xref: /freebsd/contrib/llvm-project/llvm/utils/TableGen/AsmMatcherEmitter.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This tablegen backend emits a target specifier matcher for converting parsed
10 // assembly operands in the MCInst structures. It also emits a matcher for
11 // custom operand parsing.
12 //
13 // Converting assembly operands into MCInst structures
14 // ---------------------------------------------------
15 //
16 // The input to the target specific matcher is a list of literal tokens and
17 // operands. The target specific parser should generally eliminate any syntax
18 // which is not relevant for matching; for example, comma tokens should have
19 // already been consumed and eliminated by the parser. Most instructions will
20 // end up with a single literal token (the instruction name) and some number of
21 // operands.
22 //
23 // Some example inputs, for X86:
24 //   'addl' (immediate ...) (register ...)
25 //   'add' (immediate ...) (memory ...)
26 //   'call' '*' %epc
27 //
28 // The assembly matcher is responsible for converting this input into a precise
29 // machine instruction (i.e., an instruction with a well defined encoding). This
30 // mapping has several properties which complicate matching:
31 //
32 //  - It may be ambiguous; many architectures can legally encode particular
33 //    variants of an instruction in different ways (for example, using a smaller
34 //    encoding for small immediates). Such ambiguities should never be
35 //    arbitrarily resolved by the assembler, the assembler is always responsible
36 //    for choosing the "best" available instruction.
37 //
38 //  - It may depend on the subtarget or the assembler context. Instructions
39 //    which are invalid for the current mode, but otherwise unambiguous (e.g.,
40 //    an SSE instruction in a file being assembled for i486) should be accepted
41 //    and rejected by the assembler front end. However, if the proper encoding
42 //    for an instruction is dependent on the assembler context then the matcher
43 //    is responsible for selecting the correct machine instruction for the
44 //    current mode.
45 //
46 // The core matching algorithm attempts to exploit the regularity in most
47 // instruction sets to quickly determine the set of possibly matching
48 // instructions, and the simplify the generated code. Additionally, this helps
49 // to ensure that the ambiguities are intentionally resolved by the user.
50 //
51 // The matching is divided into two distinct phases:
52 //
53 //   1. Classification: Each operand is mapped to the unique set which (a)
54 //      contains it, and (b) is the largest such subset for which a single
55 //      instruction could match all members.
56 //
57 //      For register classes, we can generate these subgroups automatically. For
58 //      arbitrary operands, we expect the user to define the classes and their
59 //      relations to one another (for example, 8-bit signed immediates as a
60 //      subset of 32-bit immediates).
61 //
62 //      By partitioning the operands in this way, we guarantee that for any
63 //      tuple of classes, any single instruction must match either all or none
64 //      of the sets of operands which could classify to that tuple.
65 //
66 //      In addition, the subset relation amongst classes induces a partial order
67 //      on such tuples, which we use to resolve ambiguities.
68 //
69 //   2. The input can now be treated as a tuple of classes (static tokens are
70 //      simple singleton sets). Each such tuple should generally map to a single
71 //      instruction (we currently ignore cases where this isn't true, whee!!!),
72 //      which we can emit a simple matcher for.
73 //
74 // Custom Operand Parsing
75 // ----------------------
76 //
77 //  Some targets need a custom way to parse operands, some specific instructions
78 //  can contain arguments that can represent processor flags and other kinds of
79 //  identifiers that need to be mapped to specific values in the final encoded
80 //  instructions. The target specific custom operand parsing works in the
81 //  following way:
82 //
83 //   1. A operand match table is built, each entry contains a mnemonic, an
84 //      operand class, a mask for all operand positions for that same
85 //      class/mnemonic and target features to be checked while trying to match.
86 //
87 //   2. The operand matcher will try every possible entry with the same
88 //      mnemonic and will check if the target feature for this mnemonic also
89 //      matches. After that, if the operand to be matched has its index
90 //      present in the mask, a successful match occurs. Otherwise, fallback
91 //      to the regular operand parsing.
92 //
93 //   3. For a match success, each operand class that has a 'ParserMethod'
94 //      becomes part of a switch from where the custom method is called.
95 //
96 //===----------------------------------------------------------------------===//
97 
98 #include "Common/CodeGenInstAlias.h"
99 #include "Common/CodeGenInstruction.h"
100 #include "Common/CodeGenRegisters.h"
101 #include "Common/CodeGenTarget.h"
102 #include "Common/SubtargetFeatureInfo.h"
103 #include "Common/Types.h"
104 #include "llvm/ADT/CachedHashString.h"
105 #include "llvm/ADT/PointerUnion.h"
106 #include "llvm/ADT/STLExtras.h"
107 #include "llvm/ADT/SmallPtrSet.h"
108 #include "llvm/ADT/SmallVector.h"
109 #include "llvm/ADT/StringExtras.h"
110 #include "llvm/Support/CommandLine.h"
111 #include "llvm/Support/Debug.h"
112 #include "llvm/Support/ErrorHandling.h"
113 #include "llvm/TableGen/Error.h"
114 #include "llvm/TableGen/Record.h"
115 #include "llvm/TableGen/StringMatcher.h"
116 #include "llvm/TableGen/StringToOffsetTable.h"
117 #include "llvm/TableGen/TableGenBackend.h"
118 #include <cassert>
119 #include <cctype>
120 #include <forward_list>
121 #include <map>
122 #include <set>
123 
124 using namespace llvm;
125 
126 #define DEBUG_TYPE "asm-matcher-emitter"
127 
128 cl::OptionCategory AsmMatcherEmitterCat("Options for -gen-asm-matcher");
129 
130 static cl::opt<std::string>
131     MatchPrefix("match-prefix", cl::init(""),
132                 cl::desc("Only match instructions with the given prefix"),
133                 cl::cat(AsmMatcherEmitterCat));
134 
135 namespace {
136 class AsmMatcherInfo;
137 
138 // Register sets are used as keys in some second-order sets TableGen creates
139 // when generating its data structures. This means that the order of two
140 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and
141 // can even affect compiler output (at least seen in diagnostics produced when
142 // all matches fail). So we use a type that sorts them consistently.
143 typedef std::set<Record *, LessRecordByID> RegisterSet;
144 
145 class AsmMatcherEmitter {
146   RecordKeeper &Records;
147 
148 public:
AsmMatcherEmitter(RecordKeeper & R)149   AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
150 
151   void run(raw_ostream &o);
152 };
153 
154 /// ClassInfo - Helper class for storing the information about a particular
155 /// class of operands which can be matched.
156 struct ClassInfo {
157   enum ClassInfoKind {
158     /// Invalid kind, for use as a sentinel value.
159     Invalid = 0,
160 
161     /// The class for a particular token.
162     Token,
163 
164     /// The (first) register class, subsequent register classes are
165     /// RegisterClass0+1, and so on.
166     RegisterClass0,
167 
168     /// The (first) user defined class, subsequent user defined classes are
169     /// UserClass0+1, and so on.
170     UserClass0 = 1 << 16
171   };
172 
173   /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
174   /// N) for the Nth user defined class.
175   unsigned Kind;
176 
177   /// SuperClasses - The super classes of this class. Note that for simplicities
178   /// sake user operands only record their immediate super class, while register
179   /// operands include all superclasses.
180   std::vector<ClassInfo *> SuperClasses;
181 
182   /// Name - The full class name, suitable for use in an enum.
183   std::string Name;
184 
185   /// ClassName - The unadorned generic name for this class (e.g., Token).
186   std::string ClassName;
187 
188   /// ValueName - The name of the value this class represents; for a token this
189   /// is the literal token string, for an operand it is the TableGen class (or
190   /// empty if this is a derived class).
191   std::string ValueName;
192 
193   /// PredicateMethod - The name of the operand method to test whether the
194   /// operand matches this class; this is not valid for Token or register kinds.
195   std::string PredicateMethod;
196 
197   /// RenderMethod - The name of the operand method to add this operand to an
198   /// MCInst; this is not valid for Token or register kinds.
199   std::string RenderMethod;
200 
201   /// ParserMethod - The name of the operand method to do a target specific
202   /// parsing on the operand.
203   std::string ParserMethod;
204 
205   /// For register classes: the records for all the registers in this class.
206   RegisterSet Registers;
207 
208   /// For custom match classes: the diagnostic kind for when the predicate
209   /// fails.
210   std::string DiagnosticType;
211 
212   /// For custom match classes: the diagnostic string for when the predicate
213   /// fails.
214   std::string DiagnosticString;
215 
216   /// Is this operand optional and not always required.
217   bool IsOptional;
218 
219   /// DefaultMethod - The name of the method that returns the default operand
220   /// for optional operand
221   std::string DefaultMethod;
222 
223 public:
224   /// isRegisterClass() - Check if this is a register class.
isRegisterClass__anon44b0c7da0111::ClassInfo225   bool isRegisterClass() const {
226     return Kind >= RegisterClass0 && Kind < UserClass0;
227   }
228 
229   /// isUserClass() - Check if this is a user defined class.
isUserClass__anon44b0c7da0111::ClassInfo230   bool isUserClass() const { return Kind >= UserClass0; }
231 
232   /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
233   /// are related if they are in the same class hierarchy.
isRelatedTo__anon44b0c7da0111::ClassInfo234   bool isRelatedTo(const ClassInfo &RHS) const {
235     // Tokens are only related to tokens.
236     if (Kind == Token || RHS.Kind == Token)
237       return Kind == Token && RHS.Kind == Token;
238 
239     // Registers classes are only related to registers classes, and only if
240     // their intersection is non-empty.
241     if (isRegisterClass() || RHS.isRegisterClass()) {
242       if (!isRegisterClass() || !RHS.isRegisterClass())
243         return false;
244 
245       std::vector<Record *> Tmp;
246       std::set_intersection(Registers.begin(), Registers.end(),
247                             RHS.Registers.begin(), RHS.Registers.end(),
248                             std::back_inserter(Tmp), LessRecordByID());
249 
250       return !Tmp.empty();
251     }
252 
253     // Otherwise we have two users operands; they are related if they are in the
254     // same class hierarchy.
255     //
256     // FIXME: This is an oversimplification, they should only be related if they
257     // intersect, however we don't have that information.
258     assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
259     const ClassInfo *Root = this;
260     while (!Root->SuperClasses.empty())
261       Root = Root->SuperClasses.front();
262 
263     const ClassInfo *RHSRoot = &RHS;
264     while (!RHSRoot->SuperClasses.empty())
265       RHSRoot = RHSRoot->SuperClasses.front();
266 
267     return Root == RHSRoot;
268   }
269 
270   /// isSubsetOf - Test whether this class is a subset of \p RHS.
isSubsetOf__anon44b0c7da0111::ClassInfo271   bool isSubsetOf(const ClassInfo &RHS) const {
272     // This is a subset of RHS if it is the same class...
273     if (this == &RHS)
274       return true;
275 
276     // ... or if any of its super classes are a subset of RHS.
277     SmallVector<const ClassInfo *, 16> Worklist(SuperClasses.begin(),
278                                                 SuperClasses.end());
279     SmallPtrSet<const ClassInfo *, 16> Visited;
280     while (!Worklist.empty()) {
281       auto *CI = Worklist.pop_back_val();
282       if (CI == &RHS)
283         return true;
284       for (auto *Super : CI->SuperClasses)
285         if (Visited.insert(Super).second)
286           Worklist.push_back(Super);
287     }
288 
289     return false;
290   }
291 
getTreeDepth__anon44b0c7da0111::ClassInfo292   int getTreeDepth() const {
293     int Depth = 0;
294     const ClassInfo *Root = this;
295     while (!Root->SuperClasses.empty()) {
296       Depth++;
297       Root = Root->SuperClasses.front();
298     }
299     return Depth;
300   }
301 
findRoot__anon44b0c7da0111::ClassInfo302   const ClassInfo *findRoot() const {
303     const ClassInfo *Root = this;
304     while (!Root->SuperClasses.empty())
305       Root = Root->SuperClasses.front();
306     return Root;
307   }
308 
309   /// Compare two classes. This does not produce a total ordering, but does
310   /// guarantee that subclasses are sorted before their parents, and that the
311   /// ordering is transitive.
operator <__anon44b0c7da0111::ClassInfo312   bool operator<(const ClassInfo &RHS) const {
313     if (this == &RHS)
314       return false;
315 
316     // First, enforce the ordering between the three different types of class.
317     // Tokens sort before registers, which sort before user classes.
318     if (Kind == Token) {
319       if (RHS.Kind != Token)
320         return true;
321       assert(RHS.Kind == Token);
322     } else if (isRegisterClass()) {
323       if (RHS.Kind == Token)
324         return false;
325       else if (RHS.isUserClass())
326         return true;
327       assert(RHS.isRegisterClass());
328     } else if (isUserClass()) {
329       if (!RHS.isUserClass())
330         return false;
331       assert(RHS.isUserClass());
332     } else {
333       llvm_unreachable("Unknown ClassInfoKind");
334     }
335 
336     if (Kind == Token || isUserClass()) {
337       // Related tokens and user classes get sorted by depth in the inheritence
338       // tree (so that subclasses are before their parents).
339       if (isRelatedTo(RHS)) {
340         if (getTreeDepth() > RHS.getTreeDepth())
341           return true;
342         if (getTreeDepth() < RHS.getTreeDepth())
343           return false;
344       } else {
345         // Unrelated tokens and user classes are ordered by the name of their
346         // root nodes, so that there is a consistent ordering between
347         // unconnected trees.
348         return findRoot()->ValueName < RHS.findRoot()->ValueName;
349       }
350     } else if (isRegisterClass()) {
351       // For register sets, sort by number of registers. This guarantees that
352       // a set will always sort before all of it's strict supersets.
353       if (Registers.size() != RHS.Registers.size())
354         return Registers.size() < RHS.Registers.size();
355     } else {
356       llvm_unreachable("Unknown ClassInfoKind");
357     }
358 
359     // FIXME: We should be able to just return false here, as we only need a
360     // partial order (we use stable sorts, so this is deterministic) and the
361     // name of a class shouldn't be significant. However, some of the backends
362     // accidentally rely on this behaviour, so it will have to stay like this
363     // until they are fixed.
364     return ValueName < RHS.ValueName;
365   }
366 };
367 
368 class AsmVariantInfo {
369 public:
370   StringRef RegisterPrefix;
371   StringRef TokenizingCharacters;
372   StringRef SeparatorCharacters;
373   StringRef BreakCharacters;
374   StringRef Name;
375   int AsmVariantNo;
376 };
377 
getPreferSmallerInstructions(CodeGenTarget const & Target)378 bool getPreferSmallerInstructions(CodeGenTarget const &Target) {
379   return Target.getAsmParser()->getValueAsBit("PreferSmallerInstructions");
380 }
381 
382 /// MatchableInfo - Helper class for storing the necessary information for an
383 /// instruction or alias which is capable of being matched.
384 struct MatchableInfo {
385   struct AsmOperand {
386     /// Token - This is the token that the operand came from.
387     StringRef Token;
388 
389     /// The unique class instance this operand should match.
390     ClassInfo *Class;
391 
392     /// The operand name this is, if anything.
393     StringRef SrcOpName;
394 
395     /// The operand name this is, before renaming for tied operands.
396     StringRef OrigSrcOpName;
397 
398     /// The suboperand index within SrcOpName, or -1 for the entire operand.
399     int SubOpIdx;
400 
401     /// Whether the token is "isolated", i.e., it is preceded and followed
402     /// by separators.
403     bool IsIsolatedToken;
404 
405     /// Register record if this token is singleton register.
406     Record *SingletonReg;
407 
AsmOperand__anon44b0c7da0111::MatchableInfo::AsmOperand408     explicit AsmOperand(bool IsIsolatedToken, StringRef T)
409         : Token(T), Class(nullptr), SubOpIdx(-1),
410           IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {}
411   };
412 
413   /// ResOperand - This represents a single operand in the result instruction
414   /// generated by the match.  In cases (like addressing modes) where a single
415   /// assembler operand expands to multiple MCOperands, this represents the
416   /// single assembler operand, not the MCOperand.
417   struct ResOperand {
418     enum {
419       /// RenderAsmOperand - This represents an operand result that is
420       /// generated by calling the render method on the assembly operand.  The
421       /// corresponding AsmOperand is specified by AsmOperandNum.
422       RenderAsmOperand,
423 
424       /// TiedOperand - This represents a result operand that is a duplicate of
425       /// a previous result operand.
426       TiedOperand,
427 
428       /// ImmOperand - This represents an immediate value that is dumped into
429       /// the operand.
430       ImmOperand,
431 
432       /// RegOperand - This represents a fixed register that is dumped in.
433       RegOperand
434     } Kind;
435 
436     /// Tuple containing the index of the (earlier) result operand that should
437     /// be copied from, as well as the indices of the corresponding (parsed)
438     /// operands in the asm string.
439     struct TiedOperandsTuple {
440       unsigned ResOpnd;
441       unsigned SrcOpnd1Idx;
442       unsigned SrcOpnd2Idx;
443     };
444 
445     union {
446       /// This is the operand # in the AsmOperands list that this should be
447       /// copied from.
448       unsigned AsmOperandNum;
449 
450       /// Description of tied operands.
451       TiedOperandsTuple TiedOperands;
452 
453       /// ImmVal - This is the immediate value added to the instruction.
454       int64_t ImmVal;
455 
456       /// Register - This is the register record.
457       Record *Register;
458     };
459 
460     /// MINumOperands - The number of MCInst operands populated by this
461     /// operand.
462     unsigned MINumOperands;
463 
getRenderedOp__anon44b0c7da0111::MatchableInfo::ResOperand464     static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
465       ResOperand X;
466       X.Kind = RenderAsmOperand;
467       X.AsmOperandNum = AsmOpNum;
468       X.MINumOperands = NumOperands;
469       return X;
470     }
471 
getTiedOp__anon44b0c7da0111::MatchableInfo::ResOperand472     static ResOperand getTiedOp(unsigned TiedOperandNum, unsigned SrcOperand1,
473                                 unsigned SrcOperand2) {
474       ResOperand X;
475       X.Kind = TiedOperand;
476       X.TiedOperands = {TiedOperandNum, SrcOperand1, SrcOperand2};
477       X.MINumOperands = 1;
478       return X;
479     }
480 
getImmOp__anon44b0c7da0111::MatchableInfo::ResOperand481     static ResOperand getImmOp(int64_t Val) {
482       ResOperand X;
483       X.Kind = ImmOperand;
484       X.ImmVal = Val;
485       X.MINumOperands = 1;
486       return X;
487     }
488 
getRegOp__anon44b0c7da0111::MatchableInfo::ResOperand489     static ResOperand getRegOp(Record *Reg) {
490       ResOperand X;
491       X.Kind = RegOperand;
492       X.Register = Reg;
493       X.MINumOperands = 1;
494       return X;
495     }
496   };
497 
498   /// AsmVariantID - Target's assembly syntax variant no.
499   int AsmVariantID;
500 
501   /// AsmString - The assembly string for this instruction (with variants
502   /// removed), e.g. "movsx $src, $dst".
503   std::string AsmString;
504 
505   /// TheDef - This is the definition of the instruction or InstAlias that this
506   /// matchable came from.
507   Record *const TheDef;
508 
509   // ResInstSize - The size of the resulting instruction for this matchable.
510   unsigned ResInstSize;
511 
512   /// DefRec - This is the definition that it came from.
513   PointerUnion<const CodeGenInstruction *, const CodeGenInstAlias *> DefRec;
514 
getResultInst__anon44b0c7da0111::MatchableInfo515   const CodeGenInstruction *getResultInst() const {
516     if (isa<const CodeGenInstruction *>(DefRec))
517       return cast<const CodeGenInstruction *>(DefRec);
518     return cast<const CodeGenInstAlias *>(DefRec)->ResultInst;
519   }
520 
521   /// ResOperands - This is the operand list that should be built for the result
522   /// MCInst.
523   SmallVector<ResOperand, 8> ResOperands;
524 
525   /// Mnemonic - This is the first token of the matched instruction, its
526   /// mnemonic.
527   StringRef Mnemonic;
528 
529   /// AsmOperands - The textual operands that this instruction matches,
530   /// annotated with a class and where in the OperandList they were defined.
531   /// This directly corresponds to the tokenized AsmString after the mnemonic is
532   /// removed.
533   SmallVector<AsmOperand, 8> AsmOperands;
534 
535   /// Predicates - The required subtarget features to match this instruction.
536   SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
537 
538   /// ConversionFnKind - The enum value which is passed to the generated
539   /// convertToMCInst to convert parsed operands into an MCInst for this
540   /// function.
541   std::string ConversionFnKind;
542 
543   /// If this instruction is deprecated in some form.
544   bool HasDeprecation = false;
545 
546   /// If this is an alias, this is use to determine whether or not to using
547   /// the conversion function defined by the instruction's AsmMatchConverter
548   /// or to use the function generated by the alias.
549   bool UseInstAsmMatchConverter;
550 
MatchableInfo__anon44b0c7da0111::MatchableInfo551   MatchableInfo(const CodeGenInstruction &CGI)
552       : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef),
553         ResInstSize(TheDef->getValueAsInt("Size")), DefRec(&CGI),
554         UseInstAsmMatchConverter(true) {}
555 
MatchableInfo__anon44b0c7da0111::MatchableInfo556   MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias)
557       : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef),
558         ResInstSize(Alias->ResultInst->TheDef->getValueAsInt("Size")),
559         DefRec(Alias.release()), UseInstAsmMatchConverter(TheDef->getValueAsBit(
560                                      "UseInstAsmMatchConverter")) {}
561 
562   // Could remove this and the dtor if PointerUnion supported unique_ptr
563   // elements with a dynamic failure/assertion (like the one below) in the case
564   // where it was copied while being in an owning state.
