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   /// DefaultMethod - The name of the method that returns the default operand
207   /// for optional operand
208   std::string DefaultMethod;
209 
210 public:
211   /// isRegisterClass() - Check if this is a register class.
212   bool isRegisterClass() const {
213     return Kind >= RegisterClass0 && Kind < UserClass0;
214   }
215 
216   /// isUserClass() - Check if this is a user defined class.
217   bool isUserClass() const {
218     return Kind >= UserClass0;
219   }
220 
221   /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
222   /// are related if they are in the same class hierarchy.
223   bool isRelatedTo(const ClassInfo &RHS) const {
224     // Tokens are only related to tokens.
225     if (Kind == Token || RHS.Kind == Token)
226       return Kind == Token && RHS.Kind == Token;
227 
228     // Registers classes are only related to registers classes, and only if
229     // their intersection is non-empty.
230     if (isRegisterClass() || RHS.isRegisterClass()) {
231       if (!isRegisterClass() || !RHS.isRegisterClass())
232         return false;
233 
234       RegisterSet Tmp;
235       std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin());
236       std::set_intersection(Registers.begin(), Registers.end(),
237                             RHS.Registers.begin(), RHS.Registers.end(),
238                             II, LessRecordByID());
239 
240       return !Tmp.empty();
241     }
242 
243     // Otherwise we have two users operands; they are related if they are in the
244     // same class hierarchy.
245     //
246     // FIXME: This is an oversimplification, they should only be related if they
247     // intersect, however we don't have that information.
248     assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
249     const ClassInfo *Root = this;
250     while (!Root->SuperClasses.empty())
251       Root = Root->SuperClasses.front();
252 
253     const ClassInfo *RHSRoot = &RHS;
254     while (!RHSRoot->SuperClasses.empty())
255       RHSRoot = RHSRoot->SuperClasses.front();
256 
257     return Root == RHSRoot;
258   }
259 
260   /// isSubsetOf - Test whether this class is a subset of \p RHS.
261   bool isSubsetOf(const ClassInfo &RHS) const {
262     // This is a subset of RHS if it is the same class...
263     if (this == &RHS)
264       return true;
265 
266     // ... or if any of its super classes are a subset of RHS.
267     for (const ClassInfo *CI : SuperClasses)
268       if (CI->isSubsetOf(RHS))
269         return true;
270 
271     return false;
272   }
273 
274   int getTreeDepth() const {
275     int Depth = 0;
276     const ClassInfo *Root = this;
277     while (!Root->SuperClasses.empty()) {
278       Depth++;
279       Root = Root->SuperClasses.front();
280     }
281     return Depth;
282   }
283 
284   const ClassInfo *findRoot() const {
285     const ClassInfo *Root = this;
286     while (!Root->SuperClasses.empty())
287       Root = Root->SuperClasses.front();
288     return Root;
289   }
290 
291   /// Compare two classes. This does not produce a total ordering, but does
292   /// guarantee that subclasses are sorted before their parents, and that the
293   /// ordering is transitive.
294   bool operator<(const ClassInfo &RHS) const {
295     if (this == &RHS)
296       return false;
297 
298     // First, enforce the ordering between the three different types of class.
299     // Tokens sort before registers, which sort before user classes.
300     if (Kind == Token) {
301       if (RHS.Kind != Token)
302         return true;
303       assert(RHS.Kind == Token);
304     } else if (isRegisterClass()) {
305       if (RHS.Kind == Token)
306         return false;
307       else if (RHS.isUserClass())
308         return true;
309       assert(RHS.isRegisterClass());
310     } else if (isUserClass()) {
311       if (!RHS.isUserClass())
312         return false;
313       assert(RHS.isUserClass());
314     } else {
315       llvm_unreachable("Unknown ClassInfoKind");
316     }
317 
318     if (Kind == Token || isUserClass()) {
319       // Related tokens and user classes get sorted by depth in the inheritence
320       // tree (so that subclasses are before their parents).
321       if (isRelatedTo(RHS)) {
322         if (getTreeDepth() > RHS.getTreeDepth())
323           return true;
324         if (getTreeDepth() < RHS.getTreeDepth())
325           return false;
326       } else {
327         // Unrelated tokens and user classes are ordered by the name of their
328         // root nodes, so that there is a consistent ordering between
329         // unconnected trees.
330         return findRoot()->ValueName < RHS.findRoot()->ValueName;
331       }
332     } else if (isRegisterClass()) {
333       // For register sets, sort by number of registers. This guarantees that
334       // a set will always sort before all of it's strict supersets.
335       if (Registers.size() != RHS.Registers.size())
336         return Registers.size() < RHS.Registers.size();
337     } else {
338       llvm_unreachable("Unknown ClassInfoKind");
339     }
340 
341     // FIXME: We should be able to just return false here, as we only need a
342     // partial order (we use stable sorts, so this is deterministic) and the
343     // name of a class shouldn't be significant. However, some of the backends
344     // accidentally rely on this behaviour, so it will have to stay like this
345     // until they are fixed.
346     return ValueName < RHS.ValueName;
347   }
348 };
349 
350 class AsmVariantInfo {
351 public:
352   std::string RegisterPrefix;
353   std::string TokenizingCharacters;
354   std::string SeparatorCharacters;
355   std::string BreakCharacters;
356   std::string Name;
357   int AsmVariantNo;
358 };
359 
360 /// MatchableInfo - Helper class for storing the necessary information for an
361 /// instruction or alias which is capable of being matched.
362 struct MatchableInfo {
363   struct AsmOperand {
364     /// Token - This is the token that the operand came from.
365     StringRef Token;
366 
367     /// The unique class instance this operand should match.
368     ClassInfo *Class;
369 
370     /// The operand name this is, if anything.
371     StringRef SrcOpName;
372 
373     /// The suboperand index within SrcOpName, or -1 for the entire operand.
374     int SubOpIdx;
375 
376     /// Whether the token is "isolated", i.e., it is preceded and followed
377     /// by separators.
378     bool IsIsolatedToken;
379 
380     /// Register record if this token is singleton register.
381     Record *SingletonReg;
382 
383     explicit AsmOperand(bool IsIsolatedToken, StringRef T)
384         : Token(T), Class(nullptr), SubOpIdx(-1),
385           IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {}
386   };
387 
388   /// ResOperand - This represents a single operand in the result instruction
389   /// generated by the match.  In cases (like addressing modes) where a single
390   /// assembler operand expands to multiple MCOperands, this represents the
391   /// single assembler operand, not the MCOperand.
392   struct ResOperand {
393     enum {
394       /// RenderAsmOperand - This represents an operand result that is
395       /// generated by calling the render method on the assembly operand.  The
396       /// corresponding AsmOperand is specified by AsmOperandNum.
397       RenderAsmOperand,
398 
399       /// TiedOperand - This represents a result operand that is a duplicate of
400       /// a previous result operand.
401       TiedOperand,
402 
403       /// ImmOperand - This represents an immediate value that is dumped into
404       /// the operand.
405       ImmOperand,
406 
407       /// RegOperand - This represents a fixed register that is dumped in.
408       RegOperand
409     } Kind;
410 
411     union {
412       /// This is the operand # in the AsmOperands list that this should be
413       /// copied from.
414       unsigned AsmOperandNum;
415 
416       /// TiedOperandNum - This is the (earlier) result operand that should be
417       /// copied from.
418       unsigned TiedOperandNum;
419 
420       /// ImmVal - This is the immediate value added to the instruction.
421       int64_t ImmVal;
422 
423       /// Register - This is the register record.
424       Record *Register;
425     };
426 
427     /// MINumOperands - The number of MCInst operands populated by this
428     /// operand.
429     unsigned MINumOperands;
430 
431     static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
432       ResOperand X;
433       X.Kind = RenderAsmOperand;
434       X.AsmOperandNum = AsmOpNum;
435       X.MINumOperands = NumOperands;
436       return X;
437     }
438 
439     static ResOperand getTiedOp(unsigned TiedOperandNum) {
440       ResOperand X;
441       X.Kind = TiedOperand;
442       X.TiedOperandNum = TiedOperandNum;
443       X.MINumOperands = 1;
444       return X;
445     }
446 
447     static ResOperand getImmOp(int64_t Val) {
448       ResOperand X;
449       X.Kind = ImmOperand;
450       X.ImmVal = Val;
451       X.MINumOperands = 1;
452       return X;
453     }
454 
455     static ResOperand getRegOp(Record *Reg) {
456       ResOperand X;
457       X.Kind = RegOperand;
458       X.Register = Reg;
459       X.MINumOperands = 1;
460       return X;
461     }
462   };
463 
464   /// AsmVariantID - Target's assembly syntax variant no.
465   int AsmVariantID;
466 
467   /// AsmString - The assembly string for this instruction (with variants
468   /// removed), e.g. "movsx $src, $dst".
469   std::string AsmString;
470 
471   /// TheDef - This is the definition of the instruction or InstAlias that this
472   /// matchable came from.
473   Record *const TheDef;
474 
475   /// DefRec - This is the definition that it came from.
476   PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
477 
478   const CodeGenInstruction *getResultInst() const {
479     if (DefRec.is<const CodeGenInstruction*>())
480       return DefRec.get<const CodeGenInstruction*>();
481     return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
482   }
483 
484   /// ResOperands - This is the operand list that should be built for the result
485   /// MCInst.
486   SmallVector<ResOperand, 8> ResOperands;
487 
488   /// Mnemonic - This is the first token of the matched instruction, its
489   /// mnemonic.
490   StringRef Mnemonic;
491 
492   /// AsmOperands - The textual operands that this instruction matches,
493   /// annotated with a class and where in the OperandList they were defined.
494   /// This directly corresponds to the tokenized AsmString after the mnemonic is
495   /// removed.
496   SmallVector<AsmOperand, 8> AsmOperands;
497 
498   /// Predicates - The required subtarget features to match this instruction.
499   SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
500 
501   /// ConversionFnKind - The enum value which is passed to the generated
502   /// convertToMCInst to convert parsed operands into an MCInst for this
503   /// function.
504   std::string ConversionFnKind;
505 
506   /// If this instruction is deprecated in some form.
507   bool HasDeprecation;
508 
509   /// If this is an alias, this is use to determine whether or not to using
510   /// the conversion function defined by the instruction's AsmMatchConverter
511   /// or to use the function generated by the alias.
512   bool UseInstAsmMatchConverter;
513 
514   MatchableInfo(const CodeGenInstruction &CGI)
515     : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI),
516       UseInstAsmMatchConverter(true) {
517   }
518 
519   MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias)
520     : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef),
521       DefRec(Alias.release()),
522       UseInstAsmMatchConverter(
523         TheDef->getValueAsBit("UseInstAsmMatchConverter")) {
524   }
525 
526   // Could remove this and the dtor if PointerUnion supported unique_ptr
527   // elements with a dynamic failure/assertion (like the one below) in the case
528   // where it was copied while being in an owning state.
529   MatchableInfo(const MatchableInfo &RHS)
530       : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString),
531         TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands),
532         Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands),
533         RequiredFeatures(RHS.RequiredFeatures),
534         ConversionFnKind(RHS.ConversionFnKind),
535         HasDeprecation(RHS.HasDeprecation),
536         UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) {
537     assert(!DefRec.is<const CodeGenInstAlias *>());
538   }
539 
540   ~MatchableInfo() {
541     delete DefRec.dyn_cast<const CodeGenInstAlias*>();
542   }
543 
544   // Two-operand aliases clone from the main matchable, but mark the second
545   // operand as a tied operand of the first for purposes of the assembler.
546   void formTwoOperandAlias(StringRef Constraint);
547 
548   void initialize(const AsmMatcherInfo &Info,
549                   SmallPtrSetImpl<Record*> &SingletonRegisters,
550                   AsmVariantInfo const &Variant,
551                   bool HasMnemonicFirst);
552 
553   /// validate - Return true if this matchable is a valid thing to match against
554   /// and perform a bunch of validity checking.
555   bool validate(StringRef CommentDelimiter, bool Hack) const;
556 
557   /// findAsmOperand - Find the AsmOperand with the specified name and
558   /// suboperand index.
559   int findAsmOperand(StringRef N, int SubOpIdx) const {
560     auto I = find_if(AsmOperands, [&](const AsmOperand &Op) {
561       return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx;
562     });
563     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
564   }
565 
566   /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
567   /// This does not check the suboperand index.
568   int findAsmOperandNamed(StringRef N) const {
569     auto I = find_if(AsmOperands,
570                      [&](const AsmOperand &Op) { return Op.SrcOpName == N; });
571     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
572   }
573 
574   void buildInstructionResultOperands();
575   void buildAliasResultOperands();
576 
577   /// operator< - Compare two matchables.
578   bool operator<(const MatchableInfo &RHS) const {
579     // The primary comparator is the instruction mnemonic.
580     if (int Cmp = Mnemonic.compare(RHS.Mnemonic))
581       return Cmp == -1;
582 
583     if (AsmOperands.size() != RHS.AsmOperands.size())
584       return AsmOperands.size() < RHS.AsmOperands.size();
585 
586     // Compare lexicographically by operand. The matcher validates that other
587     // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
588     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
589       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
590         return true;
591       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
592         return false;
593     }
594 
595     // Give matches that require more features higher precedence. This is useful
596     // because we cannot define AssemblerPredicates with the negation of
597     // processor features. For example, ARM v6 "nop" may be either a HINT or
598     // MOV. With v6, we want to match HINT. The assembler has no way to
599     // predicate MOV under "NoV6", but HINT will always match first because it
600     // requires V6 while MOV does not.
601     if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
602       return RequiredFeatures.size() > RHS.RequiredFeatures.size();
603 
604     return false;
605   }
606 
607   /// couldMatchAmbiguouslyWith - Check whether this matchable could
608   /// ambiguously match the same set of operands as \p RHS (without being a
609   /// strictly superior match).
610   bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const {
611     // The primary comparator is the instruction mnemonic.
612     if (Mnemonic != RHS.Mnemonic)
613       return false;
614 
615     // The number of operands is unambiguous.
616     if (AsmOperands.size() != RHS.AsmOperands.size())
617       return false;
618 
619     // Otherwise, make sure the ordering of the two instructions is unambiguous
620     // by checking that either (a) a token or operand kind discriminates them,
621     // or (b) the ordering among equivalent kinds is consistent.
622 
623     // Tokens and operand kinds are unambiguous (assuming a correct target
624     // specific parser).
