Op.SingletonReg = Reg->TheDef;
882 
883   // If there is no register prefix (i.e. "%" in "%eax"), then this may
884   // be some random non-register token, just ignore it.
885 }
886 
887 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
888                                SmallPtrSetImpl<Record*> &SingletonRegisters,
889                                AsmVariantInfo const &Variant,
890                                bool HasMnemonicFirst) {
891   AsmVariantID = Variant.AsmVariantNo;
892   AsmString =
893     CodeGenInstruction::FlattenAsmStringVariants(AsmString,
894                                                  Variant.AsmVariantNo);
895 
896   tokenizeAsmString(Info, Variant);
897 
898   // The first token of the instruction is the mnemonic, which must be a
899   // simple string, not a $foo variable or a singleton register.
900   if (AsmOperands.empty())
901     PrintFatalError(TheDef->getLoc(),
902                   "Instruction '" + TheDef->getName() + "' has no tokens");
903 
904   assert(!AsmOperands[0].Token.empty());
905   if (HasMnemonicFirst) {
906     Mnemonic = AsmOperands[0].Token;
907     if (Mnemonic[0] == '$')
908       PrintFatalError(TheDef->getLoc(),
909                       "Invalid instruction mnemonic '" + Mnemonic + "'!");
910 
911     // Remove the first operand, it is tracked in the mnemonic field.
912     AsmOperands.erase(AsmOperands.begin());
913   } else if (AsmOperands[0].Token[0] != '$')
914     Mnemonic = AsmOperands[0].Token;
915 
916   // Compute the require features.
917   for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates"))
918     if (const SubtargetFeatureInfo *Feature =
919             Info.getSubtargetFeature(Predicate))
920       RequiredFeatures.push_back(Feature);
921 
922   // Collect singleton registers, if used.
923   for (MatchableInfo::AsmOperand &Op : AsmOperands) {
924     extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix);
925     if (Record *Reg = Op.SingletonReg)
926       SingletonRegisters.insert(Reg);
927   }
928 
929   const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
930   if (!DepMask)
931     DepMask = TheDef->getValue("ComplexDeprecationPredicate");
932 
933   HasDeprecation =
934       DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
935 }
936 
937 /// Append an AsmOperand for the given substring of AsmString.
938 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) {
939   AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token));
940 }
941 
942 /// tokenizeAsmString - Tokenize a simplified assembly string.
943 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info,
944                                       AsmVariantInfo const &Variant) {
945   StringRef String = AsmString;
946   size_t Prev = 0;
947   bool InTok = false;
948   bool IsIsolatedToken = true;
949   for (size_t i = 0, e = String.size(); i != e; ++i) {
950     char Char = String[i];
951     if (Variant.BreakCharacters.find(Char) != std::string::npos) {
952       if (InTok) {
953         addAsmOperand(String.slice(Prev, i), false);
954         Prev = i;
955         IsIsolatedToken = false;
956       }
957       InTok = true;
958       continue;
959     }
960     if (Variant.TokenizingCharacters.find(Char) != std::string::npos) {
961       if (InTok) {
962         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
963         InTok = false;
964         IsIsolatedToken = false;
965       }
966       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
967       Prev = i + 1;
968       IsIsolatedToken = true;
969       continue;
970     }
971     if (Variant.SeparatorCharacters.find(Char) != std::string::npos) {
972       if (InTok) {
973         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
974         InTok = false;
975       }
976       Prev = i + 1;
977       IsIsolatedToken = true;
978       continue;
979     }
980 
981     switch (Char) {
982     case '\\':
983       if (InTok) {
984         addAsmOperand(String.slice(Prev, i), false);
985         InTok = false;
986         IsIsolatedToken = false;
987       }
988       ++i;
989       assert(i != String.size() && "Invalid quoted character");
990       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
991       Prev = i + 1;
992       IsIsolatedToken = false;
993       break;
994 
995     case '$': {
996       if (InTok) {
997         addAsmOperand(String.slice(Prev, i), false);
998         InTok = false;
999         IsIsolatedToken = false;
1000       }
1001 
1002       // If this isn't "${", start new identifier looking like "$xxx"
1003       if (i + 1 == String.size() || String[i + 1] != '{') {
1004         Prev = i;
1005         break;
1006       }
1007 
1008       size_t EndPos = String.find('}', i);
1009       assert(EndPos != StringRef::npos &&
1010              "Missing brace in operand reference!");
1011       addAsmOperand(String.slice(i, EndPos+1), IsIsolatedToken);
1012       Prev = EndPos + 1;
1013       i = EndPos;
1014       IsIsolatedToken = false;
1015       break;
1016     }
1017 
1018     default:
1019       InTok = true;
1020       break;
1021     }
1022   }
1023   if (InTok && Prev != String.size())
1024     addAsmOperand(String.substr(Prev), IsIsolatedToken);
1025 }
1026 
1027 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const {
1028   // Reject matchables with no .s string.
1029   if (AsmString.empty())
1030     PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
1031 
1032   // Reject any matchables with a newline in them, they should be marked
1033   // isCodeGenOnly if they are pseudo instructions.
1034   if (AsmString.find('\n') != std::string::npos)
1035     PrintFatalError(TheDef->getLoc(),
1036                   "multiline instruction is not valid for the asmparser, "
1037                   "mark it isCodeGenOnly");
1038 
1039   // Remove comments from the asm string.  We know that the asmstring only
1040   // has one line.
1041   if (!CommentDelimiter.empty() &&
1042       StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
1043     PrintFatalError(TheDef->getLoc(),
1044                   "asmstring for instruction has comment character in it, "
1045                   "mark it isCodeGenOnly");
1046 
1047   // Reject matchables with operand modifiers, these aren't something we can
1048   // handle, the target should be refactored to use operands instead of
1049   // modifiers.
1050   //
1051   // Also, check for instructions which reference the operand multiple times;
1052   // this implies a constraint we would not honor.
1053   std::set<std::string> OperandNames;
1054   for (const AsmOperand &Op : AsmOperands) {
1055     StringRef Tok = Op.Token;
1056     if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
1057       PrintFatalError(TheDef->getLoc(),
1058                       "matchable with operand modifier '" + Tok +
1059                       "' not supported by asm matcher.  Mark isCodeGenOnly!");
1060 
1061     // Verify that any operand is only mentioned once.
1062     // We reject aliases and ignore instructions for now.
1063     if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
1064       if (!Hack)
1065         PrintFatalError(TheDef->getLoc(),
1066                         "ERROR: matchable with tied operand '" + Tok +
1067                         "' can never be matched!");
1068       // FIXME: Should reject these.  The ARM backend hits this with $lane in a
1069       // bunch of instructions.  It is unclear what the right answer is.
1070       DEBUG({
1071         errs() << "warning: '" << TheDef->getName() << "': "
1072                << "ignoring instruction with tied operand '"
1073                << Tok << "'\n";
1074       });
1075       return false;
1076     }
1077   }
1078 
1079   return true;
1080 }
1081 
1082 static std::string getEnumNameForToken(StringRef Str) {
1083   std::string Res;
1084 
1085   for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
1086     switch (*it) {
1087     case '*': Res += "_STAR_"; break;
1088     case '%': Res += "_PCT_"; break;
1089     case ':': Res += "_COLON_"; break;
1090     case '!': Res += "_EXCLAIM_"; break;
1091     case '.': Res += "_DOT_"; break;
1092     case '<': Res += "_LT_"; break;
1093     case '>': Res += "_GT_"; break;
1094     case '-': Res += "_MINUS_"; break;
1095     default:
1096       if ((*it >= 'A' && *it <= 'Z') ||
1097           (*it >= 'a' && *it <= 'z') ||
1098           (*it >= '0' && *it <= '9'))
1099         Res += *it;
1100       else
1101         Res += "_" + utostr((unsigned) *it) + "_";
1102     }
1103   }
1104 
1105   return Res;
1106 }
1107 
1108 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
1109   ClassInfo *&Entry = TokenClasses[Token];
1110 
1111   if (!Entry) {
1112     Classes.emplace_front();
1113     Entry = &Classes.front();
1114     Entry->Kind = ClassInfo::Token;
1115     Entry->ClassName = "Token";
1116     Entry->Name = "MCK_" + getEnumNameForToken(Token);
1117     Entry->ValueName = Token;
1118     Entry->PredicateMethod = "<invalid>";
1119     Entry->RenderMethod = "<invalid>";
1120     Entry->ParserMethod = "";
1121     Entry->DiagnosticType = "";
1122     Entry->IsOptional = false;
1123   }
1124 
1125   return Entry;
1126 }
1127 
1128 ClassInfo *
1129 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1130                                 int SubOpIdx) {
1131   Record *Rec = OI.Rec;
1132   if (SubOpIdx != -1)
1133     Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1134   return getOperandClass(Rec, SubOpIdx);
1135 }
1136 
1137 ClassInfo *
1138 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1139   if (Rec->isSubClassOf("RegisterOperand")) {
1140     // RegisterOperand may have an associated ParserMatchClass. If it does,
1141     // use it, else just fall back to the underlying register class.
1142     const RecordVal *R = Rec->getValue("ParserMatchClass");
1143     if (!R || !R->getValue())
1144       PrintFatalError("Record `" + Rec->getName() +
1145         "' does not have a ParserMatchClass!\n");
1146 
1147     if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) {
1148       Record *MatchClass = DI->getDef();
1149       if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1150         return CI;
1151     }
1152 
1153     // No custom match class. Just use the register class.
1154     Record *ClassRec = Rec->getValueAsDef("RegClass");
1155     if (!ClassRec)
1156       PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1157                     "' has no associated register class!\n");
1158     if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1159       return CI;
1160     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1161   }
1162 
1163   if (Rec->isSubClassOf("RegisterClass")) {
1164     if (ClassInfo *CI = RegisterClassClasses[Rec])
1165       return CI;
1166     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1167   }
1168 
1169   if (!Rec->isSubClassOf("Operand"))
1170     PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() +
1171                   "' does not derive from class Operand!\n");
1172   Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1173   if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1174     return CI;
1175 
1176   PrintFatalError(Rec->getLoc(), "operand has no match class!");
1177 }
1178 
1179 struct LessRegisterSet {
1180   bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const {
1181     // std::set<T> defines its own compariso "operator<", but it
1182     // performs a lexicographical comparison by T's innate comparison
1183     // for some reason. We don't want non-deterministic pointer
1184     // comparisons so use this instead.
1185     return std::lexicographical_compare(LHS.begin(), LHS.end(),
1186                                         RHS.begin(), RHS.end(),
1187                                         LessRecordByID());
1188   }
1189 };
1190 
1191 void AsmMatcherInfo::
1192 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) {
1193   const auto &Registers = Target.getRegBank().getRegisters();
1194   auto &RegClassList = Target.getRegBank().getRegClasses();
1195 
1196   typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1197 
1198   // The register sets used for matching.
1199   RegisterSetSet RegisterSets;
1200 
1201   // Gather the defined sets.
1202   for (const CodeGenRegisterClass &RC : RegClassList)
1203     RegisterSets.insert(
1204         RegisterSet(RC.getOrder().begin(), RC.getOrder().end()));
1205 
1206   // Add any required singleton sets.