MatchableInfo__anon44b0c7da0111::MatchableInfo565   MatchableInfo(const MatchableInfo &RHS)
566       : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString),
567         TheDef(RHS.TheDef), ResInstSize(RHS.ResInstSize), DefRec(RHS.DefRec),
568         ResOperands(RHS.ResOperands), Mnemonic(RHS.Mnemonic),
569         AsmOperands(RHS.AsmOperands), RequiredFeatures(RHS.RequiredFeatures),
570         ConversionFnKind(RHS.ConversionFnKind),
571         HasDeprecation(RHS.HasDeprecation),
572         UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) {
573     assert(!isa<const CodeGenInstAlias *>(DefRec));
574   }
575 
~MatchableInfo__anon44b0c7da0111::MatchableInfo576   ~MatchableInfo() {
577     delete dyn_cast_if_present<const CodeGenInstAlias *>(DefRec);
578   }
579 
580   // Two-operand aliases clone from the main matchable, but mark the second
581   // operand as a tied operand of the first for purposes of the assembler.
582   void formTwoOperandAlias(StringRef Constraint);
583 
584   void initialize(const AsmMatcherInfo &Info,
585                   SmallPtrSetImpl<Record *> &SingletonRegisters,
586                   AsmVariantInfo const &Variant, bool HasMnemonicFirst);
587 
588   /// validate - Return true if this matchable is a valid thing to match against
589   /// and perform a bunch of validity checking.
590   bool validate(StringRef CommentDelimiter, bool IsAlias) const;
591 
592   /// findAsmOperand - Find the AsmOperand with the specified name and
593   /// suboperand index.
findAsmOperand__anon44b0c7da0111::MatchableInfo594   int findAsmOperand(StringRef N, int SubOpIdx) const {
595     auto I = find_if(AsmOperands, [&](const AsmOperand &Op) {
596       return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx;
597     });
598     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
599   }
600 
601   /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
602   /// This does not check the suboperand index.
findAsmOperandNamed__anon44b0c7da0111::MatchableInfo603   int findAsmOperandNamed(StringRef N, int LastIdx = -1) const {
604     auto I =
605         llvm::find_if(llvm::drop_begin(AsmOperands, LastIdx + 1),
606                       [&](const AsmOperand &Op) { return Op.SrcOpName == N; });
607     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
608   }
609 
findAsmOperandOriginallyNamed__anon44b0c7da0111::MatchableInfo610   int findAsmOperandOriginallyNamed(StringRef N) const {
611     auto I = find_if(AsmOperands, [&](const AsmOperand &Op) {
612       return Op.OrigSrcOpName == N;
613     });
614     return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1;
615   }
616 
617   void buildInstructionResultOperands();
618   void buildAliasResultOperands(bool AliasConstraintsAreChecked);
619 
620   /// shouldBeMatchedBefore - Compare two matchables for ordering.
shouldBeMatchedBefore__anon44b0c7da0111::MatchableInfo621   bool shouldBeMatchedBefore(const MatchableInfo &RHS,
622                              bool PreferSmallerInstructions) const {
623     // The primary comparator is the instruction mnemonic.
624     if (int Cmp = Mnemonic.compare_insensitive(RHS.Mnemonic))
625       return Cmp == -1;
626 
627     // (Optionally) Order by the resultant instuctions size.
628     // eg. for ARM thumb instructions smaller encodings should be preferred.
629     if (PreferSmallerInstructions && ResInstSize != RHS.ResInstSize)
630       return ResInstSize < RHS.ResInstSize;
631 
632     if (AsmOperands.size() != RHS.AsmOperands.size())
633       return AsmOperands.size() < RHS.AsmOperands.size();
634 
635     // Compare lexicographically by operand. The matcher validates that other
636     // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
637     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
638       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
639         return true;
640       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
641         return false;
642     }
643 
644     // For X86 AVX/AVX512 instructions, we prefer vex encoding because the
645     // vex encoding size is smaller. Since X86InstrSSE.td is included ahead
646     // of X86InstrAVX512.td, the AVX instruction ID is less than AVX512 ID.
647     // We use the ID to sort AVX instruction before AVX512 instruction in
648     // matching table. As well as InstAlias.
649     if (getResultInst()->TheDef->isSubClassOf("Instruction") &&
650         getResultInst()->TheDef->getValueAsBit("HasPositionOrder") &&
651         RHS.getResultInst()->TheDef->isSubClassOf("Instruction") &&
652         RHS.getResultInst()->TheDef->getValueAsBit("HasPositionOrder"))
653       return getResultInst()->TheDef->getID() <
654              RHS.getResultInst()->TheDef->getID();
655 
656     // Give matches that require more features higher precedence. This is useful
657     // because we cannot define AssemblerPredicates with the negation of
658     // processor features. For example, ARM v6 "nop" may be either a HINT or
659     // MOV. With v6, we want to match HINT. The assembler has no way to
660     // predicate MOV under "NoV6", but HINT will always match first because it
661     // requires V6 while MOV does not.
662     if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
663       return RequiredFeatures.size() > RHS.RequiredFeatures.size();
664 
665     return false;
666   }
667 
668   /// couldMatchAmbiguouslyWith - Check whether this matchable could
669   /// ambiguously match the same set of operands as \p RHS (without being a
670   /// strictly superior match).
couldMatchAmbiguouslyWith__anon44b0c7da0111::MatchableInfo671   bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS,
672                                  bool PreferSmallerInstructions) const {
673     // The primary comparator is the instruction mnemonic.
674     if (Mnemonic != RHS.Mnemonic)
675       return false;
676 
677     // Different variants can't conflict.
678     if (AsmVariantID != RHS.AsmVariantID)
679       return false;
680 
681     // The size of instruction is unambiguous.
682     if (PreferSmallerInstructions && ResInstSize != RHS.ResInstSize)
683       return false;
684 
685     // The number of operands is unambiguous.
686     if (AsmOperands.size() != RHS.AsmOperands.size())
687       return false;
688 
689     // Otherwise, make sure the ordering of the two instructions is unambiguous
690     // by checking that either (a) a token or operand kind discriminates them,
691     // or (b) the ordering among equivalent kinds is consistent.
692 
693     // Tokens and operand kinds are unambiguous (assuming a correct target
694     // specific parser).
695     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
696       if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
697           AsmOperands[i].Class->Kind == ClassInfo::Token)
698         if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
699             *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
700           return false;
701 
702     // Otherwise, this operand could commute if all operands are equivalent, or
703     // there is a pair of operands that compare less than and a pair that
704     // compare greater than.
705     bool HasLT = false, HasGT = false;
706     for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
707       if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
708         HasLT = true;
709       if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
710         HasGT = true;
711     }
712 
713     return HasLT == HasGT;
714   }
715 
716   void dump() const;
717 
718 private:
719   void tokenizeAsmString(AsmMatcherInfo const &Info,
720                          AsmVariantInfo const &Variant);
721   void addAsmOperand(StringRef Token, bool IsIsolatedToken = false);
722 };
723 
724 struct OperandMatchEntry {
725   unsigned OperandMask;
726   const MatchableInfo *MI;
727   ClassInfo *CI;
728 
create__anon44b0c7da0111::OperandMatchEntry729   static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
730                                   unsigned opMask) {
731     OperandMatchEntry X;
732     X.OperandMask = opMask;
733     X.CI = ci;
734     X.MI = mi;
735     return X;
736   }
737 };
738 
739 class AsmMatcherInfo {
740 public:
741   /// Tracked Records
742   RecordKeeper &Records;
743 
744   /// The tablegen AsmParser record.
745   Record *AsmParser;
746 
747   /// Target - The target information.
748   CodeGenTarget &Target;
749 
750   /// The classes which are needed for matching.
751   std::forward_list<ClassInfo> Classes;
752 
753   /// The information on the matchables to match.
754   std::vector<std::unique_ptr<MatchableInfo>> Matchables;
755 
756   /// Info for custom matching operands by user defined methods.
757   std::vector<OperandMatchEntry> OperandMatchInfo;
758 
759   /// Map of Register records to their class information.
760   typedef std::map<Record *, ClassInfo *, LessRecordByID> RegisterClassesTy;
761   RegisterClassesTy RegisterClasses;
762 
763   /// Map of Predicate records to their subtarget information.
764   std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
765 
766   /// Map of AsmOperandClass records to their class information.
767   std::map<Record *, ClassInfo *> AsmOperandClasses;
768 
769   /// Map of RegisterClass records to their class information.
770   std::map<Record *, ClassInfo *> RegisterClassClasses;
771 
772 private:
773   /// Map of token to class information which has already been constructed.
774   std::map<std::string, ClassInfo *> TokenClasses;
775 
776 private:
777   /// getTokenClass - Lookup or create the class for the given token.
778   ClassInfo *getTokenClass(StringRef Token);
779 
780   /// getOperandClass - Lookup or create the class for the given operand.
781   ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
782                              int SubOpIdx);
783   ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
784 
785   /// buildRegisterClasses - Build the ClassInfo* instances for register
786   /// classes.
787   void buildRegisterClasses(SmallPtrSetImpl<Record *> &SingletonRegisters);
788 
789   /// buildOperandClasses - Build the ClassInfo* instances for user defined
790   /// operand classes.
791   void buildOperandClasses();
792 
793   void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
794                                         unsigned AsmOpIdx);
795   void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
796                                   MatchableInfo::AsmOperand &Op);
797 
798 public:
799   AsmMatcherInfo(Record *AsmParser, CodeGenTarget &Target,
800                  RecordKeeper &Records);
801 
802   /// Construct the various tables used during matching.
803   void buildInfo();
804 
805   /// buildOperandMatchInfo - Build the necessary information to handle user
806   /// defined operand parsing methods.
807   void buildOperandMatchInfo();
808 
809   /// getSubtargetFeature - Lookup or create the subtarget feature info for the
810   /// given operand.
getSubtargetFeature(Record * Def) const811   const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
812     assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
813     const auto &I = SubtargetFeatures.find(Def);
814     return I == SubtargetFeatures.end() ? nullptr : &I->second;
815   }
816 
getRecords() const817   RecordKeeper &getRecords() const { return Records; }
818 
hasOptionalOperands() const819   bool hasOptionalOperands() const {
820     return any_of(Classes,
821                   [](const ClassInfo &Class) { return Class.IsOptional; });
822   }
823 };
824 
825 } // end anonymous namespace
826 
827 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump() const828 LLVM_DUMP_METHOD void MatchableInfo::dump() const {
829   errs() << TheDef->getName() << " -- "
830          << "flattened:\"" << AsmString << "\"\n";
831 
832   errs() << "  variant: " << AsmVariantID << "\n";
833 
834   for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
835     const AsmOperand &Op = AsmOperands[i];
836     errs() << "  op[" << i << "] = " << Op.Class->ClassName << " - ";
837     errs() << '\"' << Op.Token << "\"\n";
838   }
839 }
840 #endif
841 
842 static std::pair<StringRef, StringRef>
parseTwoOperandConstraint(StringRef S,ArrayRef<SMLoc> Loc)843 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
844   // Split via the '='.
845   std::pair<StringRef, StringRef> Ops = S.split('=');
846   if (Ops.second == "")
847     PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
848   // Trim whitespace and the leading '$' on the operand names.
849   size_t start = Ops.first.find_first_of('$');
850   if (start == std::string::npos)
851     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
852   Ops.first = Ops.first.slice(start + 1, std::string::npos);
853   size_t end = Ops.first.find_last_of(" \t");
854   Ops.first = Ops.first.slice(0, end);
855   // Now the second operand.
856   start = Ops.second.find_first_of('$');
857   if (start == std::string::npos)
858     PrintFatalError(Loc, "expected '$' prefix on asm operand name");
859   Ops.second = Ops.second.slice(start + 1, std::string::npos);
860   end = Ops.second.find_last_of(" \t");
861   Ops.first = Ops.first.slice(0, end);
862   return Ops;
863 }
864 
formTwoOperandAlias(StringRef Constraint)865 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
866   // Figure out which operands are aliased and mark them as tied.
867   std::pair<StringRef, StringRef> Ops =
868       parseTwoOperandConstraint(Constraint, TheDef->getLoc());
869 
870   // Find the AsmOperands that refer to the operands we're aliasing.
871   int SrcAsmOperand = findAsmOperandNamed(Ops.first);
872   int DstAsmOperand = findAsmOperandNamed(Ops.second);
873   if (SrcAsmOperand == -1)
874     PrintFatalError(TheDef->getLoc(),
875                     "unknown source two-operand alias operand '" + Ops.first +
876                         "'.");
877   if (DstAsmOperand == -1)
878     PrintFatalError(TheDef->getLoc(),
879                     "unknown destination two-operand alias operand '" +
880                         Ops.second + "'.");
881 
882   // Find the ResOperand that refers to the operand we're aliasing away
883   // and update it to refer to the combined operand instead.
884   for (ResOperand &Op : ResOperands) {
885     if (Op.Kind == ResOperand::RenderAsmOperand &&
886         Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
887       Op.AsmOperandNum = DstAsmOperand;
888       break;
889     }
890   }
891   // Remove the AsmOperand for the alias operand.
892   AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
893   // Adjust the ResOperand references to any AsmOperands that followed
894   // the one we just deleted.
895   for (ResOperand &Op : ResOperands) {
896     switch (Op.Kind) {
897     default:
898       // Nothing to do for operands that don't reference AsmOperands.
899       break;
900     case ResOperand::RenderAsmOperand:
901       if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
902         --Op.AsmOperandNum;
903       break;
904     }
905   }
906 }
907 
908 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
909 /// if present, from specified token.
extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand & Op,const AsmMatcherInfo & Info,StringRef RegisterPrefix)910 static void extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op,
911                                                   const AsmMatcherInfo &Info,
912                                                   StringRef RegisterPrefix) {
913   StringRef Tok = Op.Token;
914 
915   // If this token is not an isolated token, i.e., it isn't separated from
916   // other tokens (e.g. with whitespace), don't interpret it as a register name.
917   if (!Op.IsIsolatedToken)
918     return;
919 
920   if (RegisterPrefix.empty()) {
921     std::string LoweredTok = Tok.lower();
922     if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
923       Op.SingletonReg = Reg->TheDef;
924     return;
925   }
926 
927   if (!Tok.starts_with(RegisterPrefix))
928     return;
929 
930   StringRef RegName = Tok.substr(RegisterPrefix.size());
931   if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
932     Op.SingletonReg = Reg->TheDef;
933 
934   // If there is no register prefix (i.e. "%" in "%eax"), then this may
935   // be some random non-register token, just ignore it.
936 }
937 
initialize(const AsmMatcherInfo & Info,SmallPtrSetImpl<Record * > & SingletonRegisters,AsmVariantInfo const & Variant,bool HasMnemonicFirst)938 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
939                                SmallPtrSetImpl<Record *> &SingletonRegisters,
940                                AsmVariantInfo const &Variant,
941                                bool HasMnemonicFirst) {
942   AsmVariantID = Variant.AsmVariantNo;
943   AsmString = CodeGenInstruction::FlattenAsmStringVariants(
944       AsmString, Variant.AsmVariantNo);
945 
946   tokenizeAsmString(Info, Variant);
947 
948   // The first token of the instruction is the mnemonic, which must be a
949   // simple string, not a $foo variable or a singleton register.
950   if (AsmOperands.empty())
951     PrintFatalError(TheDef->getLoc(),
952                     "Instruction '" + TheDef->getName() + "' has no tokens");
953 
954   assert(!AsmOperands[0].Token.empty());
955   if (HasMnemonicFirst) {
956     Mnemonic = AsmOperands[0].Token;
957     if (Mnemonic[0] == '$')
958       PrintFatalError(TheDef->getLoc(),
959                       "Invalid instruction mnemonic '" + Mnemonic + "'!");
960 
961     // Remove the first operand, it is tracked in the mnemonic field.
962     AsmOperands.erase(AsmOperands.begin());
963   } else if (AsmOperands[0].Token[0] != '$')
964     Mnemonic = AsmOperands[0].Token;
965 
966   // Compute the require features.
967   for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates"))
968     if (const SubtargetFeatureInfo *Feature =
969             Info.getSubtargetFeature(Predicate))
970       RequiredFeatures.push_back(Feature);
971 
972   // Collect singleton registers, if used.
973   for (MatchableInfo::AsmOperand &Op : AsmOperands) {
974     extractSingletonRegisterForAsmOperand(Op, Info, Variant.RegisterPrefix);
975     if (Record *Reg = Op.SingletonReg)
976       SingletonRegisters.insert(Reg);
977   }
978 
979   const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
980   if (!DepMask)
981     DepMask = TheDef->getValue("ComplexDeprecationPredicate");
982 
983   HasDeprecation =
984       DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
985 }
986 
987 /// Append an AsmOperand for the given substring of AsmString.
addAsmOperand(StringRef Token,bool IsIsolatedToken)988 void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) {
989   AsmOperands.push_back(AsmOperand(IsIsolatedToken, Token));
990 }
991 
992 /// tokenizeAsmString - Tokenize a simplified assembly string.
tokenizeAsmString(const AsmMatcherInfo & Info,AsmVariantInfo const & Variant)993 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info,
994                                       AsmVariantInfo const &Variant) {
995   StringRef String = AsmString;
996   size_t Prev = 0;
997   bool InTok = false;
998   bool IsIsolatedToken = true;
999   for (size_t i = 0, e = String.size(); i != e; ++i) {
1000     char Char = String[i];
1001     if (Variant.BreakCharacters.contains(Char)) {
1002       if (InTok) {
1003         addAsmOperand(String.slice(Prev, i), false);
1004         Prev = i;
1005         IsIsolatedToken = false;
1006       }
1007       InTok = true;
1008       continue;
1009     }
1010     if (Variant.TokenizingCharacters.contains(Char)) {
1011       if (InTok) {
1012         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
1013         InTok = false;
1014         IsIsolatedToken = false;
1015       }
1016       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
1017       Prev = i + 1;
1018       IsIsolatedToken = true;
1019       continue;
1020     }
1021     if (Variant.SeparatorCharacters.contains(Char)) {
1022       if (InTok) {
1023         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
1024         InTok = false;
1025       }
1026       Prev = i + 1;
1027       IsIsolatedToken = true;
1028       continue;
1029     }
1030 
1031     switch (Char) {
1032     case '\\':
1033       if (InTok) {
1034         addAsmOperand(String.slice(Prev, i), false);
1035         InTok = false;
1036         IsIsolatedToken = false;
1037       }
1038       ++i;
1039       assert(i != String.size() && "Invalid quoted character");
1040       addAsmOperand(String.slice(i, i + 1), IsIsolatedToken);
1041       Prev = i + 1;
1042       IsIsolatedToken = false;
1043       break;
1044 
1045     case '$': {
1046       if (InTok) {
1047         addAsmOperand(String.slice(Prev, i), IsIsolatedToken);
1048         InTok = false;
1049         IsIsolatedToken = false;
1050       }
1051 
1052       // If this isn't "${", start new identifier looking like "$xxx"
1053       if (i + 1 == String.size() || String[i + 1] != '{') {
1054         Prev = i;
1055         break;
1056       }
1057 
1058       size_t EndPos = String.find('}', i);
1059       assert(EndPos != StringRef::npos &&
1060              "Missing brace in operand reference!");
1061       addAsmOperand(String.slice(i, EndPos + 1), IsIsolatedToken);
1062       Prev = EndPos + 1;
1063       i = EndPos;
1064       IsIsolatedToken = false;
1065       break;
1066     }
1067 
1068     default:
1069       InTok = true;
1070       break;
1071     }
1072   }
1073   if (InTok && Prev != String.size())
1074     addAsmOperand(String.substr(Prev), IsIsolatedToken);
1075 }
1076 
validate(StringRef CommentDelimiter,bool IsAlias) const1077 bool MatchableInfo::validate(StringRef CommentDelimiter, bool IsAlias) const {
1078   // Reject matchables with no .s string.
1079   if (AsmString.empty())
1080     PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
1081 
1082   // Reject any matchables with a newline in them, they should be marked
1083   // isCodeGenOnly if they are pseudo instructions.
1084   if (AsmString.find('\n') != std::string::npos)
1085     PrintFatalError(TheDef->getLoc(),
1086                     "multiline instruction is not valid for the asmparser, "
1087                     "mark it isCodeGenOnly");
1088 
1089   // Remove comments from the asm string.  We know that the asmstring only
1090   // has one line.
1091   if (!CommentDelimiter.empty() &&
1092       StringRef(AsmString).contains(CommentDelimiter))
1093     PrintFatalError(TheDef->getLoc(),
1094                     "asmstring for instruction has comment character in it, "
1095                     "mark it isCodeGenOnly");
1096 
1097   // Reject matchables with operand modifiers, these aren't something we can
1098   // handle, the target should be refactored to use operands instead of
1099   // modifiers.
1100   //
1101   // Also, check for instructions which reference the operand multiple times,
1102   // if they don't define a custom AsmMatcher: this implies a constraint that
1103   // the built-in matching code would not honor.
1104   std::set<std::string> OperandNames;
1105   for (const AsmOperand &Op : AsmOperands) {
1106     StringRef Tok = Op.Token;
1107     if (Tok[0] == '$' && Tok.contains(':'))
1108       PrintFatalError(
1109           TheDef->getLoc(),
1110           "matchable with operand modifier '" + Tok +
1111               "' not supported by asm matcher.  Mark isCodeGenOnly!");
1112     // Verify that any operand is only mentioned once.
1113     // We reject aliases and ignore instructions for now.
