1 //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This tablegen backend emits a target specifier matcher for converting parsed
11 // assembly operands in the MCInst structures. It also emits a matcher for
12 // custom operand parsing.
13 //
14 // Converting assembly operands into MCInst structures
15 // ---------------------------------------------------
16 //
17 // The input to the target specific matcher is a list of literal tokens and
18 // operands. The target specific parser should generally eliminate any syntax
19 // which is not relevant for matching; for example, comma tokens should have
20 // already been consumed and eliminated by the parser. Most instructions will
21 // end up with a single literal token (the instruction name) and some number of
22 // operands.
23 //
24 // Some example inputs, for X86:
25 // 'addl' (immediate ...) (register ...)
26 // 'add' (immediate ...) (memory ...)
27 // 'call' '*' %epc
28 //
29 // The assembly matcher is responsible for converting this input into a precise
30 // machine instruction (i.e., an instruction with a well defined encoding). This
31 // mapping has several properties which complicate matching:
32 //
33 // - It may be ambiguous; many architectures can legally encode particular
34 // variants of an instruction in different ways (for example, using a smaller
35 // encoding for small immediates). Such ambiguities should never be
36 // arbitrarily resolved by the assembler, the assembler is always responsible
37 // for choosing the "best" available instruction.
38 //
39 // - It may depend on the subtarget or the assembler context. Instructions
40 // which are invalid for the current mode, but otherwise unambiguous (e.g.,
41 // an SSE instruction in a file being assembled for i486) should be accepted
42 // and rejected by the assembler front end. However, if the proper encoding
43 // for an instruction is dependent on the assembler context then the matcher
44 // is responsible for selecting the correct machine instruction for the
45 // current mode.
46 //
47 // The core matching algorithm attempts to exploit the regularity in most
48 // instruction sets to quickly determine the set of possibly matching
49 // instructions, and the simplify the generated code. Additionally, this helps
50 // to ensure that the ambiguities are intentionally resolved by the user.
51 //
52 // The matching is divided into two distinct phases:
53 //
54 // 1. Classification: Each operand is mapped to the unique set which (a)
55 // contains it, and (b) is the largest such subset for which a single
56 // instruction could match all members.
57 //
58 // For register classes, we can generate these subgroups automatically. For
59 // arbitrary operands, we expect the user to define the classes and their
60 // relations to one another (for example, 8-bit signed immediates as a
61 // subset of 32-bit immediates).
62 //
63 // By partitioning the operands in this way, we guarantee that for any
64 // tuple of classes, any single instruction must match either all or none
65 // of the sets of operands which could classify to that tuple.
66 //
67 // In addition, the subset relation amongst classes induces a partial order
68 // on such tuples, which we use to resolve ambiguities.
69 //
70 // 2. The input can now be treated as a tuple of classes (static tokens are
71 // simple singleton sets). Each such tuple should generally map to a single
72 // instruction (we currently ignore cases where this isn't true, whee!!!),
73 // which we can emit a simple matcher for.
74 //
75 // Custom Operand Parsing
76 // ----------------------
77 //
78 // Some targets need a custom way to parse operands, some specific instructions
79 // can contain arguments that can represent processor flags and other kinds of
80 // identifiers that need to be mapped to specific values in the final encoded
81 // instructions. The target specific custom operand parsing works in the
82 // following way:
83 //
84 // 1. A operand match table is built, each entry contains a mnemonic, an
85 // operand class, a mask for all operand positions for that same
86 // class/mnemonic and target features to be checked while trying to match.
87 //
88 // 2. The operand matcher will try every possible entry with the same
89 // mnemonic and will check if the target feature for this mnemonic also
90 // matches. After that, if the operand to be matched has its index
91 // present in the mask, a successful match occurs. Otherwise, fallback
92 // to the regular operand parsing.
93 //
94 // 3. For a match success, each operand class that has a 'ParserMethod'
95 // becomes part of a switch from where the custom method is called.
96 //
97 //===----------------------------------------------------------------------===//
98
99 #include "CodeGenTarget.h"
100 #include "llvm/ADT/PointerUnion.h"
101 #include "llvm/ADT/STLExtras.h"
102 #include "llvm/ADT/SmallPtrSet.h"
103 #include "llvm/ADT/SmallVector.h"
104 #include "llvm/ADT/StringExtras.h"
105 #include "llvm/Support/CommandLine.h"
106 #include "llvm/Support/Debug.h"
107 #include "llvm/Support/ErrorHandling.h"
108 #include "llvm/TableGen/Error.h"
109 #include "llvm/TableGen/Record.h"
110 #include "llvm/TableGen/StringMatcher.h"
111 #include "llvm/TableGen/StringToOffsetTable.h"
112 #include "llvm/TableGen/TableGenBackend.h"
113 #include <cassert>
114 #include <cctype>
115 #include <map>
116 #include <set>
117 #include <sstream>
118 #include <forward_list>
119 using namespace llvm;
120
121 #define DEBUG_TYPE "asm-matcher-emitter"
122
123 static cl::opt<std::string>
124 MatchPrefix("match-prefix", cl::init(""),
125 cl::desc("Only match instructions with the given prefix"));
126
127 namespace {
128 class AsmMatcherInfo;
129 struct SubtargetFeatureInfo;
130
131 // Register sets are used as keys in some second-order sets TableGen creates
132 // when generating its data structures. This means that the order of two
133 // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and
134 // can even affect compiler output (at least seen in diagnostics produced when
135 // all matches fail). So we use a type that sorts them consistently.
136 typedef std::set<Record*, LessRecordByID> RegisterSet;
137
138 class AsmMatcherEmitter {
139 RecordKeeper &Records;
140 public:
AsmMatcherEmitter(RecordKeeper & R)141 AsmMatcherEmitter(RecordKeeper &R) : Records(R) {}
142
143 void run(raw_ostream &o);
144 };
145
146 /// ClassInfo - Helper class for storing the information about a particular
147 /// class of operands which can be matched.
148 struct ClassInfo {
149 enum ClassInfoKind {
150 /// Invalid kind, for use as a sentinel value.
151 Invalid = 0,
152
153 /// The class for a particular token.
154 Token,
155
156 /// The (first) register class, subsequent register classes are
157 /// RegisterClass0+1, and so on.
158 RegisterClass0,
159
160 /// The (first) user defined class, subsequent user defined classes are
161 /// UserClass0+1, and so on.
162 UserClass0 = 1<<16
163 };
164
165 /// Kind - The class kind, which is either a predefined kind, or (UserClass0 +
166 /// N) for the Nth user defined class.
167 unsigned Kind;
168
169 /// SuperClasses - The super classes of this class. Note that for simplicities
170 /// sake user operands only record their immediate super class, while register
171 /// operands include all superclasses.
172 std::vector<ClassInfo*> SuperClasses;
173
174 /// Name - The full class name, suitable for use in an enum.
175 std::string Name;
176
177 /// ClassName - The unadorned generic name for this class (e.g., Token).
178 std::string ClassName;
179
180 /// ValueName - The name of the value this class represents; for a token this
181 /// is the literal token string, for an operand it is the TableGen class (or
182 /// empty if this is a derived class).
183 std::string ValueName;
184
185 /// PredicateMethod - The name of the operand method to test whether the
186 /// operand matches this class; this is not valid for Token or register kinds.
187 std::string PredicateMethod;
188
189 /// RenderMethod - The name of the operand method to add this operand to an
190 /// MCInst; this is not valid for Token or register kinds.
191 std::string RenderMethod;
192
193 /// ParserMethod - The name of the operand method to do a target specific
194 /// parsing on the operand.
195 std::string ParserMethod;
196
197 /// For register classes: the records for all the registers in this class.
198 RegisterSet Registers;
199
200 /// For custom match classes: the diagnostic kind for when the predicate fails.
201 std::string DiagnosticType;
202 public:
203 /// isRegisterClass() - Check if this is a register class.
isRegisterClass__anon51f7b9bc0111::ClassInfo204 bool isRegisterClass() const {
205 return Kind >= RegisterClass0 && Kind < UserClass0;
206 }
207
208 /// isUserClass() - Check if this is a user defined class.
isUserClass__anon51f7b9bc0111::ClassInfo209 bool isUserClass() const {
210 return Kind >= UserClass0;
211 }
212
213 /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes
214 /// are related if they are in the same class hierarchy.
isRelatedTo__anon51f7b9bc0111::ClassInfo215 bool isRelatedTo(const ClassInfo &RHS) const {
216 // Tokens are only related to tokens.
217 if (Kind == Token || RHS.Kind == Token)
218 return Kind == Token && RHS.Kind == Token;
219
220 // Registers classes are only related to registers classes, and only if
221 // their intersection is non-empty.
222 if (isRegisterClass() || RHS.isRegisterClass()) {
223 if (!isRegisterClass() || !RHS.isRegisterClass())
224 return false;
225
226 RegisterSet Tmp;
227 std::insert_iterator<RegisterSet> II(Tmp, Tmp.begin());
228 std::set_intersection(Registers.begin(), Registers.end(),
229 RHS.Registers.begin(), RHS.Registers.end(),
230 II, LessRecordByID());
231
232 return !Tmp.empty();
233 }
234
235 // Otherwise we have two users operands; they are related if they are in the
236 // same class hierarchy.
237 //
238 // FIXME: This is an oversimplification, they should only be related if they
239 // intersect, however we don't have that information.
240 assert(isUserClass() && RHS.isUserClass() && "Unexpected class!");
241 const ClassInfo *Root = this;
242 while (!Root->SuperClasses.empty())
243 Root = Root->SuperClasses.front();
244
245 const ClassInfo *RHSRoot = &RHS;
246 while (!RHSRoot->SuperClasses.empty())
247 RHSRoot = RHSRoot->SuperClasses.front();
248
249 return Root == RHSRoot;
250 }
251
252 /// isSubsetOf - Test whether this class is a subset of \p RHS.
isSubsetOf__anon51f7b9bc0111::ClassInfo253 bool isSubsetOf(const ClassInfo &RHS) const {
254 // This is a subset of RHS if it is the same class...
255 if (this == &RHS)
256 return true;
257
258 // ... or if any of its super classes are a subset of RHS.
259 for (const ClassInfo *CI : SuperClasses)
260 if (CI->isSubsetOf(RHS))
261 return true;
262
263 return false;
264 }
265
266 /// operator< - Compare two classes.
267 // FIXME: This ordering seems to be broken. For example:
268 // u64 < i64, i64 < s8, s8 < u64, forming a cycle
269 // u64 is a subset of i64
270 // i64 and s8 are not subsets of each other, so are ordered by name
271 // s8 and u64 are not subsets of each other, so are ordered by name
operator <__anon51f7b9bc0111::ClassInfo272 bool operator<(const ClassInfo &RHS) const {
273 if (this == &RHS)
274 return false;
275
276 // Unrelated classes can be ordered by kind.
277 if (!isRelatedTo(RHS))
278 return Kind < RHS.Kind;
279
280 switch (Kind) {
281 case Invalid:
282 llvm_unreachable("Invalid kind!");
283
284 default:
285 // This class precedes the RHS if it is a proper subset of the RHS.
286 if (isSubsetOf(RHS))
287 return true;
288 if (RHS.isSubsetOf(*this))
289 return false;
290
291 // Otherwise, order by name to ensure we have a total ordering.
292 return ValueName < RHS.ValueName;
293 }
294 }
295 };
296
297 class AsmVariantInfo {
298 public:
299 std::string TokenizingCharacters;
300 std::string SeparatorCharacters;
301 std::string BreakCharacters;
302 };
303
304 /// MatchableInfo - Helper class for storing the necessary information for an
305 /// instruction or alias which is capable of being matched.