1207   for (Record *Rec : SingletonRegisters) {
1208     RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1209   }
1210 
1211   // Introduce derived sets where necessary (when a register does not determine
1212   // a unique register set class), and build the mapping of registers to the set
1213   // they should classify to.
1214   std::map<Record*, RegisterSet> RegisterMap;
1215   for (const CodeGenRegister &CGR : Registers) {
1216     // Compute the intersection of all sets containing this register.
1217     RegisterSet ContainingSet;
1218 
1219     for (const RegisterSet &RS : RegisterSets) {
1220       if (!RS.count(CGR.TheDef))
1221         continue;
1222 
1223       if (ContainingSet.empty()) {
1224         ContainingSet = RS;
1225         continue;
1226       }
1227 
1228       RegisterSet Tmp;
1229       std::swap(Tmp, ContainingSet);
1230       std::insert_iterator<RegisterSet> II(ContainingSet,
1231                                            ContainingSet.begin());
1232       std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II,
1233                             LessRecordByID());
1234     }
1235 
1236     if (!ContainingSet.empty()) {
1237       RegisterSets.insert(ContainingSet);
1238       RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
1239     }
1240   }
1241 
1242   // Construct the register classes.
1243   std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses;
1244   unsigned Index = 0;
1245   for (const RegisterSet &RS : RegisterSets) {
1246     Classes.emplace_front();
1247     ClassInfo *CI = &Classes.front();
1248     CI->Kind = ClassInfo::RegisterClass0 + Index;
1249     CI->ClassName = "Reg" + utostr(Index);
1250     CI->Name = "MCK_Reg" + utostr(Index);
1251     CI->ValueName = "";
1252     CI->PredicateMethod = ""; // unused
1253     CI->RenderMethod = "addRegOperands";
1254     CI->Registers = RS;
1255     // FIXME: diagnostic type.
1256     CI->DiagnosticType = "";
1257     CI->IsOptional = false;
1258     RegisterSetClasses.insert(std::make_pair(RS, CI));
1259     ++Index;
1260   }
1261 
1262   // Find the superclasses; we could compute only the subgroup lattice edges,
1263   // but there isn't really a point.
1264   for (const RegisterSet &RS : RegisterSets) {
1265     ClassInfo *CI = RegisterSetClasses[RS];
1266     for (const RegisterSet &RS2 : RegisterSets)
1267       if (RS != RS2 &&
1268           std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(),
1269                         LessRecordByID()))
1270         CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1271   }
1272 
1273   // Name the register classes which correspond to a user defined RegisterClass.
1274   for (const CodeGenRegisterClass &RC : RegClassList) {
1275     // Def will be NULL for non-user defined register classes.
1276     Record *Def = RC.getDef();
1277     if (!Def)
1278       continue;
1279     ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(),
1280                                                    RC.getOrder().end())];
1281     if (CI->ValueName.empty()) {
1282       CI->ClassName = RC.getName();
1283       CI->Name = "MCK_" + RC.getName();
1284       CI->ValueName = RC.getName();
1285     } else
1286       CI->ValueName = CI->ValueName + "," + RC.getName();
1287 
1288     RegisterClassClasses.insert(std::make_pair(Def, CI));
1289   }
1290 
1291   // Populate the map for individual registers.
1292   for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(),
1293          ie = RegisterMap.end(); it != ie; ++it)
1294     RegisterClasses[it->first] = RegisterSetClasses[it->second];
1295 
1296   // Name the register classes which correspond to singleton registers.
1297   for (Record *Rec : SingletonRegisters) {
1298     ClassInfo *CI = RegisterClasses[Rec];
1299     assert(CI && "Missing singleton register class info!");
1300 
1301     if (CI->ValueName.empty()) {
1302       CI->ClassName = Rec->getName();
1303       CI->Name = "MCK_" + Rec->getName();
1304       CI->ValueName = Rec->getName();
1305     } else
1306       CI->ValueName = CI->ValueName + "," + Rec->getName();
1307   }
1308 }
1309 
1310 void AsmMatcherInfo::buildOperandClasses() {
1311   std::vector<Record*> AsmOperands =
1312     Records.getAllDerivedDefinitions("AsmOperandClass");
1313 
1314   // Pre-populate AsmOperandClasses map.
1315   for (Record *Rec : AsmOperands) {
1316     Classes.emplace_front();
1317     AsmOperandClasses[Rec] = &Classes.front();
1318   }
1319 
1320   unsigned Index = 0;
1321   for (Record *Rec : AsmOperands) {
1322     ClassInfo *CI = AsmOperandClasses[Rec];
1323     CI->Kind = ClassInfo::UserClass0 + Index;
1324 
1325     ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1326     for (Init *I : Supers->getValues()) {
1327       DefInit *DI = dyn_cast<DefInit>(I);
1328       if (!DI) {
1329         PrintError(Rec->getLoc(), "Invalid super class reference!");
1330         continue;
1331       }
1332 
1333       ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1334       if (!SC)
1335         PrintError(Rec->getLoc(), "Invalid super class reference!");
1336       else
1337         CI->SuperClasses.push_back(SC);
1338     }
1339     CI->ClassName = Rec->getValueAsString("Name");
1340     CI->Name = "MCK_" + CI->ClassName;
1341     CI->ValueName = Rec->getName();
1342 
1343     // Get or construct the predicate method name.
1344     Init *PMName = Rec->getValueInit("PredicateMethod");
1345     if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1346       CI->PredicateMethod = SI->getValue();
1347     } else {
1348       assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1349       CI->PredicateMethod = "is" + CI->ClassName;
1350     }
1351 
1352     // Get or construct the render method name.
1353     Init *RMName = Rec->getValueInit("RenderMethod");
1354     if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1355       CI->RenderMethod = SI->getValue();
1356     } else {
1357       assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1358       CI->RenderMethod = "add" + CI->ClassName + "Operands";
1359     }
1360 
1361     // Get the parse method name or leave it as empty.
1362     Init *PRMName = Rec->getValueInit("ParserMethod");
1363     if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1364       CI->ParserMethod = SI->getValue();
1365 
1366     // Get the diagnostic type or leave it as empty.
1367     // Get the parse method name or leave it as empty.
1368     Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1369     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1370       CI->DiagnosticType = SI->getValue();
1371 
1372     Init *IsOptional = Rec->getValueInit("IsOptional");
1373     if (BitInit *BI = dyn_cast<BitInit>(IsOptional))
1374       CI->IsOptional = BI->getValue();
1375 
1376     ++Index;
1377   }
1378 }
1379 
1380 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1381                                CodeGenTarget &target,
1382                                RecordKeeper &records)
1383   : Records(records), AsmParser(asmParser), Target(target) {
1384 }
1385 
1386 /// buildOperandMatchInfo - Build the necessary information to handle user
1387 /// defined operand parsing methods.
1388 void AsmMatcherInfo::buildOperandMatchInfo() {
1389 
1390   /// Map containing a mask with all operands indices that can be found for
1391   /// that class inside a instruction.
1392   typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy;
1393   OpClassMaskTy OpClassMask;
1394 
1395   for (const auto &MI : Matchables) {
1396     OpClassMask.clear();
1397 
1398     // Keep track of all operands of this instructions which belong to the
1399     // same class.
1400     for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
1401       const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
1402       if (Op.Class->ParserMethod.empty())
1403         continue;
1404       unsigned &OperandMask = OpClassMask[Op.Class];
1405       OperandMask |= (1 << i);
1406     }
1407 
1408     // Generate operand match info for each mnemonic/operand class pair.
1409     for (const auto &OCM : OpClassMask) {
1410       unsigned OpMask = OCM.second;
1411       ClassInfo *CI = OCM.first;
1412       OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI,
1413                                                            OpMask));
1414     }
1415   }
1416 }
1417 
1418 void AsmMatcherInfo::buildInfo() {
1419   // Build information about all of the AssemblerPredicates.
1420   std::vector<Record*> AllPredicates =
1421     Records.getAllDerivedDefinitions("Predicate");
1422   for (Record *Pred : AllPredicates) {
1423     // Ignore predicates that are not intended for the assembler.
1424     if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1425       continue;
1426 
1427     if (Pred->getName().empty())
1428       PrintFatalError(Pred->getLoc(), "Predicate has no name!");
1429 
1430     SubtargetFeatures.insert(std::make_pair(
1431         Pred, SubtargetFeatureInfo(Pred, SubtargetFeatures.size())));
1432     DEBUG(SubtargetFeatures.find(Pred)->second.dump());
1433     assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!");
1434   }
1435 
1436   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
1437 
1438   // Parse the instructions; we need to do this first so that we can gather the
1439   // singleton register classes.
1440   SmallPtrSet<Record*, 16> SingletonRegisters;
1441   unsigned VariantCount = Target.getAsmParserVariantCount();
1442   for (unsigned VC = 0; VC != VariantCount; ++VC) {
1443     Record *AsmVariant = Target.getAsmParserVariant(VC);
1444     std::string CommentDelimiter =
1445       AsmVariant->getValueAsString("CommentDelimiter");
1446     AsmVariantInfo Variant;
1447     Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1448     Variant.TokenizingCharacters =
1449         AsmVariant->getValueAsString("TokenizingCharacters");
1450     Variant.SeparatorCharacters =
1451         AsmVariant->getValueAsString("SeparatorCharacters");
1452     Variant.BreakCharacters =
1453         AsmVariant->getValueAsString("BreakCharacters");
1454     Variant.AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1455 
1456     for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) {
1457 
1458       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1459       // filter the set of instructions we consider.
1460       if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix))
1461         continue;
1462 
1463       // Ignore "codegen only" instructions.
1464       if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1465         continue;
1466 
1467       auto II = llvm::make_unique<MatchableInfo>(*CGI);
1468 
1469       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1470 
1471       // Ignore instructions which shouldn't be matched and diagnose invalid
1472       // instruction definitions with an error.
1473       if (!II->validate(CommentDelimiter, true))
1474         continue;
1475 
1476       Matchables.push_back(std::move(II));
1477     }
1478 
1479     // Parse all of the InstAlias definitions and stick them in the list of
1480     // matchables.
1481     std::vector<Record*> AllInstAliases =
1482       Records.getAllDerivedDefinitions("InstAlias");
1483     for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1484       auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i],
1485                                                        Variant.AsmVariantNo,
1486                                                        Target);
1487 
1488       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1489       // filter the set of instruction aliases we consider, based on the target
1490       // instruction.
1491       if (!StringRef(Alias->ResultInst->TheDef->getName())
1492             .startswith( MatchPrefix))
1493         continue;
1494 
1495       auto II = llvm::make_unique<MatchableInfo>(std::move(Alias));
1496 
1497       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1498 
1499       // Validate the alias definitions.
1500       II->validate(CommentDelimiter, false);
1501 
1502       Matchables.push_back(std::move(II));
1503     }
1504   }
1505 
1506   // Build info for the register classes.
1507   buildRegisterClasses(SingletonRegisters);
1508 
1509   // Build info for the user defined assembly operand classes.