1114     if (!IsAlias && TheDef->getValueAsString("AsmMatchConverter").empty() &&
1115         Tok[0] == '$' && !OperandNames.insert(std::string(Tok)).second) {
1116       LLVM_DEBUG({
1117         errs() << "warning: '" << TheDef->getName() << "': "
1118                << "ignoring instruction with tied operand '" << Tok << "'\n";
1119       });
1120       return false;
1121     }
1122   }
1123 
1124   return true;
1125 }
1126 
getEnumNameForToken(StringRef Str)1127 static std::string getEnumNameForToken(StringRef Str) {
1128   std::string Res;
1129 
1130   for (char C : Str) {
1131     switch (C) {
1132     case '*':
1133       Res += "_STAR_";
1134       break;
1135     case '%':
1136       Res += "_PCT_";
1137       break;
1138     case ':':
1139       Res += "_COLON_";
1140       break;
1141     case '!':
1142       Res += "_EXCLAIM_";
1143       break;
1144     case '.':
1145       Res += "_DOT_";
1146       break;
1147     case '<':
1148       Res += "_LT_";
1149       break;
1150     case '>':
1151       Res += "_GT_";
1152       break;
1153     case '-':
1154       Res += "_MINUS_";
1155       break;
1156     case '#':
1157       Res += "_HASH_";
1158       break;
1159     default:
1160       if (isAlnum(C))
1161         Res += C;
1162       else
1163         Res += "_" + utostr((unsigned)C) + "_";
1164     }
1165   }
1166 
1167   return Res;
1168 }
1169 
getTokenClass(StringRef Token)1170 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
1171   ClassInfo *&Entry = TokenClasses[std::string(Token)];
1172 
1173   if (!Entry) {
1174     Classes.emplace_front();
1175     Entry = &Classes.front();
1176     Entry->Kind = ClassInfo::Token;
1177     Entry->ClassName = "Token";
1178     Entry->Name = "MCK_" + getEnumNameForToken(Token);
1179     Entry->ValueName = std::string(Token);
1180     Entry->PredicateMethod = "<invalid>";
1181     Entry->RenderMethod = "<invalid>";
1182     Entry->ParserMethod = "";
1183     Entry->DiagnosticType = "";
1184     Entry->IsOptional = false;
1185     Entry->DefaultMethod = "<invalid>";
1186   }
1187 
1188   return Entry;
1189 }
1190 
1191 ClassInfo *
getOperandClass(const CGIOperandList::OperandInfo & OI,int SubOpIdx)1192 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1193                                 int SubOpIdx) {
1194   Record *Rec = OI.Rec;
1195   if (SubOpIdx != -1)
1196     Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1197   return getOperandClass(Rec, SubOpIdx);
1198 }
1199 
getOperandClass(Record * Rec,int SubOpIdx)1200 ClassInfo *AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1201   if (Rec->isSubClassOf("RegisterOperand")) {
1202     // RegisterOperand may have an associated ParserMatchClass. If it does,
1203     // use it, else just fall back to the underlying register class.
1204     const RecordVal *R = Rec->getValue("ParserMatchClass");
1205     if (!R || !R->getValue())
1206       PrintFatalError(Rec->getLoc(),
1207                       "Record `" + Rec->getName() +
1208                           "' does not have a ParserMatchClass!\n");
1209 
1210     if (DefInit *DI = dyn_cast<DefInit>(R->getValue())) {
1211       Record *MatchClass = DI->getDef();
1212       if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1213         return CI;
1214     }
1215 
1216     // No custom match class. Just use the register class.
1217     Record *ClassRec = Rec->getValueAsDef("RegClass");
1218     if (!ClassRec)
1219       PrintFatalError(Rec->getLoc(),
1220                       "RegisterOperand `" + Rec->getName() +
1221                           "' has no associated register class!\n");
1222     if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1223       return CI;
1224     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1225   }
1226 
1227   if (Rec->isSubClassOf("RegisterClass")) {
1228     if (ClassInfo *CI = RegisterClassClasses[Rec])
1229       return CI;
1230     PrintFatalError(Rec->getLoc(), "register class has no class info!");
1231   }
1232 
1233   if (!Rec->isSubClassOf("Operand"))
1234     PrintFatalError(Rec->getLoc(),
1235                     "Operand `" + Rec->getName() +
1236                         "' does not derive from class Operand!\n");
1237   Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1238   if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1239     return CI;
1240 
1241   PrintFatalError(Rec->getLoc(), "operand has no match class!");
1242 }
1243 
1244 struct LessRegisterSet {
operator ()LessRegisterSet1245   bool operator()(const RegisterSet &LHS, const RegisterSet &RHS) const {
1246     // std::set<T> defines its own compariso "operator<", but it
1247     // performs a lexicographical comparison by T's innate comparison
1248     // for some reason. We don't want non-deterministic pointer
1249     // comparisons so use this instead.
1250     return std::lexicographical_compare(LHS.begin(), LHS.end(), RHS.begin(),
1251                                         RHS.end(), LessRecordByID());
1252   }
1253 };
1254 
buildRegisterClasses(SmallPtrSetImpl<Record * > & SingletonRegisters)1255 void AsmMatcherInfo::buildRegisterClasses(
1256     SmallPtrSetImpl<Record *> &SingletonRegisters) {
1257   const auto &Registers = Target.getRegBank().getRegisters();
1258   auto &RegClassList = Target.getRegBank().getRegClasses();
1259 
1260   typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1261 
1262   // The register sets used for matching.
1263   RegisterSetSet RegisterSets;
1264 
1265   // Gather the defined sets.
1266   for (const CodeGenRegisterClass &RC : RegClassList)
1267     RegisterSets.insert(
1268         RegisterSet(RC.getOrder().begin(), RC.getOrder().end()));
1269 
1270   // Add any required singleton sets.
1271   for (Record *Rec : SingletonRegisters) {
1272     RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1273   }
1274 
1275   // Introduce derived sets where necessary (when a register does not determine
1276   // a unique register set class), and build the mapping of registers to the set
1277   // they should classify to.
1278   std::map<Record *, RegisterSet> RegisterMap;
1279   for (const CodeGenRegister &CGR : Registers) {
1280     // Compute the intersection of all sets containing this register.
1281     RegisterSet ContainingSet;
1282 
1283     for (const RegisterSet &RS : RegisterSets) {
1284       if (!RS.count(CGR.TheDef))
1285         continue;
1286 
1287       if (ContainingSet.empty()) {
1288         ContainingSet = RS;
1289         continue;
1290       }
1291 
1292       RegisterSet Tmp;
1293       std::set_intersection(ContainingSet.begin(), ContainingSet.end(),
1294                             RS.begin(), RS.end(),
1295                             std::inserter(Tmp, Tmp.begin()), LessRecordByID());
1296       ContainingSet = std::move(Tmp);
1297     }
1298 
1299     if (!ContainingSet.empty()) {
1300       RegisterSets.insert(ContainingSet);
1301       RegisterMap.insert(std::pair(CGR.TheDef, ContainingSet));
1302     }
1303   }
1304 
1305   // Construct the register classes.
1306   std::map<RegisterSet, ClassInfo *, LessRegisterSet> RegisterSetClasses;
1307   unsigned Index = 0;
1308   for (const RegisterSet &RS : RegisterSets) {
1309     Classes.emplace_front();
1310     ClassInfo *CI = &Classes.front();
1311     CI->Kind = ClassInfo::RegisterClass0 + Index;
1312     CI->ClassName = "Reg" + utostr(Index);
1313     CI->Name = "MCK_Reg" + utostr(Index);
1314     CI->ValueName = "";
1315     CI->PredicateMethod = ""; // unused
1316     CI->RenderMethod = "addRegOperands";
1317     CI->Registers = RS;
1318     // FIXME: diagnostic type.
1319     CI->DiagnosticType = "";
1320     CI->IsOptional = false;
1321     CI->DefaultMethod = ""; // unused
1322     RegisterSetClasses.insert(std::pair(RS, CI));
1323     ++Index;
1324   }
1325 
1326   // Find the superclasses; we could compute only the subgroup lattice edges,
1327   // but there isn't really a point.
1328   for (const RegisterSet &RS : RegisterSets) {
1329     ClassInfo *CI = RegisterSetClasses[RS];
1330     for (const RegisterSet &RS2 : RegisterSets)
1331       if (RS != RS2 && std::includes(RS2.begin(), RS2.end(), RS.begin(),
1332                                      RS.end(), LessRecordByID()))
1333         CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1334   }
1335 
1336   // Name the register classes which correspond to a user defined RegisterClass.
1337   for (const CodeGenRegisterClass &RC : RegClassList) {
1338     // Def will be NULL for non-user defined register classes.
1339     Record *Def = RC.getDef();
1340     if (!Def)
1341       continue;
1342     ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(),
1343                                                    RC.getOrder().end())];
1344     if (CI->ValueName.empty()) {
1345       CI->ClassName = RC.getName();
1346       CI->Name = "MCK_" + RC.getName();
1347       CI->ValueName = RC.getName();
1348     } else
1349       CI->ValueName = CI->ValueName + "," + RC.getName();
1350 
1351     Init *DiagnosticType = Def->getValueInit("DiagnosticType");
1352     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1353       CI->DiagnosticType = std::string(SI->getValue());
1354 
1355     Init *DiagnosticString = Def->getValueInit("DiagnosticString");
1356     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticString))
1357       CI->DiagnosticString = std::string(SI->getValue());
1358 
1359     // If we have a diagnostic string but the diagnostic type is not specified
1360     // explicitly, create an anonymous diagnostic type.
1361     if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty())
1362       CI->DiagnosticType = RC.getName();
1363 
1364     RegisterClassClasses.insert(std::pair(Def, CI));
1365   }
1366 
1367   // Populate the map for individual registers.
1368   for (auto &It : RegisterMap)
1369     RegisterClasses[It.first] = RegisterSetClasses[It.second];
1370 
1371   // Name the register classes which correspond to singleton registers.
1372   for (Record *Rec : SingletonRegisters) {
1373     ClassInfo *CI = RegisterClasses[Rec];
1374     assert(CI && "Missing singleton register class info!");
1375 
1376     if (CI->ValueName.empty()) {
1377       CI->ClassName = std::string(Rec->getName());
1378       CI->Name = "MCK_" + Rec->getName().str();
1379       CI->ValueName = std::string(Rec->getName());
1380     } else
1381       CI->ValueName = CI->ValueName + "," + Rec->getName().str();
1382   }
1383 }
1384 
buildOperandClasses()1385 void AsmMatcherInfo::buildOperandClasses() {
1386   std::vector<Record *> AsmOperands =
1387       Records.getAllDerivedDefinitions("AsmOperandClass");
1388 
1389   // Pre-populate AsmOperandClasses map.
1390   for (Record *Rec : AsmOperands) {
1391     Classes.emplace_front();
1392     AsmOperandClasses[Rec] = &Classes.front();
1393   }
1394 
1395   unsigned Index = 0;
1396   for (Record *Rec : AsmOperands) {
1397     ClassInfo *CI = AsmOperandClasses[Rec];
1398     CI->Kind = ClassInfo::UserClass0 + Index;
1399 
1400     ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1401     for (Init *I : Supers->getValues()) {
1402       DefInit *DI = dyn_cast<DefInit>(I);
1403       if (!DI) {
1404         PrintError(Rec->getLoc(), "Invalid super class reference!");
1405         continue;
1406       }
1407 
1408       ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1409       if (!SC)
1410         PrintError(Rec->getLoc(), "Invalid super class reference!");
1411       else
1412         CI->SuperClasses.push_back(SC);
1413     }
1414     CI->ClassName = std::string(Rec->getValueAsString("Name"));
1415     CI->Name = "MCK_" + CI->ClassName;
1416     CI->ValueName = std::string(Rec->getName());
1417 
1418     // Get or construct the predicate method name.
1419     Init *PMName = Rec->getValueInit("PredicateMethod");
1420     if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1421       CI->PredicateMethod = std::string(SI->getValue());
1422     } else {
1423       assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1424       CI->PredicateMethod = "is" + CI->ClassName;
1425     }
1426 
1427     // Get or construct the render method name.
1428     Init *RMName = Rec->getValueInit("RenderMethod");
1429     if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1430       CI->RenderMethod = std::string(SI->getValue());
1431     } else {
1432       assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1433       CI->RenderMethod = "add" + CI->ClassName + "Operands";
1434     }
1435 
1436     // Get the parse method name or leave it as empty.
1437     Init *PRMName = Rec->getValueInit("ParserMethod");
1438     if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1439       CI->ParserMethod = std::string(SI->getValue());
1440 
1441     // Get the diagnostic type and string or leave them as empty.
1442     Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1443     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1444       CI->DiagnosticType = std::string(SI->getValue());
1445     Init *DiagnosticString = Rec->getValueInit("DiagnosticString");
1446     if (StringInit *SI = dyn_cast<StringInit>(DiagnosticString))
1447       CI->DiagnosticString = std::string(SI->getValue());
1448     // If we have a DiagnosticString, we need a DiagnosticType for use within
1449     // the matcher.
1450     if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty())
1451       CI->DiagnosticType = CI->ClassName;
1452 
1453     Init *IsOptional = Rec->getValueInit("IsOptional");
1454     if (BitInit *BI = dyn_cast<BitInit>(IsOptional))
1455       CI->IsOptional = BI->getValue();
1456 
1457     // Get or construct the default method name.
1458     Init *DMName = Rec->getValueInit("DefaultMethod");
1459     if (StringInit *SI = dyn_cast<StringInit>(DMName)) {
1460       CI->DefaultMethod = std::string(SI->getValue());
1461     } else {
1462       assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!");
1463       CI->DefaultMethod = "default" + CI->ClassName + "Operands";
1464     }
1465 
1466     ++Index;
1467   }
1468 }
1469 
AsmMatcherInfo(Record * asmParser,CodeGenTarget & target,RecordKeeper & records)1470 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser, CodeGenTarget &target,
1471                                RecordKeeper &records)
1472     : Records(records), AsmParser(asmParser), Target(target) {}
1473 
1474 /// buildOperandMatchInfo - Build the necessary information to handle user
1475 /// defined operand parsing methods.
buildOperandMatchInfo()1476 void AsmMatcherInfo::buildOperandMatchInfo() {
1477 
1478   /// Map containing a mask with all operands indices that can be found for
1479   /// that class inside a instruction.
1480   typedef std::map<ClassInfo *, unsigned, deref<std::less<>>> OpClassMaskTy;
1481   OpClassMaskTy OpClassMask;
1482 
1483   bool CallCustomParserForAllOperands =
1484       AsmParser->getValueAsBit("CallCustomParserForAllOperands");
1485   for (const auto &MI : Matchables) {
1486     OpClassMask.clear();
1487 
1488     // Keep track of all operands of this instructions which belong to the
1489     // same class.
1490     unsigned NumOptionalOps = 0;
1491     for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
1492       const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
1493       if (CallCustomParserForAllOperands || !Op.Class->ParserMethod.empty()) {
1494         unsigned &OperandMask = OpClassMask[Op.Class];
1495         OperandMask |= maskTrailingOnes<unsigned>(NumOptionalOps + 1)
1496                        << (i - NumOptionalOps);
1497       }
1498       if (Op.Class->IsOptional)
1499         ++NumOptionalOps;
1500     }
1501 
1502     // Generate operand match info for each mnemonic/operand class pair.
1503     for (const auto &OCM : OpClassMask) {
1504       unsigned OpMask = OCM.second;
1505       ClassInfo *CI = OCM.first;
1506       OperandMatchInfo.push_back(
1507           OperandMatchEntry::create(MI.get(), CI, OpMask));
1508     }
1509   }
1510 }
1511 
buildInfo()1512 void AsmMatcherInfo::buildInfo() {
1513   // Build information about all of the AssemblerPredicates.
1514   const std::vector<std::pair<Record *, SubtargetFeatureInfo>>
1515       &SubtargetFeaturePairs = SubtargetFeatureInfo::getAll(Records);
1516   SubtargetFeatures.insert(SubtargetFeaturePairs.begin(),
1517                            SubtargetFeaturePairs.end());
1518 #ifndef NDEBUG
1519   for (const auto &Pair : SubtargetFeatures)
1520     LLVM_DEBUG(Pair.second.dump());
1521 #endif // NDEBUG
1522 
1523   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
1524   bool ReportMultipleNearMisses =
1525       AsmParser->getValueAsBit("ReportMultipleNearMisses");
1526 
1527   // Parse the instructions; we need to do this first so that we can gather the
1528   // singleton register classes.
1529   SmallPtrSet<Record *, 16> SingletonRegisters;
1530   unsigned VariantCount = Target.getAsmParserVariantCount();
1531   for (unsigned VC = 0; VC != VariantCount; ++VC) {
1532     Record *AsmVariant = Target.getAsmParserVariant(VC);
1533     StringRef CommentDelimiter =
1534         AsmVariant->getValueAsString("CommentDelimiter");
1535     AsmVariantInfo Variant;
1536     Variant.RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1537     Variant.TokenizingCharacters =
1538         AsmVariant->getValueAsString("TokenizingCharacters");
1539     Variant.SeparatorCharacters =
1540         AsmVariant->getValueAsString("SeparatorCharacters");
1541     Variant.BreakCharacters = AsmVariant->getValueAsString("BreakCharacters");
1542     Variant.Name = AsmVariant->getValueAsString("Name");
1543     Variant.AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1544 
1545     for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) {
1546 
1547       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1548       // filter the set of instructions we consider.
1549       if (!StringRef(CGI->TheDef->getName()).starts_with(MatchPrefix))
1550         continue;
1551 
1552       // Ignore "codegen only" instructions.
1553       if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1554         continue;
1555 
1556       // Ignore instructions for different instructions
1557       StringRef V = CGI->TheDef->getValueAsString("AsmVariantName");
1558       if (!V.empty() && V != Variant.Name)
1559         continue;
1560 
1561       auto II = std::make_unique<MatchableInfo>(*CGI);
1562 
1563       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1564 
1565       // Ignore instructions which shouldn't be matched and diagnose invalid
1566       // instruction definitions with an error.
1567       if (!II->validate(CommentDelimiter, false))
1568         continue;
1569 
1570       Matchables.push_back(std::move(II));
1571     }
1572 
1573     // Parse all of the InstAlias definitions and stick them in the list of
1574     // matchables.
1575     std::vector<Record *> AllInstAliases =
1576         Records.getAllDerivedDefinitions("InstAlias");
1577     for (Record *InstAlias : AllInstAliases) {
1578       auto Alias = std::make_unique<CodeGenInstAlias>(InstAlias, Target);
1579 
1580       // If the tblgen -match-prefix option is specified (for tblgen hackers),
1581       // filter the set of instruction aliases we consider, based on the target
1582       // instruction.
1583       if (!StringRef(Alias->ResultInst->TheDef->getName())
1584                .starts_with(MatchPrefix))
1585         continue;
1586 
1587       StringRef V = Alias->TheDef->getValueAsString("AsmVariantName");
1588       if (!V.empty() && V != Variant.Name)
1589         continue;
1590 
1591       auto II = std::make_unique<MatchableInfo>(std::move(Alias));
1592 
1593       II->initialize(*this, SingletonRegisters, Variant, HasMnemonicFirst);
1594 
1595       // Validate the alias definitions.
1596       II->validate(CommentDelimiter, true);
1597 
1598       Matchables.push_back(std::move(II));
1599     }
1600   }
1601 
1602   // Build info for the register classes.
1603   buildRegisterClasses(SingletonRegisters);
1604 
1605   // Build info for the user defined assembly operand classes.
1606   buildOperandClasses();
1607 
1608   // Build the information about matchables, now that we have fully formed
1609   // classes.
1610   std::vector<std::unique_ptr<MatchableInfo>> NewMatchables;
1611   for (auto &II : Matchables) {
1612     // Parse the tokens after the mnemonic.
1613     // Note: buildInstructionOperandReference may insert new AsmOperands, so
1614     // don't precompute the loop bound.
1615     for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1616       MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1617       StringRef Token = Op.Token;
1618 
1619       // Check for singleton registers.
1620       if (Record *RegRecord = Op.SingletonReg) {
1621         Op.Class = RegisterClasses[RegRecord];
1622         assert(Op.Class && Op.Class->Registers.size() == 1 &&
1623                "Unexpected class for singleton register");
1624         continue;
1625       }
1626 
1627       // Check for simple tokens.
1628       if (Token[0] != '$') {
1629         Op.Class = getTokenClass(Token);
1630         continue;
1631       }
1632 
1633       if (Token.size() > 1 && isdigit(Token[1])) {
1634         Op.Class = getTokenClass(Token);
1635         continue;
1636       }
1637 
1638       // Otherwise this is an operand reference.
1639       StringRef OperandName;
1640       if (Token[1] == '{')
1641         OperandName = Token.substr(2, Token.size() - 3);
1642       else
1643         OperandName = Token.substr(1);
1644 
1645       if (isa<const CodeGenInstruction *>(II->DefRec))
1646         buildInstructionOperandReference(II.get(), OperandName, i);
1647       else
1648         buildAliasOperandReference(II.get(), OperandName, Op);
1649     }
1650 
1651     if (isa<const CodeGenInstruction *>(II->DefRec)) {
1652       II->buildInstructionResultOperands();
1653       // If the instruction has a two-operand alias, build up the
1654       // matchable here. We'll add them in bulk at the end to avoid
1655       // confusing this loop.
1656       StringRef Constraint =
1657           II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1658       if (Constraint != "") {
1659         // Start by making a copy of the original matchable.
1660         auto AliasII = std::make_unique<MatchableInfo>(*II);
1661 
1662         // Adjust it to be a two-operand alias.
1663         AliasII->formTwoOperandAlias(Constraint);
1664 
1665         // Add the alias to the matchables list.
1666         NewMatchables.push_back(std::move(AliasII));
1667       }
1668     } else
1669       // FIXME: The tied operands checking is not yet integrated with the
1670       // framework for reporting multiple near misses. To prevent invalid
1671       // formats from being matched with an alias if a tied-operands check
1672       // would otherwise have disallowed it, we just disallow such constructs
1673       // in TableGen completely.
1674       II->buildAliasResultOperands(!ReportMultipleNearMisses);
1675   }
1676   if (!NewMatchables.empty())
1677     Matchables.insert(Matchables.end(),
1678                       std::make_move_iterator(NewMatchables.begin()),
1679                       std::make_move_iterator(NewMatchables.end()));
1680 
1681   // Process token alias definitions and set up the associated superclass
1682   // information.
1683   std::vector<Record *> AllTokenAliases =
1684       Records.getAllDerivedDefinitions("TokenAlias");
1685   for (Record *Rec : AllTokenAliases) {
1686     ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1687     ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1688     if (FromClass == ToClass)
1689       PrintFatalError(Rec->getLoc(),
1690                       "error: Destination value identical to source value.");
1691     FromClass->SuperClasses.push_back(ToClass);
1692   }
1693 
1694   // Reorder classes so that classes precede super classes.
1695   Classes.sort();
1696 
1697 #ifdef EXPENSIVE_CHECKS
1698   // Verify that the table is sorted and operator < works transitively.
