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