1510   buildOperandClasses();
1511 
1512   // Build the information about matchables, now that we have fully formed
1513   // classes.
1514   std::vector<std::unique_ptr<MatchableInfo>> NewMatchables;
1515   for (auto &II : Matchables) {
1516     // Parse the tokens after the mnemonic.
1517     // Note: buildInstructionOperandReference may insert new AsmOperands, so
1518     // don't precompute the loop bound.
1519     for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1520       MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1521       StringRef Token = Op.Token;
1522 
1523       // Check for singleton registers.
1524       if (Record *RegRecord = Op.SingletonReg) {
1525         Op.Class = RegisterClasses[RegRecord];
1526         assert(Op.Class && Op.Class->Registers.size() == 1 &&
1527                "Unexpected class for singleton register");
1528         continue;
1529       }
1530 
1531       // Check for simple tokens.
1532       if (Token[0] != '$') {
1533         Op.Class = getTokenClass(Token);
1534         continue;
1535       }
1536 
1537       if (Token.size() > 1 && isdigit(Token[1])) {
1538         Op.Class = getTokenClass(Token);
1539         continue;
1540       }
1541 
1542       // Otherwise this is an operand reference.
1543       StringRef OperandName;
1544       if (Token[1] == '{')
1545         OperandName = Token.substr(2, Token.size() - 3);
1546       else
1547         OperandName = Token.substr(1);
1548 
1549       if (II->DefRec.is<const CodeGenInstruction*>())
1550         buildInstructionOperandReference(II.get(), OperandName, i);
1551       else
1552         buildAliasOperandReference(II.get(), OperandName, Op);
1553     }
1554 
1555     if (II->DefRec.is<const CodeGenInstruction*>()) {
1556       II->buildInstructionResultOperands();
1557       // If the instruction has a two-operand alias, build up the
1558       // matchable here. We'll add them in bulk at the end to avoid
1559       // confusing this loop.
1560       std::string Constraint =
1561         II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1562       if (Constraint != "") {
1563         // Start by making a copy of the original matchable.
1564         auto AliasII = llvm::make_unique<MatchableInfo>(*II);
1565 
1566         // Adjust it to be a two-operand alias.
1567         AliasII->formTwoOperandAlias(Constraint);
1568 
1569         // Add the alias to the matchables list.
1570         NewMatchables.push_back(std::move(AliasII));
1571       }
1572     } else
1573       II->buildAliasResultOperands();
1574   }
1575   if (!NewMatchables.empty())
1576     Matchables.insert(Matchables.end(),
1577                       std::make_move_iterator(NewMatchables.begin()),
1578                       std::make_move_iterator(NewMatchables.end()));
1579 
1580   // Process token alias definitions and set up the associated superclass
1581   // information.
1582   std::vector<Record*> AllTokenAliases =
1583     Records.getAllDerivedDefinitions("TokenAlias");
1584   for (Record *Rec : AllTokenAliases) {
1585     ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1586     ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1587     if (FromClass == ToClass)
1588       PrintFatalError(Rec->getLoc(),
1589                     "error: Destination value identical to source value.");
1590     FromClass->SuperClasses.push_back(ToClass);
1591   }
1592 
1593   // Reorder classes so that classes precede super classes.
1594   Classes.sort();
1595 
1596 #ifndef NDEBUG
1597   // Verify that the table is now sorted
1598   for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) {
1599     for (auto J = I; J != E; ++J) {
1600       assert(!(*J < *I));
1601       assert(I == J || !J->isSubsetOf(*I));
1602     }
1603   }
1604 #endif // NDEBUG
1605 }
1606 
1607 /// buildInstructionOperandReference - The specified operand is a reference to a
1608 /// named operand such as $src.  Resolve the Class and OperandInfo pointers.
1609 void AsmMatcherInfo::
1610 buildInstructionOperandReference(MatchableInfo *II,
1611                                  StringRef OperandName,
1612                                  unsigned AsmOpIdx) {
1613   const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1614   const CGIOperandList &Operands = CGI.Operands;
1615   MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1616 
1617   // Map this token to an operand.
1618   unsigned Idx;
1619   if (!Operands.hasOperandNamed(OperandName, Idx))
1620     PrintFatalError(II->TheDef->getLoc(),
1621                     "error: unable to find operand: '" + OperandName + "'");
1622 
1623   // If the instruction operand has multiple suboperands, but the parser
1624   // match class for the asm operand is still the default "ImmAsmOperand",
1625   // then handle each suboperand separately.
1626   if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1627     Record *Rec = Operands[Idx].Rec;
1628     assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1629     Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1630     if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1631       // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1632       StringRef Token = Op->Token; // save this in case Op gets moved
1633       for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1634         MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token);
1635         NewAsmOp.SubOpIdx = SI;
1636         II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1637       }
1638       // Replace Op with first suboperand.
1639       Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1640       Op->SubOpIdx = 0;
1641     }
1642   }
1643 
1644   // Set up the operand class.
1645   Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1646 
1647   // If the named operand is tied, canonicalize it to the untied operand.
1648   // For example, something like:
1649   //   (outs GPR:$dst), (ins GPR:$src)
1650   // with an asmstring of
1651   //   "inc $src"
1652   // we want to canonicalize to:
1653   //   "inc $dst"
1654   // so that we know how to provide the $dst operand when filling in the result.
1655   int OITied = -1;
1656   if (Operands[Idx].MINumOperands == 1)
1657     OITied = Operands[Idx].getTiedRegister();
1658   if (OITied != -1) {
1659     // The tied operand index is an MIOperand index, find the operand that
1660     // contains it.
1661     std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1662     OperandName = Operands[Idx.first].Name;
1663     Op->SubOpIdx = Idx.second;
1664   }
1665 
1666   Op->SrcOpName = OperandName;
1667 }
1668 
1669 /// buildAliasOperandReference - When parsing an operand reference out of the
1670 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1671 /// operand reference is by looking it up in the result pattern definition.
1672 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1673                                                 StringRef OperandName,
1674                                                 MatchableInfo::AsmOperand &Op) {
1675   const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1676 
1677   // Set up the operand class.
1678   for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1679     if (CGA.ResultOperands[i].isRecord() &&
1680         CGA.ResultOperands[i].getName() == OperandName) {
1681       // It's safe to go with the first one we find, because CodeGenInstAlias
1682       // validates that all operands with the same name have the same record.
1683       Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1684       // Use the match class from the Alias definition, not the
1685       // destination instruction, as we may have an immediate that's
1686       // being munged by the match class.
1687       Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1688                                  Op.SubOpIdx);
1689       Op.SrcOpName = OperandName;
1690       return;
1691     }
1692 
1693   PrintFatalError(II->TheDef->getLoc(),
1694                   "error: unable to find operand: '" + OperandName + "'");
1695 }
1696 
1697 void MatchableInfo::buildInstructionResultOperands() {
1698   const CodeGenInstruction *ResultInst = getResultInst();
1699 
1700   // Loop over all operands of the result instruction, determining how to
1701   // populate them.
1702   for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) {
1703     // If this is a tied operand, just copy from the previously handled operand.
1704     int TiedOp = -1;
1705     if (OpInfo.MINumOperands == 1)
1706       TiedOp = OpInfo.getTiedRegister();
1707     if (TiedOp != -1) {
1708       ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1709       continue;
1710     }
1711 
1712     // Find out what operand from the asmparser this MCInst operand comes from.
1713     int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1714     if (OpInfo.Name.empty() || SrcOperand == -1) {
1715       // This may happen for operands that are tied to a suboperand of a
1716       // complex operand.  Simply use a dummy value here; nobody should
1717       // use this operand slot.
1718       // FIXME: The long term goal is for the MCOperand list to not contain
1719       // tied operands at all.
1720       ResOperands.push_back(ResOperand::getImmOp(0));
1721       continue;
1722     }
1723 
1724     // Check if the one AsmOperand populates the entire operand.
1725     unsigned NumOperands = OpInfo.MINumOperands;
1726     if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1727       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1728       continue;
1729     }
1730 
1731     // Add a separate ResOperand for each suboperand.
1732     for (unsigned AI = 0; AI < NumOperands; ++AI) {
1733       assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1734              AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1735              "unexpected AsmOperands for suboperands");
1736       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1737     }
1738   }
1739 }
1740 
1741 void MatchableInfo::buildAliasResultOperands() {
1742   const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1743   const CodeGenInstruction *ResultInst = getResultInst();
1744 
1745   // Loop over all operands of the result instruction, determining how to
1746   // populate them.
1747   unsigned AliasOpNo = 0;
1748   unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1749   for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1750     const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1751 
1752     // If this is a tied operand, just copy from the previously handled operand.
1753     int TiedOp = -1;
1754     if (OpInfo->MINumOperands == 1)
1755       TiedOp = OpInfo->getTiedRegister();
1756     if (TiedOp != -1) {
1757       ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1758       continue;
1759     }
1760 
1761     // Handle all the suboperands for this operand.
1762     const std::string &OpName = OpInfo->Name;
1763     for ( ; AliasOpNo <  LastOpNo &&
1764             CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1765       int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1766 
1767       // Find out what operand from the asmparser that this MCInst operand
1768       // comes from.
1769       switch (CGA.ResultOperands[AliasOpNo].Kind) {
1770       case CodeGenInstAlias::ResultOperand::K_Record: {
1771         StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1772         int SrcOperand = findAsmOperand(Name, SubIdx);
1773         if (SrcOperand == -1)
1774           PrintFatalError(TheDef->getLoc(), "Instruction '" +
1775                         TheDef->getName() + "' has operand '" + OpName +
1776                         "' that doesn't appear in asm string!");
1777         unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1778         ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1779                                                         NumOperands));
1780         break;
1781       }
1782       case CodeGenInstAlias::ResultOperand::K_Imm: {
1783         int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1784         ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1785         break;
1786       }
1787       case CodeGenInstAlias::ResultOperand::K_Reg: {
1788         Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1789         ResOperands.push_back(ResOperand::getRegOp(Reg));
1790         break;
1791       }
1792       }
1793     }
1794   }
1795 }
1796 
1797 static unsigned getConverterOperandID(const std::string &Name,
1798                                       SmallSetVector<std::string, 16> &Table,
1799                                       bool &IsNew) {
1800   IsNew = Table.insert(Name);
1801 
1802   unsigned ID = IsNew ? Table.size() - 1 :
1803     std::find(Table.begin(), Table.end(), Name) - Table.begin();
1804 
1805   assert(ID < Table.size());
1806 
1807   return ID;
1808 }
1809 
1810 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1811                              std::vector<std::unique_ptr<MatchableInfo>> &Infos,
1812                              bool HasMnemonicFirst, raw_ostream &OS) {
1813   SmallSetVector<std::string, 16> OperandConversionKinds;
1814   SmallSetVector<std::string, 16> InstructionConversionKinds;
1815   std::vector<std::vector<uint8_t> > ConversionTable;
1816   size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1817 
1818   // TargetOperandClass - This is the target's operand class, like X86Operand.
1819   std::string TargetOperandClass = Target.getName() + "Operand";
1820 
1821   // Write the convert function to a separate stream, so we can drop it after
1822   // the enum. We'll build up the conversion handlers for the individual
1823   // operand types opportunistically as we encounter them.