1699   for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) {
1700     for (auto J = I; J != E; ++J) {
1701       assert(!(*J < *I));
1702       assert(I == J || !J->isSubsetOf(*I));
1703     }
1704   }
1705 #endif
1706 }
1707 
1708 /// buildInstructionOperandReference - The specified operand is a reference to a
1709 /// named operand such as $src.  Resolve the Class and OperandInfo pointers.
buildInstructionOperandReference(MatchableInfo * II,StringRef OperandName,unsigned AsmOpIdx)1710 void AsmMatcherInfo::buildInstructionOperandReference(MatchableInfo *II,
1711                                                       StringRef OperandName,
1712                                                       unsigned AsmOpIdx) {
1713   const CodeGenInstruction &CGI = *cast<const CodeGenInstruction *>(II->DefRec);
1714   const CGIOperandList &Operands = CGI.Operands;
1715   MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1716 
1717   // Map this token to an operand.
1718   unsigned Idx;
1719   if (!Operands.hasOperandNamed(OperandName, Idx))
1720     PrintFatalError(II->TheDef->getLoc(),
1721                     "error: unable to find operand: '" + OperandName + "'");
1722 
1723   // If the instruction operand has multiple suboperands, but the parser
1724   // match class for the asm operand is still the default "ImmAsmOperand",
1725   // then handle each suboperand separately.
1726   if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1727     Record *Rec = Operands[Idx].Rec;
1728     assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1729     Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1730     if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1731       // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1732       StringRef Token = Op->Token; // save this in case Op gets moved
1733       for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1734         MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token);
1735         NewAsmOp.SubOpIdx = SI;
1736         II->AsmOperands.insert(II->AsmOperands.begin() + AsmOpIdx + SI,
1737                                NewAsmOp);
1738       }
1739       // Replace Op with first suboperand.
1740       Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1741       Op->SubOpIdx = 0;
1742     }
1743   }
1744 
1745   // Set up the operand class.
1746   Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1747   Op->OrigSrcOpName = OperandName;
1748 
1749   // If the named operand is tied, canonicalize it to the untied operand.
1750   // For example, something like:
1751   //   (outs GPR:$dst), (ins GPR:$src)
1752   // with an asmstring of
1753   //   "inc $src"
1754   // we want to canonicalize to:
1755   //   "inc $dst"
1756   // so that we know how to provide the $dst operand when filling in the result.
1757   int OITied = -1;
1758   if (Operands[Idx].MINumOperands == 1)
1759     OITied = Operands[Idx].getTiedRegister();
1760   if (OITied != -1) {
1761     // The tied operand index is an MIOperand index, find the operand that
1762     // contains it.
1763     std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1764     OperandName = Operands[Idx.first].Name;
1765     Op->SubOpIdx = Idx.second;
1766   }
1767 
1768   Op->SrcOpName = OperandName;
1769 }
1770 
1771 /// buildAliasOperandReference - When parsing an operand reference out of the
1772 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1773 /// operand reference is by looking it up in the result pattern definition.
buildAliasOperandReference(MatchableInfo * II,StringRef OperandName,MatchableInfo::AsmOperand & Op)1774 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1775                                                 StringRef OperandName,
1776                                                 MatchableInfo::AsmOperand &Op) {
1777   const CodeGenInstAlias &CGA = *cast<const CodeGenInstAlias *>(II->DefRec);
1778 
1779   // Set up the operand class.
1780   for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1781     if (CGA.ResultOperands[i].isRecord() &&
1782         CGA.ResultOperands[i].getName() == OperandName) {
1783       // It's safe to go with the first one we find, because CodeGenInstAlias
1784       // validates that all operands with the same name have the same record.
1785       Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1786       // Use the match class from the Alias definition, not the
1787       // destination instruction, as we may have an immediate that's
1788       // being munged by the match class.
1789       Op.Class =
1790           getOperandClass(CGA.ResultOperands[i].getRecord(), Op.SubOpIdx);
1791       Op.SrcOpName = OperandName;
1792       Op.OrigSrcOpName = OperandName;
1793       return;
1794     }
1795 
1796   PrintFatalError(II->TheDef->getLoc(),
1797                   "error: unable to find operand: '" + OperandName + "'");
1798 }
1799 
buildInstructionResultOperands()1800 void MatchableInfo::buildInstructionResultOperands() {
1801   const CodeGenInstruction *ResultInst = getResultInst();
1802 
1803   // Loop over all operands of the result instruction, determining how to
1804   // populate them.
1805   for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) {
1806     // If this is a tied operand, just copy from the previously handled operand.
1807     int TiedOp = -1;
1808     if (OpInfo.MINumOperands == 1)
1809       TiedOp = OpInfo.getTiedRegister();
1810     if (TiedOp != -1) {
1811       int TiedSrcOperand = findAsmOperandOriginallyNamed(OpInfo.Name);
1812       if (TiedSrcOperand != -1 &&
1813           ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand)
1814         ResOperands.push_back(ResOperand::getTiedOp(
1815             TiedOp, ResOperands[TiedOp].AsmOperandNum, TiedSrcOperand));
1816       else
1817         ResOperands.push_back(ResOperand::getTiedOp(TiedOp, 0, 0));
1818       continue;
1819     }
1820 
1821     int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1822     if (OpInfo.Name.empty() || SrcOperand == -1) {
1823       // This may happen for operands that are tied to a suboperand of a
1824       // complex operand.  Simply use a dummy value here; nobody should
1825       // use this operand slot.
1826       // FIXME: The long term goal is for the MCOperand list to not contain
1827       // tied operands at all.
1828       ResOperands.push_back(ResOperand::getImmOp(0));
1829       continue;
1830     }
1831 
1832     // Check if the one AsmOperand populates the entire operand.
1833     unsigned NumOperands = OpInfo.MINumOperands;
1834     if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1835       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1836       continue;
1837     }
1838 
1839     // Add a separate ResOperand for each suboperand.
1840     for (unsigned AI = 0; AI < NumOperands; ++AI) {
1841       assert(AsmOperands[SrcOperand + AI].SubOpIdx == (int)AI &&
1842              AsmOperands[SrcOperand + AI].SrcOpName == OpInfo.Name &&
1843              "unexpected AsmOperands for suboperands");
1844       ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1845     }
1846   }
1847 }
1848 
buildAliasResultOperands(bool AliasConstraintsAreChecked)1849 void MatchableInfo::buildAliasResultOperands(bool AliasConstraintsAreChecked) {
1850   const CodeGenInstAlias &CGA = *cast<const CodeGenInstAlias *>(DefRec);
1851   const CodeGenInstruction *ResultInst = getResultInst();
1852 
1853   // Map of:  $reg -> #lastref
1854   //   where $reg is the name of the operand in the asm string
1855   //   where #lastref is the last processed index where $reg was referenced in
1856   //   the asm string.
1857   SmallDenseMap<StringRef, int> OperandRefs;
1858 
1859   // Loop over all operands of the result instruction, determining how to
1860   // populate them.
1861   unsigned AliasOpNo = 0;
1862   unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1863   for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1864     const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1865 
1866     // If this is a tied operand, just copy from the previously handled operand.
1867     int TiedOp = -1;
1868     if (OpInfo->MINumOperands == 1)
1869       TiedOp = OpInfo->getTiedRegister();
1870     if (TiedOp != -1) {
1871       unsigned SrcOp1 = 0;
1872       unsigned SrcOp2 = 0;
1873 
1874       // If an operand has been specified twice in the asm string,
1875       // add the two source operand's indices to the TiedOp so that
1876       // at runtime the 'tied' constraint is checked.
1877       if (ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand) {
1878         SrcOp1 = ResOperands[TiedOp].AsmOperandNum;
1879 
1880         // Find the next operand (similarly named operand) in the string.
1881         StringRef Name = AsmOperands[SrcOp1].SrcOpName;
1882         auto Insert = OperandRefs.try_emplace(Name, SrcOp1);
1883         SrcOp2 = findAsmOperandNamed(Name, Insert.first->second);
1884 
1885         // Not updating the record in OperandRefs will cause TableGen
1886         // to fail with an error at the end of this function.
1887         if (AliasConstraintsAreChecked)
1888           Insert.first->second = SrcOp2;
1889 
1890         // In case it only has one reference in the asm string,
1891         // it doesn't need to be checked for tied constraints.
1892         SrcOp2 = (SrcOp2 == (unsigned)-1) ? SrcOp1 : SrcOp2;
1893       }
1894 
1895       // If the alias operand is of a different operand class, we only want
1896       // to benefit from the tied-operands check and just match the operand
1897       // as a normal, but not copy the original (TiedOp) to the result
1898       // instruction. We do this by passing -1 as the tied operand to copy.
1899       if (ResultInst->Operands[i].Rec->getName() !=
1900           ResultInst->Operands[TiedOp].Rec->getName()) {
1901         SrcOp1 = ResOperands[TiedOp].AsmOperandNum;
1902         int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1903         StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1904         SrcOp2 = findAsmOperand(Name, SubIdx);
1905         ResOperands.push_back(
1906             ResOperand::getTiedOp((unsigned)-1, SrcOp1, SrcOp2));
1907       } else {
1908         ResOperands.push_back(ResOperand::getTiedOp(TiedOp, SrcOp1, SrcOp2));
1909         continue;
1910       }
1911     }
1912 
1913     // Handle all the suboperands for this operand.
1914     const std::string &OpName = OpInfo->Name;
1915     for (; AliasOpNo < LastOpNo &&
1916            CGA.ResultInstOperandIndex[AliasOpNo].first == i;
1917          ++AliasOpNo) {
1918       int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1919 
1920       // Find out what operand from the asmparser that this MCInst operand
1921       // comes from.
1922       switch (CGA.ResultOperands[AliasOpNo].Kind) {
1923       case CodeGenInstAlias::ResultOperand::K_Record: {
1924         StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1925         int SrcOperand = findAsmOperand(Name, SubIdx);
1926         if (SrcOperand == -1)
1927           PrintFatalError(TheDef->getLoc(),
1928                           "Instruction '" + TheDef->getName() +
1929                               "' has operand '" + OpName +
1930                               "' that doesn't appear in asm string!");
1931 
1932         // Add it to the operand references. If it is added a second time, the
1933         // record won't be updated and it will fail later on.
1934         OperandRefs.try_emplace(Name, SrcOperand);
1935 
1936         unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1937         ResOperands.push_back(
1938             ResOperand::getRenderedOp(SrcOperand, NumOperands));
1939         break;
1940       }
1941       case CodeGenInstAlias::ResultOperand::K_Imm: {
1942         int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1943         ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1944         break;
1945       }
1946       case CodeGenInstAlias::ResultOperand::K_Reg: {
1947         Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1948         ResOperands.push_back(ResOperand::getRegOp(Reg));
1949         break;
1950       }
1951       }
1952     }
1953   }
1954 
1955   // Check that operands are not repeated more times than is supported.
1956   for (auto &T : OperandRefs) {
1957     if (T.second != -1 && findAsmOperandNamed(T.first, T.second) != -1)
1958       PrintFatalError(TheDef->getLoc(),
1959                       "Operand '" + T.first + "' can never be matched");
1960   }
1961 }
1962 
1963 static unsigned
getConverterOperandID(const std::string & Name,SmallSetVector<CachedHashString,16> & Table,bool & IsNew)1964 getConverterOperandID(const std::string &Name,
1965                       SmallSetVector<CachedHashString, 16> &Table,
1966                       bool &IsNew) {
1967   IsNew = Table.insert(CachedHashString(Name));
1968 
1969   unsigned ID = IsNew ? Table.size() - 1 : find(Table, Name) - Table.begin();
1970 
1971   assert(ID < Table.size());
1972 
1973   return ID;
1974 }
1975 
1976 static unsigned
emitConvertFuncs(CodeGenTarget & Target,StringRef ClassName,std::vector<std::unique_ptr<MatchableInfo>> & Infos,bool HasMnemonicFirst,bool HasOptionalOperands,raw_ostream & OS)1977 emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1978                  std::vector<std::unique_ptr<MatchableInfo>> &Infos,
1979                  bool HasMnemonicFirst, bool HasOptionalOperands,
1980                  raw_ostream &OS) {
1981   SmallSetVector<CachedHashString, 16> OperandConversionKinds;
1982   SmallSetVector<CachedHashString, 16> InstructionConversionKinds;
1983   std::vector<std::vector<uint8_t>> ConversionTable;
1984   size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1985 
1986   // TargetOperandClass - This is the target's operand class, like X86Operand.
1987   std::string TargetOperandClass = Target.getName().str() + "Operand";
1988 
1989   // Write the convert function to a separate stream, so we can drop it after
1990   // the enum. We'll build up the conversion handlers for the individual
1991   // operand types opportunistically as we encounter them.
1992   std::string ConvertFnBody;
1993   raw_string_ostream CvtOS(ConvertFnBody);
1994   // Start the unified conversion function.
1995   if (HasOptionalOperands) {
1996     CvtOS << "void " << Target.getName() << ClassName << "::\n"
1997           << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1998           << "unsigned Opcode,\n"
1999           << "                const OperandVector &Operands,\n"
2000           << "                const SmallBitVector &OptionalOperandsMask,\n"
2001           << "                ArrayRef<unsigned> DefaultsOffset) {\n";
2002   } else {
2003     CvtOS << "void " << Target.getName() << ClassName << "::\n"
2004           << "convertToMCInst(unsigned Kind, MCInst &Inst, "
2005           << "unsigned Opcode,\n"
2006           << "                const OperandVector &Operands) {\n";
2007   }
2008   CvtOS << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n";
2009   CvtOS << "  const uint8_t *Converter = ConversionTable[Kind];\n";
2010   CvtOS << "  Inst.setOpcode(Opcode);\n";
2011   CvtOS << "  for (const uint8_t *p = Converter; *p; p += 2) {\n";
2012   if (HasOptionalOperands) {
2013     // When optional operands are involved, formal and actual operand indices
2014     // may differ. Map the former to the latter by subtracting the number of
2015     // absent optional operands.
2016     // FIXME: This is not an operand index in the CVT_Tied case
2017     CvtOS << "    unsigned OpIdx = *(p + 1) - DefaultsOffset[*(p + 1)];\n";
2018   } else {
2019     CvtOS << "    unsigned OpIdx = *(p + 1);\n";
2020   }
2021   CvtOS << "    switch (*p) {\n";
2022   CvtOS << "    default: llvm_unreachable(\"invalid conversion entry!\");\n";
2023   CvtOS << "    case CVT_Reg:\n";
2024   CvtOS << "      static_cast<" << TargetOperandClass
2025         << " &>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n";
2026   CvtOS << "      break;\n";
2027   CvtOS << "    case CVT_Tied: {\n";
2028   CvtOS << "      assert(*(p + 1) < (size_t)(std::end(TiedAsmOperandTable) -\n";
2029   CvtOS
2030       << "                              std::begin(TiedAsmOperandTable)) &&\n";
2031   CvtOS << "             \"Tied operand not found\");\n";
2032   CvtOS << "      unsigned TiedResOpnd = TiedAsmOperandTable[*(p + 1)][0];\n";
2033   CvtOS << "      if (TiedResOpnd != (uint8_t)-1)\n";
2034   CvtOS << "        Inst.addOperand(Inst.getOperand(TiedResOpnd));\n";
2035   CvtOS << "      break;\n";
2036   CvtOS << "    }\n";
2037 
2038   std::string OperandFnBody;
2039   raw_string_ostream OpOS(OperandFnBody);
2040   // Start the operand number lookup function.
2041   OpOS << "void " << Target.getName() << ClassName << "::\n"
2042        << "convertToMapAndConstraints(unsigned Kind,\n";
2043   OpOS.indent(27);
2044   OpOS << "const OperandVector &Operands) {\n"
2045        << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
2046        << "  unsigned NumMCOperands = 0;\n"
2047        << "  const uint8_t *Converter = ConversionTable[Kind];\n"
2048        << "  for (const uint8_t *p = Converter; *p; p += 2) {\n"
2049        << "    switch (*p) {\n"
2050        << "    default: llvm_unreachable(\"invalid conversion entry!\");\n"
2051        << "    case CVT_Reg:\n"
2052        << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2053        << "      Operands[*(p + 1)]->setConstraint(\"r\");\n"
2054        << "      ++NumMCOperands;\n"
2055        << "      break;\n"
2056        << "    case CVT_Tied:\n"
2057        << "      ++NumMCOperands;\n"
2058        << "      break;\n";
2059 
2060   // Pre-populate the operand conversion kinds with the standard always
2061   // available entries.
2062   OperandConversionKinds.insert(CachedHashString("CVT_Done"));
2063   OperandConversionKinds.insert(CachedHashString("CVT_Reg"));
2064   OperandConversionKinds.insert(CachedHashString("CVT_Tied"));
2065   enum { CVT_Done, CVT_Reg, CVT_Tied };
2066 
2067   // Map of e.g. <0, 2, 3> -> "Tie_0_2_3" enum label.
2068   std::map<std::tuple<uint8_t, uint8_t, uint8_t>, std::string>
2069       TiedOperandsEnumMap;
2070 
2071   for (auto &II : Infos) {
2072     // Check if we have a custom match function.
2073     StringRef AsmMatchConverter =
2074         II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
2075     if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) {
2076       std::string Signature = ("ConvertCustom_" + AsmMatchConverter).str();
2077       II->ConversionFnKind = Signature;
2078 
2079       // Check if we have already generated this signature.
2080       if (!InstructionConversionKinds.insert(CachedHashString(Signature)))
2081         continue;
2082 
2083       // Remember this converter for the kind enum.
2084       unsigned KindID = OperandConversionKinds.size();
2085       OperandConversionKinds.insert(
2086           CachedHashString("CVT_" + getEnumNameForToken(AsmMatchConverter)));
2087 
2088       // Add the converter row for this instruction.
2089       ConversionTable.emplace_back();
2090       ConversionTable.back().push_back(KindID);
2091       ConversionTable.back().push_back(CVT_Done);
2092 
2093       // Add the handler to the conversion driver function.
2094       CvtOS << "    case CVT_" << getEnumNameForToken(AsmMatchConverter)
2095             << ":\n"
2096             << "      " << AsmMatchConverter << "(Inst, Operands);\n"
2097             << "      break;\n";
2098 
2099       // FIXME: Handle the operand number lookup for custom match functions.
2100       continue;
2101     }
2102 
2103     // Build the conversion function signature.
2104     std::string Signature = "Convert";
2105 
2106     std::vector<uint8_t> ConversionRow;
2107 
2108     // Compute the convert enum and the case body.
2109     MaxRowLength = std::max(MaxRowLength, II->ResOperands.size() * 2 + 1);
2110 
2111     for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) {
2112       const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i];
2113 
2114       // Generate code to populate each result operand.
2115       switch (OpInfo.Kind) {
2116       case MatchableInfo::ResOperand::RenderAsmOperand: {
2117         // This comes from something we parsed.
2118         const MatchableInfo::AsmOperand &Op =
2119             II->AsmOperands[OpInfo.AsmOperandNum];
2120 
2121         // Registers are always converted the same, don't duplicate the
2122         // conversion function based on them.
2123         Signature += "__";
2124         std::string Class;
2125         Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
2126         Signature += Class;
2127         Signature += utostr(OpInfo.MINumOperands);
2128         Signature += "_" + itostr(OpInfo.AsmOperandNum);
2129 
2130         // Add the conversion kind, if necessary, and get the associated ID
2131         // the index of its entry in the vector).
2132         std::string Name =
2133             "CVT_" +
2134             (Op.Class->isRegisterClass() ? "Reg" : Op.Class->RenderMethod);
2135         if (Op.Class->IsOptional) {
2136           // For optional operands we must also care about DefaultMethod
2137           assert(HasOptionalOperands);
2138           Name += "_" + Op.Class->DefaultMethod;
2139         }
2140         Name = getEnumNameForToken(Name);
2141 
2142         bool IsNewConverter = false;
2143         unsigned ID =
2144             getConverterOperandID(Name, OperandConversionKinds, IsNewConverter);
2145 
2146         // Add the operand entry to the instruction kind conversion row.
2147         ConversionRow.push_back(ID);
2148         ConversionRow.push_back(OpInfo.AsmOperandNum + HasMnemonicFirst);
2149 
2150         if (!IsNewConverter)
2151           break;
2152 
2153         // This is a new operand kind. Add a handler for it to the
2154         // converter driver.
2155         CvtOS << "    case " << Name << ":\n";
2156         if (Op.Class->IsOptional) {
2157           // If optional operand is not present in actual instruction then we
2158           // should call its DefaultMethod before RenderMethod
2159           assert(HasOptionalOperands);
2160           CvtOS << "      if (OptionalOperandsMask[*(p + 1) - 1]) {\n"
2161                 << "        " << Op.Class->DefaultMethod << "()"
2162                 << "->" << Op.Class->RenderMethod << "(Inst, "
2163                 << OpInfo.MINumOperands << ");\n"
2164                 << "      } else {\n"
2165                 << "        static_cast<" << TargetOperandClass
2166                 << " &>(*Operands[OpIdx])." << Op.Class->RenderMethod
2167                 << "(Inst, " << OpInfo.MINumOperands << ");\n"
2168                 << "      }\n";
2169         } else {
2170           CvtOS << "      static_cast<" << TargetOperandClass
2171                 << " &>(*Operands[OpIdx])." << Op.Class->RenderMethod
2172                 << "(Inst, " << OpInfo.MINumOperands << ");\n";
2173         }
2174         CvtOS << "      break;\n";
2175 
2176         // Add a handler for the operand number lookup.
2177         OpOS << "    case " << Name << ":\n"
2178              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
2179 
2180         if (Op.Class->isRegisterClass())
2181           OpOS << "      Operands[*(p + 1)]->setConstraint(\"r\");\n";
2182         else
2183           OpOS << "      Operands[*(p + 1)]->setConstraint(\"m\");\n";
2184         OpOS << "      NumMCOperands += " << OpInfo.MINumOperands << ";\n"
2185              << "      break;\n";
2186         break;
2187       }
2188       case MatchableInfo::ResOperand::TiedOperand: {
2189         // If this operand is tied to a previous one, just copy the MCInst
2190         // operand from the earlier one.We can only tie single MCOperand values.