625     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
626       if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
627           AsmOperands[i].Class->Kind == ClassInfo::Token)
628         if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
629             *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
630           return false;
631 
632     // Otherwise, this operand could commute if all operands are equivalent, or
633     // there is a pair of operands that compare less than and a pair that
634     // compare greater than.
635     bool HasLT = false, HasGT = false;
636     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
637       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
638         HasLT = true;
639       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
640         HasGT = true;
641     }
642 
643     return HasLT == HasGT;
644   }
645 
646   void dump() const;
647 
648 private:
649   void tokenizeAsmString(AsmMatcherInfo const &Info,
650                          AsmVariantInfo const &Variant);
651   void addAsmOperand(StringRef Token, bool IsIsolatedToken = false);
652 };
653 
654 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
655 /// feature which participates in instruction matching.
656 struct SubtargetFeatureInfo {
657   /// \brief The predicate record for this feature.
658   Record *TheDef;
659 
660   /// \brief An unique index assigned to represent this feature.
661   uint64_t Index;
662 
663   SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {}
664 
665   /// \brief The name of the enumerated constant identifying this feature.
666   std::string getEnumName() const {
667     return "Feature_" + TheDef->getName();
668   }
669 
670   void dump() const {
671     errs() << getEnumName() << " " << Index << "\n";
672     TheDef->dump();
673   }
674 };
675 
676 struct OperandMatchEntry {
677   unsigned OperandMask;
678   const MatchableInfo* MI;
679   ClassInfo *CI;
680 
681   static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
682                                   unsigned opMask) {
683     OperandMatchEntry X;
684     X.OperandMask = opMask;
685     X.CI = ci;
686     X.MI = mi;
687     return X;
688   }
689 };
690 
691 class AsmMatcherInfo {
692 public:
693   /// Tracked Records
694   RecordKeeper &Records;
695 
696   /// The tablegen AsmParser record.
697   Record *AsmParser;
698 
699   /// Target - The target information.
700   CodeGenTarget &Target;
701 
702   /// The classes which are needed for matching.
703   std::forward_list<ClassInfo> Classes;
704 
705   /// The information on the matchables to match.
706   std::vector<std::unique_ptr<MatchableInfo>> Matchables;
707 
708   /// Info for custom matching operands by user defined methods.
709   std::vector<OperandMatchEntry> OperandMatchInfo;
710 
711   /// Map of Register records to their class information.
712   typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy;
713   RegisterClassesTy RegisterClasses;
714 
715   /// Map of Predicate records to their subtarget information.
716   std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
717 
718   /// Map of AsmOperandClass records to their class information.
719   std::map<Record*, ClassInfo*> AsmOperandClasses;
720 
721 private:
722   /// Map of token to class information which has already been constructed.
723   std::map<std::string, ClassInfo*> TokenClasses;
724 
725   /// Map of RegisterClass records to their class information.
726   std::map<Record*, ClassInfo*> RegisterClassClasses;
727 
728 private:
729   /// getTokenClass - Lookup or create the class for the given token.
730   ClassInfo *getTokenClass(StringRef Token);
731 
732   /// getOperandClass - Lookup or create the class for the given operand.
733   ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
734                              int SubOpIdx);
735   ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
736 
737   /// buildRegisterClasses - Build the ClassInfo* instances for register
738   /// classes.
739   void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters);
740 
741   /// buildOperandClasses - Build the ClassInfo* instances for user defined
742   /// operand classes.
743   void buildOperandClasses();
744 
745   void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
746                                         unsigned AsmOpIdx);
747   void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
748                                   MatchableInfo::AsmOperand &Op);
749 
750 public:
751   AsmMatcherInfo(Record *AsmParser,
752                  CodeGenTarget &Target,
753                  RecordKeeper &Records);
754 
755   /// buildInfo - Construct the various tables used during matching.
756   void buildInfo();
757 
758   /// buildOperandMatchInfo - Build the necessary information to handle user
759   /// defined operand parsing methods.
760   void buildOperandMatchInfo();
761 
762   /// getSubtargetFeature - Lookup or create the subtarget feature info for the
763   /// given operand.
764   const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
765     assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
766     const auto &I = SubtargetFeatures.find(Def);
767     return I == SubtargetFeatures.end() ? nullptr : &I->second;
768   }
769 
770   RecordKeeper &getRecords() const {
771     return Records;
772   }
773 
774   bool hasOptionalOperands() const {
775     return find_if(Classes, [](const ClassInfo &Class) {
776              return Class.IsOptional;
777            }) != Classes.end();
778   }
779 };
780 
781 } // end anonymous namespace
782 
783 void MatchableInfo::dump() const {
784   errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
785 
786   for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
787     const AsmOperand &Op = AsmOperands[i];
788     errs() << "  op[" << i << "] = " << Op.Class->ClassName << " - ";
789     errs() << '\"' << Op.Token << "\"\n";
790   }
791 }
792 
793 static std::pair<StringRef, StringRef>
794 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
795   // Split via the '='.
796   std::pair<StringRef, StringRef> Ops = S.split('=');
797   if (Ops.second == "")
798     PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
799   // Trim whitespace and the leading '$' on the operand names.
800   size_t start = Ops.first.find_first_of('$');
801   if (start == std::string::npos)
802     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
803   Ops.first = Ops.first.slice(start + 1, std::string::npos);
804   size_t end = Ops.first.find_last_of(" \t");
805   Ops.first = Ops.first.slice(0, end);
806   // Now the second operand.
807   start = Ops.second.find_first_of('$');
808   if (start == std::string::npos)
809     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
810   Ops.second = Ops.second.slice(start + 1, std::string::npos);
811   end = Ops.second.find_last_of(" \t");
812   Ops.first = Ops.first.slice(0, end);
813   return Ops;
814 }
815 
816 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
817   // Figure out which operands are aliased and mark them as tied.
818   std::pair<StringRef, StringRef> Ops =
819     parseTwoOperandConstraint(Constraint, TheDef->getLoc());
820 
821   // Find the AsmOperands that refer to the operands we're aliasing.
822   int SrcAsmOperand = findAsmOperandNamed(Ops.first);
823   int DstAsmOperand = findAsmOperandNamed(Ops.second);
824   if (SrcAsmOperand == -1)
825     PrintFatalError(TheDef->getLoc(),
826                     "unknown source two-operand alias operand '" + Ops.first +
827                     "'.");
828   if (DstAsmOperand == -1)
829     PrintFatalError(TheDef->getLoc(),
830                     "unknown destination two-operand alias operand '" +
831                     Ops.second + "'.");
832 
833   // Find the ResOperand that refers to the operand we're aliasing away
834   // and update it to refer to the combined operand instead.
835   for (ResOperand &Op : ResOperands) {
836     if (Op.Kind == ResOperand::RenderAsmOperand &&
837         Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
838       Op.AsmOperandNum = DstAsmOperand;
839       break;
840     }
841   }
842   // Remove the AsmOperand for the alias operand.
843   AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
844   // Adjust the ResOperand references to any AsmOperands that followed
845   // the one we just deleted.
846   for (ResOperand &Op : ResOperands) {
847     switch(Op.Kind) {
848     default:
849       // Nothing to do for operands that don't reference AsmOperands.
850       break;
851     case ResOperand::RenderAsmOperand:
852       if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
853         --Op.AsmOperandNum;
854       break;
855     case ResOperand::TiedOperand:
856       if (Op.TiedOperandNum > (unsigned)SrcAsmOperand)
857         --Op.TiedOperandNum;
858       break;
859     }
860   }
861 }
862 
863 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
864 /// if present, from specified token.
865 static void
866 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op,
867                                       const AsmMatcherInfo &Info,
868                                       StringRef RegisterPrefix) {
869   StringRef Tok = Op.Token;
870 
871   // If this token is not an isolated token, i.e., it isn't separated from
872   // other tokens (e.g. with whitespace), don't interpret it as a register name.
873   if (!Op.IsIsolatedToken)
874     return;
875 
876   if (RegisterPrefix.empty()) {
877     std::string LoweredTok = Tok.lower();
878     if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
879       Op.SingletonReg = Reg->TheDef;
880     return;
881   }
882 
883   if (!Tok.startswith(RegisterPrefix))
884     return;
885 
886   StringRef RegName = Tok.substr(RegisterPrefix.size());
887   if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
888     Op.SingletonReg = Reg->TheDef;
889 
890   // If there is no register prefix (i.e. "%" in "%eax"), then this may
891   // be some random non-register token, just ignore it.
892 }
893 
894 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
895                                SmallPtrSetImpl<Record*> &SingletonRegisters,
896                                AsmVariantInfo const &Variant,
897                                bool HasMnemonicFirst) {
898   AsmVariantID = Variant.AsmVariantNo;
899   AsmString =
900     CodeGenInstruction::FlattenAsmStringVariants(AsmString,
901                                                  Variant.AsmVariantNo);
902 
903   tokenizeAsmString(Info, Variant);
904 
905   // The first token of the instruction is the mnemonic, which must be a
906   // simple string, not a $foo variable or a singleton register.
907   if (AsmOperands.empty())
908     PrintFatalError(TheDef->getLoc(),
909                   "Instruction '" + TheDef->getName() + "' has no tokens");
910 
911   assert(!AsmOperands[0].Token.empty());
912   if (HasMnemonicFirst) {
913     Mnemonic = AsmOperands[0].Token;
914     if (Mnemonic[0] == '$')
915       PrintFatalError(TheDef->getLoc(),
916                       "Invalid instruction mnemonic '" + Mnemonic + "'!");
917 
918     // Remove the first operand, it is tracked in the mnemonic field.
919     AsmOperands.erase(AsmOperands.begin());
920   } else if (AsmOperands[0].Token[0] != '$')
921     Mnemonic = AsmOperands[0].Token;
922 
923   // Compute the require features.
924   for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates"))
925     if (const SubtargetFeatureInfo *Feature =
926             Info.getSubtargetFeature(Predicate))
927       RequiredFeatures.push_back(Feature);
928 
929   // Collect singleton registers, if used.
930   for (MatchableInfo::AsmOperand &Op : AsmOperands) {
931     extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix);
932     if (Record *Reg = Op.SingletonReg)
933       SingletonRegisters.insert(Reg);
934   }
935 
936   const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
937   if (!DepMask)
938     DepMask = TheDef->getValue("ComplexDeprecationPredicate");
939 
940   HasDeprecation =
941       DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
942 }
943 
944 /// Append an AsmOperand for the given substring of AsmString.
945 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) {
946   AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token));
947 }
948 
949 /// tokenizeAsmString - Tokenize a simplified assembly string.
950 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info,
951                                       AsmVariantInfo const &Variant) {
952   StringRef String = AsmString;
953   size_t Prev = 0;
954   bool InTok = false;
955   bool IsIsolatedToken = true;
956   for (size_t i = 0, e = String.size(); i != e; ++i) {
957     char Char = String[i];
958     if (Variant.BreakCharacters.find(Char) != std::string::npos) {
959       if (InTok) {
960         addAsmOperand(String.slice(Prev, i), false);
961         Prev = i;
962         IsIsolatedToken = false;
963       }
964       InTok = true;
965       continue;
966     }
967     if (Variant.TokenizingCharacters.find(Char) != std::string::npos) {
968       if (InTok) {
969         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
970         InTok = false;
971         IsIsolatedToken = false;
972       }
973       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
974       Prev = i + 1;
975       IsIsolatedToken = true;
976       continue;
977     }
978     if (Variant.SeparatorCharacters.find(Char) != std::string::npos) {
979       if (InTok) {
980         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
981         InTok = false;
982       }
983       Prev = i + 1;
984       IsIsolatedToken = true;
985       continue;
986     }
987 
988     switch (Char) {
989     case '\\':
990       if (InTok) {
991         addAsmOperand(String.slice(Prev, i), false);
992         InTok = false;
993         IsIsolatedToken = false;
994       }
995       ++i;
996       assert(i != String.size() && "Invalid quoted character");
997       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
998       Prev = i + 1;
999       IsIsolatedToken = false;
1000       break;
1001 
1002     case '$': {
1003       if (InTok) {
1004         addAsmOperand(String.slice(Prev, i), false);
1005         InTok = false;
1006         IsIsolatedToken = false;
1007       }
1008 
1009       // If this isn't "${", start new identifier looking like "$xxx"
1010       if (i + 1 == String.size() || String[i + 1] != '{') {
1011         Prev = i;
1012         break;
1013       }
1014 
1015       size_t EndPos = String.find('}', i);
1016       assert(EndPos != StringRef::npos &&
1017              "Missing brace in operand reference!");
1018       addAsmOperand(String.slice(i, EndPos+1), IsIsolatedToken);
1019       Prev = EndPos + 1;
1020       i = EndPos;
1021       IsIsolatedToken = false;
1022       break;
1023     }
1024 
1025     default:
1026       InTok = true;
1027       break;
1028     }
1029   }
1030   if (InTok && Prev != String.size())
1031     addAsmOperand(String.substr(Prev), IsIsolatedToken);
1032 }
1033 
1034 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const {
1035   // Reject matchables with no .s string.
1036   if (AsmString.empty())
1037     PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
1038 
1039   // Reject any matchables with a newline in them, they should be marked
1040   // isCodeGenOnly if they are pseudo instructions.
1041   if (AsmString.find('\n') != std::string::npos)
1042     PrintFatalError(TheDef->getLoc(),
1043                   "multiline instruction is not valid for the asmparser, "
1044                   "mark it isCodeGenOnly");
1045 
1046   // Remove comments from the asm string.  We know that the asmstring only
1047   // has one line.
1048   if (!CommentDelimiter.empty() &&
1049       StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
1050     PrintFatalError(TheDef->getLoc(),
1051                   "asmstring for instruction has comment character in it, "
1052                   "mark it isCodeGenOnly");
1053 
1054   // Reject matchables with operand modifiers, these aren't something we can
1055   // handle, the target should be refactored to use operands instead of
1056   // modifiers.
1057   //
1058   // Also, check for instructions which reference the operand multiple times;
1059   // this implies a constraint we would not honor.
1060   std::set<std::string> OperandNames;
1061   for (const AsmOperand &Op : AsmOperands) {
1062     StringRef Tok = Op.Token;
1063     if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
1064       PrintFatalError(TheDef->getLoc(),
1065                       "matchable with operand modifier '" + Tok +
1066                       "' not supported by asm matcher.  Mark isCodeGenOnly!");
1067 
1068     // Verify that any operand is only mentioned once.
1069     // We reject aliases and ignore instructions for now.