1824   std::string ConvertFnBody;
1825   raw_string_ostream CvtOS(ConvertFnBody);
1826   // Start the unified conversion function.
1827   CvtOS << "void " << Target.getName() << ClassName << "::\n"
1828         << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1829         << "unsigned Opcode,\n"
1830         << "                const OperandVector"
1831         << " &Operands) {\n"
1832         << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1833         << "  const uint8_t *Converter = ConversionTable[Kind];\n"
1834         << "  Inst.setOpcode(Opcode);\n"
1835         << "  for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1836         << "    switch (*p) {\n"
1837         << "    default: llvm_unreachable(\"invalid conversion entry!\");\n"
1838         << "    case CVT_Reg:\n"
1839         << "      static_cast<" << TargetOperandClass
1840         << "&>(*Operands[*(p + 1)]).addRegOperands(Inst, 1);\n"
1841         << "      break;\n"
1842         << "    case CVT_Tied:\n"
1843         << "      Inst.addOperand(Inst.getOperand(*(p + 1)));\n"
1844         << "      break;\n";
1845 
1846   std::string OperandFnBody;
1847   raw_string_ostream OpOS(OperandFnBody);
1848   // Start the operand number lookup function.
1849   OpOS << "void " << Target.getName() << ClassName << "::\n"
1850        << "convertToMapAndConstraints(unsigned Kind,\n";
1851   OpOS.indent(27);
1852   OpOS << "const OperandVector &Operands) {\n"
1853        << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1854        << "  unsigned NumMCOperands = 0;\n"
1855        << "  const uint8_t *Converter = ConversionTable[Kind];\n"
1856        << "  for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1857        << "    switch (*p) {\n"
1858        << "    default: llvm_unreachable(\"invalid conversion entry!\");\n"
1859        << "    case CVT_Reg:\n"
1860        << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1861        << "      Operands[*(p + 1)]->setConstraint(\"r\");\n"
1862        << "      ++NumMCOperands;\n"
1863        << "      break;\n"
1864        << "    case CVT_Tied:\n"
1865        << "      ++NumMCOperands;\n"
1866        << "      break;\n";
1867 
1868   // Pre-populate the operand conversion kinds with the standard always
1869   // available entries.
1870   OperandConversionKinds.insert("CVT_Done");
1871   OperandConversionKinds.insert("CVT_Reg");
1872   OperandConversionKinds.insert("CVT_Tied");
1873   enum { CVT_Done, CVT_Reg, CVT_Tied };
1874 
1875   for (auto &II : Infos) {
1876     // Check if we have a custom match function.
1877     std::string AsmMatchConverter =
1878       II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1879     if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) {
1880       std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1881       II->ConversionFnKind = Signature;
1882 
1883       // Check if we have already generated this signature.
1884       if (!InstructionConversionKinds.insert(Signature))
1885         continue;
1886 
1887       // Remember this converter for the kind enum.
1888       unsigned KindID = OperandConversionKinds.size();
1889       OperandConversionKinds.insert("CVT_" +
1890                                     getEnumNameForToken(AsmMatchConverter));
1891 
1892       // Add the converter row for this instruction.
1893       ConversionTable.emplace_back();
1894       ConversionTable.back().push_back(KindID);
1895       ConversionTable.back().push_back(CVT_Done);
1896 
1897       // Add the handler to the conversion driver function.
1898       CvtOS << "    case CVT_"
1899             << getEnumNameForToken(AsmMatchConverter) << ":\n"
1900             << "      " << AsmMatchConverter << "(Inst, Operands);\n"
1901             << "      break;\n";
1902 
1903       // FIXME: Handle the operand number lookup for custom match functions.
1904       continue;
1905     }
1906 
1907     // Build the conversion function signature.
1908     std::string Signature = "Convert";
1909 
1910     std::vector<uint8_t> ConversionRow;
1911 
1912     // Compute the convert enum and the case body.
1913     MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 );
1914 
1915     for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) {
1916       const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i];
1917 
1918       // Generate code to populate each result operand.
1919       switch (OpInfo.Kind) {
1920       case MatchableInfo::ResOperand::RenderAsmOperand: {
1921         // This comes from something we parsed.
1922         const MatchableInfo::AsmOperand &Op =
1923           II->AsmOperands[OpInfo.AsmOperandNum];
1924 
1925         // Registers are always converted the same, don't duplicate the
1926         // conversion function based on them.
1927         Signature += "__";
1928         std::string Class;
1929         Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
1930         Signature += Class;
1931         Signature += utostr(OpInfo.MINumOperands);
1932         Signature += "_" + itostr(OpInfo.AsmOperandNum);
1933 
1934         // Add the conversion kind, if necessary, and get the associated ID
1935         // the index of its entry in the vector).
1936         std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" :
1937                                      Op.Class->RenderMethod);
1938         Name = getEnumNameForToken(Name);
1939 
1940         bool IsNewConverter = false;
1941         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1942                                             IsNewConverter);
1943 
1944         // Add the operand entry to the instruction kind conversion row.
1945         ConversionRow.push_back(ID);
1946         ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst);
1947 
1948         if (!IsNewConverter)
1949           break;
1950 
1951         // This is a new operand kind. Add a handler for it to the
1952         // converter driver.
1953         CvtOS << "    case " << Name << ":\n"
1954               << "      static_cast<" << TargetOperandClass
1955               << "&>(*Operands[*(p + 1)])." << Op.Class->RenderMethod
1956               << "(Inst, " << OpInfo.MINumOperands << ");\n"
1957               << "      break;\n";
1958 
1959         // Add a handler for the operand number lookup.
1960         OpOS << "    case " << Name << ":\n"
1961              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
1962 
1963         if (Op.Class->isRegisterClass())
1964           OpOS << "      Operands[*(p + 1)]->setConstraint(\"r\");\n";
1965         else
1966           OpOS << "      Operands[*(p + 1)]->setConstraint(\"m\");\n";
1967         OpOS << "      NumMCOperands += " << OpInfo.MINumOperands << ";\n"
1968              << "      break;\n";
1969         break;
1970       }
1971       case MatchableInfo::ResOperand::TiedOperand: {
1972         // If this operand is tied to a previous one, just copy the MCInst
1973         // operand from the earlier one.We can only tie single MCOperand values.
1974         assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
1975         unsigned TiedOp = OpInfo.TiedOperandNum;
1976         assert(i > TiedOp && "Tied operand precedes its target!");
1977         Signature += "__Tie" + utostr(TiedOp);
1978         ConversionRow.push_back(CVT_Tied);
1979         ConversionRow.push_back(TiedOp);
1980         break;
1981       }
1982       case MatchableInfo::ResOperand::ImmOperand: {
1983         int64_t Val = OpInfo.ImmVal;
1984         std::string Ty = "imm_" + itostr(Val);
1985         Ty = getEnumNameForToken(Ty);
1986         Signature += "__" + Ty;
1987 
1988         std::string Name = "CVT_" + Ty;
1989         bool IsNewConverter = false;
1990         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1991                                             IsNewConverter);
1992         // Add the operand entry to the instruction kind conversion row.
1993         ConversionRow.push_back(ID);
1994         ConversionRow.push_back(0);
1995 
1996         if (!IsNewConverter)
1997           break;
1998 
1999         CvtOS << "    case " << Name << ":\n"
2000               << "      Inst.addOperand(MCOperand::createImm(" << Val << "));\n"
2001               << "      break;\n";
2002 
2003         OpOS << "    case " << Name << ":\n"
2004              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2005              << "      Operands[*(p + 1)]->setConstraint(\"\");\n"
2006              << "      ++NumMCOperands;\n"
2007              << "      break;\n";
2008         break;
2009       }
2010       case MatchableInfo::ResOperand::RegOperand: {
2011         std::string Reg, Name;
2012         if (!OpInfo.Register) {
2013           Name = "reg0";
2014           Reg = "0";
2015         } else {
2016           Reg = getQualifiedName(OpInfo.Register);
2017           Name = "reg" + OpInfo.Register->getName();
2018         }
2019         Signature += "__" + Name;
2020         Name = "CVT_" + Name;
2021         bool IsNewConverter = false;
2022         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2023                                             IsNewConverter);
2024         // Add the operand entry to the instruction kind conversion row.
2025         ConversionRow.push_back(ID);
2026         ConversionRow.push_back(0);
2027 
2028         if (!IsNewConverter)
2029           break;
2030         CvtOS << "    case " << Name << ":\n"
2031               << "      Inst.addOperand(MCOperand::createReg(" << Reg << "));\n"
2032               << "      break;\n";
2033 
2034         OpOS << "    case " << Name << ":\n"
2035              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2036              << "      Operands[*(p + 1)]->setConstraint(\"m\");\n"
2037              << "      ++NumMCOperands;\n"
2038              << "      break;\n";
2039       }
2040       }
2041     }
2042 
2043     // If there were no operands, add to the signature to that effect
2044     if (Signature == "Convert")
2045       Signature += "_NoOperands";
2046 
2047     II->ConversionFnKind = Signature;
2048 
2049     // Save the signature. If we already have it, don't add a new row
2050     // to the table.
2051     if (!InstructionConversionKinds.insert(Signature))
2052       continue;
2053 
2054     // Add the row to the table.
2055     ConversionTable.push_back(std::move(ConversionRow));
2056   }
2057 
2058   // Finish up the converter driver function.
2059   CvtOS << "    }\n  }\n}\n\n";
2060 
2061   // Finish up the operand number lookup function.
2062   OpOS << "    }\n  }\n}\n\n";
2063 
2064   OS << "namespace {\n";
2065 
2066   // Output the operand conversion kind enum.
2067   OS << "enum OperatorConversionKind {\n";
2068   for (const std::string &Converter : OperandConversionKinds)
2069     OS << "  " << Converter << ",\n";
2070   OS << "  CVT_NUM_CONVERTERS\n";
2071   OS << "};\n\n";
2072 
2073   // Output the instruction conversion kind enum.
2074   OS << "enum InstructionConversionKind {\n";
2075   for (const std::string &Signature : InstructionConversionKinds)
2076     OS << "  " << Signature << ",\n";
2077   OS << "  CVT_NUM_SIGNATURES\n";
2078   OS << "};\n\n";
2079 
2080   OS << "} // end anonymous namespace\n\n";
2081 
2082   // Output the conversion table.
2083   OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
2084      << MaxRowLength << "] = {\n";
2085 
2086   for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
2087     assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
2088     OS << "  // " << InstructionConversionKinds[Row] << "\n";
2089     OS << "  { ";
2090     for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2)
2091       OS << OperandConversionKinds[ConversionTable[Row][i]] << ", "
2092          << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
2093     OS << "CVT_Done },\n";
2094   }
2095 
2096   OS << "};\n\n";
2097 
2098   // Spit out the conversion driver function.
2099   OS << CvtOS.str();
2100 
2101   // Spit out the operand number lookup function.