306 struct MatchableInfo {
307 struct AsmOperand {
308 /// Token - This is the token that the operand came from.
309 StringRef Token;
310
311 /// The unique class instance this operand should match.
312 ClassInfo *Class;
313
314 /// The operand name this is, if anything.
315 StringRef SrcOpName;
316
317 /// The suboperand index within SrcOpName, or -1 for the entire operand.
318 int SubOpIdx;
319
320 /// Whether the token is "isolated", i.e., it is preceded and followed
321 /// by separators.
322 bool IsIsolatedToken;
323
324 /// Register record if this token is singleton register.
325 Record *SingletonReg;
326
AsmOperand__anon51f7b9bc0111::MatchableInfo::AsmOperand327 explicit AsmOperand(bool IsIsolatedToken, StringRef T)
328 : Token(T), Class(nullptr), SubOpIdx(-1),
329 IsIsolatedToken(IsIsolatedToken), SingletonReg(nullptr) {}
330 };
331
332 /// ResOperand - This represents a single operand in the result instruction
333 /// generated by the match. In cases (like addressing modes) where a single
334 /// assembler operand expands to multiple MCOperands, this represents the
335 /// single assembler operand, not the MCOperand.
336 struct ResOperand {
337 enum {
338 /// RenderAsmOperand - This represents an operand result that is
339 /// generated by calling the render method on the assembly operand. The
340 /// corresponding AsmOperand is specified by AsmOperandNum.
341 RenderAsmOperand,
342
343 /// TiedOperand - This represents a result operand that is a duplicate of
344 /// a previous result operand.
345 TiedOperand,
346
347 /// ImmOperand - This represents an immediate value that is dumped into
348 /// the operand.
349 ImmOperand,
350
351 /// RegOperand - This represents a fixed register that is dumped in.
352 RegOperand
353 } Kind;
354
355 union {
356 /// This is the operand # in the AsmOperands list that this should be
357 /// copied from.
358 unsigned AsmOperandNum;
359
360 /// TiedOperandNum - This is the (earlier) result operand that should be
361 /// copied from.
362 unsigned TiedOperandNum;
363
364 /// ImmVal - This is the immediate value added to the instruction.
365 int64_t ImmVal;
366
367 /// Register - This is the register record.
368 Record *Register;
369 };
370
371 /// MINumOperands - The number of MCInst operands populated by this
372 /// operand.
373 unsigned MINumOperands;
374
getRenderedOp__anon51f7b9bc0111::MatchableInfo::ResOperand375 static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) {
376 ResOperand X;
377 X.Kind = RenderAsmOperand;
378 X.AsmOperandNum = AsmOpNum;
379 X.MINumOperands = NumOperands;
380 return X;
381 }
382
getTiedOp__anon51f7b9bc0111::MatchableInfo::ResOperand383 static ResOperand getTiedOp(unsigned TiedOperandNum) {
384 ResOperand X;
385 X.Kind = TiedOperand;
386 X.TiedOperandNum = TiedOperandNum;
387 X.MINumOperands = 1;
388 return X;
389 }
390
getImmOp__anon51f7b9bc0111::MatchableInfo::ResOperand391 static ResOperand getImmOp(int64_t Val) {
392 ResOperand X;
393 X.Kind = ImmOperand;
394 X.ImmVal = Val;
395 X.MINumOperands = 1;
396 return X;
397 }
398
getRegOp__anon51f7b9bc0111::MatchableInfo::ResOperand399 static ResOperand getRegOp(Record *Reg) {
400 ResOperand X;
401 X.Kind = RegOperand;
402 X.Register = Reg;
403 X.MINumOperands = 1;
404 return X;
405 }
406 };
407
408 /// AsmVariantID - Target's assembly syntax variant no.
409 int AsmVariantID;
410
411 /// AsmString - The assembly string for this instruction (with variants
412 /// removed), e.g. "movsx $src, $dst".
413 std::string AsmString;
414
415 /// TheDef - This is the definition of the instruction or InstAlias that this
416 /// matchable came from.
417 Record *const TheDef;
418
419 /// DefRec - This is the definition that it came from.
420 PointerUnion<const CodeGenInstruction*, const CodeGenInstAlias*> DefRec;
421
getResultInst__anon51f7b9bc0111::MatchableInfo422 const CodeGenInstruction *getResultInst() const {
423 if (DefRec.is<const CodeGenInstruction*>())
424 return DefRec.get<const CodeGenInstruction*>();
425 return DefRec.get<const CodeGenInstAlias*>()->ResultInst;
426 }
427
428 /// ResOperands - This is the operand list that should be built for the result
429 /// MCInst.
430 SmallVector<ResOperand, 8> ResOperands;
431
432 /// Mnemonic - This is the first token of the matched instruction, its
433 /// mnemonic.
434 StringRef Mnemonic;
435
436 /// AsmOperands - The textual operands that this instruction matches,
437 /// annotated with a class and where in the OperandList they were defined.
438 /// This directly corresponds to the tokenized AsmString after the mnemonic is
439 /// removed.
440 SmallVector<AsmOperand, 8> AsmOperands;
441
442 /// Predicates - The required subtarget features to match this instruction.
443 SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures;
444
445 /// ConversionFnKind - The enum value which is passed to the generated
446 /// convertToMCInst to convert parsed operands into an MCInst for this
447 /// function.
448 std::string ConversionFnKind;
449
450 /// If this instruction is deprecated in some form.
451 bool HasDeprecation;
452
453 /// If this is an alias, this is use to determine whether or not to using
454 /// the conversion function defined by the instruction's AsmMatchConverter
455 /// or to use the function generated by the alias.
456 bool UseInstAsmMatchConverter;
457
MatchableInfo__anon51f7b9bc0111::MatchableInfo458 MatchableInfo(const CodeGenInstruction &CGI)
459 : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), DefRec(&CGI),
460 UseInstAsmMatchConverter(true) {
461 }
462
MatchableInfo__anon51f7b9bc0111::MatchableInfo463 MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias)
464 : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef),
465 DefRec(Alias.release()),
466 UseInstAsmMatchConverter(
467 TheDef->getValueAsBit("UseInstAsmMatchConverter")) {
468 }
469
470 // Could remove this and the dtor if PointerUnion supported unique_ptr
471 // elements with a dynamic failure/assertion (like the one below) in the case
472 // where it was copied while being in an owning state.
MatchableInfo__anon51f7b9bc0111::MatchableInfo473 MatchableInfo(const MatchableInfo &RHS)
474 : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString),
475 TheDef(RHS.TheDef), DefRec(RHS.DefRec), ResOperands(RHS.ResOperands),
476 Mnemonic(RHS.Mnemonic), AsmOperands(RHS.AsmOperands),
477 RequiredFeatures(RHS.RequiredFeatures),
478 ConversionFnKind(RHS.ConversionFnKind),
479 HasDeprecation(RHS.HasDeprecation),
480 UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) {
481 assert(!DefRec.is<const CodeGenInstAlias *>());
482 }
483
~MatchableInfo__anon51f7b9bc0111::MatchableInfo484 ~MatchableInfo() {
485 delete DefRec.dyn_cast<const CodeGenInstAlias*>();
486 }
487
488 // Two-operand aliases clone from the main matchable, but mark the second
489 // operand as a tied operand of the first for purposes of the assembler.
490 void formTwoOperandAlias(StringRef Constraint);
491
492 void initialize(const AsmMatcherInfo &Info,
493 SmallPtrSetImpl<Record*> &SingletonRegisters,
494 int AsmVariantNo, StringRef RegisterPrefix,
495 AsmVariantInfo const &Variant);
496
497 /// validate - Return true if this matchable is a valid thing to match against
498 /// and perform a bunch of validity checking.
499 bool validate(StringRef CommentDelimiter, bool Hack) const;
500
501 /// findAsmOperand - Find the AsmOperand with the specified name and
502 /// suboperand index.
findAsmOperand__anon51f7b9bc0111::MatchableInfo503 int findAsmOperand(StringRef N, int SubOpIdx) const {
504 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
505 if (N == AsmOperands[i].SrcOpName &&
506 SubOpIdx == AsmOperands[i].SubOpIdx)
507 return i;
508 return -1;
509 }
510
511 /// findAsmOperandNamed - Find the first AsmOperand with the specified name.
512 /// This does not check the suboperand index.
findAsmOperandNamed__anon51f7b9bc0111::MatchableInfo513 int findAsmOperandNamed(StringRef N) const {
514 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
515 if (N == AsmOperands[i].SrcOpName)
516 return i;
517 return -1;
518 }
519
520 void buildInstructionResultOperands();
521 void buildAliasResultOperands();
522
523 /// operator< - Compare two matchables.
operator <__anon51f7b9bc0111::MatchableInfo524 bool operator<(const MatchableInfo &RHS) const {
525 // The primary comparator is the instruction mnemonic.
526 if (Mnemonic != RHS.Mnemonic)
527 return Mnemonic < RHS.Mnemonic;
528
529 if (AsmOperands.size() != RHS.AsmOperands.size())
530 return AsmOperands.size() < RHS.AsmOperands.size();
531
532 // Compare lexicographically by operand. The matcher validates that other
533 // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith().
534 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
535 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
536 return true;
537 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
538 return false;
539 }
540
541 // Give matches that require more features higher precedence. This is useful
542 // because we cannot define AssemblerPredicates with the negation of
543 // processor features. For example, ARM v6 "nop" may be either a HINT or
544 // MOV. With v6, we want to match HINT. The assembler has no way to
545 // predicate MOV under "NoV6", but HINT will always match first because it
546 // requires V6 while MOV does not.
547 if (RequiredFeatures.size() != RHS.RequiredFeatures.size())
548 return RequiredFeatures.size() > RHS.RequiredFeatures.size();
549
550 return false;
551 }
552
553 /// couldMatchAmbiguouslyWith - Check whether this matchable could
554 /// ambiguously match the same set of operands as \p RHS (without being a
555 /// strictly superior match).
couldMatchAmbiguouslyWith__anon51f7b9bc0111::MatchableInfo556 bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS) const {
557 // The primary comparator is the instruction mnemonic.
558 if (Mnemonic != RHS.Mnemonic)
559 return false;
560
561 // The number of operands is unambiguous.
562 if (AsmOperands.size() != RHS.AsmOperands.size())
563 return false;
564
565 // Otherwise, make sure the ordering of the two instructions is unambiguous
566 // by checking that either (a) a token or operand kind discriminates them,
567 // or (b) the ordering among equivalent kinds is consistent.
568
569 // Tokens and operand kinds are unambiguous (assuming a correct target
570 // specific parser).
571 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i)
572 if (AsmOperands[i].Class->Kind != RHS.AsmOperands[i].Class->Kind ||
573 AsmOperands[i].Class->Kind == ClassInfo::Token)
574 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class ||
575 *RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
576 return false;
577
578 // Otherwise, this operand could commute if all operands are equivalent, or
579 // there is a pair of operands that compare less than and a pair that
580 // compare greater than.
581 bool HasLT = false, HasGT = false;
582 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
583 if (*AsmOperands[i].Class < *RHS.AsmOperands[i].Class)
584 HasLT = true;
585 if (*RHS.AsmOperands[i].Class < *AsmOperands[i].Class)
586 HasGT = true;
587 }
588
589 return !(HasLT ^ HasGT);
590 }
591
592 void dump() const;
593
594 private:
595 void tokenizeAsmString(AsmMatcherInfo const &Info,
596 AsmVariantInfo const &Variant);
597 void addAsmOperand(size_t Start, size_t End,
598 std::string const &SeparatorCharacters);
599 };
600
601 /// SubtargetFeatureInfo - Helper class for storing information on a subtarget
602 /// feature which participates in instruction matching.