1070     if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
1071       if (!Hack)
1072         PrintFatalError(TheDef->getLoc(),
1073                         "ERROR: matchable with tied operand '" + Tok +
1074                         "' can never be matched!");
1075       // FIXME: Should reject these.  The ARM backend hits this with $lane in a
1076       // bunch of instructions.  It is unclear what the right answer is.
1077       DEBUG({
1078         errs() << "warning: '" << TheDef->getName() << "': "
1079                << "ignoring instruction with tied operand '"
1080                << Tok << "'\n";
1081       });
1082       return false;
1083     }
1084   }
1085 
1086   return true;
1087 }
1088 
1089 static std::string getEnumNameForToken(StringRef Str) {
1090   std::string Res;
1091 
1092   for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
1093     switch (*it) {
1094     case '*': Res += "_STAR_"; break;
1095     case '%': Res += "_PCT_"; break;
1096     case ':': Res += "_COLON_"; break;
1097     case '!': Res += "_EXCLAIM_"; break;
1098     case '.': Res += "_DOT_"; break;
1099     case '<': Res += "_LT_"; break;
1100     case '>': Res += "_GT_"; break;
1101     case '-': Res += "_MINUS_"; break;
1102     default:
1103       if ((*it >= 'A' && *it <= 'Z') ||
1104           (*it >= 'a' && *it <= 'z') ||
1105           (*it >= '0' && *it <= '9'))
1106         Res += *it;
1107       else
1108         Res += "_" + utostr((unsigned) *it) + "_";
1109     }
1110   }
1111 
1112   return Res;
1113 }
1114 
1115 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
1116   ClassInfo *&Entry = TokenClasses[Token];
1117 
1118   if (!Entry) {
1119     Classes.emplace_front();
1120     Entry = &Classes.front();
1121     Entry->Kind = ClassInfo::Token;
1122     Entry->ClassName = "Token";
1123     Entry->Name = "MCK_" + getEnumNameForToken(Token);
1124     Entry->ValueName = Token;
1125     Entry->PredicateMethod = "<invalid>";
1126     Entry->RenderMethod = "<invalid>";
1127     Entry->ParserMethod = "";
1128     Entry->DiagnosticType = "";
1129     Entry->IsOptional = false;
1130     Entry->DefaultMethod = "<invalid>";
1131   }
1132 
1133   return Entry;
1134 }
1135 
1136 ClassInfo *
1137 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1138                                 int SubOpIdx) {
1139   Record *Rec = OI.Rec;
1140   if (SubOpIdx != -1)
1141     Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1142   return getOperandClass(Rec, SubOpIdx);
1143 }
1144 
1145 ClassInfo *
1146 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1147   if (Rec->isSubClassOf("RegisterOperand")) {
1148     // RegisterOperand may have an associated ParserMatchClass. If it does,
1149     // use it, else just fall back to the underlying register class.
1150     const RecordVal *R = Rec->getValue("ParserMatchClass");
1151     if (!R || !R->getValue())
1152       PrintFatalError("Record `" + Rec->getName() +
1153         "' does not have a ParserMatchClass!\n");
1154 
1155     if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) {
1156       Record *MatchClass = DI->getDef();
1157       if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1158         return CI;
1159     }
1160 
1161     // No custom match class. Just use the register class.
1162     Record *ClassRec = Rec->getValueAsDef("RegClass");
1163     if (!ClassRec)
1164       PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1165                     "' has no associated register class!\n");
1166     if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1167       return CI;
1168     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1169   }
1170 
1171   if (Rec->isSubClassOf("RegisterClass")) {
1172     if (ClassInfo *CI = RegisterClassClasses[Rec])
1173       return CI;
1174     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1175   }
1176 
1177   if (!Rec->isSubClassOf("Operand"))
1178     PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() +
1179                   "' does not derive from class Operand!\n");
1180   Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1181   if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1182     return CI;
1183 
1184   PrintFatalError(Rec->getLoc(), "operand has no match class!");
1185 }
1186 
1187 struct LessRegisterSet {
1188   bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const {
1189     // std::set<T> defines its own compariso "operator<", but it
1190     // performs a lexicographical comparison by T's innate comparison
1191     // for some reason. We don't want non-deterministic pointer
1192     // comparisons so use this instead.
1193     return std::lexicographical_compare(LHS.begin(), LHS.end(),
1194                                         RHS.begin(), RHS.end(),
1195                                         LessRecordByID());
1196   }
1197 };
1198 
1199 void AsmMatcherInfo::
1200 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) {
1201   const auto &Registers = Target.getRegBank().getRegisters();
1202   auto &RegClassList = Target.getRegBank().getRegClasses();
1203 
1204   typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1205 
1206   // The register sets used for matching.
1207   RegisterSetSet RegisterSets;
1208 
1209   // Gather the defined sets.
1210   for (const CodeGenRegisterClass &RC : RegClassList)
1211     RegisterSets.insert(
1212         RegisterSet(RC.getOrder().begin(), RC.getOrder().end()));
1213 
1214   // Add any required singleton sets.
1215   for (Record *Rec : SingletonRegisters) {
1216     RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1217   }
1218 
1219   // Introduce derived sets where necessary (when a register does not determine
1220   // a unique register set class), and build the mapping of registers to the set
1221   // they should classify to.
1222   std::map<Record*, RegisterSet> RegisterMap;
1223   for (const CodeGenRegister &CGR : Registers) {
1224     // Compute the intersection of all sets containing this register.
1225     RegisterSet ContainingSet;
1226 
1227     for (const RegisterSet &RS : RegisterSets) {
1228       if (!RS.count(CGR.TheDef))
1229         continue;
1230 
1231       if (ContainingSet.empty()) {
1232         ContainingSet = RS;
1233         continue;
1234       }
1235 
1236       RegisterSet Tmp;
1237       std::swap(Tmp, ContainingSet);
1238       std::insert_iterator<RegisterSet> II(ContainingSet,
1239                                            ContainingSet.begin());
1240       std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II,
1241                             LessRecordByID());
1242     }
1243 
1244     if (!ContainingSet.empty()) {
1245       RegisterSets.insert(ContainingSet);
1246       RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
1247     }
1248   }
1249 
1250   // Construct the register classes.
1251   std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses;
1252   unsigned Index = 0;
1253   for (const RegisterSet &RS : RegisterSets) {
1254     Classes.emplace_front();
1255     ClassInfo *CI = &Classes.front();
1256     CI->Kind = ClassInfo::RegisterClass0 + Index;
1257     CI->ClassName = "Reg" + utostr(Index);
1258     CI->Name = "MCK_Reg" + utostr(Index);
1259     CI->ValueName = "";
1260     CI->PredicateMethod = ""; // unused
1261     CI->RenderMethod = "addRegOperands";
1262     CI->Registers = RS;
1263     // FIXME: diagnostic type.
1264     CI->DiagnosticType = "";
1265     CI->IsOptional = false;
1266     CI->DefaultMethod = ""; // unused
1267     RegisterSetClasses.insert(std::make_pair(RS, CI));
1268     ++Index;
1269   }
1270 
1271   // Find the superclasses; we could compute only the subgroup lattice edges,
1272   // but there isn't really a point.
1273   for (const RegisterSet &RS : RegisterSets) {
1274     ClassInfo *CI = RegisterSetClasses[RS];
1275     for (const RegisterSet &RS2 : RegisterSets)
1276       if (RS != RS2 &&
1277           std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(),
1278                         LessRecordByID()))
1279         CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1280   }
1281 
1282   // Name the register classes which correspond to a user defined RegisterClass.
1283   for (const CodeGenRegisterClass &RC : RegClassList) {
1284     // Def will be NULL for non-user defined register classes.
1285     Record *Def = RC.getDef();
1286     if (!Def)
1287       continue;
1288     ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(),
1289                                                    RC.getOrder().end())];
1290     if (CI->ValueName.empty()) {
1291       CI->ClassName = RC.getName();
1292       CI->Name = "MCK_" + RC.getName();
1293       CI->ValueName = RC.getName();
1294     } else
1295       CI->ValueName = CI->ValueName + "," + RC.getName();
1296 
1297     RegisterClassClasses.insert(std::make_pair(Def, CI));
1298   }
1299 
1300   // Populate the map for individual registers.
1301   for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(),
1302          ie = RegisterMap.end(); it != ie; ++it)
1303     RegisterClasses[it->first] = RegisterSetClasses[it->second];
1304 
1305   // Name the register classes which correspond to singleton registers.
1306   for (Record *Rec : SingletonRegisters) {
1307     ClassInfo *CI = RegisterClasses[Rec];
1308     assert(CI && "Missing singleton register class info!");
1309 
1310     if (CI->ValueName.empty()) {
1311       CI->ClassName = Rec->getName();
1312       CI->Name = "MCK_" + Rec->getName();
1313       CI->ValueName = Rec->getName();
1314     } else
1315       CI->ValueName = CI->ValueName + "," + Rec->getName();
1316   }
1317 }
1318 
1319 void AsmMatcherInfo::buildOperandClasses() {
1320   std::vector<Record*> AsmOperands =
1321     Records.getAllDerivedDefinitions("AsmOperandClass");
1322 
1323   // Pre-populate AsmOperandClasses map.
1324   for (Record *Rec : AsmOperands) {
1325     Classes.emplace_front();
1326     AsmOperandClasses[Rec] = &Classes.front();
1327   }
1328 
1329   unsigned Index = 0;
1330   for (Record *Rec : AsmOperands) {
1331     ClassInfo *CI = AsmOperandClasses[Rec];
1332     CI->Kind = ClassInfo::UserClass0 + Index;
1333 
1334     ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1335     for (Init *I : Supers->getValues()) {
1336       DefInit *DI = dyn_cast<DefInit>(I);
1337       if (!DI) {
1338         PrintError(Rec->getLoc(), "Invalid super class reference!");
1339         continue;
1340       }
1341 
1342       ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1343       if (!SC)
1344         PrintError(Rec->getLoc(), "Invalid super class reference!");
1345       else
1346         CI->SuperClasses.push_back(SC);
1347     }
1348     CI->ClassName = Rec->getValueAsString("Name");
1349     CI->Name = "MCK_" + CI->ClassName;
1350     CI->ValueName = Rec->getName();
1351 
1352     // Get or construct the predicate method name.
1353     Init *PMName = Rec->getValueInit("PredicateMethod");
1354     if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1355       CI->PredicateMethod = SI->getValue();
1356     } else {
1357       assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1358       CI->PredicateMethod = "is" + CI->ClassName;
1359     }
1360 
1361     // Get or construct the render method name.
1362     Init *RMName = Rec->getValueInit("RenderMethod");
1363     if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1364       CI->RenderMethod = SI->getValue();
1365     } else {
1366       assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1367       CI->RenderMethod = "add" + CI->ClassName + "Operands";
1368     }
1369 
1370     // Get the parse method name or leave it as empty.
1371     Init *PRMName = Rec->getValueInit("ParserMethod");
1372     if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1373       CI->ParserMethod = SI->getValue();
1374 
1375     // Get the diagnostic type or leave it as empty.
1376     // Get the parse method name or leave it as empty.
1377     Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1378     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1379       CI->DiagnosticType = SI->getValue();
1380 
1381     Init *IsOptional = Rec->getValueInit("IsOptional");
1382     if (BitInit *BI = dyn_cast<BitInit>(IsOptional))
1383       CI->IsOptional = BI->getValue();
1384 
1385     // Get or construct the default method name.
1386     Init *DMName = Rec->getValueInit("DefaultMethod");
1387     if (StringInit *SI = dyn_cast<StringInit>(DMName)) {
1388       CI->DefaultMethod = SI->getValue();
1389     } else {
1390       assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!");
1391       CI->DefaultMethod = "default" + CI->ClassName + "Operands";
1392     }
1393 
1394     ++Index;
1395   }
1396 }
1397 
1398 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1399                                CodeGenTarget &target,
1400                                RecordKeeper &records)
1401   : Records(records), AsmParser(asmParser), Target(target) {
1402 }
1403 
1404 /// buildOperandMatchInfo - Build the necessary information to handle user
1405 /// defined operand parsing methods.
1406 void AsmMatcherInfo::buildOperandMatchInfo() {
1407 
1408   /// Map containing a mask with all operands indices that can be found for
1409   /// that class inside a instruction.
1410   typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy;
1411   OpClassMaskTy OpClassMask;
1412 
1413   for (const auto &MI : Matchables) {
1414     OpClassMask.clear();
1415 
1416     // Keep track of all operands of this instructions which belong to the
1417     // same class.
1418     for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
1419       const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
1420       if (Op.Class->ParserMethod.empty())
1421         continue;
1422       unsigned &OperandMask = OpClassMask[Op.Class];
1423       OperandMask |= (1 << i);
1424     }
1425 
1426     // Generate operand match info for each mnemonic/operand class pair.
1427     for (const auto &OCM : OpClassMask) {
1428       unsigned OpMask = OCM.second;
1429       ClassInfo *CI = OCM.first;
1430       OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI,
1431                                                            OpMask));
1432     }
1433   }
1434 }
1435 
1436 void AsmMatcherInfo::buildInfo() {
1437   // Build information about all of the AssemblerPredicates.
1438   std::vector<Record*> AllPredicates =
1439     Records.getAllDerivedDefinitions("Predicate");
1440   for (Record *Pred : AllPredicates) {
1441     // Ignore predicates that are not intended for the assembler.
1442     if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1443       continue;
1444 
1445     if (Pred->getName().empty())
1446       PrintFatalError(Pred->getLoc(), "Predicate has no name!");
1447 
1448     SubtargetFeatures.insert(std::make_pair(
1449         Pred, SubtargetFeatureInfo(Pred, SubtargetFeatures.size())));
1450     DEBUG(SubtargetFeatures.find(Pred)->second.dump());
1451     assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!");
1452   }
1453 
1454   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
1455 
1456   // Parse the instructions; we need to do this first so that we can gather the
1457   // singleton register classes.
1458   SmallPtrSet<Record*, 16> SingletonRegisters;
1459   unsigned VariantCount = Target.getAsmParserVariantCount();
1460   for (unsigned VC = 0; VC != VariantCount; ++VC) {
1461     Record *AsmVariant = Target.getAsmParserVariant(VC);
1462     std::string CommentDelimiter =
1463       AsmVariant->getValueAsString("CommentDelimiter");
1464     AsmVariantInfo Variant;
1465     Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1466     Variant.TokenizingCharacters =
1467         AsmVariant->getValueAsString("TokenizingCharacters");
1468     Variant.SeparatorCharacters =
1469         AsmVariant->getValueAsString("SeparatorCharacters");
1470     Variant.BreakCharacters =
1471         AsmVariant->getValueAsString("BreakCharacters");
1472     Variant.Name = AsmVariant->getValueAsString("Name");
1473     Variant.AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1474 
1475     for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) {
1476 
1477       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1478       // filter the set of instructions we consider.