2191         assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
2192         uint8_t TiedOp = OpInfo.TiedOperands.ResOpnd;
2193         uint8_t SrcOp1 = OpInfo.TiedOperands.SrcOpnd1Idx + HasMnemonicFirst;
2194         uint8_t SrcOp2 = OpInfo.TiedOperands.SrcOpnd2Idx + HasMnemonicFirst;
2195         assert((i > TiedOp || TiedOp == (uint8_t)-1) &&
2196                "Tied operand precedes its target!");
2197         auto TiedTupleName = std::string("Tie") + utostr(TiedOp) + '_' +
2198                              utostr(SrcOp1) + '_' + utostr(SrcOp2);
2199         Signature += "__" + TiedTupleName;
2200         ConversionRow.push_back(CVT_Tied);
2201         ConversionRow.push_back(TiedOp);
2202         ConversionRow.push_back(SrcOp1);
2203         ConversionRow.push_back(SrcOp2);
2204 
2205         // Also create an 'enum' for this combination of tied operands.
2206         auto Key = std::tuple(TiedOp, SrcOp1, SrcOp2);
2207         TiedOperandsEnumMap.emplace(Key, TiedTupleName);
2208         break;
2209       }
2210       case MatchableInfo::ResOperand::ImmOperand: {
2211         int64_t Val = OpInfo.ImmVal;
2212         std::string Ty = "imm_" + itostr(Val);
2213         Ty = getEnumNameForToken(Ty);
2214         Signature += "__" + Ty;
2215 
2216         std::string Name = "CVT_" + Ty;
2217         bool IsNewConverter = false;
2218         unsigned ID =
2219             getConverterOperandID(Name, OperandConversionKinds, IsNewConverter);
2220         // Add the operand entry to the instruction kind conversion row.
2221         ConversionRow.push_back(ID);
2222         ConversionRow.push_back(0);
2223 
2224         if (!IsNewConverter)
2225           break;
2226 
2227         CvtOS << "    case " << Name << ":\n"
2228               << "      Inst.addOperand(MCOperand::createImm(" << Val << "));\n"
2229               << "      break;\n";
2230 
2231         OpOS << "    case " << Name << ":\n"
2232              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2233              << "      Operands[*(p + 1)]->setConstraint(\"\");\n"
2234              << "      ++NumMCOperands;\n"
2235              << "      break;\n";
2236         break;
2237       }
2238       case MatchableInfo::ResOperand::RegOperand: {
2239         std::string Reg, Name;
2240         if (!OpInfo.Register) {
2241           Name = "reg0";
2242           Reg = "0";
2243         } else {
2244           Reg = getQualifiedName(OpInfo.Register);
2245           Name = "reg" + OpInfo.Register->getName().str();
2246         }
2247         Signature += "__" + Name;
2248         Name = "CVT_" + Name;
2249         bool IsNewConverter = false;
2250         unsigned ID =
2251             getConverterOperandID(Name, OperandConversionKinds, IsNewConverter);
2252         // Add the operand entry to the instruction kind conversion row.
2253         ConversionRow.push_back(ID);
2254         ConversionRow.push_back(0);
2255 
2256         if (!IsNewConverter)
2257           break;
2258         CvtOS << "    case " << Name << ":\n"
2259               << "      Inst.addOperand(MCOperand::createReg(" << Reg << "));\n"
2260               << "      break;\n";
2261 
2262         OpOS << "    case " << Name << ":\n"
2263              << "      Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
2264              << "      Operands[*(p + 1)]->setConstraint(\"m\");\n"
2265              << "      ++NumMCOperands;\n"
2266              << "      break;\n";
2267       }
2268       }
2269     }
2270 
2271     // If there were no operands, add to the signature to that effect
2272     if (Signature == "Convert")
2273       Signature += "_NoOperands";
2274 
2275     II->ConversionFnKind = Signature;
2276 
2277     // Save the signature. If we already have it, don't add a new row
2278     // to the table.
2279     if (!InstructionConversionKinds.insert(CachedHashString(Signature)))
2280       continue;
2281 
2282     // Add the row to the table.
2283     ConversionTable.push_back(std::move(ConversionRow));
2284   }
2285 
2286   // Finish up the converter driver function.
2287   CvtOS << "    }\n  }\n}\n\n";
2288 
2289   // Finish up the operand number lookup function.
2290   OpOS << "    }\n  }\n}\n\n";
2291 
2292   // Output a static table for tied operands.
2293   if (TiedOperandsEnumMap.size()) {
2294     // The number of tied operand combinations will be small in practice,
2295     // but just add the assert to be sure.
2296     assert(TiedOperandsEnumMap.size() <= 254 &&
2297            "Too many tied-operand combinations to reference with "
2298            "an 8bit offset from the conversion table, where index "
2299            "'255' is reserved as operand not to be copied.");
2300 
2301     OS << "enum {\n";
2302     for (auto &KV : TiedOperandsEnumMap) {
2303       OS << "  " << KV.second << ",\n";
2304     }
2305     OS << "};\n\n";
2306 
2307     OS << "static const uint8_t TiedAsmOperandTable[][3] = {\n";
2308     for (auto &KV : TiedOperandsEnumMap) {
2309       OS << "  /* " << KV.second << " */ { " << utostr(std::get<0>(KV.first))
2310          << ", " << utostr(std::get<1>(KV.first)) << ", "
2311          << utostr(std::get<2>(KV.first)) << " },\n";
2312     }
2313     OS << "};\n\n";
2314   } else
2315     OS << "static const uint8_t TiedAsmOperandTable[][3] = "
2316           "{ /* empty  */ {0, 0, 0} };\n\n";
2317 
2318   OS << "namespace {\n";
2319 
2320   // Output the operand conversion kind enum.
2321   OS << "enum OperatorConversionKind {\n";
2322   for (const auto &Converter : OperandConversionKinds)
2323     OS << "  " << Converter << ",\n";
2324   OS << "  CVT_NUM_CONVERTERS\n";
2325   OS << "};\n\n";
2326 
2327   // Output the instruction conversion kind enum.
2328   OS << "enum InstructionConversionKind {\n";
2329   for (const auto &Signature : InstructionConversionKinds)
2330     OS << "  " << Signature << ",\n";
2331   OS << "  CVT_NUM_SIGNATURES\n";
2332   OS << "};\n\n";
2333 
2334   OS << "} // end anonymous namespace\n\n";
2335 
2336   // Output the conversion table.
2337   OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
2338      << MaxRowLength << "] = {\n";
2339 
2340   for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
2341     assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
2342     OS << "  // " << InstructionConversionKinds[Row] << "\n";
2343     OS << "  { ";
2344     for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2) {
2345       OS << OperandConversionKinds[ConversionTable[Row][i]] << ", ";
2346       if (OperandConversionKinds[ConversionTable[Row][i]] !=
2347           CachedHashString("CVT_Tied")) {
2348         OS << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
2349         continue;
2350       }
2351 
2352       // For a tied operand, emit a reference to the TiedAsmOperandTable
2353       // that contains the operand to copy, and the parsed operands to
2354       // check for their tied constraints.
2355       auto Key = std::tuple((uint8_t)ConversionTable[Row][i + 1],
2356                             (uint8_t)ConversionTable[Row][i + 2],
2357                             (uint8_t)ConversionTable[Row][i + 3]);
2358       auto TiedOpndEnum = TiedOperandsEnumMap.find(Key);
2359       assert(TiedOpndEnum != TiedOperandsEnumMap.end() &&
2360              "No record for tied operand pair");
2361       OS << TiedOpndEnum->second << ", ";
2362       i += 2;
2363     }
2364     OS << "CVT_Done },\n";
2365   }
2366 
2367   OS << "};\n\n";
2368 
2369   // Spit out the conversion driver function.
2370   OS << ConvertFnBody;
2371 
2372   // Spit out the operand number lookup function.
2373   OS << OperandFnBody;
2374 
2375   return ConversionTable.size();
2376 }
2377 
2378 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
emitMatchClassEnumeration(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2379 static void emitMatchClassEnumeration(CodeGenTarget &Target,
2380                                       std::forward_list<ClassInfo> &Infos,
2381                                       raw_ostream &OS) {
2382   OS << "namespace {\n\n";
2383 
2384   OS << "/// MatchClassKind - The kinds of classes which participate in\n"
2385      << "/// instruction matching.\n";
2386   OS << "enum MatchClassKind {\n";
2387   OS << "  InvalidMatchClass = 0,\n";
2388   OS << "  OptionalMatchClass = 1,\n";
2389   ClassInfo::ClassInfoKind LastKind = ClassInfo::Token;
2390   StringRef LastName = "OptionalMatchClass";
2391   for (const auto &CI : Infos) {
2392     if (LastKind == ClassInfo::Token && CI.Kind != ClassInfo::Token) {
2393       OS << "  MCK_LAST_TOKEN = " << LastName << ",\n";
2394     } else if (LastKind < ClassInfo::UserClass0 &&
2395                CI.Kind >= ClassInfo::UserClass0) {
2396       OS << "  MCK_LAST_REGISTER = " << LastName << ",\n";
2397     }
2398     LastKind = (ClassInfo::ClassInfoKind)CI.Kind;
2399     LastName = CI.Name;
2400 
2401     OS << "  " << CI.Name << ", // ";
2402     if (CI.Kind == ClassInfo::Token) {
2403       OS << "'" << CI.ValueName << "'\n";
2404     } else if (CI.isRegisterClass()) {
2405       if (!CI.ValueName.empty())
2406         OS << "register class '" << CI.ValueName << "'\n";
2407       else
2408         OS << "derived register class\n";
2409     } else {
2410       OS << "user defined class '" << CI.ValueName << "'\n";
2411     }
2412   }
2413   OS << "  NumMatchClassKinds\n";
2414   OS << "};\n\n";
2415 
2416   OS << "} // end anonymous namespace\n\n";
2417 }
2418 
2419 /// emitMatchClassDiagStrings - Emit a function to get the diagnostic text to be
2420 /// used when an assembly operand does not match the expected operand class.
emitOperandMatchErrorDiagStrings(AsmMatcherInfo & Info,raw_ostream & OS)2421 static void emitOperandMatchErrorDiagStrings(AsmMatcherInfo &Info,
2422                                              raw_ostream &OS) {
2423   // If the target does not use DiagnosticString for any operands, don't emit
2424   // an unused function.
2425   if (llvm::all_of(Info.Classes, [](const ClassInfo &CI) {
2426         return CI.DiagnosticString.empty();
2427       }))
2428     return;
2429 
2430   OS << "static const char *getMatchKindDiag(" << Info.Target.getName()
2431      << "AsmParser::" << Info.Target.getName()
2432      << "MatchResultTy MatchResult) {\n";
2433   OS << "  switch (MatchResult) {\n";
2434 
2435   for (const auto &CI : Info.Classes) {
2436     if (!CI.DiagnosticString.empty()) {
2437       assert(!CI.DiagnosticType.empty() &&
2438              "DiagnosticString set without DiagnosticType");
2439       OS << "  case " << Info.Target.getName() << "AsmParser::Match_"
2440          << CI.DiagnosticType << ":\n";
2441       OS << "    return \"" << CI.DiagnosticString << "\";\n";
2442     }
2443   }
2444 
2445   OS << "  default:\n";
2446   OS << "    return nullptr;\n";
2447 
2448   OS << "  }\n";
2449   OS << "}\n\n";
2450 }
2451 
emitRegisterMatchErrorFunc(AsmMatcherInfo & Info,raw_ostream & OS)2452 static void emitRegisterMatchErrorFunc(AsmMatcherInfo &Info, raw_ostream &OS) {
2453   OS << "static unsigned getDiagKindFromRegisterClass(MatchClassKind "
2454         "RegisterClass) {\n";
2455   if (none_of(Info.Classes, [](const ClassInfo &CI) {
2456         return CI.isRegisterClass() && !CI.DiagnosticType.empty();
2457       })) {
2458     OS << "  return MCTargetAsmParser::Match_InvalidOperand;\n";
2459   } else {
2460     OS << "  switch (RegisterClass) {\n";
2461     for (const auto &CI : Info.Classes) {
2462       if (CI.isRegisterClass() && !CI.DiagnosticType.empty()) {
2463         OS << "  case " << CI.Name << ":\n";
2464         OS << "    return " << Info.Target.getName() << "AsmParser::Match_"
2465            << CI.DiagnosticType << ";\n";
2466       }
2467     }
2468 
2469     OS << "  default:\n";
2470     OS << "    return MCTargetAsmParser::Match_InvalidOperand;\n";
2471 
2472     OS << "  }\n";
2473   }
2474   OS << "}\n\n";
2475 }
2476 
2477 /// emitValidateOperandClass - Emit the function to validate an operand class.
emitValidateOperandClass(AsmMatcherInfo & Info,raw_ostream & OS)2478 static void emitValidateOperandClass(AsmMatcherInfo &Info, raw_ostream &OS) {
2479   OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
2480      << "MatchClassKind Kind) {\n";
2481   OS << "  " << Info.Target.getName() << "Operand &Operand = ("
2482      << Info.Target.getName() << "Operand &)GOp;\n";
2483 
2484   // The InvalidMatchClass is not to match any operand.
2485   OS << "  if (Kind == InvalidMatchClass)\n";
2486   OS << "    return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2487 
2488   // Check for Token operands first.
2489   // FIXME: Use a more specific diagnostic type.
2490   OS << "  if (Operand.isToken() && Kind <= MCK_LAST_TOKEN)\n";
2491   OS << "    return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2492      << "             MCTargetAsmParser::Match_Success :\n"
2493      << "             MCTargetAsmParser::Match_InvalidOperand;\n\n";
2494 
2495   // Check the user classes. We don't care what order since we're only
2496   // actually matching against one of them.
2497   OS << "  switch (Kind) {\n"
2498         "  default: break;\n";
2499   for (const auto &CI : Info.Classes) {
2500     if (!CI.isUserClass())
2501       continue;
2502 
2503     OS << "  // '" << CI.ClassName << "' class\n";
2504     OS << "  case " << CI.Name << ": {\n";
2505     OS << "    DiagnosticPredicate DP(Operand." << CI.PredicateMethod
2506        << "());\n";
2507     OS << "    if (DP.isMatch())\n";
2508     OS << "      return MCTargetAsmParser::Match_Success;\n";
2509     if (!CI.DiagnosticType.empty()) {
2510       OS << "    if (DP.isNearMatch())\n";
2511       OS << "      return " << Info.Target.getName() << "AsmParser::Match_"
2512          << CI.DiagnosticType << ";\n";
2513       OS << "    break;\n";
2514     } else
2515       OS << "    break;\n";
2516     OS << "    }\n";
2517   }
2518   OS << "  } // end switch (Kind)\n\n";
2519 
2520   // Check for register operands, including sub-classes.
2521   OS << "  if (Operand.isReg()) {\n";
2522   OS << "    MatchClassKind OpKind;\n";
2523   OS << "    switch (Operand.getReg().id()) {\n";
2524   OS << "    default: OpKind = InvalidMatchClass; break;\n";
2525   for (const auto &RC : Info.RegisterClasses)
2526     OS << "    case " << RC.first->getValueAsString("Namespace")
2527        << "::" << RC.first->getName() << ": OpKind = " << RC.second->Name
2528        << "; break;\n";
2529   OS << "    }\n";
2530   OS << "    return isSubclass(OpKind, Kind) ? "
2531      << "(unsigned)MCTargetAsmParser::Match_Success :\n                     "
2532      << "                 getDiagKindFromRegisterClass(Kind);\n  }\n\n";
2533 
2534   // Expected operand is a register, but actual is not.
2535   OS << "  if (Kind > MCK_LAST_TOKEN && Kind <= MCK_LAST_REGISTER)\n";
2536   OS << "    return getDiagKindFromRegisterClass(Kind);\n\n";
2537 
2538   // Generic fallthrough match failure case for operands that don't have
2539   // specialized diagnostic types.
2540   OS << "  return MCTargetAsmParser::Match_InvalidOperand;\n";
2541   OS << "}\n\n";
2542 }
2543 
2544 /// emitIsSubclass - Emit the subclass predicate function.
emitIsSubclass(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2545 static void emitIsSubclass(CodeGenTarget &Target,
2546                            std::forward_list<ClassInfo> &Infos,
2547                            raw_ostream &OS) {
2548   OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2549   OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2550   OS << "  if (A == B)\n";
2551   OS << "    return true;\n\n";
2552 
2553   bool EmittedSwitch = false;
2554   for (const auto &A : Infos) {
2555     std::vector<StringRef> SuperClasses;
2556     if (A.IsOptional)
2557       SuperClasses.push_back("OptionalMatchClass");
2558     for (const auto &B : Infos) {
2559       if (&A != &B && A.isSubsetOf(B))
2560         SuperClasses.push_back(B.Name);
2561     }
2562 
2563     if (SuperClasses.empty())
2564       continue;
2565 
2566     // If this is the first SuperClass, emit the switch header.
2567     if (!EmittedSwitch) {
2568       OS << "  switch (A) {\n";
2569       OS << "  default:\n";
2570       OS << "    return false;\n";
2571       EmittedSwitch = true;
2572     }
2573 
2574     OS << "\n  case " << A.Name << ":\n";
2575 
2576     if (SuperClasses.size() == 1) {
2577       OS << "    return B == " << SuperClasses.back() << ";\n";
2578       continue;
2579     }
2580 
2581     if (!SuperClasses.empty()) {
2582       OS << "    switch (B) {\n";
2583       OS << "    default: return false;\n";
2584       for (StringRef SC : SuperClasses)
2585         OS << "    case " << SC << ": return true;\n";
2586       OS << "    }\n";
2587     } else {
2588       // No case statement to emit
2589       OS << "    return false;\n";
2590     }
2591   }
2592 
2593   // If there were case statements emitted into the string stream write the
2594   // default.
2595   if (EmittedSwitch)
2596     OS << "  }\n";
2597   else
2598     OS << "  return false;\n";
2599 
2600   OS << "}\n\n";
2601 }
2602 
2603 /// emitMatchTokenString - Emit the function to match a token string to the
2604 /// appropriate match class value.
emitMatchTokenString(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2605 static void emitMatchTokenString(CodeGenTarget &Target,
2606                                  std::forward_list<ClassInfo> &Infos,
2607                                  raw_ostream &OS) {
2608   // Construct the match list.
2609   std::vector<StringMatcher::StringPair> Matches;
2610   for (const auto &CI : Infos) {
2611     if (CI.Kind == ClassInfo::Token)
2612       Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";");
2613   }
2614 
2615   OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2616 
2617   StringMatcher("Name", Matches, OS).Emit();
2618 
2619   OS << "  return InvalidMatchClass;\n";
2620   OS << "}\n\n";
2621 }
2622 
2623 /// emitMatchRegisterName - Emit the function to match a string to the target
2624 /// specific register enum.
emitMatchRegisterName(CodeGenTarget & Target,Record * AsmParser,raw_ostream & OS)2625 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2626                                   raw_ostream &OS) {
2627   // Construct the match list.
2628   std::vector<StringMatcher::StringPair> Matches;
2629   const auto &Regs = Target.getRegBank().getRegisters();
2630   std::string Namespace =
2631       Regs.front().TheDef->getValueAsString("Namespace").str();
2632   for (const CodeGenRegister &Reg : Regs) {
2633     StringRef AsmName = Reg.TheDef->getValueAsString("AsmName");
2634     if (AsmName.empty())
2635       continue;
2636 
2637     Matches.emplace_back(AsmName.str(), "return " + Namespace +
2638                                             "::" + Reg.getName().str() + ';');
2639   }
2640 
2641   OS << "static MCRegister MatchRegisterName(StringRef Name) {\n";
2642 
2643   bool IgnoreDuplicates =
2644       AsmParser->getValueAsBit("AllowDuplicateRegisterNames");
2645   StringMatcher("Name", Matches, OS).Emit(0, IgnoreDuplicates);
2646 
2647   OS << "  return " << Namespace << "::NoRegister;\n";
2648   OS << "}\n\n";
2649 }
2650 
2651 /// Emit the function to match a string to the target
2652 /// specific register enum.
emitMatchRegisterAltName(CodeGenTarget & Target,Record * AsmParser,raw_ostream & OS)2653 static void emitMatchRegisterAltName(CodeGenTarget &Target, Record *AsmParser,
2654                                      raw_ostream &OS) {
2655   // Construct the match list.
2656   std::vector<StringMatcher::StringPair> Matches;
2657   const auto &Regs = Target.getRegBank().getRegisters();
2658   std::string Namespace =
2659       Regs.front().TheDef->getValueAsString("Namespace").str();
2660   for (const CodeGenRegister &Reg : Regs) {
2661 
2662     auto AltNames = Reg.TheDef->getValueAsListOfStrings("AltNames");
2663 
2664     for (auto AltName : AltNames) {
2665       AltName = StringRef(AltName).trim();
2666 
2667       // don't handle empty alternative names
2668       if (AltName.empty())
2669         continue;
2670 
2671       Matches.emplace_back(AltName.str(), "return " + Namespace +
2672                                               "::" + Reg.getName().str() + ';');
2673     }
2674   }
2675 
2676   OS << "static MCRegister MatchRegisterAltName(StringRef Name) {\n";
2677 
2678   bool IgnoreDuplicates =
2679       AsmParser->getValueAsBit("AllowDuplicateRegisterNames");
2680   StringMatcher("Name", Matches, OS).Emit(0, IgnoreDuplicates);
2681 
2682   OS << "  return " << Namespace << "::NoRegister;\n";
2683   OS << "}\n\n";
2684 }
2685 
2686 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
emitOperandDiagnosticTypes(AsmMatcherInfo & Info,raw_ostream & OS)2687 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2688   // Get the set of diagnostic types from all of the operand classes.
2689   std::set<StringRef> Types;
2690   for (const auto &OpClassEntry : Info.AsmOperandClasses) {
2691     if (!OpClassEntry.second->DiagnosticType.empty())
2692       Types.insert(OpClassEntry.second->DiagnosticType);
2693   }
2694   for (const auto &OpClassEntry : Info.RegisterClassClasses) {
2695     if (!OpClassEntry.second->DiagnosticType.empty())
2696       Types.insert(OpClassEntry.second->DiagnosticType);
2697   }
2698 
2699   if (Types.empty())
2700     return;
2701 
2702   // Now emit the enum entries.