603 struct SubtargetFeatureInfo {
604 /// \brief The predicate record for this feature.
605 Record *TheDef;
606
607 /// \brief An unique index assigned to represent this feature.
608 uint64_t Index;
609
SubtargetFeatureInfo__anon51f7b9bc0111::SubtargetFeatureInfo610 SubtargetFeatureInfo(Record *D, uint64_t Idx) : TheDef(D), Index(Idx) {}
611
612 /// \brief The name of the enumerated constant identifying this feature.
getEnumName__anon51f7b9bc0111::SubtargetFeatureInfo613 std::string getEnumName() const {
614 return "Feature_" + TheDef->getName();
615 }
616
dump__anon51f7b9bc0111::SubtargetFeatureInfo617 void dump() const {
618 errs() << getEnumName() << " " << Index << "\n";
619 TheDef->dump();
620 }
621 };
622
623 struct OperandMatchEntry {
624 unsigned OperandMask;
625 const MatchableInfo* MI;
626 ClassInfo *CI;
627
create__anon51f7b9bc0111::OperandMatchEntry628 static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci,
629 unsigned opMask) {
630 OperandMatchEntry X;
631 X.OperandMask = opMask;
632 X.CI = ci;
633 X.MI = mi;
634 return X;
635 }
636 };
637
638
639 class AsmMatcherInfo {
640 public:
641 /// Tracked Records
642 RecordKeeper &Records;
643
644 /// The tablegen AsmParser record.
645 Record *AsmParser;
646
647 /// Target - The target information.
648 CodeGenTarget &Target;
649
650 /// The classes which are needed for matching.
651 std::forward_list<ClassInfo> Classes;
652
653 /// The information on the matchables to match.
654 std::vector<std::unique_ptr<MatchableInfo>> Matchables;
655
656 /// Info for custom matching operands by user defined methods.
657 std::vector<OperandMatchEntry> OperandMatchInfo;
658
659 /// Map of Register records to their class information.
660 typedef std::map<Record*, ClassInfo*, LessRecordByID> RegisterClassesTy;
661 RegisterClassesTy RegisterClasses;
662
663 /// Map of Predicate records to their subtarget information.
664 std::map<Record *, SubtargetFeatureInfo, LessRecordByID> SubtargetFeatures;
665
666 /// Map of AsmOperandClass records to their class information.
667 std::map<Record*, ClassInfo*> AsmOperandClasses;
668
669 private:
670 /// Map of token to class information which has already been constructed.
671 std::map<std::string, ClassInfo*> TokenClasses;
672
673 /// Map of RegisterClass records to their class information.
674 std::map<Record*, ClassInfo*> RegisterClassClasses;
675
676 private:
677 /// getTokenClass - Lookup or create the class for the given token.
678 ClassInfo *getTokenClass(StringRef Token);
679
680 /// getOperandClass - Lookup or create the class for the given operand.
681 ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI,
682 int SubOpIdx);
683 ClassInfo *getOperandClass(Record *Rec, int SubOpIdx);
684
685 /// buildRegisterClasses - Build the ClassInfo* instances for register
686 /// classes.
687 void buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters);
688
689 /// buildOperandClasses - Build the ClassInfo* instances for user defined
690 /// operand classes.
691 void buildOperandClasses();
692
693 void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName,
694 unsigned AsmOpIdx);
695 void buildAliasOperandReference(MatchableInfo *II, StringRef OpName,
696 MatchableInfo::AsmOperand &Op);
697
698 public:
699 AsmMatcherInfo(Record *AsmParser,
700 CodeGenTarget &Target,
701 RecordKeeper &Records);
702
703 /// buildInfo - Construct the various tables used during matching.
704 void buildInfo();
705
706 /// buildOperandMatchInfo - Build the necessary information to handle user
707 /// defined operand parsing methods.
708 void buildOperandMatchInfo();
709
710 /// getSubtargetFeature - Lookup or create the subtarget feature info for the
711 /// given operand.
getSubtargetFeature(Record * Def) const712 const SubtargetFeatureInfo *getSubtargetFeature(Record *Def) const {
713 assert(Def->isSubClassOf("Predicate") && "Invalid predicate type!");
714 const auto &I = SubtargetFeatures.find(Def);
715 return I == SubtargetFeatures.end() ? nullptr : &I->second;
716 }
717
getRecords() const718 RecordKeeper &getRecords() const {
719 return Records;
720 }
721 };
722
723 } // End anonymous namespace
724
dump() const725 void MatchableInfo::dump() const {
726 errs() << TheDef->getName() << " -- " << "flattened:\"" << AsmString <<"\"\n";
727
728 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
729 const AsmOperand &Op = AsmOperands[i];
730 errs() << " op[" << i << "] = " << Op.Class->ClassName << " - ";
731 errs() << '\"' << Op.Token << "\"\n";
732 }
733 }
734
735 static std::pair<StringRef, StringRef>
parseTwoOperandConstraint(StringRef S,ArrayRef<SMLoc> Loc)736 parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) {
737 // Split via the '='.
738 std::pair<StringRef, StringRef> Ops = S.split('=');
739 if (Ops.second == "")
740 PrintFatalError(Loc, "missing '=' in two-operand alias constraint");
741 // Trim whitespace and the leading '$' on the operand names.
742 size_t start = Ops.first.find_first_of('$');
743 if (start == std::string::npos)
744 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
745 Ops.first = Ops.first.slice(start + 1, std::string::npos);
746 size_t end = Ops.first.find_last_of(" \t");
747 Ops.first = Ops.first.slice(0, end);
748 // Now the second operand.
749 start = Ops.second.find_first_of('$');
750 if (start == std::string::npos)
751 PrintFatalError(Loc, "expected '$' prefix on asm operand name");
752 Ops.second = Ops.second.slice(start + 1, std::string::npos);
753 end = Ops.second.find_last_of(" \t");
754 Ops.first = Ops.first.slice(0, end);
755 return Ops;
756 }
757
formTwoOperandAlias(StringRef Constraint)758 void MatchableInfo::formTwoOperandAlias(StringRef Constraint) {
759 // Figure out which operands are aliased and mark them as tied.
760 std::pair<StringRef, StringRef> Ops =
761 parseTwoOperandConstraint(Constraint, TheDef->getLoc());
762
763 // Find the AsmOperands that refer to the operands we're aliasing.
764 int SrcAsmOperand = findAsmOperandNamed(Ops.first);
765 int DstAsmOperand = findAsmOperandNamed(Ops.second);
766 if (SrcAsmOperand == -1)
767 PrintFatalError(TheDef->getLoc(),
768 "unknown source two-operand alias operand '" + Ops.first +
769 "'.");
770 if (DstAsmOperand == -1)
771 PrintFatalError(TheDef->getLoc(),
772 "unknown destination two-operand alias operand '" +
773 Ops.second + "'.");
774
775 // Find the ResOperand that refers to the operand we're aliasing away
776 // and update it to refer to the combined operand instead.
777 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
778 ResOperand &Op = ResOperands[i];
779 if (Op.Kind == ResOperand::RenderAsmOperand &&
780 Op.AsmOperandNum == (unsigned)SrcAsmOperand) {
781 Op.AsmOperandNum = DstAsmOperand;
782 break;
783 }
784 }
785 // Remove the AsmOperand for the alias operand.
786 AsmOperands.erase(AsmOperands.begin() + SrcAsmOperand);
787 // Adjust the ResOperand references to any AsmOperands that followed
788 // the one we just deleted.
789 for (unsigned i = 0, e = ResOperands.size(); i != e; ++i) {
790 ResOperand &Op = ResOperands[i];
791 switch(Op.Kind) {
792 default:
793 // Nothing to do for operands that don't reference AsmOperands.
794 break;
795 case ResOperand::RenderAsmOperand:
796 if (Op.AsmOperandNum > (unsigned)SrcAsmOperand)
797 --Op.AsmOperandNum;
798 break;
799 case ResOperand::TiedOperand:
800 if (Op.TiedOperandNum > (unsigned)SrcAsmOperand)
801 --Op.TiedOperandNum;
802 break;
803 }
804 }
805 }
806
807 /// extractSingletonRegisterForAsmOperand - Extract singleton register,
808 /// if present, from specified token.
809 static void
extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand & Op,const AsmMatcherInfo & Info,StringRef RegisterPrefix)810 extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op,
811 const AsmMatcherInfo &Info,
812 StringRef RegisterPrefix) {
813 StringRef Tok = Op.Token;
814
815 // If this token is not an isolated token, i.e., it isn't separated from
816 // other tokens (e.g. with whitespace), don't interpret it as a register name.
817 if (!Op.IsIsolatedToken)
818 return;
819
820 if (RegisterPrefix.empty()) {
821 std::string LoweredTok = Tok.lower();
822 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(LoweredTok))
823 Op.SingletonReg = Reg->TheDef;
824 return;
825 }
826
827 if (!Tok.startswith(RegisterPrefix))
828 return;
829
830 StringRef RegName = Tok.substr(RegisterPrefix.size());
831 if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(RegName))
832 Op.SingletonReg = Reg->TheDef;
833
834 // If there is no register prefix (i.e. "%" in "%eax"), then this may
835 // be some random non-register token, just ignore it.
836 return;
837 }
838
initialize(const AsmMatcherInfo & Info,SmallPtrSetImpl<Record * > & SingletonRegisters,int AsmVariantNo,StringRef RegisterPrefix,AsmVariantInfo const & Variant)839 void MatchableInfo::initialize(const AsmMatcherInfo &Info,
840 SmallPtrSetImpl<Record*> &SingletonRegisters,
841 int AsmVariantNo, StringRef RegisterPrefix,
842 AsmVariantInfo const &Variant) {
843 AsmVariantID = AsmVariantNo;
844 AsmString =
845 CodeGenInstruction::FlattenAsmStringVariants(AsmString, AsmVariantNo);
846
847 tokenizeAsmString(Info, Variant);
848
849 // Compute the require features.
850 for (Record *Predicate : TheDef->getValueAsListOfDefs("Predicates"))
851 if (const SubtargetFeatureInfo *Feature =
852 Info.getSubtargetFeature(Predicate))
853 RequiredFeatures.push_back(Feature);
854
855 // Collect singleton registers, if used.
856 for (MatchableInfo::AsmOperand &Op : AsmOperands) {
857 extractSingletonRegisterForAsmOperand(Op, Info, RegisterPrefix);
858 if (Record *Reg = Op.SingletonReg)
859 SingletonRegisters.insert(Reg);
860 }
861
862 const RecordVal *DepMask = TheDef->getValue("DeprecatedFeatureMask");
863 if (!DepMask)
864 DepMask = TheDef->getValue("ComplexDeprecationPredicate");
865
866 HasDeprecation =
867 DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false;
868 }
869
870 /// Append an AsmOperand for the given substring of AsmString.
addAsmOperand(size_t Start,size_t End,std::string const & Separators)871 void MatchableInfo::addAsmOperand(size_t Start, size_t End,
872 std::string const &Separators) {
873 StringRef String = AsmString;
874 // Look for separators before and after to figure out is this token is
875 // isolated. Accept '$$' as that's how we escape '$'.