1479       if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix))
1480         continue;
1481 
1482       // Ignore "codegen only" instructions.
1483       if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1484         continue;
1485 
1486       // Ignore instructions for different instructions
1487       const std::string V = CGI->TheDef->getValueAsString("AsmVariantName");
1488       if (!V.empty() && V != Variant.Name)
1489         continue;
1490 
1491       auto II = llvm::make_unique<MatchableInfo>(*CGI);
1492 
1493       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1494 
1495       // Ignore instructions which shouldn't be matched and diagnose invalid
1496       // instruction definitions with an error.
1497       if (!II->validate(CommentDelimiter, true))
1498         continue;
1499 
1500       Matchables.push_back(std::move(II));
1501     }
1502 
1503     // Parse all of the InstAlias definitions and stick them in the list of
1504     // matchables.
1505     std::vector<Record*> AllInstAliases =
1506       Records.getAllDerivedDefinitions("InstAlias");
1507     for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1508       auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i],
1509                                                        Variant.AsmVariantNo,
1510                                                        Target);
1511 
1512       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1513       // filter the set of instruction aliases we consider, based on the target
1514       // instruction.
1515       if (!StringRef(Alias->ResultInst->TheDef->getName())
1516             .startswith( MatchPrefix))
1517         continue;
1518 
1519       const std::string V = Alias->TheDef->getValueAsString("AsmVariantName");
1520       if (!V.empty() && V != Variant.Name)
1521         continue;
1522 
1523       auto II = llvm::make_unique<MatchableInfo>(std::move(Alias));
1524 
1525       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1526 
1527       // Validate the alias definitions.
1528       II->validate(CommentDelimiter, false);
1529 
1530       Matchables.push_back(std::move(II));
1531     }
1532   }
1533 
1534   // Build info for the register classes.
1535   buildRegisterClasses(SingletonRegisters);
1536 
1537   // Build info for the user defined assembly operand classes.
1538   buildOperandClasses();
1539 
1540   // Build the information about matchables, now that we have fully formed
1541   // classes.
1542   std::vector<std::unique_ptr<MatchableInfo>> NewMatchables;
1543   for (auto &II : Matchables) {
1544     // Parse the tokens after the mnemonic.
1545     // Note: buildInstructionOperandReference may insert new AsmOperands, so
1546     // don't precompute the loop bound.
1547     for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1548       MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1549       StringRef Token = Op.Token;
1550 
1551       // Check for singleton registers.
1552       if (Record *RegRecord = Op.SingletonReg) {
1553         Op.Class = RegisterClasses[RegRecord];
1554         assert(Op.Class && Op.Class->Registers.size() == 1 &&
1555                "Unexpected class for singleton register");
1556         continue;
1557       }
1558 
1559       // Check for simple tokens.
1560       if (Token[0] != '$') {
1561         Op.Class = getTokenClass(Token);
1562         continue;
1563       }
1564 
1565       if (Token.size() > 1 && isdigit(Token[1])) {
1566         Op.Class = getTokenClass(Token);
1567         continue;
1568       }
1569 
1570       // Otherwise this is an operand reference.
1571       StringRef OperandName;
1572       if (Token[1] == '{')
1573         OperandName = Token.substr(2, Token.size() - 3);
1574       else
1575         OperandName = Token.substr(1);
1576 
1577       if (II->DefRec.is<const CodeGenInstruction*>())
1578         buildInstructionOperandReference(II.get(), OperandName, i);
1579       else
1580         buildAliasOperandReference(II.get(), OperandName, Op);
1581     }
1582 
1583     if (II->DefRec.is<const CodeGenInstruction*>()) {
1584       II->buildInstructionResultOperands();
1585       // If the instruction has a two-operand alias, build up the
1586       // matchable here. We'll add them in bulk at the end to avoid
1587       // confusing this loop.
1588       std::string Constraint =
1589         II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1590       if (Constraint != "") {
1591         // Start by making a copy of the original matchable.
1592         auto AliasII = llvm::make_unique<MatchableInfo>(*II);
1593 
1594         // Adjust it to be a two-operand alias.
1595         AliasII->formTwoOperandAlias(Constraint);
1596 
1597         // Add the alias to the matchables list.
1598         NewMatchables.push_back(std::move(AliasII));
1599       }
1600     } else
1601       II->buildAliasResultOperands();
1602   }
1603   if (!NewMatchables.empty())
1604     Matchables.insert(Matchables.end(),
1605                       std::make_move_iterator(NewMatchables.begin()),
1606                       std::make_move_iterator(NewMatchables.end()));
1607 
1608   // Process token alias definitions and set up the associated superclass
1609   // information.
1610   std::vector<Record*> AllTokenAliases =
1611     Records.getAllDerivedDefinitions("TokenAlias");
1612   for (Record *Rec : AllTokenAliases) {
1613     ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1614     ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1615     if (FromClass == ToClass)
1616       PrintFatalError(Rec->getLoc(),
1617                     "error: Destination value identical to source value.");
1618     FromClass->SuperClasses.push_back(ToClass);
1619   }
1620 
1621   // Reorder classes so that classes precede super classes.
1622   Classes.sort();
1623 
1624 #ifndef NDEBUG
1625   // Verify that the table is now sorted
1626   for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) {
1627     for (auto J = I; J != E; ++J) {
1628       assert(!(*J < *I));
1629       assert(I == J || !J->isSubsetOf(*I));
1630     }
1631   }
1632 #endif // NDEBUG
1633 }
1634 
1635 /// buildInstructionOperandReference - The specified operand is a reference to a
1636 /// named operand such as $src.  Resolve the Class and OperandInfo pointers.
1637 void AsmMatcherInfo::
1638 buildInstructionOperandReference(MatchableInfo *II,
1639                                  StringRef OperandName,
1640                                  unsigned AsmOpIdx) {
1641   const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1642   const CGIOperandList &Operands = CGI.Operands;
1643   MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1644 
1645   // Map this token to an operand.
1646   unsigned Idx;
1647   if (!Operands.hasOperandNamed(OperandName, Idx))
1648     PrintFatalError(II->TheDef->getLoc(),
1649                     "error: unable to find operand: '" + OperandName + "'");
1650 
1651   // If the instruction operand has multiple suboperands, but the parser
1652   // match class for the asm operand is still the default "ImmAsmOperand",
1653   // then handle each suboperand separately.
1654   if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1655     Record *Rec = Operands[Idx].Rec;
1656     assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1657     Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1658     if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1659       // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1660       StringRef Token = Op->Token; // save this in case Op gets moved
1661       for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1662         MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token);
1663         NewAsmOp.SubOpIdx = SI;
1664         II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1665       }
1666       // Replace Op with first suboperand.
1667       Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1668       Op->SubOpIdx = 0;
1669     }
1670   }
1671 
1672   // Set up the operand class.
1673   Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1674 
1675   // If the named operand is tied, canonicalize it to the untied operand.
1676   // For example, something like:
1677   //   (outs GPR:$dst), (ins GPR:$src)
1678   // with an asmstring of
1679   //   "inc $src"
1680   // we want to canonicalize to:
1681   //   "inc $dst"
1682   // so that we know how to provide the $dst operand when filling in the result.
1683   int OITied = -1;
1684   if (Operands[Idx].MINumOperands == 1)
1685     OITied = Operands[Idx].getTiedRegister();
1686   if (OITied != -1) {
1687     // The tied operand index is an MIOperand index, find the operand that
1688     // contains it.
1689     std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1690     OperandName = Operands[Idx.first].Name;
1691     Op->SubOpIdx = Idx.second;
1692   }
1693 
1694   Op->SrcOpName = OperandName;
1695 }
1696 
1697 /// buildAliasOperandReference - When parsing an operand reference out of the
1698 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1699 /// operand reference is by looking it up in the result pattern definition.
1700 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1701                                                 StringRef OperandName,
1702                                                 MatchableInfo::AsmOperand &Op) {
1703   const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1704 
1705   // Set up the operand class.
1706   for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1707     if (CGA.ResultOperands[i].isRecord() &&
1708         CGA.ResultOperands[i].getName() == OperandName) {
1709       // It's safe to go with the first one we find, because CodeGenInstAlias
1710       // validates that all operands with the same name have the same record.
1711       Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1712       // Use the match class from the Alias definition, not the
1713       // destination instruction, as we may have an immediate that's
1714       // being munged by the match class.
1715       Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1716                                  Op.SubOpIdx);
1717       Op.SrcOpName = OperandName;
1718       return;
1719     }
1720 
1721   PrintFatalError(II->TheDef->getLoc(),
1722                   "error: unable to find operand: '" + OperandName + "'");
1723 }
1724 
1725 void MatchableInfo::buildInstructionResultOperands() {
1726   const CodeGenInstruction *ResultInst = getResultInst();
1727 
1728   // Loop over all operands of the result instruction, determining how to
1729   // populate them.
1730   for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) {
1731     // If this is a tied operand, just copy from the previously handled operand.
1732     int TiedOp = -1;
1733     if (OpInfo.MINumOperands == 1)
1734       TiedOp = OpInfo.getTiedRegister();
1735     if (TiedOp != -1) {
1736       ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1737       continue;
1738     }
1739 
1740     // Find out what operand from the asmparser this MCInst operand comes from.
1741     int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1742     if (OpInfo.Name.empty() || SrcOperand == -1) {
1743       // This may happen for operands that are tied to a suboperand of a
1744       // complex operand.  Simply use a dummy value here; nobody should
1745       // use this operand slot.
1746       // FIXME: The long term goal is for the MCOperand list to not contain
1747       // tied operands at all.
1748       ResOperands.push_back(ResOperand::getImmOp(0));
1749       continue;
1750     }
1751 
1752     // Check if the one AsmOperand populates the entire operand.
1753     unsigned NumOperands = OpInfo.MINumOperands;
1754     if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1755       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1756       continue;
1757     }
1758 
1759     // Add a separate ResOperand for each suboperand.
1760     for (unsigned AI = 0; AI < NumOperands; ++AI) {
1761       assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1762              AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1763              "unexpected AsmOperands for suboperands");
1764       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1765     }
1766   }
1767 }
1768 
1769 void MatchableInfo::buildAliasResultOperands() {
1770   const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1771   const CodeGenInstruction *ResultInst = getResultInst();
1772 
1773   // Loop over all operands of the result instruction, determining how to
1774   // populate them.
1775   unsigned AliasOpNo = 0;
1776   unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1777   for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1778     const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1779 
1780     // If this is a tied operand, just copy from the previously handled operand.
1781     int TiedOp = -1;
1782     if (OpInfo->MINumOperands == 1)
1783       TiedOp = OpInfo->getTiedRegister();
1784     if (TiedOp != -1) {
1785       ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1786       continue;
1787     }
1788 
1789     // Handle all the suboperands for this operand.
1790     const std::string &OpName = OpInfo->Name;
1791     for ( ; AliasOpNo <  LastOpNo &&
1792             CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1793       int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1794 
1795       // Find out what operand from the asmparser that this MCInst operand
1796       // comes from.
1797       switch (CGA.ResultOperands[AliasOpNo].Kind) {
1798       case CodeGenInstAlias::ResultOperand::K_Record: {
1799         StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1800         int SrcOperand = findAsmOperand(Name, SubIdx);
1801         if (SrcOperand == -1)
1802           PrintFatalError(TheDef->getLoc(), "Instruction '" +
1803                         TheDef->getName() + "' has operand '" + OpName +
1804                         "' that doesn't appear in asm string!");
1805         unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1806         ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1807                                                         NumOperands));
1808         break;
1809       }
1810       case CodeGenInstAlias::ResultOperand::K_Imm: {
1811         int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1812         ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1813         break;
1814       }
1815       case CodeGenInstAlias::ResultOperand::K_Reg: {
1816         Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1817         ResOperands.push_back(ResOperand::getRegOp(Reg));
1818         break;
1819       }
1820       }
1821     }
1822   }
1823 }
1824 
1825 static unsigned getConverterOperandID(const std::string &Name,
1826                                       SmallSetVector<std::string, 16> &Table,
1827                                       bool &IsNew) {
1828   IsNew = Table.insert(Name);
1829 
1830   unsigned ID = IsNew ? Table.size() - 1 : find(Table, Name) - Table.begin();
1831 
1832   assert(ID < Table.size());
1833 
1834   return ID;
1835 }
1836 
1837 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1838                              std::vector<std::unique_ptr<MatchableInfo>> &Infos,
1839                              bool HasMnemonicFirst, bool HasOptionalOperands,
1840                              raw_ostream &OS) {
1841   SmallSetVector<std::string, 16> OperandConversionKinds;
1842   SmallSetVector<std::string, 16> InstructionConversionKinds;
1843   std::vector<std::vector<uint8_t> > ConversionTable;
1844   size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1845 
1846   // TargetOperandClass - This is the target's operand class, like X86Operand.
1847   std::string TargetOperandClass = Target.getName() + "Operand";
1848 
1849   // Write the convert function to a separate stream, so we can drop it after
1850   // the enum. We'll build up the conversion handlers for the individual
1851   // operand types opportunistically as we encounter them.
1852   std::string ConvertFnBody;
1853   raw_string_ostream CvtOS(ConvertFnBody);
1854   // Start the unified conversion function.
1855   if (HasOptionalOperands) {
1856     CvtOS << "void " << Target.getName() << ClassName << "::\n"
1857           << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1858           << "unsigned Opcode,\n"
1859           << "                const OperandVector &Operands,\n"
1860           << "                const SmallBitVector &OptionalOperandsMask) {\n";
1861   } else {
1862     CvtOS << "void " << Target.getName() << ClassName << "::\n"
1863           << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1864           << "unsigned Opcode,\n"
1865           << "                const OperandVector &Operands) {\n";
1866   }
1867   CvtOS << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n";
1868   CvtOS << "  const uint8_t *Converter = ConversionTable[Kind];\n";
1869   if (HasOptionalOperands) {
1870     CvtOS << "  unsigned NumDefaults = 0;\n";
1871   }
1872   CvtOS << "  unsigned OpIdx;\n";
1873   CvtOS << "  Inst.setOpcode(Opcode);\n";
1874   CvtOS << "  for (const uint8_t *p = Converter; *p; p+= 2) {\n";
1875   if (HasOptionalOperands) {
1876     CvtOS << "    OpIdx = *(p + 1) - NumDefaults;\n";
1877   } else {
1878     CvtOS << "    OpIdx = *(p + 1);\n";
1879   }
1880   CvtOS << "    switch (*p) {\n";
1881   CvtOS << "    default: llvm_unreachable(\"invalid conversion entry!\");\n";
1882   CvtOS << "    case CVT_Reg:\n";
1883   CvtOS << "      static_cast<" << TargetOperandClass
1884         << "&>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n";
1885   CvtOS << "      break;\n";
1886   CvtOS << "    case CVT_Tied:\n";
1887   CvtOS << "      Inst.addOperand(Inst.getOperand(OpIdx));\n";
1888   CvtOS << "      break;\n";
1889 
1890   std::string OperandFnBody;
1891   raw_string_ostream OpOS(OperandFnBody);
1892   // Start the operand number lookup function.