2102   OS << OpOS.str();
2103 }
2104 
2105 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
2106 static void emitMatchClassEnumeration(CodeGenTarget &Target,
2107                                       std::forward_list<ClassInfo> &Infos,
2108                                       raw_ostream &OS) {
2109   OS << "namespace {\n\n";
2110 
2111   OS << "/// MatchClassKind - The kinds of classes which participate in\n"
2112      << "/// instruction matching.\n";
2113   OS << "enum MatchClassKind {\n";
2114   OS << "  InvalidMatchClass = 0,\n";
2115   OS << "  OptionalMatchClass = 1,\n";
2116   for (const auto &CI : Infos) {
2117     OS << "  " << CI.Name << ", // ";
2118     if (CI.Kind == ClassInfo::Token) {
2119       OS << "'" << CI.ValueName << "'\n";
2120     } else if (CI.isRegisterClass()) {
2121       if (!CI.ValueName.empty())
2122         OS << "register class '" << CI.ValueName << "'\n";
2123       else
2124         OS << "derived register class\n";
2125     } else {
2126       OS << "user defined class '" << CI.ValueName << "'\n";
2127     }
2128   }
2129   OS << "  NumMatchClassKinds\n";
2130   OS << "};\n\n";
2131 
2132   OS << "}\n\n";
2133 }
2134 
2135 /// emitValidateOperandClass - Emit the function to validate an operand class.
2136 static void emitValidateOperandClass(AsmMatcherInfo &Info,
2137                                      raw_ostream &OS) {
2138   OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
2139      << "MatchClassKind Kind) {\n";
2140   OS << "  " << Info.Target.getName() << "Operand &Operand = ("
2141      << Info.Target.getName() << "Operand&)GOp;\n";
2142 
2143   // The InvalidMatchClass is not to match any operand.
2144   OS << "  if (Kind == InvalidMatchClass)\n";
2145   OS << "    return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2146 
2147   // Check for Token operands first.
2148   // FIXME: Use a more specific diagnostic type.
2149   OS << "  if (Operand.isToken())\n";
2150   OS << "    return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2151      << "             MCTargetAsmParser::Match_Success :\n"
2152      << "             MCTargetAsmParser::Match_InvalidOperand;\n\n";
2153 
2154   // Check the user classes. We don't care what order since we're only
2155   // actually matching against one of them.
2156   OS << "  switch (Kind) {\n"
2157         "  default: break;\n";
2158   for (const auto &CI : Info.Classes) {
2159     if (!CI.isUserClass())
2160       continue;
2161 
2162     OS << "  // '" << CI.ClassName << "' class\n";
2163     OS << "  case " << CI.Name << ":\n";
2164     OS << "    if (Operand." << CI.PredicateMethod << "())\n";
2165     OS << "      return MCTargetAsmParser::Match_Success;\n";
2166     if (!CI.DiagnosticType.empty())
2167       OS << "    return " << Info.Target.getName() << "AsmParser::Match_"
2168          << CI.DiagnosticType << ";\n";
2169     else
2170       OS << "    break;\n";
2171   }
2172   OS << "  } // end switch (Kind)\n\n";
2173 
2174   // Check for register operands, including sub-classes.
2175   OS << "  if (Operand.isReg()) {\n";
2176   OS << "    MatchClassKind OpKind;\n";
2177   OS << "    switch (Operand.getReg()) {\n";
2178   OS << "    default: OpKind = InvalidMatchClass; break;\n";
2179   for (const auto &RC : Info.RegisterClasses)
2180     OS << "    case " << Info.Target.getName() << "::"
2181        << RC.first->getName() << ": OpKind = " << RC.second->Name
2182        << "; break;\n";
2183   OS << "    }\n";
2184   OS << "    return isSubclass(OpKind, Kind) ? "
2185      << "MCTargetAsmParser::Match_Success :\n                             "
2186      << "         MCTargetAsmParser::Match_InvalidOperand;\n  }\n\n";
2187 
2188   // Generic fallthrough match failure case for operands that don't have
2189   // specialized diagnostic types.
2190   OS << "  return MCTargetAsmParser::Match_InvalidOperand;\n";
2191   OS << "}\n\n";
2192 }
2193 
2194 /// emitIsSubclass - Emit the subclass predicate function.
2195 static void emitIsSubclass(CodeGenTarget &Target,
2196                            std::forward_list<ClassInfo> &Infos,
2197                            raw_ostream &OS) {
2198   OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2199   OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2200   OS << "  if (A == B)\n";
2201   OS << "    return true;\n\n";
2202 
2203   bool EmittedSwitch = false;
2204   for (const auto &A : Infos) {
2205     std::vector<StringRef> SuperClasses;
2206     if (A.IsOptional)
2207       SuperClasses.push_back("OptionalMatchClass");
2208     for (const auto &B : Infos) {
2209       if (&A != &B && A.isSubsetOf(B))
2210         SuperClasses.push_back(B.Name);
2211     }
2212 
2213     if (SuperClasses.empty())
2214       continue;
2215 
2216     // If this is the first SuperClass, emit the switch header.
2217     if (!EmittedSwitch) {
2218       OS << "  switch (A) {\n";
2219       OS << "  default:\n";
2220       OS << "    return false;\n";
2221       EmittedSwitch = true;
2222     }
2223 
2224     OS << "\n  case " << A.Name << ":\n";
2225 
2226     if (SuperClasses.size() == 1) {
2227       OS << "    return B == " << SuperClasses.back() << ";\n";
2228       continue;
2229     }
2230 
2231     if (!SuperClasses.empty()) {
2232       OS << "    switch (B) {\n";
2233       OS << "    default: return false;\n";
2234       for (StringRef SC : SuperClasses)
2235         OS << "    case " << SC << ": return true;\n";
2236       OS << "    }\n";
2237     } else {
2238       // No case statement to emit
2239       OS << "    return false;\n";
2240     }
2241   }
2242 
2243   // If there were case statements emitted into the string stream write the
2244   // default.
2245   if (EmittedSwitch)
2246     OS << "  }\n";
2247   else
2248     OS << "  return false;\n";
2249 
2250   OS << "}\n\n";
2251 }
2252 
2253 /// emitMatchTokenString - Emit the function to match a token string to the
2254 /// appropriate match class value.
2255 static void emitMatchTokenString(CodeGenTarget &Target,
2256                                  std::forward_list<ClassInfo> &Infos,
2257                                  raw_ostream &OS) {
2258   // Construct the match list.
2259   std::vector<StringMatcher::StringPair> Matches;
2260   for (const auto &CI : Infos) {
2261     if (CI.Kind == ClassInfo::Token)
2262       Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";");
2263   }
2264 
2265   OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2266 
2267   StringMatcher("Name", Matches, OS).Emit();
2268 
2269   OS << "  return InvalidMatchClass;\n";
2270   OS << "}\n\n";
2271 }
2272 
2273 /// emitMatchRegisterName - Emit the function to match a string to the target
2274 /// specific register enum.
2275 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2276                                   raw_ostream &OS) {
2277   // Construct the match list.
2278   std::vector<StringMatcher::StringPair> Matches;
2279   const auto &Regs = Target.getRegBank().getRegisters();
2280   for (const CodeGenRegister &Reg : Regs) {
2281     if (Reg.TheDef->getValueAsString("AsmName").empty())
2282       continue;
2283 
2284     Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"),
2285                          "return " + utostr(Reg.EnumValue) + ";");
2286   }
2287 
2288   OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
2289 
2290   StringMatcher("Name", Matches, OS).Emit();
2291 
2292   OS << "  return 0;\n";
2293   OS << "}\n\n";
2294 }
2295 
2296 /// Emit the function to match a string to the target
2297 /// specific register enum.
2298 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser,
2299                                      raw_ostream &OS) {
2300   // Construct the match list.
2301   std::vector<StringMatcher::StringPair> Matches;
2302   const auto &Regs = Target.getRegBank().getRegisters();
2303   for (const CodeGenRegister &Reg : Regs) {
2304 
2305     auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames");
2306 
2307     for (auto AltName : AltNames) {
2308       AltName = StringRef(AltName).trim();
2309 
2310       // don't handle empty alternative names
2311       if (AltName.empty())
2312         continue;
2313 
2314       Matches.emplace_back(AltName,
2315                            "return " + utostr(Reg.EnumValue) + ";");
2316     }
2317   }
2318 
2319   OS << "static unsigned MatchRegisterAltName(StringRef Name) {\n";
2320 
2321   StringMatcher("Name", Matches, OS).Emit();
2322 
2323   OS << "  return 0;\n";
2324   OS << "}\n\n";
2325 }
2326 
2327 static const char *getMinimalTypeForRange(uint64_t Range) {
2328   assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large");
2329   if (Range > 0xFFFFFFFFULL)
2330     return "uint64_t";
2331   if (Range > 0xFFFF)
2332     return "uint32_t";
2333   if (Range > 0xFF)
2334     return "uint16_t";
2335   return "uint8_t";
2336 }
2337 
2338 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) {
2339   uint64_t MaxIndex = Info.SubtargetFeatures.size();
2340   if (MaxIndex > 0)
2341     MaxIndex--;
2342   return getMinimalTypeForRange(1ULL << MaxIndex);
2343 }
2344 
2345 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
2346 /// definitions.
2347 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
2348                                                 raw_ostream &OS) {
2349   OS << "// Flags for subtarget features that participate in "
2350      << "instruction matching.\n";
2351   OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info)
2352      << " {\n";
2353   for (const auto &SF : Info.SubtargetFeatures) {
2354     const SubtargetFeatureInfo &SFI = SF.second;
2355     OS << "  " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n";
2356   }
2357   OS << "  Feature_None = 0\n";
2358   OS << "};\n\n";
2359 }
2360 
2361 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
2362 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2363   // Get the set of diagnostic types from all of the operand classes.
2364   std::set<StringRef> Types;
2365   for (const auto &OpClassEntry : Info.AsmOperandClasses) {
2366     if (!OpClassEntry.second->DiagnosticType.empty())
2367       Types.insert(OpClassEntry.second->DiagnosticType);
2368   }
2369 
2370   if (Types.empty()) return;
2371 
2372   // Now emit the enum entries.
2373   for (StringRef Type : Types)
2374     OS << "  Match_" << Type << ",\n";
2375   OS << "  END_OPERAND_DIAGNOSTIC_TYPES\n";
2376 }
2377 
2378 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2379 /// user-level name for a subtarget feature.
2380 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2381   OS << "// User-level names for subtarget features that participate in\n"
2382      << "// instruction matching.\n"
2383      << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2384   if (!Info.SubtargetFeatures.empty()) {
2385     OS << "  switch(Val) {\n";
2386     for (const auto &SF : Info.SubtargetFeatures) {
2387       const SubtargetFeatureInfo &SFI = SF.second;
2388       // FIXME: Totally just a placeholder name to get the algorithm working.
2389       OS << "  case " << SFI.getEnumName() << ": return \""
2390          << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2391     }
2392     OS << "  default: return \"(unknown)\";\n";
2393     OS << "  }\n";
2394   } else {
2395     // Nothing to emit, so skip the switch
2396     OS << "  return \"(unknown)\";\n";
2397   }
2398   OS << "}\n\n";
2399 }
2400 
2401 /// emitComputeAvailableFeatures - Emit the function to compute the list of
2402 /// available features given a subtarget.