876 bool IsIsolatedToken =
877 (!Start || Separators.find(String[Start - 1]) != StringRef::npos ||
878 String.substr(Start - 1, 2) == "$$") &&
879 (End >= String.size() || Separators.find(String[End]) != StringRef::npos);
880 AsmOperands.push_back(AsmOperand(IsIsolatedToken, String.slice(Start, End)));
881 }
882
883 /// tokenizeAsmString - Tokenize a simplified assembly string.
tokenizeAsmString(const AsmMatcherInfo & Info,AsmVariantInfo const & Variant)884 void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info,
885 AsmVariantInfo const &Variant) {
886 StringRef String = AsmString;
887 unsigned Prev = 0;
888 bool InTok = false;
889 std::string Separators = Variant.TokenizingCharacters +
890 Variant.SeparatorCharacters;
891 for (unsigned i = 0, e = String.size(); i != e; ++i) {
892 if(Variant.BreakCharacters.find(String[i]) != std::string::npos) {
893 if(InTok) {
894 addAsmOperand(Prev, i, Separators);
895 Prev = i;
896 }
897 InTok = true;
898 continue;
899 }
900 if(Variant.TokenizingCharacters.find(String[i]) != std::string::npos) {
901 if(InTok) {
902 addAsmOperand(Prev, i, Separators);
903 InTok = false;
904 }
905 addAsmOperand(i, i + 1, Separators);
906 Prev = i + 1;
907 continue;
908 }
909 if(Variant.SeparatorCharacters.find(String[i]) != std::string::npos) {
910 if(InTok) {
911 addAsmOperand(Prev, i, Separators);
912 InTok = false;
913 }
914 Prev = i + 1;
915 continue;
916 }
917 switch (String[i]) {
918 case '\\':
919 if (InTok) {
920 addAsmOperand(Prev, i, Separators);
921 InTok = false;
922 }
923 ++i;
924 assert(i != String.size() && "Invalid quoted character");
925 addAsmOperand(i, i + 1, Separators);
926 Prev = i + 1;
927 break;
928
929 case '$': {
930 if (InTok && Prev != i) {
931 addAsmOperand(Prev, i, Separators);
932 InTok = false;
933 }
934
935 // If this isn't "${", start new identifier looking like "$xxx"
936 if (i + 1 == String.size() || String[i + 1] != '{') {
937 Prev = i;
938 break;
939 }
940
941 StringRef::iterator End = std::find(String.begin() + i, String.end(),'}');
942 assert(End != String.end() && "Missing brace in operand reference!");
943 size_t EndPos = End - String.begin();
944 addAsmOperand(i, EndPos+1, Separators);
945 Prev = EndPos + 1;
946 i = EndPos;
947 break;
948 }
949 default:
950 InTok = true;
951 }
952 }
953 if (InTok && Prev != String.size())
954 addAsmOperand(Prev, StringRef::npos, Separators);
955
956 // The first token of the instruction is the mnemonic, which must be a
957 // simple string, not a $foo variable or a singleton register.
958 if (AsmOperands.empty())
959 PrintFatalError(TheDef->getLoc(),
960 "Instruction '" + TheDef->getName() + "' has no tokens");
961 assert(!AsmOperands[0].Token.empty());
962 if (AsmOperands[0].Token[0] != '$')
963 Mnemonic = AsmOperands[0].Token;
964 }
965
validate(StringRef CommentDelimiter,bool Hack) const966 bool MatchableInfo::validate(StringRef CommentDelimiter, bool Hack) const {
967 // Reject matchables with no .s string.
968 if (AsmString.empty())
969 PrintFatalError(TheDef->getLoc(), "instruction with empty asm string");
970
971 // Reject any matchables with a newline in them, they should be marked
972 // isCodeGenOnly if they are pseudo instructions.
973 if (AsmString.find('\n') != std::string::npos)
974 PrintFatalError(TheDef->getLoc(),
975 "multiline instruction is not valid for the asmparser, "
976 "mark it isCodeGenOnly");
977
978 // Remove comments from the asm string. We know that the asmstring only
979 // has one line.
980 if (!CommentDelimiter.empty() &&
981 StringRef(AsmString).find(CommentDelimiter) != StringRef::npos)
982 PrintFatalError(TheDef->getLoc(),
983 "asmstring for instruction has comment character in it, "
984 "mark it isCodeGenOnly");
985
986 // Reject matchables with operand modifiers, these aren't something we can
987 // handle, the target should be refactored to use operands instead of
988 // modifiers.
989 //
990 // Also, check for instructions which reference the operand multiple times;
991 // this implies a constraint we would not honor.
992 std::set<std::string> OperandNames;
993 for (unsigned i = 0, e = AsmOperands.size(); i != e; ++i) {
994 StringRef Tok = AsmOperands[i].Token;
995 if (Tok[0] == '$' && Tok.find(':') != StringRef::npos)
996 PrintFatalError(TheDef->getLoc(),
997 "matchable with operand modifier '" + Tok +
998 "' not supported by asm matcher. Mark isCodeGenOnly!");
999
1000 // Verify that any operand is only mentioned once.
1001 // We reject aliases and ignore instructions for now.
1002 if (Tok[0] == '$' && !OperandNames.insert(Tok).second) {
1003 if (!Hack)
1004 PrintFatalError(TheDef->getLoc(),
1005 "ERROR: matchable with tied operand '" + Tok +
1006 "' can never be matched!");
1007 // FIXME: Should reject these. The ARM backend hits this with $lane in a
1008 // bunch of instructions. It is unclear what the right answer is.
1009 DEBUG({
1010 errs() << "warning: '" << TheDef->getName() << "': "
1011 << "ignoring instruction with tied operand '"
1012 << Tok << "'\n";
1013 });
1014 return false;
1015 }
1016 }
1017
1018 return true;
1019 }
1020
getEnumNameForToken(StringRef Str)1021 static std::string getEnumNameForToken(StringRef Str) {
1022 std::string Res;
1023
1024 for (StringRef::iterator it = Str.begin(), ie = Str.end(); it != ie; ++it) {
1025 switch (*it) {
1026 case '*': Res += "_STAR_"; break;
1027 case '%': Res += "_PCT_"; break;
1028 case ':': Res += "_COLON_"; break;
1029 case '!': Res += "_EXCLAIM_"; break;
1030 case '.': Res += "_DOT_"; break;
1031 case '<': Res += "_LT_"; break;
1032 case '>': Res += "_GT_"; break;
1033 case '-': Res += "_MINUS_"; break;
1034 default:
1035 if ((*it >= 'A' && *it <= 'Z') ||
1036 (*it >= 'a' && *it <= 'z') ||
1037 (*it >= '0' && *it <= '9'))
1038 Res += *it;
1039 else
1040 Res += "_" + utostr((unsigned) *it) + "_";
1041 }
1042 }
1043
1044 return Res;
1045 }
1046
getTokenClass(StringRef Token)1047 ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) {
1048 ClassInfo *&Entry = TokenClasses[Token];
1049
1050 if (!Entry) {
1051 Classes.emplace_front();
1052 Entry = &Classes.front();
1053 Entry->Kind = ClassInfo::Token;
1054 Entry->ClassName = "Token";
1055 Entry->Name = "MCK_" + getEnumNameForToken(Token);
1056 Entry->ValueName = Token;
1057 Entry->PredicateMethod = "<invalid>";
1058 Entry->RenderMethod = "<invalid>";
1059 Entry->ParserMethod = "";
1060 Entry->DiagnosticType = "";
1061 }
1062
1063 return Entry;
1064 }
1065
1066 ClassInfo *
getOperandClass(const CGIOperandList::OperandInfo & OI,int SubOpIdx)1067 AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI,
1068 int SubOpIdx) {
1069 Record *Rec = OI.Rec;
1070 if (SubOpIdx != -1)
1071 Rec = cast<DefInit>(OI.MIOperandInfo->getArg(SubOpIdx))->getDef();
1072 return getOperandClass(Rec, SubOpIdx);
1073 }
1074
1075 ClassInfo *
getOperandClass(Record * Rec,int SubOpIdx)1076 AsmMatcherInfo::getOperandClass(Record *Rec, int SubOpIdx) {
1077 if (Rec->isSubClassOf("RegisterOperand")) {
1078 // RegisterOperand may have an associated ParserMatchClass. If it does,
1079 // use it, else just fall back to the underlying register class.
1080 const RecordVal *R = Rec->getValue("ParserMatchClass");
1081 if (!R || !R->getValue())
1082 PrintFatalError("Record `" + Rec->getName() +
1083 "' does not have a ParserMatchClass!\n");
1084
1085 if (DefInit *DI= dyn_cast<DefInit>(R->getValue())) {
1086 Record *MatchClass = DI->getDef();
1087 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1088 return CI;
1089 }
1090
1091 // No custom match class. Just use the register class.
1092 Record *ClassRec = Rec->getValueAsDef("RegClass");
1093 if (!ClassRec)
1094 PrintFatalError(Rec->getLoc(), "RegisterOperand `" + Rec->getName() +
1095 "' has no associated register class!\n");
1096 if (ClassInfo *CI = RegisterClassClasses[ClassRec])
1097 return CI;
1098 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1099 }
1100
1101
1102 if (Rec->isSubClassOf("RegisterClass")) {
1103 if (ClassInfo *CI = RegisterClassClasses[Rec])
1104 return CI;
1105 PrintFatalError(Rec->getLoc(), "register class has no class info!");
1106 }
1107
1108 if (!Rec->isSubClassOf("Operand"))
1109 PrintFatalError(Rec->getLoc(), "Operand `" + Rec->getName() +
1110 "' does not derive from class Operand!\n");
1111 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1112 if (ClassInfo *CI = AsmOperandClasses[MatchClass])
1113 return CI;
1114
1115 PrintFatalError(Rec->getLoc(), "operand has no match class!");
1116 }
1117
1118 struct LessRegisterSet {
operator ()LessRegisterSet1119 bool operator() (const RegisterSet &LHS, const RegisterSet & RHS) const {
1120 // std::set<T> defines its own compariso "operator<", but it
1121 // performs a lexicographical comparison by T's innate comparison
1122 // for some reason. We don't want non-deterministic pointer
1123 // comparisons so use this instead.
1124 return std::lexicographical_compare(LHS.begin(), LHS.end(),
1125 RHS.begin(), RHS.end(),
1126 LessRecordByID());
1127 }
1128 };
1129
1130 void AsmMatcherInfo::
buildRegisterClasses(SmallPtrSetImpl<Record * > & SingletonRegisters)1131 buildRegisterClasses(SmallPtrSetImpl<Record*> &SingletonRegisters) {
1132 const auto &Registers = Target.getRegBank().getRegisters();
1133 auto &RegClassList = Target.getRegBank().getRegClasses();
1134
1135 typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet;
1136
1137 // The register sets used for matching.
1138 RegisterSetSet RegisterSets;
1139
1140 // Gather the defined sets.
1141 for (const CodeGenRegisterClass &RC : RegClassList)
1142 RegisterSets.insert(
1143 RegisterSet(RC.getOrder().begin(), RC.getOrder().end()));
1144
1145 // Add any required singleton sets.
1146 for (Record *Rec : SingletonRegisters) {
1147 RegisterSets.insert(RegisterSet(&Rec, &Rec + 1));
1148 }
1149
1150 // Introduce derived sets where necessary (when a register does not determine
1151 // a unique register set class), and build the mapping of registers to the set
1152 // they should classify to.
1153 std::map<Record*, RegisterSet> RegisterMap;
1154 for (const CodeGenRegister &CGR : Registers) {
1155 // Compute the intersection of all sets containing this register.
1156 RegisterSet ContainingSet;
1157
1158 for (const RegisterSet &RS : RegisterSets) {
1159 if (!RS.count(CGR.TheDef))
1160 continue;
1161
1162 if (ContainingSet.empty()) {
1163 ContainingSet = RS;
1164 continue;
1165 }
1166
1167 RegisterSet Tmp;
1168 std::swap(Tmp, ContainingSet);
1169 std::insert_iterator<RegisterSet> II(ContainingSet,
1170 ContainingSet.begin());
1171 std::set_intersection(Tmp.begin(), Tmp.end(), RS.begin(), RS.end(), II,
1172 LessRecordByID());
1173 }
1174
1175 if (!ContainingSet.empty()) {
1176 RegisterSets.insert(ContainingSet);
1177 RegisterMap.insert(std::make_pair(CGR.TheDef, ContainingSet));
1178 }
1179 }
1180
1181 // Construct the register classes.