1893   OpOS << "void " << Target.getName() << ClassName << "::\n"
1894        << "convertToMapAndConstraints(unsigned Kind,\n";
1895   OpOS.indent(27);
1896   OpOS << "const OperandVector &Operands) {\n"
1897        << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1898        << "  unsigned NumMCOperands = 0;\n"
1899        << "  const uint8_t *Converter = ConversionTable[Kind];\n"
1900        << "  for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1901        << "    switch (*p) {\n"
1902        << "    default: llvm_unreachable(\"invalid conversion entry!\");\n"
1903        << "    case CVT_Reg:\n"
1904        << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1905        << "      Operands[*(p + 1)]->setConstraint(\"r\");\n"
1906        << "      ++NumMCOperands;\n"
1907        << "      break;\n"
1908        << "    case CVT_Tied:\n"
1909        << "      ++NumMCOperands;\n"
1910        << "      break;\n";
1911 
1912   // Pre-populate the operand conversion kinds with the standard always
1913   // available entries.
1914   OperandConversionKinds.insert("CVT_Done");
1915   OperandConversionKinds.insert("CVT_Reg");
1916   OperandConversionKinds.insert("CVT_Tied");
1917   enum { CVT_Done, CVT_Reg, CVT_Tied };
1918 
1919   for (auto &II : Infos) {
1920     // Check if we have a custom match function.
1921     std::string AsmMatchConverter =
1922       II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1923     if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) {
1924       std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1925       II->ConversionFnKind = Signature;
1926 
1927       // Check if we have already generated this signature.
1928       if (!InstructionConversionKinds.insert(Signature))
1929         continue;
1930 
1931       // Remember this converter for the kind enum.
1932       unsigned KindID = OperandConversionKinds.size();
1933       OperandConversionKinds.insert("CVT_" +
1934                                     getEnumNameForToken(AsmMatchConverter));
1935 
1936       // Add the converter row for this instruction.
1937       ConversionTable.emplace_back();
1938       ConversionTable.back().push_back(KindID);
1939       ConversionTable.back().push_back(CVT_Done);
1940 
1941       // Add the handler to the conversion driver function.
1942       CvtOS << "    case CVT_"
1943             << getEnumNameForToken(AsmMatchConverter) << ":\n"
1944             << "      " << AsmMatchConverter << "(Inst, Operands);\n"
1945             << "      break;\n";
1946 
1947       // FIXME: Handle the operand number lookup for custom match functions.
1948       continue;
1949     }
1950 
1951     // Build the conversion function signature.
1952     std::string Signature = "Convert";
1953 
1954     std::vector<uint8_t> ConversionRow;
1955 
1956     // Compute the convert enum and the case body.
1957     MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 );
1958 
1959     for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) {
1960       const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i];
1961 
1962       // Generate code to populate each result operand.
1963       switch (OpInfo.Kind) {
1964       case MatchableInfo::ResOperand::RenderAsmOperand: {
1965         // This comes from something we parsed.
1966         const MatchableInfo::AsmOperand &Op =
1967           II->AsmOperands[OpInfo.AsmOperandNum];
1968 
1969         // Registers are always converted the same, don't duplicate the
1970         // conversion function based on them.
1971         Signature += "__";
1972         std::string Class;
1973         Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
1974         Signature += Class;
1975         Signature += utostr(OpInfo.MINumOperands);
1976         Signature += "_" + itostr(OpInfo.AsmOperandNum);
1977 
1978         // Add the conversion kind, if necessary, and get the associated ID
1979         // the index of its entry in the vector).
1980         std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" :
1981                                      Op.Class->RenderMethod);
1982         if (Op.Class->IsOptional) {
1983           // For optional operands we must also care about DefaultMethod
1984           assert(HasOptionalOperands);
1985           Name += "_" + Op.Class->DefaultMethod;
1986         }
1987         Name = getEnumNameForToken(Name);
1988 
1989         bool IsNewConverter = false;
1990         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1991                                             IsNewConverter);
1992 
1993         // Add the operand entry to the instruction kind conversion row.
1994         ConversionRow.push_back(ID);
1995         ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst);
1996 
1997         if (!IsNewConverter)
1998           break;
1999 
2000         // This is a new operand kind. Add a handler for it to the
2001         // converter driver.
2002         CvtOS << "    case " << Name << ":\n";
2003         if (Op.Class->IsOptional) {
2004           // If optional operand is not present in actual instruction then we
2005           // should call its DefaultMethod before RenderMethod
2006           assert(HasOptionalOperands);
2007           CvtOS << "      if (OptionalOperandsMask[*(p + 1) - 1]) {\n"
2008                 << "        " << Op.Class->DefaultMethod << "()"
2009                 << "->" << Op.Class->RenderMethod << "(Inst, "
2010                 << OpInfo.MINumOperands << ");\n"
2011                 << "        ++NumDefaults;\n"
2012                 << "      } else {\n"
2013                 << "        static_cast<" << TargetOperandClass
2014                 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod
2015                 << "(Inst, " << OpInfo.MINumOperands << ");\n"
2016                 << "      }\n";
2017         } else {
2018           CvtOS << "      static_cast<" << TargetOperandClass
2019                 << "&>(*Operands[OpIdx])." << Op.Class->RenderMethod
2020                 << "(Inst, " << OpInfo.MINumOperands << ");\n";
2021         }
2022         CvtOS << "      break;\n";
2023 
2024         // Add a handler for the operand number lookup.
2025         OpOS << "    case " << Name << ":\n"
2026              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
2027 
2028         if (Op.Class->isRegisterClass())
2029           OpOS << "      Operands[*(p + 1)]->setConstraint(\"r\");\n";
2030         else
2031           OpOS << "      Operands[*(p + 1)]->setConstraint(\"m\");\n";
2032         OpOS << "      NumMCOperands += " << OpInfo.MINumOperands << ";\n"
2033              << "      break;\n";
2034         break;
2035       }
2036       case MatchableInfo::ResOperand::TiedOperand: {
2037         // If this operand is tied to a previous one, just copy the MCInst
2038         // operand from the earlier one.We can only tie single MCOperand values.
2039         assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
2040         unsigned TiedOp = OpInfo.TiedOperandNum;
2041         assert(i > TiedOp && "Tied operand precedes its target!");
2042         Signature += "__Tie" + utostr(TiedOp);
2043         ConversionRow.push_back(CVT_Tied);
2044         ConversionRow.push_back(TiedOp);
2045         break;
2046       }
2047       case MatchableInfo::ResOperand::ImmOperand: {
2048         int64_t Val = OpInfo.ImmVal;
2049         std::string Ty = "imm_" + itostr(Val);
2050         Ty = getEnumNameForToken(Ty);
2051         Signature += "__" + Ty;
2052 
2053         std::string Name = "CVT_" + Ty;
2054         bool IsNewConverter = false;
2055         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2056                                             IsNewConverter);
2057         // Add the operand entry to the instruction kind conversion row.
2058         ConversionRow.push_back(ID);
2059         ConversionRow.push_back(0);
2060 
2061         if (!IsNewConverter)
2062           break;
2063 
2064         CvtOS << "    case " << Name << ":\n"
2065               << "      Inst.addOperand(MCOperand::createImm(" << Val << "));\n"
2066               << "      break;\n";
2067 
2068         OpOS << "    case " << Name << ":\n"
2069              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2070              << "      Operands[*(p + 1)]->setConstraint(\"\");\n"
2071              << "      ++NumMCOperands;\n"
2072              << "      break;\n";
2073         break;
2074       }
2075       case MatchableInfo::ResOperand::RegOperand: {
2076         std::string Reg, Name;
2077         if (!OpInfo.Register) {
2078           Name = "reg0";
2079           Reg = "0";
2080         } else {
2081           Reg = getQualifiedName(OpInfo.Register);
2082           Name = "reg" + OpInfo.Register->getName();
2083         }
2084         Signature += "__" + Name;
2085         Name = "CVT_" + Name;
2086         bool IsNewConverter = false;
2087         unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
2088                                             IsNewConverter);
2089         // Add the operand entry to the instruction kind conversion row.
2090         ConversionRow.push_back(ID);
2091         ConversionRow.push_back(0);
2092 
2093         if (!IsNewConverter)
2094           break;
2095         CvtOS << "    case " << Name << ":\n"
2096               << "      Inst.addOperand(MCOperand::createReg(" << Reg << "));\n"
2097               << "      break;\n";
2098 
2099         OpOS << "    case " << Name << ":\n"
2100              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2101              << "      Operands[*(p + 1)]->setConstraint(\"m\");\n"
2102              << "      ++NumMCOperands;\n"
2103              << "      break;\n";
2104       }
2105       }
2106     }
2107 
2108     // If there were no operands, add to the signature to that effect
2109     if (Signature == "Convert")
2110       Signature += "_NoOperands";
2111 
2112     II->ConversionFnKind = Signature;
2113 
2114     // Save the signature. If we already have it, don't add a new row
2115     // to the table.
2116     if (!InstructionConversionKinds.insert(Signature))
2117       continue;
2118 
2119     // Add the row to the table.
2120     ConversionTable.push_back(std::move(ConversionRow));
2121   }
2122 
2123   // Finish up the converter driver function.
2124   CvtOS << "    }\n  }\n}\n\n";
2125 
2126   // Finish up the operand number lookup function.
2127   OpOS << "    }\n  }\n}\n\n";
2128 
2129   OS << "namespace {\n";
2130 
2131   // Output the operand conversion kind enum.
2132   OS << "enum OperatorConversionKind {\n";
2133   for (const std::string &Converter : OperandConversionKinds)
2134     OS << "  " << Converter << ",\n";
2135   OS << "  CVT_NUM_CONVERTERS\n";
2136   OS << "};\n\n";
2137 
2138   // Output the instruction conversion kind enum.
2139   OS << "enum InstructionConversionKind {\n";
2140   for (const std::string &Signature : InstructionConversionKinds)
2141     OS << "  " << Signature << ",\n";
2142   OS << "  CVT_NUM_SIGNATURES\n";
2143   OS << "};\n\n";
2144 
2145   OS << "} // end anonymous namespace\n\n";
2146 
2147   // Output the conversion table.
2148   OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
2149      << MaxRowLength << "] = {\n";
2150 
2151   for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
2152     assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
2153     OS << "  // " << InstructionConversionKinds[Row] << "\n";
2154     OS << "  { ";
2155     for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2)
2156       OS << OperandConversionKinds[ConversionTable[Row][i]] << ", "
2157          << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
2158     OS << "CVT_Done },\n";
2159   }
2160 
2161   OS << "};\n\n";
2162 
2163   // Spit out the conversion driver function.
2164   OS << CvtOS.str();
2165 
2166   // Spit out the operand number lookup function.
2167   OS << OpOS.str();
2168 }
2169 
2170 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
2171 static void emitMatchClassEnumeration(CodeGenTarget &Target,
2172                                       std::forward_list<ClassInfo> &Infos,
2173                                       raw_ostream &OS) {
2174   OS << "namespace {\n\n";
2175 
2176   OS << "/// MatchClassKind - The kinds of classes which participate in\n"
2177      << "/// instruction matching.\n";
2178   OS << "enum MatchClassKind {\n";
2179   OS << "  InvalidMatchClass = 0,\n";
2180   OS << "  OptionalMatchClass = 1,\n";
2181   for (const auto &CI : Infos) {
2182     OS << "  " << CI.Name << ", // ";
2183     if (CI.Kind == ClassInfo::Token) {
2184       OS << "'" << CI.ValueName << "'\n";
2185     } else if (CI.isRegisterClass()) {
2186       if (!CI.ValueName.empty())
2187         OS << "register class '" << CI.ValueName << "'\n";
2188       else
2189         OS << "derived register class\n";
2190     } else {
2191       OS << "user defined class '" << CI.ValueName << "'\n";
2192     }
2193   }
2194   OS << "  NumMatchClassKinds\n";
2195   OS << "};\n\n";
2196 
2197   OS << "}\n\n";
2198 }
2199 
2200 /// emitValidateOperandClass - Emit the function to validate an operand class.
2201 static void emitValidateOperandClass(AsmMatcherInfo &Info,
2202                                      raw_ostream &OS) {
2203   OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
2204      << "MatchClassKind Kind) {\n";
2205   OS << "  " << Info.Target.getName() << "Operand &Operand = ("
2206      << Info.Target.getName() << "Operand&)GOp;\n";
2207 
2208   // The InvalidMatchClass is not to match any operand.
2209   OS << "  if (Kind == InvalidMatchClass)\n";
2210   OS << "    return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2211 
2212   // Check for Token operands first.
2213   // FIXME: Use a more specific diagnostic type.
2214   OS << "  if (Operand.isToken())\n";
2215   OS << "    return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2216      << "             MCTargetAsmParser::Match_Success :\n"
2217      << "             MCTargetAsmParser::Match_InvalidOperand;\n\n";
2218 
2219   // Check the user classes. We don't care what order since we're only
2220   // actually matching against one of them.
2221   OS << "  switch (Kind) {\n"
2222         "  default: break;\n";
2223   for (const auto &CI : Info.Classes) {
2224     if (!CI.isUserClass())
2225       continue;
2226 
2227     OS << "  // '" << CI.ClassName << "' class\n";
2228     OS << "  case " << CI.Name << ":\n";
2229     OS << "    if (Operand." << CI.PredicateMethod << "())\n";
2230     OS << "      return MCTargetAsmParser::Match_Success;\n";
2231     if (!CI.DiagnosticType.empty())
2232       OS << "    return " << Info.Target.getName() << "AsmParser::Match_"
2233          << CI.DiagnosticType << ";\n";
2234     else
2235       OS << "    break;\n";
2236   }
2237   OS << "  } // end switch (Kind)\n\n";
2238 
2239   // Check for register operands, including sub-classes.