2403 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info,
2404                                          raw_ostream &OS) {
2405   std::string ClassName =
2406     Info.AsmParser->getValueAsString("AsmParserClassName");
2407 
2408   OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n"
2409      << "ComputeAvailableFeatures(const FeatureBitset& FB) const {\n";
2410   OS << "  uint64_t Features = 0;\n";
2411   for (const auto &SF : Info.SubtargetFeatures) {
2412     const SubtargetFeatureInfo &SFI = SF.second;
2413 
2414     OS << "  if (";
2415     std::string CondStorage =
2416       SFI.TheDef->getValueAsString("AssemblerCondString");
2417     StringRef Conds = CondStorage;
2418     std::pair<StringRef,StringRef> Comma = Conds.split(',');
2419     bool First = true;
2420     do {
2421       if (!First)
2422         OS << " && ";
2423 
2424       bool Neg = false;
2425       StringRef Cond = Comma.first;
2426       if (Cond[0] == '!') {
2427         Neg = true;
2428         Cond = Cond.substr(1);
2429       }
2430 
2431       OS << "(";
2432       if (Neg)
2433         OS << "!";
2434       OS << "FB[" << Info.Target.getName() << "::" << Cond << "])";
2435 
2436       if (Comma.second.empty())
2437         break;
2438 
2439       First = false;
2440       Comma = Comma.second.split(',');
2441     } while (true);
2442 
2443     OS << ")\n";
2444     OS << "    Features |= " << SFI.getEnumName() << ";\n";
2445   }
2446   OS << "  return Features;\n";
2447   OS << "}\n\n";
2448 }
2449 
2450 static std::string GetAliasRequiredFeatures(Record *R,
2451                                             const AsmMatcherInfo &Info) {
2452   std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2453   std::string Result;
2454   unsigned NumFeatures = 0;
2455   for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2456     const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2457 
2458     if (!F)
2459       PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2460                     "' is not marked as an AssemblerPredicate!");
2461 
2462     if (NumFeatures)
2463       Result += '|';
2464 
2465     Result += F->getEnumName();
2466     ++NumFeatures;
2467   }
2468 
2469   if (NumFeatures > 1)
2470     Result = '(' + Result + ')';
2471   return Result;
2472 }
2473 
2474 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info,
2475                                      std::vector<Record*> &Aliases,
2476                                      unsigned Indent = 0,
2477                                   StringRef AsmParserVariantName = StringRef()){
2478   // Keep track of all the aliases from a mnemonic.  Use an std::map so that the
2479   // iteration order of the map is stable.
2480   std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2481 
2482   for (Record *R : Aliases) {
2483     // FIXME: Allow AssemblerVariantName to be a comma separated list.
2484     std::string AsmVariantName = R->getValueAsString("AsmVariantName");
2485     if (AsmVariantName != AsmParserVariantName)
2486       continue;
2487     AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2488   }
2489   if (AliasesFromMnemonic.empty())
2490     return;
2491 
2492   // Process each alias a "from" mnemonic at a time, building the code executed
2493   // by the string remapper.
2494   std::vector<StringMatcher::StringPair> Cases;
2495   for (const auto &AliasEntry : AliasesFromMnemonic) {
2496     const std::vector<Record*> &ToVec = AliasEntry.second;
2497 
2498     // Loop through each alias and emit code that handles each case.  If there
2499     // are two instructions without predicates, emit an error.  If there is one,
2500     // emit it last.
2501     std::string MatchCode;
2502     int AliasWithNoPredicate = -1;
2503 
2504     for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2505       Record *R = ToVec[i];
2506       std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2507 
2508       // If this unconditionally matches, remember it for later and diagnose
2509       // duplicates.
2510       if (FeatureMask.empty()) {
2511         if (AliasWithNoPredicate != -1) {
2512           // We can't have two aliases from the same mnemonic with no predicate.
2513           PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2514                      "two MnemonicAliases with the same 'from' mnemonic!");
2515           PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2516         }
2517 
2518         AliasWithNoPredicate = i;
2519         continue;
2520       }
2521       if (R->getValueAsString("ToMnemonic") == AliasEntry.first)
2522         PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2523 
2524       if (!MatchCode.empty())
2525         MatchCode += "else ";
2526       MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
2527       MatchCode += "  Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
2528     }
2529 
2530     if (AliasWithNoPredicate != -1) {
2531       Record *R = ToVec[AliasWithNoPredicate];
2532       if (!MatchCode.empty())
2533         MatchCode += "else\n  ";
2534       MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2535     }
2536 
2537     MatchCode += "return;";
2538 
2539     Cases.push_back(std::make_pair(AliasEntry.first, MatchCode));
2540   }
2541   StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2542 }
2543 
2544 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2545 /// emit a function for them and return true, otherwise return false.
2546 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2547                                 CodeGenTarget &Target) {
2548   // Ignore aliases when match-prefix is set.
2549   if (!MatchPrefix.empty())
2550     return false;
2551 
2552   std::vector<Record*> Aliases =
2553     Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2554   if (Aliases.empty()) return false;
2555 
2556   OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2557     "uint64_t Features, unsigned VariantID) {\n";
2558   OS << "  switch (VariantID) {\n";
2559   unsigned VariantCount = Target.getAsmParserVariantCount();
2560   for (unsigned VC = 0; VC != VariantCount; ++VC) {
2561     Record *AsmVariant = Target.getAsmParserVariant(VC);
2562     int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2563     std::string AsmParserVariantName = AsmVariant->getValueAsString("Name");
2564     OS << "    case " << AsmParserVariantNo << ":\n";
2565     emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2566                              AsmParserVariantName);
2567     OS << "    break;\n";
2568   }
2569   OS << "  }\n";
2570 
2571   // Emit aliases that apply to all variants.
2572   emitMnemonicAliasVariant(OS, Info, Aliases);
2573 
2574   OS << "}\n\n";
2575 
2576   return true;
2577 }
2578 
2579 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2580                               const AsmMatcherInfo &Info, StringRef ClassName,
2581                               StringToOffsetTable &StringTable,
2582                               unsigned MaxMnemonicIndex, bool HasMnemonicFirst) {
2583   unsigned MaxMask = 0;
2584   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2585     MaxMask |= OMI.OperandMask;
2586   }
2587 
2588   // Emit the static custom operand parsing table;
2589   OS << "namespace {\n";
2590   OS << "  struct OperandMatchEntry {\n";
2591   OS << "    " << getMinimalRequiredFeaturesType(Info)
2592                << " RequiredFeatures;\n";
2593   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex)
2594                << " Mnemonic;\n";
2595   OS << "    " << getMinimalTypeForRange(std::distance(
2596                       Info.Classes.begin(), Info.Classes.end())) << " Class;\n";
2597   OS << "    " << getMinimalTypeForRange(MaxMask)
2598                << " OperandMask;\n\n";
2599   OS << "    StringRef getMnemonic() const {\n";
2600   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2601   OS << "                       MnemonicTable[Mnemonic]);\n";
2602   OS << "    }\n";
2603   OS << "  };\n\n";
2604 
2605   OS << "  // Predicate for searching for an opcode.\n";
2606   OS << "  struct LessOpcodeOperand {\n";
2607   OS << "    bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2608   OS << "      return LHS.getMnemonic()  < RHS;\n";
2609   OS << "    }\n";
2610   OS << "    bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2611   OS << "      return LHS < RHS.getMnemonic();\n";
2612   OS << "    }\n";
2613   OS << "    bool operator()(const OperandMatchEntry &LHS,";
2614   OS << " const OperandMatchEntry &RHS) {\n";
2615   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2616   OS << "    }\n";
2617   OS << "  };\n";
2618 
2619   OS << "} // end anonymous namespace.\n\n";
2620 
2621   OS << "static const OperandMatchEntry OperandMatchTable["
2622      << Info.OperandMatchInfo.size() << "] = {\n";
2623 
2624   OS << "  /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2625   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2626     const MatchableInfo &II = *OMI.MI;
2627 
2628     OS << "  { ";
2629 
2630     // Write the required features mask.
2631     if (!II.RequiredFeatures.empty()) {
2632       for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2633         if (i) OS << "|";
2634         OS << II.RequiredFeatures[i]->getEnumName();
2635       }
2636     } else
2637       OS << "0";
2638 
2639     // Store a pascal-style length byte in the mnemonic.
2640     std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2641     OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2642        << " /* " << II.Mnemonic << " */, ";
2643 
2644     OS << OMI.CI->Name;
2645 
2646     OS << ", " << OMI.OperandMask;
2647     OS << " /* ";
2648     bool printComma = false;
2649     for (int i = 0, e = 31; i !=e; ++i)
2650       if (OMI.OperandMask & (1 << i)) {
2651         if (printComma)
2652           OS << ", ";
2653         OS << i;
2654         printComma = true;
2655       }
2656     OS << " */";
2657 
2658     OS << " },\n";
2659   }
2660   OS << "};\n\n";
2661 
2662   // Emit the operand class switch to call the correct custom parser for
2663   // the found operand class.
2664   OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2665      << Target.getName() << ClassName << "::\n"
2666      << "tryCustomParseOperand(OperandVector"
2667      << " &Operands,\n                      unsigned MCK) {\n\n"
2668      << "  switch(MCK) {\n";
2669 
2670   for (const auto &CI : Info.Classes) {
2671     if (CI.ParserMethod.empty())
2672       continue;
2673     OS << "  case " << CI.Name << ":\n"
2674        << "    return " << CI.ParserMethod << "(Operands);\n";
2675   }
2676 
2677   OS << "  default:\n";
2678   OS << "    return MatchOperand_NoMatch;\n";
2679   OS << "  }\n";
2680   OS << "  return MatchOperand_NoMatch;\n";
2681   OS << "}\n\n";
2682 
2683   // Emit the static custom operand parser. This code is very similar with
2684   // the other matcher. Also use MatchResultTy here just in case we go for
2685   // a better error handling.
2686   OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2687      << Target.getName() << ClassName << "::\n"
2688      << "MatchOperandParserImpl(OperandVector"
2689      << " &Operands,\n                       StringRef Mnemonic) {\n";
2690 
2691   // Emit code to get the available features.
2692   OS << "  // Get the current feature set.\n";
2693   OS << "  uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2694 
2695   OS << "  // Get the next operand index.\n";
2696   OS << "  unsigned NextOpNum = Operands.size()"
2697      << (HasMnemonicFirst ? " - 1" : "") << ";\n";
2698 
2699   // Emit code to search the table.
2700   OS << "  // Search the table.\n";
2701   if (HasMnemonicFirst) {
2702     OS << "  auto MnemonicRange =\n";
2703     OS << "    std::equal_range(std::begin(OperandMatchTable), "
2704           "std::end(OperandMatchTable),\n";
2705     OS << "                     Mnemonic, LessOpcodeOperand());\n\n";
2706   } else {
2707     OS << "  auto MnemonicRange = std::make_pair(std::begin(OperandMatchTable),"
2708           " std::end(OperandMatchTable));\n";
2709     OS << "  if (!Mnemonic.empty())\n";
2710     OS << "    MnemonicRange =\n";
2711     OS << "      std::equal_range(std::begin(OperandMatchTable), "
2712           "std::end(OperandMatchTable),\n";
2713     OS << "                       Mnemonic, LessOpcodeOperand());\n\n";
2714   }
2715 
2716   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2717   OS << "    return MatchOperand_NoMatch;\n\n";
2718 
2719   OS << "  for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2720      << "       *ie = MnemonicRange.second; it != ie; ++it) {\n";
2721 
2722   OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2723   OS << "    assert(Mnemonic == it->getMnemonic());\n\n";
2724 
2725   // Emit check that the required features are available.