1182 std::map<RegisterSet, ClassInfo*, LessRegisterSet> RegisterSetClasses;
1183 unsigned Index = 0;
1184 for (const RegisterSet &RS : RegisterSets) {
1185 Classes.emplace_front();
1186 ClassInfo *CI = &Classes.front();
1187 CI->Kind = ClassInfo::RegisterClass0 + Index;
1188 CI->ClassName = "Reg" + utostr(Index);
1189 CI->Name = "MCK_Reg" + utostr(Index);
1190 CI->ValueName = "";
1191 CI->PredicateMethod = ""; // unused
1192 CI->RenderMethod = "addRegOperands";
1193 CI->Registers = RS;
1194 // FIXME: diagnostic type.
1195 CI->DiagnosticType = "";
1196 RegisterSetClasses.insert(std::make_pair(RS, CI));
1197 ++Index;
1198 }
1199
1200 // Find the superclasses; we could compute only the subgroup lattice edges,
1201 // but there isn't really a point.
1202 for (const RegisterSet &RS : RegisterSets) {
1203 ClassInfo *CI = RegisterSetClasses[RS];
1204 for (const RegisterSet &RS2 : RegisterSets)
1205 if (RS != RS2 &&
1206 std::includes(RS2.begin(), RS2.end(), RS.begin(), RS.end(),
1207 LessRecordByID()))
1208 CI->SuperClasses.push_back(RegisterSetClasses[RS2]);
1209 }
1210
1211 // Name the register classes which correspond to a user defined RegisterClass.
1212 for (const CodeGenRegisterClass &RC : RegClassList) {
1213 // Def will be NULL for non-user defined register classes.
1214 Record *Def = RC.getDef();
1215 if (!Def)
1216 continue;
1217 ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(),
1218 RC.getOrder().end())];
1219 if (CI->ValueName.empty()) {
1220 CI->ClassName = RC.getName();
1221 CI->Name = "MCK_" + RC.getName();
1222 CI->ValueName = RC.getName();
1223 } else
1224 CI->ValueName = CI->ValueName + "," + RC.getName();
1225
1226 RegisterClassClasses.insert(std::make_pair(Def, CI));
1227 }
1228
1229 // Populate the map for individual registers.
1230 for (std::map<Record*, RegisterSet>::iterator it = RegisterMap.begin(),
1231 ie = RegisterMap.end(); it != ie; ++it)
1232 RegisterClasses[it->first] = RegisterSetClasses[it->second];
1233
1234 // Name the register classes which correspond to singleton registers.
1235 for (Record *Rec : SingletonRegisters) {
1236 ClassInfo *CI = RegisterClasses[Rec];
1237 assert(CI && "Missing singleton register class info!");
1238
1239 if (CI->ValueName.empty()) {
1240 CI->ClassName = Rec->getName();
1241 CI->Name = "MCK_" + Rec->getName();
1242 CI->ValueName = Rec->getName();
1243 } else
1244 CI->ValueName = CI->ValueName + "," + Rec->getName();
1245 }
1246 }
1247
buildOperandClasses()1248 void AsmMatcherInfo::buildOperandClasses() {
1249 std::vector<Record*> AsmOperands =
1250 Records.getAllDerivedDefinitions("AsmOperandClass");
1251
1252 // Pre-populate AsmOperandClasses map.
1253 for (Record *Rec : AsmOperands) {
1254 Classes.emplace_front();
1255 AsmOperandClasses[Rec] = &Classes.front();
1256 }
1257
1258 unsigned Index = 0;
1259 for (Record *Rec : AsmOperands) {
1260 ClassInfo *CI = AsmOperandClasses[Rec];
1261 CI->Kind = ClassInfo::UserClass0 + Index;
1262
1263 ListInit *Supers = Rec->getValueAsListInit("SuperClasses");
1264 for (Init *I : Supers->getValues()) {
1265 DefInit *DI = dyn_cast<DefInit>(I);
1266 if (!DI) {
1267 PrintError(Rec->getLoc(), "Invalid super class reference!");
1268 continue;
1269 }
1270
1271 ClassInfo *SC = AsmOperandClasses[DI->getDef()];
1272 if (!SC)
1273 PrintError(Rec->getLoc(), "Invalid super class reference!");
1274 else
1275 CI->SuperClasses.push_back(SC);
1276 }
1277 CI->ClassName = Rec->getValueAsString("Name");
1278 CI->Name = "MCK_" + CI->ClassName;
1279 CI->ValueName = Rec->getName();
1280
1281 // Get or construct the predicate method name.
1282 Init *PMName = Rec->getValueInit("PredicateMethod");
1283 if (StringInit *SI = dyn_cast<StringInit>(PMName)) {
1284 CI->PredicateMethod = SI->getValue();
1285 } else {
1286 assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!");
1287 CI->PredicateMethod = "is" + CI->ClassName;
1288 }
1289
1290 // Get or construct the render method name.
1291 Init *RMName = Rec->getValueInit("RenderMethod");
1292 if (StringInit *SI = dyn_cast<StringInit>(RMName)) {
1293 CI->RenderMethod = SI->getValue();
1294 } else {
1295 assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!");
1296 CI->RenderMethod = "add" + CI->ClassName + "Operands";
1297 }
1298
1299 // Get the parse method name or leave it as empty.
1300 Init *PRMName = Rec->getValueInit("ParserMethod");
1301 if (StringInit *SI = dyn_cast<StringInit>(PRMName))
1302 CI->ParserMethod = SI->getValue();
1303
1304 // Get the diagnostic type or leave it as empty.
1305 // Get the parse method name or leave it as empty.
1306 Init *DiagnosticType = Rec->getValueInit("DiagnosticType");
1307 if (StringInit *SI = dyn_cast<StringInit>(DiagnosticType))
1308 CI->DiagnosticType = SI->getValue();
1309
1310 ++Index;
1311 }
1312 }
1313
AsmMatcherInfo(Record * asmParser,CodeGenTarget & target,RecordKeeper & records)1314 AsmMatcherInfo::AsmMatcherInfo(Record *asmParser,
1315 CodeGenTarget &target,
1316 RecordKeeper &records)
1317 : Records(records), AsmParser(asmParser), Target(target) {
1318 }
1319
1320 /// buildOperandMatchInfo - Build the necessary information to handle user
1321 /// defined operand parsing methods.
buildOperandMatchInfo()1322 void AsmMatcherInfo::buildOperandMatchInfo() {
1323
1324 /// Map containing a mask with all operands indices that can be found for
1325 /// that class inside a instruction.
1326 typedef std::map<ClassInfo *, unsigned, less_ptr<ClassInfo>> OpClassMaskTy;
1327 OpClassMaskTy OpClassMask;
1328
1329 for (const auto &MI : Matchables) {
1330 OpClassMask.clear();
1331
1332 // Keep track of all operands of this instructions which belong to the
1333 // same class.
1334 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
1335 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
1336 if (Op.Class->ParserMethod.empty())
1337 continue;
1338 unsigned &OperandMask = OpClassMask[Op.Class];
1339 OperandMask |= (1 << i);
1340 }
1341
1342 // Generate operand match info for each mnemonic/operand class pair.
1343 for (const auto &OCM : OpClassMask) {
1344 unsigned OpMask = OCM.second;
1345 ClassInfo *CI = OCM.first;
1346 OperandMatchInfo.push_back(OperandMatchEntry::create(MI.get(), CI,
1347 OpMask));
1348 }
1349 }
1350 }
1351
buildInfo()1352 void AsmMatcherInfo::buildInfo() {
1353 // Build information about all of the AssemblerPredicates.
1354 std::vector<Record*> AllPredicates =
1355 Records.getAllDerivedDefinitions("Predicate");
1356 for (unsigned i = 0, e = AllPredicates.size(); i != e; ++i) {
1357 Record *Pred = AllPredicates[i];
1358 // Ignore predicates that are not intended for the assembler.
1359 if (!Pred->getValueAsBit("AssemblerMatcherPredicate"))
1360 continue;
1361
1362 if (Pred->getName().empty())
1363 PrintFatalError(Pred->getLoc(), "Predicate has no name!");
1364
1365 SubtargetFeatures.insert(std::make_pair(
1366 Pred, SubtargetFeatureInfo(Pred, SubtargetFeatures.size())));
1367 DEBUG(SubtargetFeatures.find(Pred)->second.dump());
1368 assert(SubtargetFeatures.size() <= 64 && "Too many subtarget features!");
1369 }
1370
1371 // Parse the instructions; we need to do this first so that we can gather the
1372 // singleton register classes.
1373 SmallPtrSet<Record*, 16> SingletonRegisters;
1374 unsigned VariantCount = Target.getAsmParserVariantCount();
1375 for (unsigned VC = 0; VC != VariantCount; ++VC) {
1376 Record *AsmVariant = Target.getAsmParserVariant(VC);
1377 std::string CommentDelimiter =
1378 AsmVariant->getValueAsString("CommentDelimiter");
1379 std::string RegisterPrefix = AsmVariant->getValueAsString("RegisterPrefix");
1380 AsmVariantInfo Variant;
1381 Variant.TokenizingCharacters =
1382 AsmVariant->getValueAsString("TokenizingCharacters");
1383 Variant.SeparatorCharacters =
1384 AsmVariant->getValueAsString("SeparatorCharacters");
1385 Variant.BreakCharacters =
1386 AsmVariant->getValueAsString("BreakCharacters");
1387 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
1388
1389 for (const CodeGenInstruction *CGI : Target.instructions()) {
1390
1391 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1392 // filter the set of instructions we consider.
1393 if (!StringRef(CGI->TheDef->getName()).startswith(MatchPrefix))
1394 continue;
1395
1396 // Ignore "codegen only" instructions.
1397 if (CGI->TheDef->getValueAsBit("isCodeGenOnly"))
1398 continue;
1399
1400 auto II = llvm::make_unique<MatchableInfo>(*CGI);
1401
1402 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix,
1403 Variant);
1404
1405 // Ignore instructions which shouldn't be matched and diagnose invalid
1406 // instruction definitions with an error.
1407 if (!II->validate(CommentDelimiter, true))
1408 continue;
1409
1410 Matchables.push_back(std::move(II));
1411 }
1412
1413 // Parse all of the InstAlias definitions and stick them in the list of
1414 // matchables.
1415 std::vector<Record*> AllInstAliases =
1416 Records.getAllDerivedDefinitions("InstAlias");
1417 for (unsigned i = 0, e = AllInstAliases.size(); i != e; ++i) {
1418 auto Alias = llvm::make_unique<CodeGenInstAlias>(AllInstAliases[i],
1419 AsmVariantNo, Target);
1420
1421 // If the tblgen -match-prefix option is specified (for tblgen hackers),
1422 // filter the set of instruction aliases we consider, based on the target
1423 // instruction.
1424 if (!StringRef(Alias->ResultInst->TheDef->getName())
1425 .startswith( MatchPrefix))
1426 continue;
1427
1428 auto II = llvm::make_unique<MatchableInfo>(std::move(Alias));
1429
1430 II->initialize(*this, SingletonRegisters, AsmVariantNo, RegisterPrefix,
1431 Variant);
1432
1433 // Validate the alias definitions.
1434 II->validate(CommentDelimiter, false);
1435
1436 Matchables.push_back(std::move(II));
1437 }
1438 }
1439
1440 // Build info for the register classes.
1441 buildRegisterClasses(SingletonRegisters);
1442
1443 // Build info for the user defined assembly operand classes.