2240   OS << "  if (Operand.isReg()) {\n";
2241   OS << "    MatchClassKind OpKind;\n";
2242   OS << "    switch (Operand.getReg()) {\n";
2243   OS << "    default: OpKind = InvalidMatchClass; break;\n";
2244   for (const auto &RC : Info.RegisterClasses)
2245     OS << "    case " << Info.Target.getName() << "::"
2246        << RC.first->getName() << ": OpKind = " << RC.second->Name
2247        << "; break;\n";
2248   OS << "    }\n";
2249   OS << "    return isSubclass(OpKind, Kind) ? "
2250      << "MCTargetAsmParser::Match_Success :\n                             "
2251      << "         MCTargetAsmParser::Match_InvalidOperand;\n  }\n\n";
2252 
2253   // Generic fallthrough match failure case for operands that don't have
2254   // specialized diagnostic types.
2255   OS << "  return MCTargetAsmParser::Match_InvalidOperand;\n";
2256   OS << "}\n\n";
2257 }
2258 
2259 /// emitIsSubclass - Emit the subclass predicate function.
2260 static void emitIsSubclass(CodeGenTarget &Target,
2261                            std::forward_list<ClassInfo> &Infos,
2262                            raw_ostream &OS) {
2263   OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2264   OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2265   OS << "  if (A == B)\n";
2266   OS << "    return true;\n\n";
2267 
2268   bool EmittedSwitch = false;
2269   for (const auto &A : Infos) {
2270     std::vector<StringRef> SuperClasses;
2271     if (A.IsOptional)
2272       SuperClasses.push_back("OptionalMatchClass");
2273     for (const auto &B : Infos) {
2274       if (&A != &B && A.isSubsetOf(B))
2275         SuperClasses.push_back(B.Name);
2276     }
2277 
2278     if (SuperClasses.empty())
2279       continue;
2280 
2281     // If this is the first SuperClass, emit the switch header.
2282     if (!EmittedSwitch) {
2283       OS << "  switch (A) {\n";
2284       OS << "  default:\n";
2285       OS << "    return false;\n";
2286       EmittedSwitch = true;
2287     }
2288 
2289     OS << "\n  case " << A.Name << ":\n";
2290 
2291     if (SuperClasses.size() == 1) {
2292       OS << "    return B == " << SuperClasses.back() << ";\n";
2293       continue;
2294     }
2295 
2296     if (!SuperClasses.empty()) {
2297       OS << "    switch (B) {\n";
2298       OS << "    default: return false;\n";
2299       for (StringRef SC : SuperClasses)
2300         OS << "    case " << SC << ": return true;\n";
2301       OS << "    }\n";
2302     } else {
2303       // No case statement to emit
2304       OS << "    return false;\n";
2305     }
2306   }
2307 
2308   // If there were case statements emitted into the string stream write the
2309   // default.
2310   if (EmittedSwitch)
2311     OS << "  }\n";
2312   else
2313     OS << "  return false;\n";
2314 
2315   OS << "}\n\n";
2316 }
2317 
2318 /// emitMatchTokenString - Emit the function to match a token string to the
2319 /// appropriate match class value.
2320 static void emitMatchTokenString(CodeGenTarget &Target,
2321                                  std::forward_list<ClassInfo> &Infos,
2322                                  raw_ostream &OS) {
2323   // Construct the match list.
2324   std::vector<StringMatcher::StringPair> Matches;
2325   for (const auto &CI : Infos) {
2326     if (CI.Kind == ClassInfo::Token)
2327       Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";");
2328   }
2329 
2330   OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2331 
2332   StringMatcher("Name", Matches, OS).Emit();
2333 
2334   OS << "  return InvalidMatchClass;\n";
2335   OS << "}\n\n";
2336 }
2337 
2338 /// emitMatchRegisterName - Emit the function to match a string to the target
2339 /// specific register enum.
2340 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2341                                   raw_ostream &OS) {
2342   // Construct the match list.
2343   std::vector<StringMatcher::StringPair> Matches;
2344   const auto &Regs = Target.getRegBank().getRegisters();
2345   for (const CodeGenRegister &Reg : Regs) {
2346     if (Reg.TheDef->getValueAsString("AsmName").empty())
2347       continue;
2348 
2349     Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"),
2350                          "return " + utostr(Reg.EnumValue) + ";");
2351   }
2352 
2353   OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
2354 
2355   StringMatcher("Name", Matches, OS).Emit();
2356 
2357   OS << "  return 0;\n";
2358   OS << "}\n\n";
2359 }
2360 
2361 /// Emit the function to match a string to the target
2362 /// specific register enum.
2363 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser,
2364                                      raw_ostream &OS) {
2365   // Construct the match list.
2366   std::vector<StringMatcher::StringPair> Matches;
2367   const auto &Regs = Target.getRegBank().getRegisters();
2368   for (const CodeGenRegister &Reg : Regs) {
2369 
2370     auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames");
2371 
2372     for (auto AltName : AltNames) {
2373       AltName = StringRef(AltName).trim();
2374 
2375       // don't handle empty alternative names
2376       if (AltName.empty())
2377         continue;
2378 
2379       Matches.emplace_back(AltName,
2380                            "return " + utostr(Reg.EnumValue) + ";");
2381     }
2382   }
2383 
2384   OS << "static unsigned MatchRegisterAltName(StringRef Name) {\n";
2385 
2386   StringMatcher("Name", Matches, OS).Emit();
2387 
2388   OS << "  return 0;\n";
2389   OS << "}\n\n";
2390 }
2391 
2392 static const char *getMinimalTypeForRange(uint64_t Range) {
2393   assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large");
2394   if (Range > 0xFFFFFFFFULL)
2395     return "uint64_t";
2396   if (Range > 0xFFFF)
2397     return "uint32_t";
2398   if (Range > 0xFF)
2399     return "uint16_t";
2400   return "uint8_t";
2401 }
2402 
2403 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) {
2404   uint64_t MaxIndex = Info.SubtargetFeatures.size();
2405   if (MaxIndex > 0)
2406     MaxIndex--;
2407   return getMinimalTypeForRange(1ULL << MaxIndex);
2408 }
2409 
2410 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
2411 /// definitions.
2412 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
2413                                                 raw_ostream &OS) {
2414   OS << "// Flags for subtarget features that participate in "
2415      << "instruction matching.\n";
2416   OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info)
2417      << " {\n";
2418   for (const auto &SF : Info.SubtargetFeatures) {
2419     const SubtargetFeatureInfo &SFI = SF.second;
2420     OS << "  " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n";
2421   }
2422   OS << "  Feature_None = 0\n";
2423   OS << "};\n\n";
2424 }
2425 
2426 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
2427 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2428   // Get the set of diagnostic types from all of the operand classes.
2429   std::set<StringRef> Types;
2430   for (const auto &OpClassEntry : Info.AsmOperandClasses) {
2431     if (!OpClassEntry.second->DiagnosticType.empty())
2432       Types.insert(OpClassEntry.second->DiagnosticType);
2433   }
2434 
2435   if (Types.empty()) return;
2436 
2437   // Now emit the enum entries.
2438   for (StringRef Type : Types)
2439     OS << "  Match_" << Type << ",\n";
2440   OS << "  END_OPERAND_DIAGNOSTIC_TYPES\n";
2441 }
2442 
2443 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2444 /// user-level name for a subtarget feature.
2445 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2446   OS << "// User-level names for subtarget features that participate in\n"
2447      << "// instruction matching.\n"
2448      << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2449   if (!Info.SubtargetFeatures.empty()) {
2450     OS << "  switch(Val) {\n";
2451     for (const auto &SF : Info.SubtargetFeatures) {
2452       const SubtargetFeatureInfo &SFI = SF.second;
2453       // FIXME: Totally just a placeholder name to get the algorithm working.
2454       OS << "  case " << SFI.getEnumName() << ": return \""
2455          << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2456     }
2457     OS << "  default: return \"(unknown)\";\n";
2458     OS << "  }\n";
2459   } else {
2460     // Nothing to emit, so skip the switch
2461     OS << "  return \"(unknown)\";\n";
2462   }
2463   OS << "}\n\n";
2464 }
2465 
2466 /// emitComputeAvailableFeatures - Emit the function to compute the list of
2467 /// available features given a subtarget.
2468 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info,
2469                                          raw_ostream &OS) {
2470   std::string ClassName =
2471     Info.AsmParser->getValueAsString("AsmParserClassName");
2472 
2473   OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n"
2474      << "ComputeAvailableFeatures(const FeatureBitset& FB) const {\n";
2475   OS << "  uint64_t Features = 0;\n";
2476   for (const auto &SF : Info.SubtargetFeatures) {
2477     const SubtargetFeatureInfo &SFI = SF.second;
2478 
2479     OS << "  if (";
2480     std::string CondStorage =
2481       SFI.TheDef->getValueAsString("AssemblerCondString");
2482     StringRef Conds = CondStorage;
2483     std::pair<StringRef,StringRef> Comma = Conds.split(',');
2484     bool First = true;
2485     do {
2486       if (!First)
2487         OS << " && ";
2488 
2489       bool Neg = false;
2490       StringRef Cond = Comma.first;
2491       if (Cond[0] == '!') {
2492         Neg = true;
2493         Cond = Cond.substr(1);
2494       }
2495 
2496       OS << "(";
2497       if (Neg)
2498         OS << "!";
2499       OS << "FB[" << Info.Target.getName() << "::" << Cond << "])";
2500 
2501       if (Comma.second.empty())
2502         break;
2503 
2504       First = false;
2505       Comma = Comma.second.split(',');
2506     } while (true);
2507 
2508     OS << ")\n";
2509     OS << "    Features |= " << SFI.getEnumName() << ";\n";
2510   }
2511   OS << "  return Features;\n";
2512   OS << "}\n\n";
2513 }
2514 
2515 static std::string GetAliasRequiredFeatures(Record *R,
2516                                             const AsmMatcherInfo &Info) {
2517   std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2518   std::string Result;
2519   unsigned NumFeatures = 0;
2520   for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2521     const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2522 
2523     if (!F)
2524       PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2525                     "' is not marked as an AssemblerPredicate!");
2526 
2527     if (NumFeatures)
2528       Result += '|';
2529 
2530     Result += F->getEnumName();
2531     ++NumFeatures;
2532   }
2533 
2534   if (NumFeatures > 1)
2535     Result = '(' + Result + ')';
2536   return Result;
2537 }
2538 
2539 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info,
2540                                      std::vector<Record*> &Aliases,
2541                                      unsigned Indent = 0,
2542                                   StringRef AsmParserVariantName = StringRef()){
2543   // Keep track of all the aliases from a mnemonic.  Use an std::map so that the
2544   // iteration order of the map is stable.
2545   std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2546 
2547   for (Record *R : Aliases) {
2548     // FIXME: Allow AssemblerVariantName to be a comma separated list.
2549     std::string AsmVariantName = R->getValueAsString("AsmVariantName");
2550     if (AsmVariantName != AsmParserVariantName)
2551       continue;
2552     AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2553   }
2554   if (AliasesFromMnemonic.empty())
2555     return;
2556 
2557   // Process each alias a "from" mnemonic at a time, building the code executed
2558   // by the string remapper.
2559   std::vector<StringMatcher::StringPair> Cases;
2560   for (const auto &AliasEntry : AliasesFromMnemonic) {
2561     const std::vector<Record*> &ToVec = AliasEntry.second;
2562 
2563     // Loop through each alias and emit code that handles each case.  If there
2564     // are two instructions without predicates, emit an error.  If there is one,
2565     // emit it last.
2566     std::string MatchCode;
2567     int AliasWithNoPredicate = -1;
2568 
2569     for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2570       Record *R = ToVec[i];
2571       std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2572 
2573       // If this unconditionally matches, remember it for later and diagnose
2574       // duplicates.
2575       if (FeatureMask.empty()) {
2576         if (AliasWithNoPredicate != -1) {
2577           // We can't have two aliases from the same mnemonic with no predicate.
2578           PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2579                      "two MnemonicAliases with the same 'from' mnemonic!");
2580           PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2581         }
2582 
2583         AliasWithNoPredicate = i;
2584         continue;
2585       }
2586       if (R->getValueAsString("ToMnemonic") == AliasEntry.first)
2587         PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2588 
2589       if (!MatchCode.empty())
2590         MatchCode += "else ";
2591       MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
2592       MatchCode += "  Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
2593     }
2594 
2595     if (AliasWithNoPredicate != -1) {
2596       Record *R = ToVec[AliasWithNoPredicate];
2597       if (!MatchCode.empty())
2598         MatchCode += "else\n  ";
2599       MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2600     }
2601 
2602     MatchCode += "return;";
2603 
2604     Cases.push_back(std::make_pair(AliasEntry.first, MatchCode));
2605   }
2606   StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2607 }
2608 
2609 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2610 /// emit a function for them and return true, otherwise return false.
2611 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2612                                 CodeGenTarget &Target) {
2613   // Ignore aliases when match-prefix is set.
2614   if (!MatchPrefix.empty())
2615     return false;
2616 
2617   std::vector<Record*> Aliases =
2618     Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2619   if (Aliases.empty()) return false;
2620 
2621   OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2622     "uint64_t Features, unsigned VariantID) {\n";
2623   OS << "  switch (VariantID) {\n";
2624   unsigned VariantCount = Target.getAsmParserVariantCount();
2625   for (unsigned VC = 0; VC != VariantCount; ++VC) {
2626     Record *AsmVariant = Target.getAsmParserVariant(VC);
2627     int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2628     std::string AsmParserVariantName = AsmVariant->getValueAsString("Name");
2629     OS << "    case " << AsmParserVariantNo << ":\n";
2630     emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2631                              AsmParserVariantName);
2632     OS << "    break;\n";
2633   }
2634   OS << "  }\n";
2635 
2636   // Emit aliases that apply to all variants.