2703   for (StringRef Type : Types)
2704     OS << "  Match_" << Type << ",\n";
2705   OS << "  END_OPERAND_DIAGNOSTIC_TYPES\n";
2706 }
2707 
2708 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2709 /// user-level name for a subtarget feature.
emitGetSubtargetFeatureName(AsmMatcherInfo & Info,raw_ostream & OS)2710 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2711   OS << "// User-level names for subtarget features that participate in\n"
2712      << "// instruction matching.\n"
2713      << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2714   if (!Info.SubtargetFeatures.empty()) {
2715     OS << "  switch(Val) {\n";
2716     for (const auto &SF : Info.SubtargetFeatures) {
2717       const SubtargetFeatureInfo &SFI = SF.second;
2718       // FIXME: Totally just a placeholder name to get the algorithm working.
2719       OS << "  case " << SFI.getEnumBitName() << ": return \""
2720          << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2721     }
2722     OS << "  default: return \"(unknown)\";\n";
2723     OS << "  }\n";
2724   } else {
2725     // Nothing to emit, so skip the switch
2726     OS << "  return \"(unknown)\";\n";
2727   }
2728   OS << "}\n\n";
2729 }
2730 
GetAliasRequiredFeatures(Record * R,const AsmMatcherInfo & Info)2731 static std::string GetAliasRequiredFeatures(Record *R,
2732                                             const AsmMatcherInfo &Info) {
2733   std::vector<Record *> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2734   std::string Result;
2735 
2736   if (ReqFeatures.empty())
2737     return Result;
2738 
2739   for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2740     const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2741 
2742     if (!F)
2743       PrintFatalError(R->getLoc(),
2744                       "Predicate '" + ReqFeatures[i]->getName() +
2745                           "' is not marked as an AssemblerPredicate!");
2746 
2747     if (i)
2748       Result += " && ";
2749 
2750     Result += "Features.test(" + F->getEnumBitName() + ')';
2751   }
2752 
2753   return Result;
2754 }
2755 
2756 static void
emitMnemonicAliasVariant(raw_ostream & OS,const AsmMatcherInfo & Info,std::vector<Record * > & Aliases,unsigned Indent=0,StringRef AsmParserVariantName=StringRef ())2757 emitMnemonicAliasVariant(raw_ostream &OS, const AsmMatcherInfo &Info,
2758                          std::vector<Record *> &Aliases, unsigned Indent = 0,
2759                          StringRef AsmParserVariantName = StringRef()) {
2760   // Keep track of all the aliases from a mnemonic.  Use an std::map so that the
2761   // iteration order of the map is stable.
2762   std::map<std::string, std::vector<Record *>> AliasesFromMnemonic;
2763 
2764   for (Record *R : Aliases) {
2765     // FIXME: Allow AssemblerVariantName to be a comma separated list.
2766     StringRef AsmVariantName = R->getValueAsString("AsmVariantName");
2767     if (AsmVariantName != AsmParserVariantName)
2768       continue;
2769     AliasesFromMnemonic[R->getValueAsString("FromMnemonic").lower()].push_back(
2770         R);
2771   }
2772   if (AliasesFromMnemonic.empty())
2773     return;
2774 
2775   // Process each alias a "from" mnemonic at a time, building the code executed
2776   // by the string remapper.
2777   std::vector<StringMatcher::StringPair> Cases;
2778   for (const auto &AliasEntry : AliasesFromMnemonic) {
2779     const std::vector<Record *> &ToVec = AliasEntry.second;
2780 
2781     // Loop through each alias and emit code that handles each case.  If there
2782     // are two instructions without predicates, emit an error.  If there is one,
2783     // emit it last.
2784     std::string MatchCode;
2785     int AliasWithNoPredicate = -1;
2786 
2787     for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2788       Record *R = ToVec[i];
2789       std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2790 
2791       // If this unconditionally matches, remember it for later and diagnose
2792       // duplicates.
2793       if (FeatureMask.empty()) {
2794         if (AliasWithNoPredicate != -1 &&
2795             R->getValueAsString("ToMnemonic") !=
2796                 ToVec[AliasWithNoPredicate]->getValueAsString("ToMnemonic")) {
2797           // We can't have two different aliases from the same mnemonic with no
2798           // predicate.
2799           PrintError(
2800               ToVec[AliasWithNoPredicate]->getLoc(),
2801               "two different MnemonicAliases with the same 'from' mnemonic!");
2802           PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2803         }
2804 
2805         AliasWithNoPredicate = i;
2806         continue;
2807       }
2808       if (R->getValueAsString("ToMnemonic") == AliasEntry.first)
2809         PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2810 
2811       if (!MatchCode.empty())
2812         MatchCode += "else ";
2813       MatchCode += "if (" + FeatureMask + ")\n";
2814       MatchCode += "  Mnemonic = \"";
2815       MatchCode += R->getValueAsString("ToMnemonic").lower();
2816       MatchCode += "\";\n";
2817     }
2818 
2819     if (AliasWithNoPredicate != -1) {
2820       Record *R = ToVec[AliasWithNoPredicate];
2821       if (!MatchCode.empty())
2822         MatchCode += "else\n  ";
2823       MatchCode += "Mnemonic = \"";
2824       MatchCode += R->getValueAsString("ToMnemonic").lower();
2825       MatchCode += "\";\n";
2826     }
2827 
2828     MatchCode += "return;";
2829 
2830     Cases.push_back(std::pair(AliasEntry.first, MatchCode));
2831   }
2832   StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2833 }
2834 
2835 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2836 /// emit a function for them and return true, otherwise return false.
emitMnemonicAliases(raw_ostream & OS,const AsmMatcherInfo & Info,CodeGenTarget & Target)2837 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2838                                 CodeGenTarget &Target) {
2839   // Ignore aliases when match-prefix is set.
2840   if (!MatchPrefix.empty())
2841     return false;
2842 
2843   std::vector<Record *> Aliases =
2844       Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2845   if (Aliases.empty())
2846     return false;
2847 
2848   OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2849         "const FeatureBitset &Features, unsigned VariantID) {\n";
2850   OS << "  switch (VariantID) {\n";
2851   unsigned VariantCount = Target.getAsmParserVariantCount();
2852   for (unsigned VC = 0; VC != VariantCount; ++VC) {
2853     Record *AsmVariant = Target.getAsmParserVariant(VC);
2854     int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2855     StringRef AsmParserVariantName = AsmVariant->getValueAsString("Name");
2856     OS << "  case " << AsmParserVariantNo << ":\n";
2857     emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2858                              AsmParserVariantName);
2859     OS << "    break;\n";
2860   }
2861   OS << "  }\n";
2862 
2863   // Emit aliases that apply to all variants.
2864   emitMnemonicAliasVariant(OS, Info, Aliases);
2865 
2866   OS << "}\n\n";
2867 
2868   return true;
2869 }
2870 
2871 static void
emitCustomOperandParsing(raw_ostream & OS,CodeGenTarget & Target,const AsmMatcherInfo & Info,StringRef ClassName,StringToOffsetTable & StringTable,unsigned MaxMnemonicIndex,unsigned MaxFeaturesIndex,bool HasMnemonicFirst,const Record & AsmParser)2872 emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2873                          const AsmMatcherInfo &Info, StringRef ClassName,
2874                          StringToOffsetTable &StringTable,
2875                          unsigned MaxMnemonicIndex, unsigned MaxFeaturesIndex,
2876                          bool HasMnemonicFirst, const Record &AsmParser) {
2877   unsigned MaxMask = 0;
2878   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2879     MaxMask |= OMI.OperandMask;
2880   }
2881 
2882   // Emit the static custom operand parsing table;
2883   OS << "namespace {\n";
2884   OS << "  struct OperandMatchEntry {\n";
2885   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex) << " Mnemonic;\n";
2886   OS << "    " << getMinimalTypeForRange(MaxMask) << " OperandMask;\n";
2887   OS << "    "
2888      << getMinimalTypeForRange(
2889             std::distance(Info.Classes.begin(), Info.Classes.end()) +
2890             2 /* Include 'InvalidMatchClass' and 'OptionalMatchClass' */)
2891      << " Class;\n";
2892   OS << "    " << getMinimalTypeForRange(MaxFeaturesIndex)
2893      << " RequiredFeaturesIdx;\n\n";
2894   OS << "    StringRef getMnemonic() const {\n";
2895   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
2896   OS << "                       MnemonicTable[Mnemonic]);\n";
2897   OS << "    }\n";
2898   OS << "  };\n\n";
2899 
2900   OS << "  // Predicate for searching for an opcode.\n";
2901   OS << "  struct LessOpcodeOperand {\n";
2902   OS << "    bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2903   OS << "      return LHS.getMnemonic()  < RHS;\n";
2904   OS << "    }\n";
2905   OS << "    bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2906   OS << "      return LHS < RHS.getMnemonic();\n";
2907   OS << "    }\n";
2908   OS << "    bool operator()(const OperandMatchEntry &LHS,";
2909   OS << " const OperandMatchEntry &RHS) {\n";
2910   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
2911   OS << "    }\n";
2912   OS << "  };\n";
2913 
2914   OS << "} // end anonymous namespace\n\n";
2915 
2916   OS << "static const OperandMatchEntry OperandMatchTable["
2917      << Info.OperandMatchInfo.size() << "] = {\n";
2918 
2919   OS << "  /* Operand List Mnemonic, Mask, Operand Class, Features */\n";
2920   for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) {
2921     const MatchableInfo &II = *OMI.MI;
2922 
2923     OS << "  { ";
2924 
2925     // Store a pascal-style length byte in the mnemonic.
2926     std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.lower();
2927     OS << StringTable.GetOrAddStringOffset(LenMnemonic, false) << " /* "
2928        << II.Mnemonic << " */, ";
2929 
2930     OS << OMI.OperandMask;
2931     OS << " /* ";
2932     ListSeparator LS;
2933     for (int i = 0, e = 31; i != e; ++i)
2934       if (OMI.OperandMask & (1 << i))
2935         OS << LS << i;
2936     OS << " */, ";
2937 
2938     OS << OMI.CI->Name;
2939 
2940     // Write the required features mask.
2941     OS << ", AMFBS";
2942     if (II.RequiredFeatures.empty())
2943       OS << "_None";
2944     else
2945       for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i)
2946         OS << '_' << II.RequiredFeatures[i]->TheDef->getName();
2947 
2948     OS << " },\n";
2949   }
2950   OS << "};\n\n";
2951 
2952   // Emit the operand class switch to call the correct custom parser for
2953   // the found operand class.
2954   OS << "ParseStatus " << Target.getName() << ClassName << "::\n"
2955      << "tryCustomParseOperand(OperandVector"
2956      << " &Operands,\n                      unsigned MCK) {\n\n"
2957      << "  switch(MCK) {\n";
2958 
2959   for (const auto &CI : Info.Classes) {
2960     if (CI.ParserMethod.empty())
2961       continue;
2962     OS << "  case " << CI.Name << ":\n"
2963        << "    return " << CI.ParserMethod << "(Operands);\n";
2964   }
2965 
2966   OS << "  default:\n";
2967   OS << "    return ParseStatus::NoMatch;\n";
2968   OS << "  }\n";
2969   OS << "  return ParseStatus::NoMatch;\n";
2970   OS << "}\n\n";
2971 
2972   // Emit the static custom operand parser. This code is very similar with
2973   // the other matcher. Also use MatchResultTy here just in case we go for
2974   // a better error handling.
2975   OS << "ParseStatus " << Target.getName() << ClassName << "::\n"
2976      << "MatchOperandParserImpl(OperandVector"
2977      << " &Operands,\n                       StringRef Mnemonic,\n"
2978      << "                       bool ParseForAllFeatures) {\n";
2979 
2980   // Emit code to get the available features.
2981   OS << "  // Get the current feature set.\n";
2982   OS << "  const FeatureBitset &AvailableFeatures = "
2983         "getAvailableFeatures();\n\n";
2984 
2985   OS << "  // Get the next operand index.\n";
2986   OS << "  unsigned NextOpNum = Operands.size()"
2987      << (HasMnemonicFirst ? " - 1" : "") << ";\n";
2988 
2989   // Emit code to search the table.
2990   OS << "  // Search the table.\n";
2991   if (HasMnemonicFirst) {
2992     OS << "  auto MnemonicRange =\n";
2993     OS << "    std::equal_range(std::begin(OperandMatchTable), "
2994           "std::end(OperandMatchTable),\n";
2995     OS << "                     Mnemonic, LessOpcodeOperand());\n\n";
2996   } else {
2997     OS << "  auto MnemonicRange = std::pair(std::begin(OperandMatchTable),"
2998           " std::end(OperandMatchTable));\n";
2999     OS << "  if (!Mnemonic.empty())\n";
3000     OS << "    MnemonicRange =\n";
3001     OS << "      std::equal_range(std::begin(OperandMatchTable), "
3002           "std::end(OperandMatchTable),\n";
3003     OS << "                       Mnemonic, LessOpcodeOperand());\n\n";
3004   }
3005 
3006   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
3007   OS << "    return ParseStatus::NoMatch;\n\n";
3008 
3009   OS << "  for (const OperandMatchEntry *it = MnemonicRange.first,\n"
3010      << "       *ie = MnemonicRange.second; it != ie; ++it) {\n";
3011 
3012   OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
3013   OS << "    assert(Mnemonic == it->getMnemonic());\n\n";
3014 
3015   // Emit check that the required features are available.
3016   OS << "    // check if the available features match\n";
3017   OS << "    const FeatureBitset &RequiredFeatures = "
3018         "FeatureBitsets[it->RequiredFeaturesIdx];\n";
3019   OS << "    if (!ParseForAllFeatures && (AvailableFeatures & "
3020         "RequiredFeatures) != RequiredFeatures)\n";
3021   OS << "      continue;\n\n";
3022 
3023   // Emit check to ensure the operand number matches.
3024   OS << "    // check if the operand in question has a custom parser.\n";
3025   OS << "    if (!(it->OperandMask & (1 << NextOpNum)))\n";
3026   OS << "      continue;\n\n";
3027 
3028   // Emit call to the custom parser method
3029   StringRef ParserName = AsmParser.getValueAsString("OperandParserMethod");
3030   if (ParserName.empty())
3031     ParserName = "tryCustomParseOperand";
3032   OS << "    // call custom parse method to handle the operand\n";
3033   OS << "    ParseStatus Result = " << ParserName << "(Operands, it->Class);\n";
3034   OS << "    if (!Result.isNoMatch())\n";
3035   OS << "      return Result;\n";
3036   OS << "  }\n\n";
3037 
3038   OS << "  // Okay, we had no match.\n";
3039   OS << "  return ParseStatus::NoMatch;\n";
3040   OS << "}\n\n";
3041 }
3042 
emitAsmTiedOperandConstraints(CodeGenTarget & Target,AsmMatcherInfo & Info,raw_ostream & OS,bool HasOptionalOperands)3043 static void emitAsmTiedOperandConstraints(CodeGenTarget &Target,
3044                                           AsmMatcherInfo &Info, raw_ostream &OS,
3045                                           bool HasOptionalOperands) {
3046   std::string AsmParserName =
3047       std::string(Info.AsmParser->getValueAsString("AsmParserClassName"));
3048   OS << "static bool ";
3049   OS << "checkAsmTiedOperandConstraints(const " << Target.getName()
3050      << AsmParserName << "&AsmParser,\n";
3051   OS << "                               unsigned Kind, const OperandVector "
3052         "&Operands,\n";
3053   if (HasOptionalOperands)
3054     OS << "                               ArrayRef<unsigned> DefaultsOffset,\n";
3055   OS << "                               uint64_t &ErrorInfo) {\n";
3056   OS << "  assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n";
3057   OS << "  const uint8_t *Converter = ConversionTable[Kind];\n";
3058   OS << "  for (const uint8_t *p = Converter; *p; p += 2) {\n";
3059   OS << "    switch (*p) {\n";
3060   OS << "    case CVT_Tied: {\n";
3061   OS << "      unsigned OpIdx = *(p + 1);\n";
3062   OS << "      assert(OpIdx < (size_t)(std::end(TiedAsmOperandTable) -\n";
3063   OS << "                              std::begin(TiedAsmOperandTable)) &&\n";
3064   OS << "             \"Tied operand not found\");\n";
3065   OS << "      unsigned OpndNum1 = TiedAsmOperandTable[OpIdx][1];\n";
3066   OS << "      unsigned OpndNum2 = TiedAsmOperandTable[OpIdx][2];\n";
3067   if (HasOptionalOperands) {
3068     // When optional operands are involved, formal and actual operand indices
3069     // may differ. Map the former to the latter by subtracting the number of
3070     // absent optional operands.
3071     OS << "      OpndNum1 = OpndNum1 - DefaultsOffset[OpndNum1];\n";
3072     OS << "      OpndNum2 = OpndNum2 - DefaultsOffset[OpndNum2];\n";
3073   }
3074   OS << "      if (OpndNum1 != OpndNum2) {\n";
3075   OS << "        auto &SrcOp1 = Operands[OpndNum1];\n";
3076   OS << "        auto &SrcOp2 = Operands[OpndNum2];\n";
3077   OS << "        if (!AsmParser.areEqualRegs(*SrcOp1, *SrcOp2)) {\n";
3078   OS << "          ErrorInfo = OpndNum2;\n";
3079   OS << "          return false;\n";
3080   OS << "        }\n";
3081   OS << "      }\n";
3082   OS << "      break;\n";
3083   OS << "    }\n";
3084   OS << "    default:\n";
3085   OS << "      break;\n";
3086   OS << "    }\n";
3087   OS << "  }\n";
3088   OS << "  return true;\n";
3089   OS << "}\n\n";
3090 }
3091 
emitMnemonicSpellChecker(raw_ostream & OS,CodeGenTarget & Target,unsigned VariantCount)3092 static void emitMnemonicSpellChecker(raw_ostream &OS, CodeGenTarget &Target,
3093                                      unsigned VariantCount) {
3094   OS << "static std::string " << Target.getName()
3095      << "MnemonicSpellCheck(StringRef S, const FeatureBitset &FBS,"
3096      << " unsigned VariantID) {\n";
3097   if (!VariantCount)
3098     OS << "  return \"\";";
3099   else {
3100     OS << "  const unsigned MaxEditDist = 2;\n";
3101     OS << "  std::vector<StringRef> Candidates;\n";
3102     OS << "  StringRef Prev = \"\";\n\n";
3103 
3104     OS << "  // Find the appropriate table for this asm variant.\n";
3105     OS << "  const MatchEntry *Start, *End;\n";
3106     OS << "  switch (VariantID) {\n";
3107     OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3108     for (unsigned VC = 0; VC != VariantCount; ++VC) {
3109       Record *AsmVariant = Target.getAsmParserVariant(VC);
3110       int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3111       OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3112          << "); End = std::end(MatchTable" << VC << "); break;\n";
3113     }
3114     OS << "  }\n\n";
3115     OS << "  for (auto I = Start; I < End; I++) {\n";
3116     OS << "    // Ignore unsupported instructions.\n";
3117     OS << "    const FeatureBitset &RequiredFeatures = "
3118           "FeatureBitsets[I->RequiredFeaturesIdx];\n";
3119     OS << "    if ((FBS & RequiredFeatures) != RequiredFeatures)\n";
3120     OS << "      continue;\n";
3121     OS << "\n";
3122     OS << "    StringRef T = I->getMnemonic();\n";
3123     OS << "    // Avoid recomputing the edit distance for the same string.\n";
3124     OS << "    if (T == Prev)\n";
3125     OS << "      continue;\n";
3126     OS << "\n";
3127     OS << "    Prev = T;\n";
3128     OS << "    unsigned Dist = S.edit_distance(T, false, MaxEditDist);\n";
3129     OS << "    if (Dist <= MaxEditDist)\n";
3130     OS << "      Candidates.push_back(T);\n";
3131     OS << "  }\n";
3132     OS << "\n";
3133     OS << "  if (Candidates.empty())\n";
3134     OS << "    return \"\";\n";
3135     OS << "\n";
3136     OS << "  std::string Res = \", did you mean: \";\n";
3137     OS << "  unsigned i = 0;\n";
3138     OS << "  for (; i < Candidates.size() - 1; i++)\n";
3139     OS << "    Res += Candidates[i].str() + \", \";\n";
3140     OS << "  return Res + Candidates[i].str() + \"?\";\n";
3141   }
3142   OS << "}\n";
3143   OS << "\n";
3144 }
3145 
emitMnemonicChecker(raw_ostream & OS,CodeGenTarget & Target,unsigned VariantCount,bool HasMnemonicFirst,bool HasMnemonicAliases)3146 static void emitMnemonicChecker(raw_ostream &OS, CodeGenTarget &Target,
3147                                 unsigned VariantCount, bool HasMnemonicFirst,
3148                                 bool HasMnemonicAliases) {
3149   OS << "static bool " << Target.getName()
3150      << "CheckMnemonic(StringRef Mnemonic,\n";
3151   OS << "                                "
3152      << "const FeatureBitset &AvailableFeatures,\n";
3153   OS << "                                "
3154      << "unsigned VariantID) {\n";
3155 
3156   if (!VariantCount) {
3157     OS << "  return false;\n";
3158   } else {
3159     if (HasMnemonicAliases) {
3160       OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
3161       OS << "  applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);";
3162       OS << "\n\n";
3163     }
3164     OS << "  // Find the appropriate table for this asm variant.\n";
3165     OS << "  const MatchEntry *Start, *End;\n";
3166     OS << "  switch (VariantID) {\n";
3167     OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3168     for (unsigned VC = 0; VC != VariantCount; ++VC) {
3169       Record *AsmVariant = Target.getAsmParserVariant(VC);
3170       int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3171       OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3172          << "); End = std::end(MatchTable" << VC << "); break;\n";
3173     }
3174     OS << "  }\n\n";
3175 
3176     OS << "  // Search the table.\n";
3177     if (HasMnemonicFirst) {
3178       OS << "  auto MnemonicRange = "
3179             "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
3180     } else {
3181       OS << "  auto MnemonicRange = std::pair(Start, End);\n";
3182       OS << "  unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
3183       OS << "  if (!Mnemonic.empty())\n";
3184       OS << "    MnemonicRange = "
3185          << "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
3186     }
3187 
3188     OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
3189     OS << "    return false;\n\n";
3190 
3191     OS << "  for (const MatchEntry *it = MnemonicRange.first, "
3192        << "*ie = MnemonicRange.second;\n";
3193     OS << "       it != ie; ++it) {\n";
3194     OS << "    const FeatureBitset &RequiredFeatures =\n";
3195     OS << "      FeatureBitsets[it->RequiredFeaturesIdx];\n";
3196     OS << "    if ((AvailableFeatures & RequiredFeatures) == ";
3197     OS << "RequiredFeatures)\n";
3198     OS << "      return true;\n";
3199     OS << "  }\n";
3200     OS << "  return false;\n";
3201   }
3202   OS << "}\n";
3203   OS << "\n";
3204 }
3205 
3206 // Emit a function mapping match classes to strings, for debugging.
emitMatchClassKindNames(std::forward_list<ClassInfo> & Infos,raw_ostream & OS)3207 static void emitMatchClassKindNames(std::forward_list<ClassInfo> &Infos,
3208                                     raw_ostream &OS) {
3209   OS << "#ifndef NDEBUG\n";
3210   OS << "const char *getMatchClassName(MatchClassKind Kind) {\n";
3211   OS << "  switch (Kind) {\n";
3212 
3213   OS << "  case InvalidMatchClass: return \"InvalidMatchClass\";\n";
3214   OS << "  case OptionalMatchClass: return \"OptionalMatchClass\";\n";
3215   for (const auto &CI : Infos) {
3216     OS << "  case " << CI.Name << ": return \"" << CI.Name << "\";\n";
3217   }
3218   OS << "  case NumMatchClassKinds: return \"NumMatchClassKinds\";\n";
3219 
3220   OS << "  }\n";
3221   OS << "  llvm_unreachable(\"unhandled MatchClassKind!\");\n";
3222   OS << "}\n\n";
3223   OS << "#endif // NDEBUG\n";
3224 }
3225 
3226 static std::string
getNameForFeatureBitset(const std::vector<Record * > & FeatureBitset)3227 getNameForFeatureBitset(const std::vector<Record *> &FeatureBitset) {
3228   std::string Name = "AMFBS";
3229   for (const auto &Feature : FeatureBitset)
3230     Name += ("_" + Feature->getName()).str();
3231   return Name;
3232 }
3233 
run(raw_ostream & OS)3234 void AsmMatcherEmitter::run(raw_ostream &OS) {
3235   CodeGenTarget Target(Records);
3236   Record *AsmParser = Target.getAsmParser();
3237   StringRef ClassName = AsmParser->getValueAsString("AsmParserClassName");
3238 
3239   emitSourceFileHeader("Assembly Matcher Source Fragment", OS, Records);
3240 
3241   // Compute the information on the instructions to match.