1444 buildOperandClasses();
1445
1446 // Build the information about matchables, now that we have fully formed
1447 // classes.
1448 std::vector<std::unique_ptr<MatchableInfo>> NewMatchables;
1449 for (auto &II : Matchables) {
1450 // Parse the tokens after the mnemonic.
1451 // Note: buildInstructionOperandReference may insert new AsmOperands, so
1452 // don't precompute the loop bound.
1453 for (unsigned i = 0; i != II->AsmOperands.size(); ++i) {
1454 MatchableInfo::AsmOperand &Op = II->AsmOperands[i];
1455 StringRef Token = Op.Token;
1456
1457 // Check for singleton registers.
1458 if (Record *RegRecord = II->AsmOperands[i].SingletonReg) {
1459 Op.Class = RegisterClasses[RegRecord];
1460 assert(Op.Class && Op.Class->Registers.size() == 1 &&
1461 "Unexpected class for singleton register");
1462 continue;
1463 }
1464
1465 // Check for simple tokens.
1466 if (Token[0] != '$') {
1467 Op.Class = getTokenClass(Token);
1468 continue;
1469 }
1470
1471 if (Token.size() > 1 && isdigit(Token[1])) {
1472 Op.Class = getTokenClass(Token);
1473 continue;
1474 }
1475
1476 // Otherwise this is an operand reference.
1477 StringRef OperandName;
1478 if (Token[1] == '{')
1479 OperandName = Token.substr(2, Token.size() - 3);
1480 else
1481 OperandName = Token.substr(1);
1482
1483 if (II->DefRec.is<const CodeGenInstruction*>())
1484 buildInstructionOperandReference(II.get(), OperandName, i);
1485 else
1486 buildAliasOperandReference(II.get(), OperandName, Op);
1487 }
1488
1489 if (II->DefRec.is<const CodeGenInstruction*>()) {
1490 II->buildInstructionResultOperands();
1491 // If the instruction has a two-operand alias, build up the
1492 // matchable here. We'll add them in bulk at the end to avoid
1493 // confusing this loop.
1494 std::string Constraint =
1495 II->TheDef->getValueAsString("TwoOperandAliasConstraint");
1496 if (Constraint != "") {
1497 // Start by making a copy of the original matchable.
1498 auto AliasII = llvm::make_unique<MatchableInfo>(*II);
1499
1500 // Adjust it to be a two-operand alias.
1501 AliasII->formTwoOperandAlias(Constraint);
1502
1503 // Add the alias to the matchables list.
1504 NewMatchables.push_back(std::move(AliasII));
1505 }
1506 } else
1507 II->buildAliasResultOperands();
1508 }
1509 if (!NewMatchables.empty())
1510 Matchables.insert(Matchables.end(),
1511 std::make_move_iterator(NewMatchables.begin()),
1512 std::make_move_iterator(NewMatchables.end()));
1513
1514 // Process token alias definitions and set up the associated superclass
1515 // information.
1516 std::vector<Record*> AllTokenAliases =
1517 Records.getAllDerivedDefinitions("TokenAlias");
1518 for (unsigned i = 0, e = AllTokenAliases.size(); i != e; ++i) {
1519 Record *Rec = AllTokenAliases[i];
1520 ClassInfo *FromClass = getTokenClass(Rec->getValueAsString("FromToken"));
1521 ClassInfo *ToClass = getTokenClass(Rec->getValueAsString("ToToken"));
1522 if (FromClass == ToClass)
1523 PrintFatalError(Rec->getLoc(),
1524 "error: Destination value identical to source value.");
1525 FromClass->SuperClasses.push_back(ToClass);
1526 }
1527
1528 // Reorder classes so that classes precede super classes.
1529 Classes.sort();
1530 }
1531
1532 /// buildInstructionOperandReference - The specified operand is a reference to a
1533 /// named operand such as $src. Resolve the Class and OperandInfo pointers.
1534 void AsmMatcherInfo::
buildInstructionOperandReference(MatchableInfo * II,StringRef OperandName,unsigned AsmOpIdx)1535 buildInstructionOperandReference(MatchableInfo *II,
1536 StringRef OperandName,
1537 unsigned AsmOpIdx) {
1538 const CodeGenInstruction &CGI = *II->DefRec.get<const CodeGenInstruction*>();
1539 const CGIOperandList &Operands = CGI.Operands;
1540 MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx];
1541
1542 // Map this token to an operand.
1543 unsigned Idx;
1544 if (!Operands.hasOperandNamed(OperandName, Idx))
1545 PrintFatalError(II->TheDef->getLoc(),
1546 "error: unable to find operand: '" + OperandName + "'");
1547
1548 // If the instruction operand has multiple suboperands, but the parser
1549 // match class for the asm operand is still the default "ImmAsmOperand",
1550 // then handle each suboperand separately.
1551 if (Op->SubOpIdx == -1 && Operands[Idx].MINumOperands > 1) {
1552 Record *Rec = Operands[Idx].Rec;
1553 assert(Rec->isSubClassOf("Operand") && "Unexpected operand!");
1554 Record *MatchClass = Rec->getValueAsDef("ParserMatchClass");
1555 if (MatchClass && MatchClass->getValueAsString("Name") == "Imm") {
1556 // Insert remaining suboperands after AsmOpIdx in II->AsmOperands.
1557 StringRef Token = Op->Token; // save this in case Op gets moved
1558 for (unsigned SI = 1, SE = Operands[Idx].MINumOperands; SI != SE; ++SI) {
1559 MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token);
1560 NewAsmOp.SubOpIdx = SI;
1561 II->AsmOperands.insert(II->AsmOperands.begin()+AsmOpIdx+SI, NewAsmOp);
1562 }
1563 // Replace Op with first suboperand.
1564 Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved
1565 Op->SubOpIdx = 0;
1566 }
1567 }
1568
1569 // Set up the operand class.
1570 Op->Class = getOperandClass(Operands[Idx], Op->SubOpIdx);
1571
1572 // If the named operand is tied, canonicalize it to the untied operand.
1573 // For example, something like:
1574 // (outs GPR:$dst), (ins GPR:$src)
1575 // with an asmstring of
1576 // "inc $src"
1577 // we want to canonicalize to:
1578 // "inc $dst"
1579 // so that we know how to provide the $dst operand when filling in the result.
1580 int OITied = -1;
1581 if (Operands[Idx].MINumOperands == 1)
1582 OITied = Operands[Idx].getTiedRegister();
1583 if (OITied != -1) {
1584 // The tied operand index is an MIOperand index, find the operand that
1585 // contains it.
1586 std::pair<unsigned, unsigned> Idx = Operands.getSubOperandNumber(OITied);
1587 OperandName = Operands[Idx.first].Name;
1588 Op->SubOpIdx = Idx.second;
1589 }
1590
1591 Op->SrcOpName = OperandName;
1592 }
1593
1594 /// buildAliasOperandReference - When parsing an operand reference out of the
1595 /// matching string (e.g. "movsx $src, $dst"), determine what the class of the
1596 /// operand reference is by looking it up in the result pattern definition.
buildAliasOperandReference(MatchableInfo * II,StringRef OperandName,MatchableInfo::AsmOperand & Op)1597 void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II,
1598 StringRef OperandName,
1599 MatchableInfo::AsmOperand &Op) {
1600 const CodeGenInstAlias &CGA = *II->DefRec.get<const CodeGenInstAlias*>();
1601
1602 // Set up the operand class.
1603 for (unsigned i = 0, e = CGA.ResultOperands.size(); i != e; ++i)
1604 if (CGA.ResultOperands[i].isRecord() &&
1605 CGA.ResultOperands[i].getName() == OperandName) {
1606 // It's safe to go with the first one we find, because CodeGenInstAlias
1607 // validates that all operands with the same name have the same record.
1608 Op.SubOpIdx = CGA.ResultInstOperandIndex[i].second;
1609 // Use the match class from the Alias definition, not the
1610 // destination instruction, as we may have an immediate that's
1611 // being munged by the match class.
1612 Op.Class = getOperandClass(CGA.ResultOperands[i].getRecord(),
1613 Op.SubOpIdx);
1614 Op.SrcOpName = OperandName;
1615 return;
1616 }
1617
1618 PrintFatalError(II->TheDef->getLoc(),
1619 "error: unable to find operand: '" + OperandName + "'");
1620 }
1621
buildInstructionResultOperands()1622 void MatchableInfo::buildInstructionResultOperands() {
1623 const CodeGenInstruction *ResultInst = getResultInst();
1624
1625 // Loop over all operands of the result instruction, determining how to
1626 // populate them.
1627 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1628 const CGIOperandList::OperandInfo &OpInfo = ResultInst->Operands[i];
1629
1630 // If this is a tied operand, just copy from the previously handled operand.
1631 int TiedOp = -1;
1632 if (OpInfo.MINumOperands == 1)
1633 TiedOp = OpInfo.getTiedRegister();
1634 if (TiedOp != -1) {
1635 ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1636 continue;
1637 }
1638
1639 // Find out what operand from the asmparser this MCInst operand comes from.
1640 int SrcOperand = findAsmOperandNamed(OpInfo.Name);
1641 if (OpInfo.Name.empty() || SrcOperand == -1) {
1642 // This may happen for operands that are tied to a suboperand of a
1643 // complex operand. Simply use a dummy value here; nobody should
1644 // use this operand slot.
1645 // FIXME: The long term goal is for the MCOperand list to not contain
1646 // tied operands at all.
1647 ResOperands.push_back(ResOperand::getImmOp(0));
1648 continue;
1649 }
1650
1651 // Check if the one AsmOperand populates the entire operand.
1652 unsigned NumOperands = OpInfo.MINumOperands;
1653 if (AsmOperands[SrcOperand].SubOpIdx == -1) {
1654 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand, NumOperands));
1655 continue;
1656 }
1657
1658 // Add a separate ResOperand for each suboperand.
1659 for (unsigned AI = 0; AI < NumOperands; ++AI) {
1660 assert(AsmOperands[SrcOperand+AI].SubOpIdx == (int)AI &&
1661 AsmOperands[SrcOperand+AI].SrcOpName == OpInfo.Name &&
1662 "unexpected AsmOperands for suboperands");
1663 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand + AI, 1));
1664 }
1665 }
1666 }
1667
buildAliasResultOperands()1668 void MatchableInfo::buildAliasResultOperands() {
1669 const CodeGenInstAlias &CGA = *DefRec.get<const CodeGenInstAlias*>();
1670 const CodeGenInstruction *ResultInst = getResultInst();
1671
1672 // Loop over all operands of the result instruction, determining how to
1673 // populate them.
1674 unsigned AliasOpNo = 0;
1675 unsigned LastOpNo = CGA.ResultInstOperandIndex.size();
1676 for (unsigned i = 0, e = ResultInst->Operands.size(); i != e; ++i) {
1677 const CGIOperandList::OperandInfo *OpInfo = &ResultInst->Operands[i];
1678
1679 // If this is a tied operand, just copy from the previously handled operand.
1680 int TiedOp = -1;
1681 if (OpInfo->MINumOperands == 1)
1682 TiedOp = OpInfo->getTiedRegister();
1683 if (TiedOp != -1) {
1684 ResOperands.push_back(ResOperand::getTiedOp(TiedOp));
1685 continue;
1686 }
1687
1688 // Handle all the suboperands for this operand.
1689 const std::string &OpName = OpInfo->Name;
1690 for ( ; AliasOpNo < LastOpNo &&
1691 CGA.ResultInstOperandIndex[AliasOpNo].first == i; ++AliasOpNo) {
1692 int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second;
1693
1694 // Find out what operand from the asmparser that this MCInst operand
1695 // comes from.