2637   emitMnemonicAliasVariant(OS, Info, Aliases);
2638 
2639   OS << "}\n\n";
2640 
2641   return true;
2642 }
2643 
2644 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2645                               const AsmMatcherInfo &Info, StringRef ClassName,
2646                               StringToOffsetTable &StringTable,
2647                               unsigned MaxMnemonicIndex, bool HasMnemonicFirst) {
2648   unsigned MaxMask = 0;
2649   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2650     MaxMask |= OMI.OperandMask;
2651   }
2652 
2653   // Emit the static custom operand parsing table;
2654   OS << "namespace {\n";
2655   OS << "  struct OperandMatchEntry {\n";
2656   OS << "    " << getMinimalRequiredFeaturesType(Info)
2657                << " RequiredFeatures;\n";
2658   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex)
2659                << " Mnemonic;\n";
2660   OS << "    " << getMinimalTypeForRange(std::distance(
2661                       Info.Classes.begin(), Info.Classes.end())) << " Class;\n";
2662   OS << "    " << getMinimalTypeForRange(MaxMask)
2663                << " OperandMask;\n\n";
2664   OS << "    StringRef getMnemonic() const {\n";
2665   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2666   OS << "                       MnemonicTable[Mnemonic]);\n";
2667   OS << "    }\n";
2668   OS << "  };\n\n";
2669 
2670   OS << "  // Predicate for searching for an opcode.\n";
2671   OS << "  struct LessOpcodeOperand {\n";
2672   OS << "    bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2673   OS << "      return LHS.getMnemonic()  < RHS;\n";
2674   OS << "    }\n";
2675   OS << "    bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2676   OS << "      return LHS < RHS.getMnemonic();\n";
2677   OS << "    }\n";
2678   OS << "    bool operator()(const OperandMatchEntry &LHS,";
2679   OS << " const OperandMatchEntry &RHS) {\n";
2680   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2681   OS << "    }\n";
2682   OS << "  };\n";
2683 
2684   OS << "} // end anonymous namespace.\n\n";
2685 
2686   OS << "static const OperandMatchEntry OperandMatchTable["
2687      << Info.OperandMatchInfo.size() << "] = {\n";
2688 
2689   OS << "  /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2690   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2691     const MatchableInfo &II = *OMI.MI;
2692 
2693     OS << "  { ";
2694 
2695     // Write the required features mask.
2696     if (!II.RequiredFeatures.empty()) {
2697       for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2698         if (i) OS << "|";
2699         OS << II.RequiredFeatures[i]->getEnumName();
2700       }
2701     } else
2702       OS << "0";
2703 
2704     // Store a pascal-style length byte in the mnemonic.
2705     std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2706     OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2707        << " /* " << II.Mnemonic << " */, ";
2708 
2709     OS << OMI.CI->Name;
2710 
2711     OS << ", " << OMI.OperandMask;
2712     OS << " /* ";
2713     bool printComma = false;
2714     for (int i = 0, e = 31; i !=e; ++i)
2715       if (OMI.OperandMask & (1 << i)) {
2716         if (printComma)
2717           OS << ", ";
2718         OS << i;
2719         printComma = true;
2720       }
2721     OS << " */";
2722 
2723     OS << " },\n";
2724   }
2725   OS << "};\n\n";
2726 
2727   // Emit the operand class switch to call the correct custom parser for
2728   // the found operand class.
2729   OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2730      << Target.getName() << ClassName << "::\n"
2731      << "tryCustomParseOperand(OperandVector"
2732      << " &Operands,\n                      unsigned MCK) {\n\n"
2733      << "  switch(MCK) {\n";
2734 
2735   for (const auto &CI : Info.Classes) {
2736     if (CI.ParserMethod.empty())
2737       continue;
2738     OS << "  case " << CI.Name << ":\n"
2739        << "    return " << CI.ParserMethod << "(Operands);\n";
2740   }
2741 
2742   OS << "  default:\n";
2743   OS << "    return MatchOperand_NoMatch;\n";
2744   OS << "  }\n";
2745   OS << "  return MatchOperand_NoMatch;\n";
2746   OS << "}\n\n";
2747 
2748   // Emit the static custom operand parser. This code is very similar with
2749   // the other matcher. Also use MatchResultTy here just in case we go for
2750   // a better error handling.
2751   OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2752      << Target.getName() << ClassName << "::\n"
2753      << "MatchOperandParserImpl(OperandVector"
2754      << " &Operands,\n                       StringRef Mnemonic) {\n";
2755 
2756   // Emit code to get the available features.
2757   OS << "  // Get the current feature set.\n";
2758   OS << "  uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2759 
2760   OS << "  // Get the next operand index.\n";
2761   OS << "  unsigned NextOpNum = Operands.size()"
2762      << (HasMnemonicFirst ? " - 1" : "") << ";\n";
2763 
2764   // Emit code to search the table.
2765   OS << "  // Search the table.\n";
2766   if (HasMnemonicFirst) {
2767     OS << "  auto MnemonicRange =\n";
2768     OS << "    std::equal_range(std::begin(OperandMatchTable), "
2769           "std::end(OperandMatchTable),\n";
2770     OS << "                     Mnemonic, LessOpcodeOperand());\n\n";
2771   } else {
2772     OS << "  auto MnemonicRange = std::make_pair(std::begin(OperandMatchTable),"
2773           " std::end(OperandMatchTable));\n";
2774     OS << "  if (!Mnemonic.empty())\n";
2775     OS << "    MnemonicRange =\n";
2776     OS << "      std::equal_range(std::begin(OperandMatchTable), "
2777           "std::end(OperandMatchTable),\n";
2778     OS << "                       Mnemonic, LessOpcodeOperand());\n\n";
2779   }
2780 
2781   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
2782   OS << "    return MatchOperand_NoMatch;\n\n";
2783 
2784   OS << "  for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2785      << "       *ie = MnemonicRange.second; it != ie; ++it) {\n";
2786 
2787   OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
2788   OS << "    assert(Mnemonic == it->getMnemonic());\n\n";
2789 
2790   // Emit check that the required features are available.
2791   OS << "    // check if the available features match\n";
2792   OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
2793      << "!= it->RequiredFeatures) {\n";
2794   OS << "      continue;\n";
2795   OS << "    }\n\n";
2796 
2797   // Emit check to ensure the operand number matches.
2798   OS << "    // check if the operand in question has a custom parser.\n";
2799   OS << "    if (!(it->OperandMask & (1 << NextOpNum)))\n";
2800   OS << "      continue;\n\n";
2801 
2802   // Emit call to the custom parser method
2803   OS << "    // call custom parse method to handle the operand\n";
2804   OS << "    OperandMatchResultTy Result = ";
2805   OS << "tryCustomParseOperand(Operands, it->Class);\n";
2806   OS << "    if (Result != MatchOperand_NoMatch)\n";
2807   OS << "      return Result;\n";
2808   OS << "  }\n\n";
2809 
2810   OS << "  // Okay, we had no match.\n";
2811   OS << "  return MatchOperand_NoMatch;\n";
2812   OS << "}\n\n";
2813 }
2814 
2815 void AsmMatcherEmitter::run(raw_ostream &OS) {
2816   CodeGenTarget Target(Records);
2817   Record *AsmParser = Target.getAsmParser();
2818   std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2819 
2820   // Compute the information on the instructions to match.
2821   AsmMatcherInfo Info(AsmParser, Target, Records);
2822   Info.buildInfo();
2823 
2824   // Sort the instruction table using the partial order on classes. We use
2825   // stable_sort to ensure that ambiguous instructions are still
2826   // deterministically ordered.
2827   std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
2828                    [](const std::unique_ptr<MatchableInfo> &a,
2829                       const std::unique_ptr<MatchableInfo> &b){
2830                      return *a < *b;});
2831 
2832 #ifndef NDEBUG
2833   // Verify that the table is now sorted
2834   for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2835        ++I) {
2836     for (auto J = I; J != E; ++J) {
2837       assert(!(**J < **I));
2838     }
2839   }
2840 #endif // NDEBUG
2841 
2842   DEBUG_WITH_TYPE("instruction_info", {
2843       for (const auto &MI : Info.Matchables)
2844         MI->dump();
2845     });
2846 
2847   // Check for ambiguous matchables.
2848   DEBUG_WITH_TYPE("ambiguous_instrs", {
2849     unsigned NumAmbiguous = 0;
2850     for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2851          ++I) {
2852       for (auto J = std::next(I); J != E; ++J) {
2853         const MatchableInfo &A = **I;
2854         const MatchableInfo &B = **J;
2855 
2856         if (A.couldMatchAmbiguouslyWith(B)) {
2857           errs() << "warning: ambiguous matchables:\n";
2858           A.dump();
2859           errs() << "\nis incomparable with:\n";
2860           B.dump();
2861           errs() << "\n\n";
2862           ++NumAmbiguous;
2863         }
2864       }
2865     }
2866     if (NumAmbiguous)
2867       errs() << "warning: " << NumAmbiguous
2868              << " ambiguous matchables!\n";
2869   });
2870 
2871   // Compute the information on the custom operand parsing.
2872   Info.buildOperandMatchInfo();
2873 
2874   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
2875   bool HasOptionalOperands = Info.hasOptionalOperands();
2876 
2877   // Write the output.
2878 
2879   // Information for the class declaration.
2880   OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2881   OS << "#undef GET_ASSEMBLER_HEADER\n";
2882   OS << "  // This should be included into the middle of the declaration of\n";
2883   OS << "  // your subclasses implementation of MCTargetAsmParser.\n";
2884   OS << "  uint64_t ComputeAvailableFeatures(const FeatureBitset& FB) const;\n";
2885   if (HasOptionalOperands) {
2886     OS << "  void convertToMCInst(unsigned Kind, MCInst &Inst, "
2887        << "unsigned Opcode,\n"
2888        << "                       const OperandVector &Operands,\n"
2889        << "                       const SmallBitVector &OptionalOperandsMask);\n";
2890   } else {
2891     OS << "  void convertToMCInst(unsigned Kind, MCInst &Inst, "
2892        << "unsigned Opcode,\n"
2893        << "                       const OperandVector &Operands);\n";
2894   }
2895   OS << "  void convertToMapAndConstraints(unsigned Kind,\n                ";
2896   OS << "           const OperandVector &Operands) override;\n";
2897   if (HasMnemonicFirst)
2898     OS << "  bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID);\n";
2899   OS << "  unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
2900      << "                                MCInst &Inst,\n"
2901      << "                                uint64_t &ErrorInfo,"
2902      << " bool matchingInlineAsm,\n"
2903      << "                                unsigned VariantID = 0);\n";
2904 
2905   if (!Info.OperandMatchInfo.empty()) {
2906     OS << "\n  enum OperandMatchResultTy {\n";
2907     OS << "    MatchOperand_Success,    // operand matched successfully\n";
2908     OS << "    MatchOperand_NoMatch,    // operand did not match\n";
2909     OS << "    MatchOperand_ParseFail   // operand matched but had errors\n";
2910     OS << "  };\n";
2911     OS << "  OperandMatchResultTy MatchOperandParserImpl(\n";
2912     OS << "    OperandVector &Operands,\n";
2913     OS << "    StringRef Mnemonic);\n";
2914 
2915     OS << "  OperandMatchResultTy tryCustomParseOperand(\n";
2916     OS << "    OperandVector &Operands,\n";
2917     OS << "    unsigned MCK);\n\n";
2918   }
2919 
2920   OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2921 
2922   // Emit the operand match diagnostic enum names.
2923   OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
2924   OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2925   emitOperandDiagnosticTypes(Info, OS);
2926   OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2927 
2928   OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2929   OS << "#undef GET_REGISTER_MATCHER\n\n";
2930 
2931   // Emit the subtarget feature enumeration.
2932   emitSubtargetFeatureFlagEnumeration(Info, OS);
2933 
2934   // Emit the function to match a register name to number.
2935   // This should be omitted for Mips target
2936   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
2937     emitMatchRegisterName(Target, AsmParser, OS);
2938 
2939   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName"))
2940     emitMatchRegisterAltName(Target, AsmParser, OS);
2941 
2942   OS << "#endif // GET_REGISTER_MATCHER\n\n";
2943 
2944   OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
2945   OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
2946 
2947   // Generate the helper function to get the names for subtarget features.
2948   emitGetSubtargetFeatureName(Info, OS);
2949 
2950   OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
2951 
2952   OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2953   OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2954 
2955   // Generate the function that remaps for mnemonic aliases.
2956   bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
2957 
2958   // Generate the convertToMCInst function to convert operands into an MCInst.
2959   // Also, generate the convertToMapAndConstraints function for MS-style inline
2960   // assembly.  The latter doesn't actually generate a MCInst.
2961   emitConvertFuncs(Target, ClassName, Info.Matchables, HasMnemonicFirst,
2962                    HasOptionalOperands, OS);
2963 
2964   // Emit the enumeration for classes which participate in matching.
2965   emitMatchClassEnumeration(Target, Info.Classes, OS);
2966 
2967   // Emit the routine to match token strings to their match class.
2968   emitMatchTokenString(Target, Info.Classes, OS);
2969 
2970   // Emit the subclass predicate routine.
2971   emitIsSubclass(Target, Info.Classes, OS);
2972 
2973   // Emit the routine to validate an operand against a match class.
2974   emitValidateOperandClass(Info, OS);
2975 
2976   // Emit the available features compute function.
2977   emitComputeAvailableFeatures(Info, OS);
2978 
2979   StringToOffsetTable StringTable;
2980 
2981   size_t MaxNumOperands = 0;
2982   unsigned MaxMnemonicIndex = 0;
2983   bool HasDeprecation = false;
2984   for (const auto &MI : Info.Matchables) {
2985     MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
2986     HasDeprecation |= MI->HasDeprecation;
2987 
2988     // Store a pascal-style length byte in the mnemonic.
2989     std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
2990     MaxMnemonicIndex = std::max(MaxMnemonicIndex,
2991                         StringTable.GetOrAddStringOffset(LenMnemonic, false));
2992   }
2993 
2994   OS << "static const char *const MnemonicTable =\n";
2995   StringTable.EmitString(OS);
2996   OS << ";\n\n";
2997 
2998   // Emit the static match table; unused classes get initalized to 0 which is
2999   // guaranteed to be InvalidMatchClass.
3000   //
3001   // FIXME: We can reduce the size of this table very easily. First, we change
3002   // it so that store the kinds in separate bit-fields for each index, which
3003   // only needs to be the max width used for classes at that index (we also need
3004   // to reject based on this during classification). If we then make sure to
3005   // order the match kinds appropriately (putting mnemonics last), then we
3006   // should only end up using a few bits for each class, especially the ones
3007   // following the mnemonic.