2726   OS << "    // check if the available features match\n";
2727   OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2728      << "!= it->RequiredFeatures) {\n";
2729   OS << "      continue;\n";
2730   OS << "    }\n\n";
2731 
2732   // Emit check to ensure the operand number matches.
2733   OS << "    // check if the operand in question has a custom parser.\n";
2734   OS << "    if (!(it->OperandMask & (1 << NextOpNum)))\n";
2735   OS << "      continue;\n\n";
2736 
2737   // Emit call to the custom parser method
2738   OS << "    // call custom parse method to handle the operand\n";
2739   OS << "    OperandMatchResultTy Result = ";
2740   OS << "tryCustomParseOperand(Operands, it->Class);\n";
2741   OS << "    if (Result != MatchOperand_NoMatch)\n";
2742   OS << "      return Result;\n";
2743   OS << "  }\n\n";
2744 
2745   OS << "  // Okay, we had no match.\n";
2746   OS << "  return MatchOperand_NoMatch;\n";
2747   OS << "}\n\n";
2748 }
2749 
2750 void AsmMatcherEmitter::run(raw_ostream &OS) {
2751   CodeGenTarget Target(Records);
2752   Record *AsmParser = Target.getAsmParser();
2753   std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2754 
2755   // Compute the information on the instructions to match.
2756   AsmMatcherInfo Info(AsmParser, Target, Records);
2757   Info.buildInfo();
2758 
2759   // Sort the instruction table using the partial order on classes. We use
2760   // stable_sort to ensure that ambiguous instructions are still
2761   // deterministically ordered.
2762   std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
2763                    [](const std::unique_ptr<MatchableInfo> &a,
2764                       const std::unique_ptr<MatchableInfo> &b){
2765                      return *a < *b;});
2766 
2767 #ifndef NDEBUG
2768   // Verify that the table is now sorted
2769   for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2770        ++I) {
2771     for (auto J = I; J != E; ++J) {
2772       assert(!(**J < **I));
2773     }
2774   }
2775 #endif // NDEBUG
2776 
2777   DEBUG_WITH_TYPE("instruction_info", {
2778       for (const auto &MI : Info.Matchables)
2779         MI->dump();
2780     });
2781 
2782   // Check for ambiguous matchables.
2783   DEBUG_WITH_TYPE("ambiguous_instrs", {
2784     unsigned NumAmbiguous = 0;
2785     for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2786          ++I) {
2787       for (auto J = std::next(I); J != E; ++J) {
2788         const MatchableInfo &A = **I;
2789         const MatchableInfo &B = **J;
2790 
2791         if (A.couldMatchAmbiguouslyWith(B)) {
2792           errs() << "warning: ambiguous matchables:\n";
2793           A.dump();
2794           errs() << "\nis incomparable with:\n";
2795           B.dump();
2796           errs() << "\n\n";
2797           ++NumAmbiguous;
2798         }
2799       }
2800     }
2801     if (NumAmbiguous)
2802       errs() << "warning: " << NumAmbiguous
2803              << " ambiguous matchables!\n";
2804   });
2805 
2806   // Compute the information on the custom operand parsing.
2807   Info.buildOperandMatchInfo();
2808 
2809   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
2810 
2811   // Write the output.
2812 
2813   // Information for the class declaration.
2814   OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2815   OS << "#undef GET_ASSEMBLER_HEADER\n";
2816   OS << "  // This should be included into the middle of the declaration of\n";
2817   OS << "  // your subclasses implementation of MCTargetAsmParser.\n";
2818   OS << "  uint64_t ComputeAvailableFeatures(const FeatureBitset& FB) const;\n";
2819   OS << "  void convertToMCInst(unsigned Kind, MCInst &Inst, "
2820      << "unsigned Opcode,\n"
2821      << "                       const OperandVector "
2822      << "&Operands);\n";
2823   OS << "  void convertToMapAndConstraints(unsigned Kind,\n                ";
2824   OS << "           const OperandVector &Operands) override;\n";
2825   if (HasMnemonicFirst)
2826     OS << "  bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID);\n";
2827   OS << "  unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
2828      << "                                MCInst &Inst,\n"
2829      << "                                uint64_t &ErrorInfo,"
2830      << " bool matchingInlineAsm,\n"
2831      << "                                unsigned VariantID = 0);\n";
2832 
2833   if (!Info.OperandMatchInfo.empty()) {
2834     OS << "\n  enum OperandMatchResultTy {\n";
2835     OS << "    MatchOperand_Success,    // operand matched successfully\n";
2836     OS << "    MatchOperand_NoMatch,    // operand did not match\n";
2837     OS << "    MatchOperand_ParseFail   // operand matched but had errors\n";
2838     OS << "  };\n";
2839     OS << "  OperandMatchResultTy MatchOperandParserImpl(\n";
2840     OS << "    OperandVector &Operands,\n";
2841     OS << "    StringRef Mnemonic);\n";
2842 
2843     OS << "  OperandMatchResultTy tryCustomParseOperand(\n";
2844     OS << "    OperandVector &Operands,\n";
2845     OS << "    unsigned MCK);\n\n";
2846   }
2847 
2848   OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2849 
2850   // Emit the operand match diagnostic enum names.
2851   OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
2852   OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2853   emitOperandDiagnosticTypes(Info, OS);
2854   OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2855 
2856   OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2857   OS << "#undef GET_REGISTER_MATCHER\n\n";
2858 
2859   // Emit the subtarget feature enumeration.
2860   emitSubtargetFeatureFlagEnumeration(Info, OS);
2861 
2862   // Emit the function to match a register name to number.
2863   // This should be omitted for Mips target
2864   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
2865     emitMatchRegisterName(Target, AsmParser, OS);
2866 
2867   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName"))
2868     emitMatchRegisterAltName(Target, AsmParser, OS);
2869 
2870   OS << "#endif // GET_REGISTER_MATCHER\n\n";
2871 
2872   OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
2873   OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
2874 
2875   // Generate the helper function to get the names for subtarget features.
2876   emitGetSubtargetFeatureName(Info, OS);
2877 
2878   OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
2879 
2880   OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2881   OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2882 
2883   // Generate the function that remaps for mnemonic aliases.
2884   bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
2885 
2886   // Generate the convertToMCInst function to convert operands into an MCInst.
2887   // Also, generate the convertToMapAndConstraints function for MS-style inline
2888   // assembly.  The latter doesn't actually generate a MCInst.
2889   emitConvertFuncs(Target, ClassName, Info.Matchables, HasMnemonicFirst, OS);
2890 
2891   // Emit the enumeration for classes which participate in matching.
2892   emitMatchClassEnumeration(Target, Info.Classes, OS);
2893 
2894   // Emit the routine to match token strings to their match class.
2895   emitMatchTokenString(Target, Info.Classes, OS);
2896 
2897   // Emit the subclass predicate routine.
2898   emitIsSubclass(Target, Info.Classes, OS);
2899 
2900   // Emit the routine to validate an operand against a match class.
2901   emitValidateOperandClass(Info, OS);
2902 
2903   // Emit the available features compute function.
2904   emitComputeAvailableFeatures(Info, OS);
2905 
2906   StringToOffsetTable StringTable;
2907 
2908   size_t MaxNumOperands = 0;
2909   unsigned MaxMnemonicIndex = 0;
2910   bool HasDeprecation = false;
2911   for (const auto &MI : Info.Matchables) {
2912     MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
2913     HasDeprecation |= MI->HasDeprecation;
2914 
2915     // Store a pascal-style length byte in the mnemonic.
2916     std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
2917     MaxMnemonicIndex = std::max(MaxMnemonicIndex,
2918                         StringTable.GetOrAddStringOffset(LenMnemonic, false));
2919   }
2920 
2921   OS << "static const char *const MnemonicTable =\n";
2922   StringTable.EmitString(OS);
2923   OS << ";\n\n";
2924 
2925   // Emit the static match table; unused classes get initalized to 0 which is
2926   // guaranteed to be InvalidMatchClass.
2927   //
2928   // FIXME: We can reduce the size of this table very easily. First, we change
2929   // it so that store the kinds in separate bit-fields for each index, which
2930   // only needs to be the max width used for classes at that index (we also need
2931   // to reject based on this during classification). If we then make sure to
2932   // order the match kinds appropriately (putting mnemonics last), then we
2933   // should only end up using a few bits for each class, especially the ones
2934   // following the mnemonic.
2935   OS << "namespace {\n";
2936   OS << "  struct MatchEntry {\n";
2937   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex)
2938                << " Mnemonic;\n";
2939   OS << "    uint16_t Opcode;\n";
2940   OS << "    " << getMinimalTypeForRange(Info.Matchables.size())
2941                << " ConvertFn;\n";
2942   OS << "    " << getMinimalRequiredFeaturesType(Info)
2943                << " RequiredFeatures;\n";
2944   OS << "    " << getMinimalTypeForRange(
2945                       std::distance(Info.Classes.begin(), Info.Classes.end()))
2946      << " Classes[" << MaxNumOperands << "];\n";
2947   OS << "    StringRef getMnemonic() const {\n";
2948   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2949   OS << "                       MnemonicTable[Mnemonic]);\n";
2950   OS << "    }\n";
2951   OS << "  };\n\n";
2952 
2953   OS << "  // Predicate for searching for an opcode.\n";
2954   OS << "  struct LessOpcode {\n";
2955   OS << "    bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
2956   OS << "      return LHS.getMnemonic() < RHS;\n";
2957   OS << "    }\n";
2958   OS << "    bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
2959   OS << "      return LHS < RHS.getMnemonic();\n";
2960   OS << "    }\n";
2961   OS << "    bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
2962   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2963   OS << "    }\n";
2964   OS << "  };\n";
2965 
2966   OS << "} // end anonymous namespace.\n\n";
2967 
2968   unsigned VariantCount = Target.getAsmParserVariantCount();
2969   for (unsigned VC = 0; VC != VariantCount; ++VC) {
2970     Record *AsmVariant = Target.getAsmParserVariant(VC);
2971     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2972 
2973     OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
2974 
2975     for (const auto &MI : Info.Matchables) {
2976       if (MI->AsmVariantID != AsmVariantNo)
2977         continue;
2978 
2979       // Store a pascal-style length byte in the mnemonic.
2980       std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
2981       OS << "  { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2982          << " /* " << MI->Mnemonic << " */, "
2983          << Target.getName() << "::"
2984          << MI->getResultInst()->TheDef->getName() << ", "
2985          << MI->ConversionFnKind << ", ";
2986 
2987       // Write the required features mask.