3242   AsmMatcherInfo Info(AsmParser, Target, Records);
3243   Info.buildInfo();
3244 
3245   bool PreferSmallerInstructions = getPreferSmallerInstructions(Target);
3246   // Sort the instruction table using the partial order on classes. We use
3247   // stable_sort to ensure that ambiguous instructions are still
3248   // deterministically ordered.
3249   llvm::stable_sort(
3250       Info.Matchables,
3251       [PreferSmallerInstructions](const std::unique_ptr<MatchableInfo> &A,
3252                                   const std::unique_ptr<MatchableInfo> &B) {
3253         return A->shouldBeMatchedBefore(*B, PreferSmallerInstructions);
3254       });
3255 
3256 #ifdef EXPENSIVE_CHECKS
3257   // Verify that the table is sorted and operator < works transitively.
3258   for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
3259        ++I) {
3260     for (auto J = I; J != E; ++J) {
3261       assert(!(*J)->shouldBeMatchedBefore(**I, PreferSmallerInstructions));
3262     }
3263   }
3264 #endif
3265 
3266   DEBUG_WITH_TYPE("instruction_info", {
3267     for (const auto &MI : Info.Matchables)
3268       MI->dump();
3269   });
3270 
3271   // Check for ambiguous matchables.
3272   DEBUG_WITH_TYPE("ambiguous_instrs", {
3273     unsigned NumAmbiguous = 0;
3274     for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
3275          ++I) {
3276       for (auto J = std::next(I); J != E; ++J) {
3277         const MatchableInfo &A = **I;
3278         const MatchableInfo &B = **J;
3279 
3280         if (A.couldMatchAmbiguouslyWith(B, PreferSmallerInstructions)) {
3281           errs() << "warning: ambiguous matchables:\n";
3282           A.dump();
3283           errs() << "\nis incomparable with:\n";
3284           B.dump();
3285           errs() << "\n\n";
3286           ++NumAmbiguous;
3287         }
3288       }
3289     }
3290     if (NumAmbiguous)
3291       errs() << "warning: " << NumAmbiguous << " ambiguous matchables!\n";
3292   });
3293 
3294   // Compute the information on the custom operand parsing.
3295   Info.buildOperandMatchInfo();
3296 
3297   bool HasMnemonicFirst = AsmParser->getValueAsBit("HasMnemonicFirst");
3298   bool HasOptionalOperands = Info.hasOptionalOperands();
3299   bool ReportMultipleNearMisses =
3300       AsmParser->getValueAsBit("ReportMultipleNearMisses");
3301 
3302   // Write the output.
3303 
3304   // Information for the class declaration.
3305   OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
3306   OS << "#undef GET_ASSEMBLER_HEADER\n";
3307   OS << "  // This should be included into the middle of the declaration of\n";
3308   OS << "  // your subclasses implementation of MCTargetAsmParser.\n";
3309   OS << "  FeatureBitset ComputeAvailableFeatures(const FeatureBitset &FB) "
3310         "const;\n";
3311   if (HasOptionalOperands) {
3312     OS << "  void convertToMCInst(unsigned Kind, MCInst &Inst, "
3313        << "unsigned Opcode,\n"
3314        << "                       const OperandVector &Operands,\n"
3315        << "                       const SmallBitVector "
3316           "&OptionalOperandsMask,\n"
3317        << "                       ArrayRef<unsigned> DefaultsOffset);\n";
3318   } else {
3319     OS << "  void convertToMCInst(unsigned Kind, MCInst &Inst, "
3320        << "unsigned Opcode,\n"
3321        << "                       const OperandVector &Operands);\n";
3322   }
3323   OS << "  void convertToMapAndConstraints(unsigned Kind,\n                ";
3324   OS << "           const OperandVector &Operands) override;\n";
3325   OS << "  unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
3326      << "                                MCInst &Inst,\n";
3327   if (ReportMultipleNearMisses)
3328     OS << "                                SmallVectorImpl<NearMissInfo> "
3329           "*NearMisses,\n";
3330   else
3331     OS << "                                uint64_t &ErrorInfo,\n"
3332        << "                                FeatureBitset &MissingFeatures,\n";
3333   OS << "                                bool matchingInlineAsm,\n"
3334      << "                                unsigned VariantID = 0);\n";
3335   if (!ReportMultipleNearMisses)
3336     OS << "  unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
3337        << "                                MCInst &Inst,\n"
3338        << "                                uint64_t &ErrorInfo,\n"
3339        << "                                bool matchingInlineAsm,\n"
3340        << "                                unsigned VariantID = 0) {\n"
3341        << "    FeatureBitset MissingFeatures;\n"
3342        << "    return MatchInstructionImpl(Operands, Inst, ErrorInfo, "
3343           "MissingFeatures,\n"
3344        << "                                matchingInlineAsm, VariantID);\n"
3345        << "  }\n\n";
3346 
3347   if (!Info.OperandMatchInfo.empty()) {
3348     OS << "  ParseStatus MatchOperandParserImpl(\n";
3349     OS << "    OperandVector &Operands,\n";
3350     OS << "    StringRef Mnemonic,\n";
3351     OS << "    bool ParseForAllFeatures = false);\n";
3352 
3353     OS << "  ParseStatus tryCustomParseOperand(\n";
3354     OS << "    OperandVector &Operands,\n";
3355     OS << "    unsigned MCK);\n\n";
3356   }
3357 
3358   OS << "#endif // GET_ASSEMBLER_HEADER\n\n";
3359 
3360   // Emit the operand match diagnostic enum names.
3361   OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
3362   OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
3363   emitOperandDiagnosticTypes(Info, OS);
3364   OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
3365 
3366   OS << "\n#ifdef GET_REGISTER_MATCHER\n";
3367   OS << "#undef GET_REGISTER_MATCHER\n\n";
3368 
3369   // Emit the subtarget feature enumeration.
3370   SubtargetFeatureInfo::emitSubtargetFeatureBitEnumeration(
3371       Info.SubtargetFeatures, OS);
3372 
3373   // Emit the function to match a register name to number.
3374   // This should be omitted for Mips target
3375   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
3376     emitMatchRegisterName(Target, AsmParser, OS);
3377 
3378   if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterAltName"))
3379     emitMatchRegisterAltName(Target, AsmParser, OS);
3380 
3381   OS << "#endif // GET_REGISTER_MATCHER\n\n";
3382 
3383   OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
3384   OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
3385 
3386   // Generate the helper function to get the names for subtarget features.
3387   emitGetSubtargetFeatureName(Info, OS);
3388 
3389   OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
3390 
3391   OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
3392   OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
3393 
3394   // Generate the function that remaps for mnemonic aliases.
3395   bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
3396 
3397   // Generate the convertToMCInst function to convert operands into an MCInst.
3398   // Also, generate the convertToMapAndConstraints function for MS-style inline
3399   // assembly.  The latter doesn't actually generate a MCInst.
3400   unsigned NumConverters =
3401       emitConvertFuncs(Target, ClassName, Info.Matchables, HasMnemonicFirst,
3402                        HasOptionalOperands, OS);
3403 
3404   // Emit the enumeration for classes which participate in matching.
3405   emitMatchClassEnumeration(Target, Info.Classes, OS);
3406 
3407   // Emit a function to get the user-visible string to describe an operand
3408   // match failure in diagnostics.
3409   emitOperandMatchErrorDiagStrings(Info, OS);
3410 
3411   // Emit a function to map register classes to operand match failure codes.
3412   emitRegisterMatchErrorFunc(Info, OS);
3413 
3414   // Emit the routine to match token strings to their match class.
3415   emitMatchTokenString(Target, Info.Classes, OS);
3416 
3417   // Emit the subclass predicate routine.
3418   emitIsSubclass(Target, Info.Classes, OS);
3419 
3420   // Emit the routine to validate an operand against a match class.
3421   emitValidateOperandClass(Info, OS);
3422 
3423   emitMatchClassKindNames(Info.Classes, OS);
3424 
3425   // Emit the available features compute function.
3426   SubtargetFeatureInfo::emitComputeAssemblerAvailableFeatures(
3427       Info.Target.getName(), ClassName, "ComputeAvailableFeatures",
3428       Info.SubtargetFeatures, OS);
3429 
3430   if (!ReportMultipleNearMisses)
3431     emitAsmTiedOperandConstraints(Target, Info, OS, HasOptionalOperands);
3432 
3433   StringToOffsetTable StringTable;
3434 
3435   size_t MaxNumOperands = 0;
3436   unsigned MaxMnemonicIndex = 0;
3437   bool HasDeprecation = false;
3438   for (const auto &MI : Info.Matchables) {
3439     MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
3440     HasDeprecation |= MI->HasDeprecation;
3441 
3442     // Store a pascal-style length byte in the mnemonic.
3443     std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.lower();
3444     MaxMnemonicIndex = std::max(
3445         MaxMnemonicIndex, StringTable.GetOrAddStringOffset(LenMnemonic, false));
3446   }
3447 
3448   OS << "static const char MnemonicTable[] =\n";
3449   StringTable.EmitString(OS);
3450   OS << ";\n\n";
3451 
3452   std::vector<std::vector<Record *>> FeatureBitsets;
3453   for (const auto &MI : Info.Matchables) {
3454     if (MI->RequiredFeatures.empty())
3455       continue;
3456     FeatureBitsets.emplace_back();
3457     for (unsigned I = 0, E = MI->RequiredFeatures.size(); I != E; ++I)
3458       FeatureBitsets.back().push_back(MI->RequiredFeatures[I]->TheDef);
3459   }
3460 
3461   llvm::sort(FeatureBitsets, [&](const std::vector<Record *> &A,
3462                                  const std::vector<Record *> &B) {
3463     if (A.size() < B.size())
3464       return true;
3465     if (A.size() > B.size())
3466       return false;
3467     for (auto Pair : zip(A, B)) {
3468       if (std::get<0>(Pair)->getName() < std::get<1>(Pair)->getName())
3469         return true;
3470       if (std::get<0>(Pair)->getName() > std::get<1>(Pair)->getName())
3471         return false;
3472     }
3473     return false;
3474   });
3475   FeatureBitsets.erase(llvm::unique(FeatureBitsets), FeatureBitsets.end());
3476   OS << "// Feature bitsets.\n"
3477      << "enum : " << getMinimalTypeForRange(FeatureBitsets.size()) << " {\n"
3478      << "  AMFBS_None,\n";
3479   for (const auto &FeatureBitset : FeatureBitsets) {
3480     if (FeatureBitset.empty())
3481       continue;
3482     OS << "  " << getNameForFeatureBitset(FeatureBitset) << ",\n";
3483   }
3484   OS << "};\n\n"
3485      << "static constexpr FeatureBitset FeatureBitsets[] = {\n"
3486      << "  {}, // AMFBS_None\n";
3487   for (const auto &FeatureBitset : FeatureBitsets) {
3488     if (FeatureBitset.empty())
3489       continue;
3490     OS << "  {";
3491     for (const auto &Feature : FeatureBitset) {
3492       const auto &I = Info.SubtargetFeatures.find(Feature);
3493       assert(I != Info.SubtargetFeatures.end() && "Didn't import predicate?");
3494       OS << I->second.getEnumBitName() << ", ";
3495     }
3496     OS << "},\n";
3497   }
3498   OS << "};\n\n";
3499 
3500   // Emit the static match table; unused classes get initialized to 0 which is
3501   // guaranteed to be InvalidMatchClass.
3502   //
3503   // FIXME: We can reduce the size of this table very easily. First, we change
3504   // it so that store the kinds in separate bit-fields for each index, which
3505   // only needs to be the max width used for classes at that index (we also need
3506   // to reject based on this during classification). If we then make sure to
3507   // order the match kinds appropriately (putting mnemonics last), then we
3508   // should only end up using a few bits for each class, especially the ones
3509   // following the mnemonic.
3510   OS << "namespace {\n";
3511   OS << "  struct MatchEntry {\n";
3512   OS << "    " << getMinimalTypeForRange(MaxMnemonicIndex) << " Mnemonic;\n";
3513   OS << "    uint16_t Opcode;\n";
3514   OS << "    " << getMinimalTypeForRange(NumConverters) << " ConvertFn;\n";
3515   OS << "    " << getMinimalTypeForRange(FeatureBitsets.size())
3516      << " RequiredFeaturesIdx;\n";
3517   OS << "    "
3518      << getMinimalTypeForRange(
3519             std::distance(Info.Classes.begin(), Info.Classes.end()) +
3520             2 /* Include 'InvalidMatchClass' and 'OptionalMatchClass' */)
3521      << " Classes[" << MaxNumOperands << "];\n";
3522   OS << "    StringRef getMnemonic() const {\n";
3523   OS << "      return StringRef(MnemonicTable + Mnemonic + 1,\n";
3524   OS << "                       MnemonicTable[Mnemonic]);\n";
3525   OS << "    }\n";
3526   OS << "  };\n\n";
3527 
3528   OS << "  // Predicate for searching for an opcode.\n";
3529   OS << "  struct LessOpcode {\n";
3530   OS << "    bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
3531   OS << "      return LHS.getMnemonic() < RHS;\n";
3532   OS << "    }\n";
3533   OS << "    bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
3534   OS << "      return LHS < RHS.getMnemonic();\n";
3535   OS << "    }\n";
3536   OS << "    bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
3537   OS << "      return LHS.getMnemonic() < RHS.getMnemonic();\n";
3538   OS << "    }\n";
3539   OS << "  };\n";
3540 
3541   OS << "} // end anonymous namespace\n\n";
3542 
3543   unsigned VariantCount = Target.getAsmParserVariantCount();
3544   for (unsigned VC = 0; VC != VariantCount; ++VC) {
3545     Record *AsmVariant = Target.getAsmParserVariant(VC);
3546     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3547 
3548     OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
3549 
3550     for (const auto &MI : Info.Matchables) {
3551       if (MI->AsmVariantID != AsmVariantNo)
3552         continue;
3553 
3554       // Store a pascal-style length byte in the mnemonic.
3555       std::string LenMnemonic =
3556           char(MI->Mnemonic.size()) + MI->Mnemonic.lower();
3557       OS << "  { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
3558          << " /* " << MI->Mnemonic << " */, " << Target.getInstNamespace()
3559          << "::" << MI->getResultInst()->TheDef->getName() << ", "
3560          << MI->ConversionFnKind << ", ";
3561 
3562       // Write the required features mask.
3563       OS << "AMFBS";
3564       if (MI->RequiredFeatures.empty())
3565         OS << "_None";
3566       else
3567         for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i)
3568           OS << '_' << MI->RequiredFeatures[i]->TheDef->getName();
3569 
3570       OS << ", { ";
3571       ListSeparator LS;
3572       for (const MatchableInfo::AsmOperand &Op : MI->AsmOperands)
3573         OS << LS << Op.Class->Name;
3574       OS << " }, },\n";
3575     }
3576 
3577     OS << "};\n\n";
3578   }
3579 
3580   OS << "#include \"llvm/Support/Debug.h\"\n";
3581   OS << "#include \"llvm/Support/Format.h\"\n\n";
3582 
3583   // Finally, build the match function.
3584   OS << "unsigned " << Target.getName() << ClassName << "::\n"
3585      << "MatchInstructionImpl(const OperandVector &Operands,\n";
3586   OS << "                     MCInst &Inst,\n";
3587   if (ReportMultipleNearMisses)
3588     OS << "                     SmallVectorImpl<NearMissInfo> *NearMisses,\n";
3589   else
3590     OS << "                     uint64_t &ErrorInfo,\n"
3591        << "                     FeatureBitset &MissingFeatures,\n";
3592   OS << "                     bool matchingInlineAsm, unsigned VariantID) {\n";
3593 
3594   if (!ReportMultipleNearMisses) {
3595     OS << "  // Eliminate obvious mismatches.\n";
3596     OS << "  if (Operands.size() > " << (MaxNumOperands + HasMnemonicFirst)
3597        << ") {\n";
3598     OS << "    ErrorInfo = " << (MaxNumOperands + HasMnemonicFirst) << ";\n";
3599     OS << "    return Match_InvalidOperand;\n";
3600     OS << "  }\n\n";
3601   }
3602 
3603   // Emit code to get the available features.
3604   OS << "  // Get the current feature set.\n";
3605   OS << "  const FeatureBitset &AvailableFeatures = "
3606         "getAvailableFeatures();\n\n";
3607 
3608   OS << "  // Get the instruction mnemonic, which is the first token.\n";
3609   if (HasMnemonicFirst) {
3610     OS << "  StringRef Mnemonic = ((" << Target.getName()
3611        << "Operand &)*Operands[0]).getToken();\n\n";
3612   } else {
3613     OS << "  StringRef Mnemonic;\n";
3614     OS << "  if (Operands[0]->isToken())\n";
3615     OS << "    Mnemonic = ((" << Target.getName()
3616        << "Operand &)*Operands[0]).getToken();\n\n";
3617   }
3618 
3619   if (HasMnemonicAliases) {
3620     OS << "  // Process all MnemonicAliases to remap the mnemonic.\n";
3621     OS << "  applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
3622   }
3623 
3624   // Emit code to compute the class list for this operand vector.
3625   if (!ReportMultipleNearMisses) {
3626     OS << "  // Some state to try to produce better error messages.\n";
3627     OS << "  bool HadMatchOtherThanFeatures = false;\n";
3628     OS << "  bool HadMatchOtherThanPredicate = false;\n";
3629     OS << "  unsigned RetCode = Match_InvalidOperand;\n";
3630     OS << "  MissingFeatures.set();\n";
3631     OS << "  // Set ErrorInfo to the operand that mismatches if it is\n";
3632     OS << "  // wrong for all instances of the instruction.\n";
3633     OS << "  ErrorInfo = ~0ULL;\n";
3634   }
3635 
3636   if (HasOptionalOperands) {
3637     OS << "  SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n";
3638   }
3639 
3640   // Emit code to search the table.
3641   OS << "  // Find the appropriate table for this asm variant.\n";
3642   OS << "  const MatchEntry *Start, *End;\n";
3643   OS << "  switch (VariantID) {\n";
3644   OS << "  default: llvm_unreachable(\"invalid variant!\");\n";
3645   for (unsigned VC = 0; VC != VariantCount; ++VC) {
3646     Record *AsmVariant = Target.getAsmParserVariant(VC);
3647     int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
3648     OS << "  case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
3649        << "); End = std::end(MatchTable" << VC << "); break;\n";
3650   }
3651   OS << "  }\n";
3652 
3653   OS << "  // Search the table.\n";
3654   if (HasMnemonicFirst) {
3655     OS << "  auto MnemonicRange = "
3656           "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n";
3657   } else {
3658     OS << "  auto MnemonicRange = std::pair(Start, End);\n";
3659     OS << "  unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
3660     OS << "  if (!Mnemonic.empty())\n";
3661     OS << "    MnemonicRange = "
3662           "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
3663   }
3664 
3665   OS << "  DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"AsmMatcher: found \" "
3666         "<<\n"
3667      << "  std::distance(MnemonicRange.first, MnemonicRange.second) <<\n"
3668      << "  \" encodings with mnemonic '\" << Mnemonic << \"'\\n\");\n\n";
3669 
3670   OS << "  // Return a more specific error code if no mnemonics match.\n";
3671   OS << "  if (MnemonicRange.first == MnemonicRange.second)\n";
3672   OS << "    return Match_MnemonicFail;\n\n";
3673 
3674   OS << "  for (const MatchEntry *it = MnemonicRange.first, "
3675      << "*ie = MnemonicRange.second;\n";
3676   OS << "       it != ie; ++it) {\n";
3677   OS << "    const FeatureBitset &RequiredFeatures = "
3678         "FeatureBitsets[it->RequiredFeaturesIdx];\n";
3679   OS << "    bool HasRequiredFeatures =\n";
3680   OS << "      (AvailableFeatures & RequiredFeatures) == RequiredFeatures;\n";
3681   OS << "    DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Trying to match "
3682         "opcode \"\n";
3683   OS << "                                          << MII.getName(it->Opcode) "
3684         "<< \"\\n\");\n";
3685 
3686   if (ReportMultipleNearMisses) {
3687     OS << "    // Some state to record ways in which this instruction did not "
3688           "match.\n";
3689     OS << "    NearMissInfo OperandNearMiss = NearMissInfo::getSuccess();\n";
3690     OS << "    NearMissInfo FeaturesNearMiss = NearMissInfo::getSuccess();\n";
3691     OS << "    NearMissInfo EarlyPredicateNearMiss = "
3692           "NearMissInfo::getSuccess();\n";
3693     OS << "    NearMissInfo LatePredicateNearMiss = "
3694           "NearMissInfo::getSuccess();\n";
3695     OS << "    bool MultipleInvalidOperands = false;\n";
3696   }
3697 
3698   if (HasMnemonicFirst) {
3699     OS << "    // equal_range guarantees that instruction mnemonic matches.\n";
3700     OS << "    assert(Mnemonic == it->getMnemonic());\n";
3701   }
3702 
3703   // Emit check that the subclasses match.