3008   OS << "namespace {\n";
3009   OS << "  struct MatchEntry {\n";
3010   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex)
3011                << " Mnemonic;\n";
3012   OS << "    uint16_t Opcode;\n";
3013   OS << "    " << getMinimalTypeForRange(Info.Matchables.size())
3014                << " ConvertFn;\n";
3015   OS << "    " << getMinimalRequiredFeaturesType(Info)
3016                << " RequiredFeatures;\n";
3017   OS << "    " << getMinimalTypeForRange(
3018                       std::distance(Info.Classes.begin(), Info.Classes.end()))
3019      << " Classes[" << MaxNumOperands << "];\n";
3020   OS << "    StringRef getMnemonic() const {\n";
3021   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
3022   OS << "                       MnemonicTable[Mnemonic]);\n";
3023   OS << "    }\n";
3024   OS << "  };\n\n";
3025 
3026   OS << "  // Predicate for searching for an opcode.\n";
3027   OS << "  struct LessOpcode {\n";
3028   OS << "    bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
3029   OS << "      return LHS.getMnemonic() < RHS;\n";
3030   OS << "    }\n";
3031   OS << "    bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
3032   OS << "      return LHS < RHS.getMnemonic();\n";
3033   OS << "    }\n";
3034   OS << "    bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
3035   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
3036   OS << "    }\n";
3037   OS << "  };\n";
3038 
3039   OS << "} // end anonymous namespace.\n\n";
3040 
3041   unsigned VariantCount = Target.getAsmParserVariantCount();
3042   for (unsigned VC = 0; VC != VariantCount; ++VC) {
3043     Record *AsmVariant = Target.getAsmParserVariant(VC);
3044     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3045 
3046     OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
3047 
3048     for (const auto &MI : Info.Matchables) {
3049       if (MI->AsmVariantID != AsmVariantNo)
3050         continue;
3051 
3052       // Store a pascal-style length byte in the mnemonic.
3053       std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
3054       OS << "  { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
3055          << " /* " << MI->Mnemonic << " */, "
3056          << Target.getName() << "::"
3057          << MI->getResultInst()->TheDef->getName() << ", "
3058          << MI->ConversionFnKind << ", ";
3059 
3060       // Write the required features mask.
3061       if (!MI->RequiredFeatures.empty()) {
3062         for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i) {
3063           if (i) OS << "|";
3064           OS << MI->RequiredFeatures[i]->getEnumName();
3065         }
3066       } else
3067         OS << "0";
3068 
3069       OS << ", { ";
3070       for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
3071         const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
3072 
3073         if (i) OS << ", ";
3074         OS << Op.Class->Name;
3075       }
3076       OS << " }, },\n";
3077     }
3078 
3079     OS << "};\n\n";
3080   }
3081 
3082   // A method to determine if a mnemonic is in the list.
3083   if (HasMnemonicFirst) {
3084     OS << "bool " << Target.getName() << ClassName << "::\n"
3085        << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n";
3086     OS << "  // Find the appropriate table for this asm variant.\n";
3087     OS << "  const MatchEntry *Start, *End;\n";
3088     OS << "  switch (VariantID) {\n";
3089     OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3090     for (unsigned VC = 0; VC != VariantCount; ++VC) {
3091       Record *AsmVariant = Target.getAsmParserVariant(VC);
3092       int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3093       OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3094          << "); End = std::end(MatchTable" << VC << "); break;\n";
3095     }
3096     OS << "  }\n";
3097     OS << "  // Search the table.\n";
3098     OS << "  auto MnemonicRange = ";
3099     OS << "std::equal_range(Start, End, Mnemonic, LessOpcode());\n";
3100     OS << "  return MnemonicRange.first != MnemonicRange.second;\n";
3101     OS << "}\n\n";
3102   }
3103 
3104   // Finally, build the match function.
3105   OS << "unsigned " << Target.getName() << ClassName << "::\n"
3106      << "MatchInstructionImpl(const OperandVector &Operands,\n";
3107   OS << "                     MCInst &Inst, uint64_t &ErrorInfo,\n"
3108      << "                     bool matchingInlineAsm, unsigned VariantID) {\n";
3109 
3110   OS << "  // Eliminate obvious mismatches.\n";
3111   OS << "  if (Operands.size() > "
3112      << (MaxNumOperands + HasMnemonicFirst) << ") {\n";
3113   OS << "    ErrorInfo = "
3114      << (MaxNumOperands + HasMnemonicFirst) << ";\n";
3115   OS << "    return Match_InvalidOperand;\n";
3116   OS << "  }\n\n";
3117 
3118   // Emit code to get the available features.
3119   OS << "  // Get the current feature set.\n";
3120   OS << "  uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
3121 
3122   OS << "  // Get the instruction mnemonic, which is the first token.\n";
3123   if (HasMnemonicFirst) {
3124     OS << "  StringRef Mnemonic = ((" << Target.getName()
3125        << "Operand&)*Operands[0]).getToken();\n\n";
3126   } else {
3127     OS << "  StringRef Mnemonic;\n";
3128     OS << "  if (Operands[0]->isToken())\n";
3129     OS << "    Mnemonic = ((" << Target.getName()
3130        << "Operand&)*Operands[0]).getToken();\n\n";
3131   }
3132 
3133   if (HasMnemonicAliases) {
3134     OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
3135     OS << "  applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
3136   }
3137 
3138   // Emit code to compute the class list for this operand vector.
3139   OS << "  // Some state to try to produce better error messages.\n";
3140   OS << "  bool HadMatchOtherThanFeatures = false;\n";
3141   OS << "  bool HadMatchOtherThanPredicate = false;\n";
3142   OS << "  unsigned RetCode = Match_InvalidOperand;\n";
3143   OS << "  uint64_t MissingFeatures = ~0ULL;\n";
3144   if (HasOptionalOperands) {
3145     OS << "  SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n";
3146   }
3147   OS << "  // Set ErrorInfo to the operand that mismatches if it is\n";
3148   OS << "  // wrong for all instances of the instruction.\n";
3149   OS << "  ErrorInfo = ~0ULL;\n";
3150 
3151   // Emit code to search the table.
3152   OS << "  // Find the appropriate table for this asm variant.\n";
3153   OS << "  const MatchEntry *Start, *End;\n";
3154   OS << "  switch (VariantID) {\n";
3155   OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3156   for (unsigned VC = 0; VC != VariantCount; ++VC) {
3157     Record *AsmVariant = Target.getAsmParserVariant(VC);
3158     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3159     OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3160        << "); End = std::end(MatchTable" << VC << "); break;\n";
3161   }
3162   OS << "  }\n";
3163 
3164   OS << "  // Search the table.\n";
3165   if (HasMnemonicFirst) {
3166     OS << "  auto MnemonicRange = "
3167           "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
3168   } else {
3169     OS << "  auto MnemonicRange = std::make_pair(Start, End);\n";
3170     OS << "  unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
3171     OS << "  if (!Mnemonic.empty())\n";
3172     OS << "    MnemonicRange = "
3173           "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
3174   }
3175 
3176   OS << "  // Return a more specific error code if no mnemonics match.\n";
3177   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
3178   OS << "    return Match_MnemonicFail;\n\n";
3179 
3180   OS << "  for (const MatchEntry *it = MnemonicRange.first, "
3181      << "*ie = MnemonicRange.second;\n";
3182   OS << "       it != ie; ++it) {\n";
3183 
3184   if (HasMnemonicFirst) {
3185     OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
3186     OS << "    assert(Mnemonic == it->getMnemonic());\n";
3187   }
3188 
3189   // Emit check that the subclasses match.
3190   OS << "    bool OperandsValid = true;\n";
3191   if (HasOptionalOperands) {
3192     OS << "    OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n";
3193   }
3194   OS << "    for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex")
3195      << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex")
3196      << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n";
3197   OS << "      auto Formal = "
3198      << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n";
3199   OS << "      if (ActualIdx >= Operands.size()) {\n";
3200   OS << "        OperandsValid = (Formal == " <<"InvalidMatchClass) || "
3201                                  "isSubclass(Formal, OptionalMatchClass);\n";
3202   OS << "        if (!OperandsValid) ErrorInfo = ActualIdx;\n";
3203   if (HasOptionalOperands) {
3204     OS << "        OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands
3205        << ");\n";
3206   }
3207   OS << "        break;\n";
3208   OS << "      }\n";
3209   OS << "      MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n";
3210   OS << "      unsigned Diag = validateOperandClass(Actual, Formal);\n";
3211   OS << "      if (Diag == Match_Success) {\n";
3212   OS << "        ++ActualIdx;\n";
3213   OS << "        continue;\n";
3214   OS << "      }\n";
3215   OS << "      // If the generic handler indicates an invalid operand\n";
3216   OS << "      // failure, check for a special case.\n";
3217   OS << "      if (Diag == Match_InvalidOperand) {\n";
3218   OS << "        Diag = validateTargetOperandClass(Actual, Formal);\n";
3219   OS << "        if (Diag == Match_Success) {\n";
3220   OS << "          ++ActualIdx;\n";
3221   OS << "          continue;\n";
3222   OS << "        }\n";
3223   OS << "      }\n";
3224   OS << "      // If current formal operand wasn't matched and it is optional\n"
3225      << "      // then try to match next formal operand\n";
3226   OS << "      if (Diag == Match_InvalidOperand "
3227      << "&& isSubclass(Formal, OptionalMatchClass)) {\n";
3228   if (HasOptionalOperands) {
3229     OS << "        OptionalOperandsMask.set(FormalIdx);\n";
3230   }
3231   OS << "        continue;\n";
3232   OS << "      }\n";
3233   OS << "      // If this operand is broken for all of the instances of this\n";
3234   OS << "      // mnemonic, keep track of it so we can report loc info.\n";
3235   OS << "      // If we already had a match that only failed due to a\n";
3236   OS << "      // target predicate, that diagnostic is preferred.\n";
3237   OS << "      if (!HadMatchOtherThanPredicate &&\n";
3238   OS << "          (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) {\n";
3239   OS << "        ErrorInfo = ActualIdx;\n";
3240   OS << "        // InvalidOperand is the default. Prefer specificity.\n";
3241   OS << "        if (Diag != Match_InvalidOperand)\n";
3242   OS << "          RetCode = Diag;\n";
3243   OS << "      }\n";
3244   OS << "      // Otherwise, just reject this instance of the mnemonic.\n";
3245   OS << "      OperandsValid = false;\n";
3246   OS << "      break;\n";
3247   OS << "    }\n\n";
3248 
3249   OS << "    if (!OperandsValid) continue;\n";
3250 
3251   // Emit check that the required features are available.
3252   OS << "    if ((AvailableFeatures & it->RequiredFeatures) "
3253      << "!= it->RequiredFeatures) {\n";
3254   OS << "      HadMatchOtherThanFeatures = true;\n";
3255   OS << "      uint64_t NewMissingFeatures = it->RequiredFeatures & "
3256         "~AvailableFeatures;\n";
3257   OS << "      if (countPopulation(NewMissingFeatures) <=\n"
3258         "          countPopulation(MissingFeatures))\n";
3259   OS << "        MissingFeatures = NewMissingFeatures;\n";
3260   OS << "      continue;\n";
3261   OS << "    }\n";
3262   OS << "\n";
3263   OS << "    Inst.clear();\n\n";
3264   OS << "    Inst.setOpcode(it->Opcode);\n";
3265   // Verify the instruction with the target-specific match predicate function.
3266   OS << "    // We have a potential match but have not rendered the operands.\n"
3267      << "    // Check the target predicate to handle any context sensitive\n"
3268         "    // constraints.\n"
3269      << "    // For example, Ties that are referenced multiple times must be\n"
3270         "    // checked here to ensure the input is the same for each match\n"
3271         "    // constraints. If we leave it any later the ties will have been\n"
3272         "    // canonicalized\n"
3273      << "    unsigned MatchResult;\n"
3274      << "    if ((MatchResult = checkEarlyTargetMatchPredicate(Inst, "
3275         "Operands)) != Match_Success) {\n"
3276      << "      Inst.clear();\n"
3277      << "      RetCode = MatchResult;\n"
3278      << "      HadMatchOtherThanPredicate = true;\n"
3279      << "      continue;\n"
3280      << "    }\n\n";
3281   OS << "    if (matchingInlineAsm) {\n";
3282   OS << "      convertToMapAndConstraints(it->ConvertFn, Operands);\n";
3283   OS << "      return Match_Success;\n";
3284   OS << "    }\n\n";
3285   OS << "    // We have selected a definite instruction, convert the parsed\n"
3286      << "    // operands into the appropriate MCInst.\n";
3287   if (HasOptionalOperands) {
3288     OS << "    convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n"
3289        << "                    OptionalOperandsMask);\n";
3290   } else {
3291     OS << "    convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
3292   }
3293   OS << "\n";
3294 
3295   // Verify the instruction with the target-specific match predicate function.
3296   OS << "    // We have a potential match. Check the target predicate to\n"
3297      << "    // handle any context sensitive constraints.\n"
3298      << "    if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
3299      << " Match_Success) {\n"
3300      << "      Inst.clear();\n"
3301      << "      RetCode = MatchResult;\n"
3302      << "      HadMatchOtherThanPredicate = true;\n"
3303      << "      continue;\n"
3304      << "    }\n\n";
3305 
3306   // Call the post-processing function, if used.
3307   std::string InsnCleanupFn =
3308     AsmParser->getValueAsString("AsmParserInstCleanup");
3309   if (!InsnCleanupFn.empty())
3310     OS << "    " << InsnCleanupFn << "(Inst);\n";
3311 
3312   if (HasDeprecation) {
3313     OS << "    std::string Info;\n";
3314     OS << "    if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n";
3315     OS << "      SMLoc Loc = ((" << Target.getName()
3316        << "Operand&)*Operands[0]).getStartLoc();\n";
3317     OS << "      getParser().Warning(Loc, Info, None);\n";
3318     OS << "    }\n";
3319   }
3320 
3321   OS << "    return Match_Success;\n";
3322   OS << "  }\n\n";
3323 
3324   OS << "  // Okay, we had no match.  Try to return a useful error code.\n";
3325   OS << "  if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
3326   OS << "    return RetCode;\n\n";
3327   OS << "  // Missing feature matches return which features were missing\n";
3328   OS << "  ErrorInfo = MissingFeatures;\n";
3329   OS << "  return Match_MissingFeature;\n";
3330   OS << "}\n\n";
3331 
3332   if (!Info.OperandMatchInfo.empty())
3333     emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
3334                              MaxMnemonicIndex, HasMnemonicFirst);
3335 
3336   OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
3337 }
3338 
3339 namespace llvm {
3340 
3341 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
3342   emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
3343   AsmMatcherEmitter(RK).run(OS);
3344 }
3345 
3346 } // end namespace llvm
3347