1696 switch (CGA.ResultOperands[AliasOpNo].Kind) {
1697 case CodeGenInstAlias::ResultOperand::K_Record: {
1698 StringRef Name = CGA.ResultOperands[AliasOpNo].getName();
1699 int SrcOperand = findAsmOperand(Name, SubIdx);
1700 if (SrcOperand == -1)
1701 PrintFatalError(TheDef->getLoc(), "Instruction '" +
1702 TheDef->getName() + "' has operand '" + OpName +
1703 "' that doesn't appear in asm string!");
1704 unsigned NumOperands = (SubIdx == -1 ? OpInfo->MINumOperands : 1);
1705 ResOperands.push_back(ResOperand::getRenderedOp(SrcOperand,
1706 NumOperands));
1707 break;
1708 }
1709 case CodeGenInstAlias::ResultOperand::K_Imm: {
1710 int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm();
1711 ResOperands.push_back(ResOperand::getImmOp(ImmVal));
1712 break;
1713 }
1714 case CodeGenInstAlias::ResultOperand::K_Reg: {
1715 Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister();
1716 ResOperands.push_back(ResOperand::getRegOp(Reg));
1717 break;
1718 }
1719 }
1720 }
1721 }
1722 }
1723
getConverterOperandID(const std::string & Name,SmallSetVector<std::string,16> & Table,bool & IsNew)1724 static unsigned getConverterOperandID(const std::string &Name,
1725 SmallSetVector<std::string, 16> &Table,
1726 bool &IsNew) {
1727 IsNew = Table.insert(Name);
1728
1729 unsigned ID = IsNew ? Table.size() - 1 :
1730 std::find(Table.begin(), Table.end(), Name) - Table.begin();
1731
1732 assert(ID < Table.size());
1733
1734 return ID;
1735 }
1736
1737
emitConvertFuncs(CodeGenTarget & Target,StringRef ClassName,std::vector<std::unique_ptr<MatchableInfo>> & Infos,raw_ostream & OS)1738 static void emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName,
1739 std::vector<std::unique_ptr<MatchableInfo>> &Infos,
1740 raw_ostream &OS) {
1741 SmallSetVector<std::string, 16> OperandConversionKinds;
1742 SmallSetVector<std::string, 16> InstructionConversionKinds;
1743 std::vector<std::vector<uint8_t> > ConversionTable;
1744 size_t MaxRowLength = 2; // minimum is custom converter plus terminator.
1745
1746 // TargetOperandClass - This is the target's operand class, like X86Operand.
1747 std::string TargetOperandClass = Target.getName() + "Operand";
1748
1749 // Write the convert function to a separate stream, so we can drop it after
1750 // the enum. We'll build up the conversion handlers for the individual
1751 // operand types opportunistically as we encounter them.
1752 std::string ConvertFnBody;
1753 raw_string_ostream CvtOS(ConvertFnBody);
1754 // Start the unified conversion function.
1755 CvtOS << "void " << Target.getName() << ClassName << "::\n"
1756 << "convertToMCInst(unsigned Kind, MCInst &Inst, "
1757 << "unsigned Opcode,\n"
1758 << " const OperandVector"
1759 << " &Operands) {\n"
1760 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1761 << " const uint8_t *Converter = ConversionTable[Kind];\n"
1762 << " Inst.setOpcode(Opcode);\n"
1763 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1764 << " switch (*p) {\n"
1765 << " default: llvm_unreachable(\"invalid conversion entry!\");\n"
1766 << " case CVT_Reg:\n"
1767 << " static_cast<" << TargetOperandClass
1768 << "&>(*Operands[*(p + 1)]).addRegOperands(Inst, 1);\n"
1769 << " break;\n"
1770 << " case CVT_Tied:\n"
1771 << " Inst.addOperand(Inst.getOperand(*(p + 1)));\n"
1772 << " break;\n";
1773
1774 std::string OperandFnBody;
1775 raw_string_ostream OpOS(OperandFnBody);
1776 // Start the operand number lookup function.
1777 OpOS << "void " << Target.getName() << ClassName << "::\n"
1778 << "convertToMapAndConstraints(unsigned Kind,\n";
1779 OpOS.indent(27);
1780 OpOS << "const OperandVector &Operands) {\n"
1781 << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n"
1782 << " unsigned NumMCOperands = 0;\n"
1783 << " const uint8_t *Converter = ConversionTable[Kind];\n"
1784 << " for (const uint8_t *p = Converter; *p; p+= 2) {\n"
1785 << " switch (*p) {\n"
1786 << " default: llvm_unreachable(\"invalid conversion entry!\");\n"
1787 << " case CVT_Reg:\n"
1788 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1789 << " Operands[*(p + 1)]->setConstraint(\"r\");\n"
1790 << " ++NumMCOperands;\n"
1791 << " break;\n"
1792 << " case CVT_Tied:\n"
1793 << " ++NumMCOperands;\n"
1794 << " break;\n";
1795
1796 // Pre-populate the operand conversion kinds with the standard always
1797 // available entries.
1798 OperandConversionKinds.insert("CVT_Done");
1799 OperandConversionKinds.insert("CVT_Reg");
1800 OperandConversionKinds.insert("CVT_Tied");
1801 enum { CVT_Done, CVT_Reg, CVT_Tied };
1802
1803 for (auto &II : Infos) {
1804 // Check if we have a custom match function.
1805 std::string AsmMatchConverter =
1806 II->getResultInst()->TheDef->getValueAsString("AsmMatchConverter");
1807 if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) {
1808 std::string Signature = "ConvertCustom_" + AsmMatchConverter;
1809 II->ConversionFnKind = Signature;
1810
1811 // Check if we have already generated this signature.
1812 if (!InstructionConversionKinds.insert(Signature))
1813 continue;
1814
1815 // Remember this converter for the kind enum.
1816 unsigned KindID = OperandConversionKinds.size();
1817 OperandConversionKinds.insert("CVT_" +
1818 getEnumNameForToken(AsmMatchConverter));
1819
1820 // Add the converter row for this instruction.
1821 ConversionTable.emplace_back();
1822 ConversionTable.back().push_back(KindID);
1823 ConversionTable.back().push_back(CVT_Done);
1824
1825 // Add the handler to the conversion driver function.
1826 CvtOS << " case CVT_"
1827 << getEnumNameForToken(AsmMatchConverter) << ":\n"
1828 << " " << AsmMatchConverter << "(Inst, Operands);\n"
1829 << " break;\n";
1830
1831 // FIXME: Handle the operand number lookup for custom match functions.
1832 continue;
1833 }
1834
1835 // Build the conversion function signature.
1836 std::string Signature = "Convert";
1837
1838 std::vector<uint8_t> ConversionRow;
1839
1840 // Compute the convert enum and the case body.
1841 MaxRowLength = std::max(MaxRowLength, II->ResOperands.size()*2 + 1 );
1842
1843 for (unsigned i = 0, e = II->ResOperands.size(); i != e; ++i) {
1844 const MatchableInfo::ResOperand &OpInfo = II->ResOperands[i];
1845
1846 // Generate code to populate each result operand.
1847 switch (OpInfo.Kind) {
1848 case MatchableInfo::ResOperand::RenderAsmOperand: {
1849 // This comes from something we parsed.
1850 const MatchableInfo::AsmOperand &Op =
1851 II->AsmOperands[OpInfo.AsmOperandNum];
1852
1853 // Registers are always converted the same, don't duplicate the
1854 // conversion function based on them.
1855 Signature += "__";
1856 std::string Class;
1857 Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName;
1858 Signature += Class;
1859 Signature += utostr(OpInfo.MINumOperands);
1860 Signature += "_" + itostr(OpInfo.AsmOperandNum);
1861
1862 // Add the conversion kind, if necessary, and get the associated ID
1863 // the index of its entry in the vector).
1864 std::string Name = "CVT_" + (Op.Class->isRegisterClass() ? "Reg" :
1865 Op.Class->RenderMethod);
1866 Name = getEnumNameForToken(Name);
1867
1868 bool IsNewConverter = false;
1869 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1870 IsNewConverter);
1871
1872 // Add the operand entry to the instruction kind conversion row.
1873 ConversionRow.push_back(ID);
1874 ConversionRow.push_back(OpInfo.AsmOperandNum);
1875
1876 if (!IsNewConverter)
1877 break;
1878
1879 // This is a new operand kind. Add a handler for it to the
1880 // converter driver.
1881 CvtOS << " case " << Name << ":\n"
1882 << " static_cast<" << TargetOperandClass
1883 << "&>(*Operands[*(p + 1)])." << Op.Class->RenderMethod
1884 << "(Inst, " << OpInfo.MINumOperands << ");\n"
1885 << " break;\n";
1886
1887 // Add a handler for the operand number lookup.
1888 OpOS << " case " << Name << ":\n"
1889 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n";
1890
1891 if (Op.Class->isRegisterClass())
1892 OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n";
1893 else
1894 OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n";
1895 OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n"
1896 << " break;\n";
1897 break;
1898 }
1899 case MatchableInfo::ResOperand::TiedOperand: {
1900 // If this operand is tied to a previous one, just copy the MCInst
1901 // operand from the earlier one.We can only tie single MCOperand values.
1902 assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand");
1903 unsigned TiedOp = OpInfo.TiedOperandNum;
1904 assert(i > TiedOp && "Tied operand precedes its target!");
1905 Signature += "__Tie" + utostr(TiedOp);
1906 ConversionRow.push_back(CVT_Tied);
1907 ConversionRow.push_back(TiedOp);
1908 break;
1909 }
1910 case MatchableInfo::ResOperand::ImmOperand: {
1911 int64_t Val = OpInfo.ImmVal;
1912 std::string Ty = "imm_" + itostr(Val);
1913 Ty = getEnumNameForToken(Ty);
1914 Signature += "__" + Ty;
1915
1916 std::string Name = "CVT_" + Ty;
1917 bool IsNewConverter = false;
1918 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1919 IsNewConverter);
1920 // Add the operand entry to the instruction kind conversion row.
1921 ConversionRow.push_back(ID);
1922 ConversionRow.push_back(0);
1923
1924 if (!IsNewConverter)
1925 break;
1926
1927 CvtOS << " case " << Name << ":\n"
1928 << " Inst.addOperand(MCOperand::createImm(" << Val << "));\n"
1929 << " break;\n";
1930
1931 OpOS << " case " << Name << ":\n"
1932 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1933 << " Operands[*(p + 1)]->setConstraint(\"\");\n"
1934 << " ++NumMCOperands;\n"
1935 << " break;\n";
1936 break;
1937 }
1938 case MatchableInfo::ResOperand::RegOperand: {
1939 std::string Reg, Name;
1940 if (!OpInfo.Register) {
1941 Name = "reg0";
1942 Reg = "0";
1943 } else {
1944 Reg = getQualifiedName(OpInfo.Register);
1945 Name = "reg" + OpInfo.Register->getName();
1946 }
1947 Signature += "__" + Name;
1948 Name = "CVT_" + Name;
1949 bool IsNewConverter = false;
1950 unsigned ID = getConverterOperandID(Name, OperandConversionKinds,
1951 IsNewConverter);
1952 // Add the operand entry to the instruction kind conversion row.
1953 ConversionRow.push_back(ID);
1954 ConversionRow.push_back(0);
1955
1956 if (!IsNewConverter)
1957 break;
1958 CvtOS << " case " << Name << ":\n"
1959 << " Inst.addOperand(MCOperand::createReg(" << Reg << "));\n"
1960 << " break;\n";
1961
1962 OpOS << " case " << Name << ":\n"
1963 << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n"
1964 << " Operands[*(p + 1)]->setConstraint(\"m\");\n"
1965 << " ++NumMCOperands;\n"
1966 << " break;\n";
1967 }
1968 }
1969 }
1970
1971 // If there were no operands, add to the signature to that effect
1972 if (Signature == "Convert")
1973 Signature += "_NoOperands";
1974
1975 II->ConversionFnKind = Signature;
1976
1977 // Save the signature. If we already have it, don't add a new row
1978 // to the table.