2988       if (!MI->RequiredFeatures.empty()) {
2989         for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i) {
2990           if (i) OS << "|";
2991           OS << MI->RequiredFeatures[i]->getEnumName();
2992         }
2993       } else
2994         OS << "0";
2995 
2996       OS << ", { ";
2997       for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
2998         const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
2999 
3000         if (i) OS << ", ";
3001         OS << Op.Class->Name;
3002       }
3003       OS << " }, },\n";
3004     }
3005 
3006     OS << "};\n\n";
3007   }
3008 
3009   // A method to determine if a mnemonic is in the list.
3010   if (HasMnemonicFirst) {
3011     OS << "bool " << Target.getName() << ClassName << "::\n"
3012        << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n";
3013     OS << "  // Find the appropriate table for this asm variant.\n";
3014     OS << "  const MatchEntry *Start, *End;\n";
3015     OS << "  switch (VariantID) {\n";
3016     OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3017     for (unsigned VC = 0; VC != VariantCount; ++VC) {
3018       Record *AsmVariant = Target.getAsmParserVariant(VC);
3019       int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3020       OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3021          << "); End = std::end(MatchTable" << VC << "); break;\n";
3022     }
3023     OS << "  }\n";
3024     OS << "  // Search the table.\n";
3025     OS << "  auto MnemonicRange = ";
3026     OS << "std::equal_range(Start, End, Mnemonic, LessOpcode());\n";
3027     OS << "  return MnemonicRange.first != MnemonicRange.second;\n";
3028     OS << "}\n\n";
3029   }
3030 
3031   // Finally, build the match function.
3032   OS << "unsigned " << Target.getName() << ClassName << "::\n"
3033      << "MatchInstructionImpl(const OperandVector &Operands,\n";
3034   OS << "                     MCInst &Inst, uint64_t &ErrorInfo,\n"
3035      << "                     bool matchingInlineAsm, unsigned VariantID) {\n";
3036 
3037   OS << "  // Eliminate obvious mismatches.\n";
3038   OS << "  if (Operands.size() > "
3039      << (MaxNumOperands + HasMnemonicFirst) << ") {\n";
3040   OS << "    ErrorInfo = "
3041      << (MaxNumOperands + HasMnemonicFirst) << ";\n";
3042   OS << "    return Match_InvalidOperand;\n";
3043   OS << "  }\n\n";
3044 
3045   // Emit code to get the available features.
3046   OS << "  // Get the current feature set.\n";
3047   OS << "  uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
3048 
3049   OS << "  // Get the instruction mnemonic, which is the first token.\n";
3050   if (HasMnemonicFirst) {
3051     OS << "  StringRef Mnemonic = ((" << Target.getName()
3052        << "Operand&)*Operands[0]).getToken();\n\n";
3053   } else {
3054     OS << "  StringRef Mnemonic;\n";
3055     OS << "  if (Operands[0]->isToken())\n";
3056     OS << "    Mnemonic = ((" << Target.getName()
3057        << "Operand&)*Operands[0]).getToken();\n\n";
3058   }
3059 
3060   if (HasMnemonicAliases) {
3061     OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
3062     OS << "  applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
3063   }
3064 
3065   // Emit code to compute the class list for this operand vector.
3066   OS << "  // Some state to try to produce better error messages.\n";
3067   OS << "  bool HadMatchOtherThanFeatures = false;\n";
3068   OS << "  bool HadMatchOtherThanPredicate = false;\n";
3069   OS << "  unsigned RetCode = Match_InvalidOperand;\n";
3070   OS << "  uint64_t MissingFeatures = ~0ULL;\n";
3071   OS << "  // Set ErrorInfo to the operand that mismatches if it is\n";
3072   OS << "  // wrong for all instances of the instruction.\n";
3073   OS << "  ErrorInfo = ~0ULL;\n";
3074 
3075   // Emit code to search the table.
3076   OS << "  // Find the appropriate table for this asm variant.\n";
3077   OS << "  const MatchEntry *Start, *End;\n";
3078   OS << "  switch (VariantID) {\n";
3079   OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3080   for (unsigned VC = 0; VC != VariantCount; ++VC) {
3081     Record *AsmVariant = Target.getAsmParserVariant(VC);
3082     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3083     OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3084        << "); End = std::end(MatchTable" << VC << "); break;\n";
3085   }
3086   OS << "  }\n";
3087 
3088   OS << "  // Search the table.\n";
3089   if (HasMnemonicFirst) {
3090     OS << "  auto MnemonicRange = "
3091           "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
3092   } else {
3093     OS << "  auto MnemonicRange = std::make_pair(Start, End);\n";
3094     OS << "  unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
3095     OS << "  if (!Mnemonic.empty())\n";
3096     OS << "    MnemonicRange = "
3097           "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
3098   }
3099 
3100   OS << "  // Return a more specific error code if no mnemonics match.\n";
3101   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
3102   OS << "    return Match_MnemonicFail;\n\n";
3103 
3104   OS << "  for (const MatchEntry *it = MnemonicRange.first, "
3105      << "*ie = MnemonicRange.second;\n";
3106   OS << "       it != ie; ++it) {\n";
3107 
3108   if (HasMnemonicFirst) {
3109     OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
3110     OS << "    assert(Mnemonic == it->getMnemonic());\n";
3111   }
3112 
3113   // Emit check that the subclasses match.
3114   OS << "    bool OperandsValid = true;\n";
3115   OS << "    for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex")
3116      << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex")
3117      << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n";
3118   OS << "      auto Formal = "
3119      << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n";
3120   OS << "      if (ActualIdx >= Operands.size()) {\n";
3121   OS << "        OperandsValid = (Formal == " <<"InvalidMatchClass) || "
3122                                  "isSubclass(Formal, OptionalMatchClass);\n";
3123   OS << "        if (!OperandsValid) ErrorInfo = ActualIdx;\n";
3124   OS << "        break;\n";
3125   OS << "      }\n";
3126   OS << "      MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n";
3127   OS << "      unsigned Diag = validateOperandClass(Actual, Formal);\n";
3128   OS << "      if (Diag == Match_Success) {\n";
3129   OS << "        ++ActualIdx;\n";
3130   OS << "        continue;\n";
3131   OS << "      }\n";
3132   OS << "      // If the generic handler indicates an invalid operand\n";
3133   OS << "      // failure, check for a special case.\n";
3134   OS << "      if (Diag == Match_InvalidOperand) {\n";
3135   OS << "        Diag = validateTargetOperandClass(Actual, Formal);\n";
3136   OS << "        if (Diag == Match_Success) {\n";
3137   OS << "          ++ActualIdx;\n";
3138   OS << "          continue;\n";
3139   OS << "        }\n";
3140   OS << "      }\n";
3141   OS << "      // If current formal operand wasn't matched and it is optional\n"
3142      << "      // then try to match next formal operand\n";
3143   OS << "      if (Diag == Match_InvalidOperand "
3144      << "&& isSubclass(Formal, OptionalMatchClass))\n";
3145   OS << "        continue;\n";
3146   OS << "      // If this operand is broken for all of the instances of this\n";
3147   OS << "      // mnemonic, keep track of it so we can report loc info.\n";
3148   OS << "      // If we already had a match that only failed due to a\n";
3149   OS << "      // target predicate, that diagnostic is preferred.\n";
3150   OS << "      if (!HadMatchOtherThanPredicate &&\n";
3151   OS << "          (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) {\n";
3152   OS << "        ErrorInfo = ActualIdx;\n";
3153   OS << "        // InvalidOperand is the default. Prefer specificity.\n";
3154   OS << "        if (Diag != Match_InvalidOperand)\n";
3155   OS << "          RetCode = Diag;\n";
3156   OS << "      }\n";
3157   OS << "      // Otherwise, just reject this instance of the mnemonic.\n";
3158   OS << "      OperandsValid = false;\n";
3159   OS << "      break;\n";
3160   OS << "    }\n\n";
3161 
3162   OS << "    if (!OperandsValid) continue;\n";
3163 
3164   // Emit check that the required features are available.
3165   OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
3166      << "!= it->RequiredFeatures) {\n";
3167   OS << "      HadMatchOtherThanFeatures = true;\n";
3168   OS << "      uint64_t NewMissingFeatures = it->RequiredFeatures & "
3169         "~AvailableFeatures;\n";
3170   OS << "      if (countPopulation(NewMissingFeatures) <=\n"
3171         "          countPopulation(MissingFeatures))\n";
3172   OS << "        MissingFeatures = NewMissingFeatures;\n";
3173   OS << "      continue;\n";
3174   OS << "    }\n";
3175   OS << "\n";
3176   OS << "    Inst.clear();\n\n";
3177   OS << "    if (matchingInlineAsm) {\n";
3178   OS << "      Inst.setOpcode(it->Opcode);\n";
3179   OS << "      convertToMapAndConstraints(it->ConvertFn, Operands);\n";
3180   OS << "      return Match_Success;\n";
3181   OS << "    }\n\n";
3182   OS << "    // We have selected a definite instruction, convert the parsed\n"
3183      << "    // operands into the appropriate MCInst.\n";
3184   OS << "    convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
3185   OS << "\n";
3186 
3187   // Verify the instruction with the target-specific match predicate function.
3188   OS << "    // We have a potential match. Check the target predicate to\n"
3189      << "    // handle any context sensitive constraints.\n"
3190      << "    unsigned MatchResult;\n"
3191      << "    if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
3192      << " Match_Success) {\n"
3193      << "      Inst.clear();\n"
3194      << "      RetCode = MatchResult;\n"
3195      << "      HadMatchOtherThanPredicate = true;\n"
3196      << "      continue;\n"
3197      << "    }\n\n";
3198 
3199   // Call the post-processing function, if used.
3200   std::string InsnCleanupFn =
3201     AsmParser->getValueAsString("AsmParserInstCleanup");
3202   if (!InsnCleanupFn.empty())
3203     OS << "    " << InsnCleanupFn << "(Inst);\n";
3204 
3205   if (HasDeprecation) {
3206     OS << "    std::string Info;\n";
3207     OS << "    if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n";
3208     OS << "      SMLoc Loc = ((" << Target.getName()
3209        << "Operand&)*Operands[0]).getStartLoc();\n";
3210     OS << "      getParser().Warning(Loc, Info, None);\n";
3211     OS << "    }\n";
3212   }
3213 
3214   OS << "    return Match_Success;\n";
3215   OS << "  }\n\n";
3216 
3217   OS << "  // Okay, we had no match.  Try to return a useful error code.\n";
3218   OS << "  if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
3219   OS << "    return RetCode;\n\n";
3220   OS << "  // Missing feature matches return which features were missing\n";
3221   OS << "  ErrorInfo = MissingFeatures;\n";
3222   OS << "  return Match_MissingFeature;\n";
3223   OS << "}\n\n";
3224 
3225   if (!Info.OperandMatchInfo.empty())
3226     emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
3227                              MaxMnemonicIndex, HasMnemonicFirst);
3228 
3229   OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
3230 }
3231 
3232 namespace llvm {
3233 
3234 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
3235   emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
3236   AsmMatcherEmitter(RK).run(OS);
3237 }
3238 
3239 } // end namespace llvm
3240