3704   if (!ReportMultipleNearMisses)
3705     OS << "    bool OperandsValid = true;\n";
3706   if (HasOptionalOperands) {
3707     OS << "    OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n";
3708   }
3709   OS << "    for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex")
3710      << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex")
3711      << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n";
3712   OS << "      auto Formal = "
3713      << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n";
3714   OS << "      DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3715   OS << "                      dbgs() << \"  Matching formal operand class \" "
3716         "<< getMatchClassName(Formal)\n";
3717   OS << "                             << \" against actual operand at index \" "
3718         "<< ActualIdx);\n";
3719   OS << "      if (ActualIdx < Operands.size())\n";
3720   OS << "        DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \" (\";\n";
3721   OS << "                        Operands[ActualIdx]->print(dbgs()); dbgs() << "
3722         "\"): \");\n";
3723   OS << "      else\n";
3724   OS << "        DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \": \");\n";
3725   OS << "      if (ActualIdx >= Operands.size()) {\n";
3726   OS << "        DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"actual operand "
3727         "index out of range\\n\");\n";
3728   if (ReportMultipleNearMisses) {
3729     OS << "        bool ThisOperandValid = (Formal == "
3730        << "InvalidMatchClass) || "
3731           "isSubclass(Formal, OptionalMatchClass);\n";
3732     OS << "        if (!ThisOperandValid) {\n";
3733     OS << "          if (!OperandNearMiss) {\n";
3734     OS << "            // Record info about match failure for later use.\n";
3735     OS << "            DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"recording "
3736           "too-few-operands near miss\\n\");\n";
3737     OS << "            OperandNearMiss =\n";
3738     OS << "                NearMissInfo::getTooFewOperands(Formal, "
3739           "it->Opcode);\n";
3740     OS << "          } else if (OperandNearMiss.getKind() != "
3741           "NearMissInfo::NearMissTooFewOperands) {\n";
3742     OS << "            // If more than one operand is invalid, give up on this "
3743           "match entry.\n";
3744     OS << "            DEBUG_WITH_TYPE(\n";
3745     OS << "                \"asm-matcher\",\n";
3746     OS << "                dbgs() << \"second invalid operand, giving up on "
3747           "this opcode\\n\");\n";
3748     OS << "            MultipleInvalidOperands = true;\n";
3749     OS << "            break;\n";
3750     OS << "          }\n";
3751     OS << "        } else {\n";
3752     OS << "          DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"but formal "
3753           "operand not required\\n\");\n";
3754     OS << "          if (isSubclass(Formal, OptionalMatchClass)) {\n";
3755     OS << "            OptionalOperandsMask.set(FormalIdx);\n";
3756     OS << "          }\n";
3757     OS << "        }\n";
3758     OS << "        continue;\n";
3759   } else {
3760     OS << "        if (Formal == InvalidMatchClass) {\n";
3761     if (HasOptionalOperands) {
3762       OS << "          OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands
3763          << ");\n";
3764     }
3765     OS << "          break;\n";
3766     OS << "        }\n";
3767     OS << "        if (isSubclass(Formal, OptionalMatchClass)) {\n";
3768     if (HasOptionalOperands) {
3769       OS << "          OptionalOperandsMask.set(FormalIdx);\n";
3770     }
3771     OS << "          continue;\n";
3772     OS << "        }\n";
3773     OS << "        OperandsValid = false;\n";
3774     OS << "        ErrorInfo = ActualIdx;\n";
3775     OS << "        break;\n";
3776   }
3777   OS << "      }\n";
3778   OS << "      MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n";
3779   OS << "      unsigned Diag = validateOperandClass(Actual, Formal);\n";
3780   OS << "      if (Diag == Match_Success) {\n";
3781   OS << "        DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3782   OS << "                        dbgs() << \"match success using generic "
3783         "matcher\\n\");\n";
3784   OS << "        ++ActualIdx;\n";
3785   OS << "        continue;\n";
3786   OS << "      }\n";
3787   OS << "      // If the generic handler indicates an invalid operand\n";
3788   OS << "      // failure, check for a special case.\n";
3789   OS << "      if (Diag != Match_Success) {\n";
3790   OS << "        unsigned TargetDiag = validateTargetOperandClass(Actual, "
3791         "Formal);\n";
3792   OS << "        if (TargetDiag == Match_Success) {\n";
3793   OS << "          DEBUG_WITH_TYPE(\"asm-matcher\",\n";
3794   OS << "                          dbgs() << \"match success using target "
3795         "matcher\\n\");\n";
3796   OS << "          ++ActualIdx;\n";
3797   OS << "          continue;\n";
3798   OS << "        }\n";
3799   OS << "        // If the target matcher returned a specific error code use\n";
3800   OS << "        // that, else use the one from the generic matcher.\n";
3801   OS << "        if (TargetDiag != Match_InvalidOperand && "
3802         "HasRequiredFeatures)\n";
3803   OS << "          Diag = TargetDiag;\n";
3804   OS << "      }\n";
3805   OS << "      // If current formal operand wasn't matched and it is optional\n"
3806      << "      // then try to match next formal operand\n";
3807   OS << "      if (Diag == Match_InvalidOperand "
3808      << "&& isSubclass(Formal, OptionalMatchClass)) {\n";
3809   if (HasOptionalOperands) {
3810     OS << "        OptionalOperandsMask.set(FormalIdx);\n";
3811   }
3812   OS << "        DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"ignoring "
3813         "optional operand\\n\");\n";
3814   OS << "        continue;\n";
3815   OS << "      }\n";
3816 
3817   if (ReportMultipleNearMisses) {
3818     OS << "      if (!OperandNearMiss) {\n";
3819     OS << "        // If this is the first invalid operand we have seen, "
3820           "record some\n";
3821     OS << "        // information about it.\n";
3822     OS << "        DEBUG_WITH_TYPE(\n";
3823     OS << "            \"asm-matcher\",\n";
3824     OS << "            dbgs()\n";
3825     OS << "                << \"operand match failed, recording near-miss with "
3826           "diag code \"\n";
3827     OS << "                << Diag << \"\\n\");\n";
3828     OS << "        OperandNearMiss =\n";
3829     OS << "            NearMissInfo::getMissedOperand(Diag, Formal, "
3830           "it->Opcode, ActualIdx);\n";
3831     OS << "        ++ActualIdx;\n";
3832     OS << "      } else {\n";
3833     OS << "        // If more than one operand is invalid, give up on this "
3834           "match entry.\n";
3835     OS << "        DEBUG_WITH_TYPE(\n";
3836     OS << "            \"asm-matcher\",\n";
3837     OS << "            dbgs() << \"second operand mismatch, skipping this "
3838           "opcode\\n\");\n";
3839     OS << "        MultipleInvalidOperands = true;\n";
3840     OS << "        break;\n";
3841     OS << "      }\n";
3842     OS << "    }\n\n";
3843   } else {
3844     OS << "      // If this operand is broken for all of the instances of "
3845           "this\n";
3846     OS << "      // mnemonic, keep track of it so we can report loc info.\n";
3847     OS << "      // If we already had a match that only failed due to a\n";
3848     OS << "      // target predicate, that diagnostic is preferred.\n";
3849     OS << "      if (!HadMatchOtherThanPredicate &&\n";
3850     OS << "          (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) "
3851           "{\n";
3852     OS << "        if (HasRequiredFeatures && (ErrorInfo != ActualIdx || Diag "
3853           "!= Match_InvalidOperand))\n";
3854     OS << "          RetCode = Diag;\n";
3855     OS << "        ErrorInfo = ActualIdx;\n";
3856     OS << "      }\n";
3857     OS << "      // Otherwise, just reject this instance of the mnemonic.\n";
3858     OS << "      OperandsValid = false;\n";
3859     OS << "      break;\n";
3860     OS << "    }\n\n";
3861   }
3862 
3863   if (ReportMultipleNearMisses)
3864     OS << "    if (MultipleInvalidOperands) {\n";
3865   else
3866     OS << "    if (!OperandsValid) {\n";
3867   OS << "      DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: "
3868         "multiple \"\n";
3869   OS << "                                               \"operand mismatches, "
3870         "ignoring \"\n";
3871   OS << "                                               \"this opcode\\n\");\n";
3872   OS << "      continue;\n";
3873   OS << "    }\n";
3874 
3875   // Emit check that the required features are available.
3876   OS << "    if (!HasRequiredFeatures) {\n";
3877   if (!ReportMultipleNearMisses)
3878     OS << "      HadMatchOtherThanFeatures = true;\n";
3879   OS << "      FeatureBitset NewMissingFeatures = RequiredFeatures & "
3880         "~AvailableFeatures;\n";
3881   OS << "      DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Missing target "
3882         "features:\";\n";
3883   OS << "                      for (unsigned I = 0, E = "
3884         "NewMissingFeatures.size(); I != E; ++I)\n";
3885   OS << "                        if (NewMissingFeatures[I])\n";
3886   OS << "                          dbgs() << ' ' << I;\n";
3887   OS << "                      dbgs() << \"\\n\");\n";
3888   if (ReportMultipleNearMisses) {
3889     OS << "      FeaturesNearMiss = "
3890           "NearMissInfo::getMissedFeature(NewMissingFeatures);\n";
3891   } else {
3892     OS << "      if (NewMissingFeatures.count() <=\n"
3893           "          MissingFeatures.count())\n";
3894     OS << "        MissingFeatures = NewMissingFeatures;\n";
3895     OS << "      continue;\n";
3896   }
3897   OS << "    }\n";
3898   OS << "\n";
3899   OS << "    Inst.clear();\n\n";
3900   OS << "    Inst.setOpcode(it->Opcode);\n";
3901   // Verify the instruction with the target-specific match predicate function.
3902   OS << "    // We have a potential match but have not rendered the operands.\n"
3903      << "    // Check the target predicate to handle any context sensitive\n"
3904         "    // constraints.\n"
3905      << "    // For example, Ties that are referenced multiple times must be\n"
3906         "    // checked here to ensure the input is the same for each match\n"
3907         "    // constraints. If we leave it any later the ties will have been\n"
3908         "    // canonicalized\n"
3909      << "    unsigned MatchResult;\n"
3910      << "    if ((MatchResult = checkEarlyTargetMatchPredicate(Inst, "
3911         "Operands)) != Match_Success) {\n"
3912      << "      Inst.clear();\n";
3913   OS << "      DEBUG_WITH_TYPE(\n";
3914   OS << "          \"asm-matcher\",\n";
3915   OS << "          dbgs() << \"Early target match predicate failed with diag "
3916         "code \"\n";
3917   OS << "                 << MatchResult << \"\\n\");\n";
3918   if (ReportMultipleNearMisses) {
3919     OS << "      EarlyPredicateNearMiss = "
3920           "NearMissInfo::getMissedPredicate(MatchResult);\n";
3921   } else {
3922     OS << "      RetCode = MatchResult;\n"
3923        << "      HadMatchOtherThanPredicate = true;\n"
3924        << "      continue;\n";
3925   }
3926   OS << "    }\n\n";
3927 
3928   if (ReportMultipleNearMisses) {
3929     OS << "    // If we did not successfully match the operands, then we can't "
3930           "convert to\n";
3931     OS << "    // an MCInst, so bail out on this instruction variant now.\n";
3932     OS << "    if (OperandNearMiss) {\n";
3933     OS << "      // If the operand mismatch was the only problem, reprrt it as "
3934           "a near-miss.\n";
3935     OS << "      if (NearMisses && !FeaturesNearMiss && "
3936           "!EarlyPredicateNearMiss) {\n";
3937     OS << "        DEBUG_WITH_TYPE(\n";
3938     OS << "            \"asm-matcher\",\n";
3939     OS << "            dbgs()\n";
3940     OS << "                << \"Opcode result: one mismatched operand, adding "
3941           "near-miss\\n\");\n";
3942     OS << "        NearMisses->push_back(OperandNearMiss);\n";
3943     OS << "      } else {\n";
3944     OS << "        DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: "
3945           "multiple \"\n";
3946     OS << "                                                 \"types of "
3947           "mismatch, so not \"\n";
3948     OS << "                                                 \"reporting "
3949           "near-miss\\n\");\n";
3950     OS << "      }\n";
3951     OS << "      continue;\n";
3952     OS << "    }\n\n";
3953   }
3954 
3955   // When converting parsed operands to MCInst we need to know whether optional
3956   // operands were parsed or not so that we can choose the correct converter
3957   // function. We also need to know this when checking tied operand constraints.
3958   // DefaultsOffset is an array of deltas between the formal (MCInst) and the
3959   // actual (parsed operand array) operand indices. When all optional operands
3960   // are present, all elements of the array are zeros. If some of the optional
3961   // operands are absent, the array might look like '0, 0, 1, 1, 1, 2, 2, 3',
3962   // where each increment in value reflects the absence of an optional operand.
3963   if (HasOptionalOperands) {
3964     OS << "    unsigned DefaultsOffset[" << (MaxNumOperands + 1)
3965        << "] = { 0 };\n";
3966     OS << "    assert(OptionalOperandsMask.size() == " << (MaxNumOperands)
3967        << ");\n";
3968     OS << "    for (unsigned i = 0, NumDefaults = 0; i < " << (MaxNumOperands)
3969        << "; ++i) {\n";
3970     OS << "      DefaultsOffset[i + 1] = NumDefaults;\n";
3971     OS << "      NumDefaults += (OptionalOperandsMask[i] ? 1 : 0);\n";
3972     OS << "    }\n\n";
3973   }
3974 
3975   OS << "    if (matchingInlineAsm) {\n";
3976   OS << "      convertToMapAndConstraints(it->ConvertFn, Operands);\n";
3977   if (!ReportMultipleNearMisses) {
3978     if (HasOptionalOperands) {
3979       OS << "      if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
3980             "Operands,\n";
3981       OS << "                                          DefaultsOffset, "
3982             "ErrorInfo))\n";
3983     } else {
3984       OS << "      if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
3985             "Operands,\n";
3986       OS << "                                          ErrorInfo))\n";
3987     }
3988     OS << "        return Match_InvalidTiedOperand;\n";
3989     OS << "\n";
3990   }
3991   OS << "      return Match_Success;\n";
3992   OS << "    }\n\n";
3993   OS << "    // We have selected a definite instruction, convert the parsed\n"
3994      << "    // operands into the appropriate MCInst.\n";
3995   if (HasOptionalOperands) {
3996     OS << "    convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n"
3997        << "                    OptionalOperandsMask, DefaultsOffset);\n";
3998   } else {
3999     OS << "    convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
4000   }
4001   OS << "\n";
4002 
4003   // Verify the instruction with the target-specific match predicate function.
4004   OS << "    // We have a potential match. Check the target predicate to\n"
4005      << "    // handle any context sensitive constraints.\n"
4006      << "    if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
4007      << " Match_Success) {\n"
4008      << "      DEBUG_WITH_TYPE(\"asm-matcher\",\n"
4009      << "                      dbgs() << \"Target match predicate failed with "
4010         "diag code \"\n"
4011      << "                             << MatchResult << \"\\n\");\n"
4012      << "      Inst.clear();\n";
4013   if (ReportMultipleNearMisses) {
4014     OS << "      LatePredicateNearMiss = "
4015           "NearMissInfo::getMissedPredicate(MatchResult);\n";
4016   } else {
4017     OS << "      RetCode = MatchResult;\n"
4018        << "      HadMatchOtherThanPredicate = true;\n"
4019        << "      continue;\n";
4020   }
4021   OS << "    }\n\n";
4022 
4023   if (ReportMultipleNearMisses) {
4024     OS << "    int NumNearMisses = ((int)(bool)OperandNearMiss +\n";
4025     OS << "                         (int)(bool)FeaturesNearMiss +\n";
4026     OS << "                         (int)(bool)EarlyPredicateNearMiss +\n";
4027     OS << "                         (int)(bool)LatePredicateNearMiss);\n";
4028     OS << "    if (NumNearMisses == 1) {\n";
4029     OS << "      // We had exactly one type of near-miss, so add that to the "
4030           "list.\n";
4031     OS << "      assert(!OperandNearMiss && \"OperandNearMiss was handled "
4032           "earlier\");\n";
4033     OS << "      DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: "
4034           "found one type of \"\n";
4035     OS << "                                            \"mismatch, so "
4036           "reporting a \"\n";
4037     OS << "                                            \"near-miss\\n\");\n";
4038     OS << "      if (NearMisses && FeaturesNearMiss)\n";
4039     OS << "        NearMisses->push_back(FeaturesNearMiss);\n";
4040     OS << "      else if (NearMisses && EarlyPredicateNearMiss)\n";
4041     OS << "        NearMisses->push_back(EarlyPredicateNearMiss);\n";
4042     OS << "      else if (NearMisses && LatePredicateNearMiss)\n";
4043     OS << "        NearMisses->push_back(LatePredicateNearMiss);\n";
4044     OS << "\n";
4045     OS << "      continue;\n";
4046     OS << "    } else if (NumNearMisses > 1) {\n";
4047     OS << "      // This instruction missed in more than one way, so ignore "
4048           "it.\n";
4049     OS << "      DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: "
4050           "multiple \"\n";
4051     OS << "                                               \"types of mismatch, "
4052           "so not \"\n";
4053     OS << "                                               \"reporting "
4054           "near-miss\\n\");\n";
4055     OS << "      continue;\n";
4056     OS << "    }\n";
4057   }
4058 
4059   // Call the post-processing function, if used.
4060   StringRef InsnCleanupFn = AsmParser->getValueAsString("AsmParserInstCleanup");
4061   if (!InsnCleanupFn.empty())
4062     OS << "    " << InsnCleanupFn << "(Inst);\n";
4063 
4064   if (HasDeprecation) {
4065     OS << "    std::string Info;\n";
4066     OS << "    if "
4067           "(!getParser().getTargetParser().getTargetOptions()."
4068           "MCNoDeprecatedWarn &&\n";
4069     OS << "        MII.getDeprecatedInfo(Inst, getSTI(), Info)) {\n";
4070     OS << "      SMLoc Loc = ((" << Target.getName()
4071        << "Operand &)*Operands[0]).getStartLoc();\n";
4072     OS << "      getParser().Warning(Loc, Info, std::nullopt);\n";
4073     OS << "    }\n";
4074   }
4075 
4076   if (!ReportMultipleNearMisses) {
4077     if (HasOptionalOperands) {
4078       OS << "    if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
4079             "Operands,\n";
4080       OS << "                                         DefaultsOffset, "
4081             "ErrorInfo))\n";
4082     } else {
4083       OS << "    if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, "
4084             "Operands,\n";
4085       OS << "                                         ErrorInfo))\n";
4086     }
4087     OS << "      return Match_InvalidTiedOperand;\n";
4088     OS << "\n";
4089   }
4090 
4091   OS << "    DEBUG_WITH_TYPE(\n";
4092   OS << "        \"asm-matcher\",\n";
4093   OS << "        dbgs() << \"Opcode result: complete match, selecting this "
4094         "opcode\\n\");\n";
4095   OS << "    return Match_Success;\n";
4096   OS << "  }\n\n";
4097 
4098   if (ReportMultipleNearMisses) {
4099     OS << "  // No instruction variants matched exactly.\n";
4100     OS << "  return Match_NearMisses;\n";
4101   } else {
4102     OS << "  // Okay, we had no match.  Try to return a useful error code.\n";
4103     OS << "  if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
4104     OS << "    return RetCode;\n\n";
4105     OS << "  ErrorInfo = 0;\n";
4106     OS << "  return Match_MissingFeature;\n";
4107   }
4108   OS << "}\n\n";
4109 
4110   if (!Info.OperandMatchInfo.empty())
4111     emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
4112                              MaxMnemonicIndex, FeatureBitsets.size(),
4113                              HasMnemonicFirst, *AsmParser);
4114 
4115   OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
4116 
4117   OS << "\n#ifdef GET_MNEMONIC_SPELL_CHECKER\n";
4118   OS << "#undef GET_MNEMONIC_SPELL_CHECKER\n\n";
4119 
4120   emitMnemonicSpellChecker(OS, Target, VariantCount);
4121 
4122   OS << "#endif // GET_MNEMONIC_SPELL_CHECKER\n\n";
4123 
4124   OS << "\n#ifdef GET_MNEMONIC_CHECKER\n";
4125   OS << "#undef GET_MNEMONIC_CHECKER\n\n";
4126 
4127   emitMnemonicChecker(OS, Target, VariantCount, HasMnemonicFirst,
4128                       HasMnemonicAliases);
4129 
4130   OS << "#endif // GET_MNEMONIC_CHECKER\n\n";
4131 }
4132 
4133 static TableGen::Emitter::OptClass<AsmMatcherEmitter>
4134     X("gen-asm-matcher", "Generate assembly instruction matcher");
4135