1979 if (!InstructionConversionKinds.insert(Signature))
1980 continue;
1981
1982 // Add the row to the table.
1983 ConversionTable.push_back(std::move(ConversionRow));
1984 }
1985
1986 // Finish up the converter driver function.
1987 CvtOS << " }\n }\n}\n\n";
1988
1989 // Finish up the operand number lookup function.
1990 OpOS << " }\n }\n}\n\n";
1991
1992 OS << "namespace {\n";
1993
1994 // Output the operand conversion kind enum.
1995 OS << "enum OperatorConversionKind {\n";
1996 for (unsigned i = 0, e = OperandConversionKinds.size(); i != e; ++i)
1997 OS << " " << OperandConversionKinds[i] << ",\n";
1998 OS << " CVT_NUM_CONVERTERS\n";
1999 OS << "};\n\n";
2000
2001 // Output the instruction conversion kind enum.
2002 OS << "enum InstructionConversionKind {\n";
2003 for (const std::string &Signature : InstructionConversionKinds)
2004 OS << " " << Signature << ",\n";
2005 OS << " CVT_NUM_SIGNATURES\n";
2006 OS << "};\n\n";
2007
2008
2009 OS << "} // end anonymous namespace\n\n";
2010
2011 // Output the conversion table.
2012 OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES]["
2013 << MaxRowLength << "] = {\n";
2014
2015 for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) {
2016 assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!");
2017 OS << " // " << InstructionConversionKinds[Row] << "\n";
2018 OS << " { ";
2019 for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2)
2020 OS << OperandConversionKinds[ConversionTable[Row][i]] << ", "
2021 << (unsigned)(ConversionTable[Row][i + 1]) << ", ";
2022 OS << "CVT_Done },\n";
2023 }
2024
2025 OS << "};\n\n";
2026
2027 // Spit out the conversion driver function.
2028 OS << CvtOS.str();
2029
2030 // Spit out the operand number lookup function.
2031 OS << OpOS.str();
2032 }
2033
2034 /// emitMatchClassEnumeration - Emit the enumeration for match class kinds.
emitMatchClassEnumeration(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2035 static void emitMatchClassEnumeration(CodeGenTarget &Target,
2036 std::forward_list<ClassInfo> &Infos,
2037 raw_ostream &OS) {
2038 OS << "namespace {\n\n";
2039
2040 OS << "/// MatchClassKind - The kinds of classes which participate in\n"
2041 << "/// instruction matching.\n";
2042 OS << "enum MatchClassKind {\n";
2043 OS << " InvalidMatchClass = 0,\n";
2044 for (const auto &CI : Infos) {
2045 OS << " " << CI.Name << ", // ";
2046 if (CI.Kind == ClassInfo::Token) {
2047 OS << "'" << CI.ValueName << "'\n";
2048 } else if (CI.isRegisterClass()) {
2049 if (!CI.ValueName.empty())
2050 OS << "register class '" << CI.ValueName << "'\n";
2051 else
2052 OS << "derived register class\n";
2053 } else {
2054 OS << "user defined class '" << CI.ValueName << "'\n";
2055 }
2056 }
2057 OS << " NumMatchClassKinds\n";
2058 OS << "};\n\n";
2059
2060 OS << "}\n\n";
2061 }
2062
2063 /// emitValidateOperandClass - Emit the function to validate an operand class.
emitValidateOperandClass(AsmMatcherInfo & Info,raw_ostream & OS)2064 static void emitValidateOperandClass(AsmMatcherInfo &Info,
2065 raw_ostream &OS) {
2066 OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, "
2067 << "MatchClassKind Kind) {\n";
2068 OS << " " << Info.Target.getName() << "Operand &Operand = ("
2069 << Info.Target.getName() << "Operand&)GOp;\n";
2070
2071 // The InvalidMatchClass is not to match any operand.
2072 OS << " if (Kind == InvalidMatchClass)\n";
2073 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n";
2074
2075 // Check for Token operands first.
2076 // FIXME: Use a more specific diagnostic type.
2077 OS << " if (Operand.isToken())\n";
2078 OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n"
2079 << " MCTargetAsmParser::Match_Success :\n"
2080 << " MCTargetAsmParser::Match_InvalidOperand;\n\n";
2081
2082 // Check the user classes. We don't care what order since we're only
2083 // actually matching against one of them.
2084 for (const auto &CI : Info.Classes) {
2085 if (!CI.isUserClass())
2086 continue;
2087
2088 OS << " // '" << CI.ClassName << "' class\n";
2089 OS << " if (Kind == " << CI.Name << ") {\n";
2090 OS << " if (Operand." << CI.PredicateMethod << "())\n";
2091 OS << " return MCTargetAsmParser::Match_Success;\n";
2092 if (!CI.DiagnosticType.empty())
2093 OS << " return " << Info.Target.getName() << "AsmParser::Match_"
2094 << CI.DiagnosticType << ";\n";
2095 OS << " }\n\n";
2096 }
2097
2098 // Check for register operands, including sub-classes.
2099 OS << " if (Operand.isReg()) {\n";
2100 OS << " MatchClassKind OpKind;\n";
2101 OS << " switch (Operand.getReg()) {\n";
2102 OS << " default: OpKind = InvalidMatchClass; break;\n";
2103 for (const auto &RC : Info.RegisterClasses)
2104 OS << " case " << Info.Target.getName() << "::"
2105 << RC.first->getName() << ": OpKind = " << RC.second->Name
2106 << "; break;\n";
2107 OS << " }\n";
2108 OS << " return isSubclass(OpKind, Kind) ? "
2109 << "MCTargetAsmParser::Match_Success :\n "
2110 << " MCTargetAsmParser::Match_InvalidOperand;\n }\n\n";
2111
2112 // Generic fallthrough match failure case for operands that don't have
2113 // specialized diagnostic types.
2114 OS << " return MCTargetAsmParser::Match_InvalidOperand;\n";
2115 OS << "}\n\n";
2116 }
2117
2118 /// emitIsSubclass - Emit the subclass predicate function.
emitIsSubclass(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2119 static void emitIsSubclass(CodeGenTarget &Target,
2120 std::forward_list<ClassInfo> &Infos,
2121 raw_ostream &OS) {
2122 OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n";
2123 OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n";
2124 OS << " if (A == B)\n";
2125 OS << " return true;\n\n";
2126
2127 std::string OStr;
2128 raw_string_ostream SS(OStr);
2129 unsigned Count = 0;
2130 SS << " switch (A) {\n";
2131 SS << " default:\n";
2132 SS << " return false;\n";
2133 for (const auto &A : Infos) {
2134 std::vector<StringRef> SuperClasses;
2135 for (const auto &B : Infos) {
2136 if (&A != &B && A.isSubsetOf(B))
2137 SuperClasses.push_back(B.Name);
2138 }
2139
2140 if (SuperClasses.empty())
2141 continue;
2142 ++Count;
2143
2144 SS << "\n case " << A.Name << ":\n";
2145
2146 if (SuperClasses.size() == 1) {
2147 SS << " return B == " << SuperClasses.back().str() << ";\n";
2148 continue;
2149 }
2150
2151 if (!SuperClasses.empty()) {
2152 SS << " switch (B) {\n";
2153 SS << " default: return false;\n";
2154 for (unsigned i = 0, e = SuperClasses.size(); i != e; ++i)
2155 SS << " case " << SuperClasses[i].str() << ": return true;\n";
2156 SS << " }\n";
2157 } else {
2158 // No case statement to emit
2159 SS << " return false;\n";
2160 }
2161 }
2162 SS << " }\n";
2163
2164 // If there were case statements emitted into the string stream, write them
2165 // to the output stream, otherwise write the default.
2166 if (Count)
2167 OS << SS.str();
2168 else
2169 OS << " return false;\n";
2170
2171 OS << "}\n\n";
2172 }
2173
2174 /// emitMatchTokenString - Emit the function to match a token string to the
2175 /// appropriate match class value.
emitMatchTokenString(CodeGenTarget & Target,std::forward_list<ClassInfo> & Infos,raw_ostream & OS)2176 static void emitMatchTokenString(CodeGenTarget &Target,
2177 std::forward_list<ClassInfo> &Infos,
2178 raw_ostream &OS) {
2179 // Construct the match list.
2180 std::vector<StringMatcher::StringPair> Matches;
2181 for (const auto &CI : Infos) {
2182 if (CI.Kind == ClassInfo::Token)
2183 Matches.emplace_back(CI.ValueName, "return " + CI.Name + ";");
2184 }
2185
2186 OS << "static MatchClassKind matchTokenString(StringRef Name) {\n";
2187
2188 StringMatcher("Name", Matches, OS).Emit();
2189
2190 OS << " return InvalidMatchClass;\n";
2191 OS << "}\n\n";
2192 }
2193
2194 /// emitMatchRegisterName - Emit the function to match a string to the target
2195 /// specific register enum.
emitMatchRegisterName(CodeGenTarget & Target,Record * AsmParser,raw_ostream & OS)2196 static void emitMatchRegisterName(CodeGenTarget &Target, Record *AsmParser,
2197 raw_ostream &OS) {
2198 // Construct the match list.
2199 std::vector<StringMatcher::StringPair> Matches;
2200 const auto &Regs = Target.getRegBank().getRegisters();
2201 for (const CodeGenRegister &Reg : Regs) {
2202 if (Reg.TheDef->getValueAsString("AsmName").empty())
2203 continue;
2204
2205 Matches.emplace_back(Reg.TheDef->getValueAsString("AsmName"),
2206 "return " + utostr(Reg.EnumValue) + ";");
2207 }
2208
2209 OS << "static unsigned MatchRegisterName(StringRef Name) {\n";
2210
2211 StringMatcher("Name", Matches, OS).Emit();
2212
2213 OS << " return 0;\n";
2214 OS << "}\n\n";
2215 }
2216
getMinimalTypeForRange(uint64_t Range)2217 static const char *getMinimalTypeForRange(uint64_t Range) {
2218 assert(Range <= 0xFFFFFFFFFFFFFFFFULL && "Enum too large");
2219 if (Range > 0xFFFFFFFFULL)
2220 return "uint64_t";
2221 if (Range > 0xFFFF)
2222 return "uint32_t";
2223 if (Range > 0xFF)
2224 return "uint16_t";
2225 return "uint8_t";
2226 }
2227
getMinimalRequiredFeaturesType(const AsmMatcherInfo & Info)2228 static const char *getMinimalRequiredFeaturesType(const AsmMatcherInfo &Info) {
2229 uint64_t MaxIndex = Info.SubtargetFeatures.size();
2230 if (MaxIndex > 0)
2231 MaxIndex--;
2232 return getMinimalTypeForRange(1ULL << MaxIndex);
2233 }
2234
2235 /// emitSubtargetFeatureFlagEnumeration - Emit the subtarget feature flag
2236 /// definitions.
emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo & Info,raw_ostream & OS)2237 static void emitSubtargetFeatureFlagEnumeration(AsmMatcherInfo &Info,
2238 raw_ostream &OS) {
2239 OS << "// Flags for subtarget features that participate in "
2240 << "instruction matching.\n";
2241 OS << "enum SubtargetFeatureFlag : " << getMinimalRequiredFeaturesType(Info)
2242 << " {\n";
2243 for (const auto &SF : Info.SubtargetFeatures) {
2244 const SubtargetFeatureInfo &SFI = SF.second;
2245 OS << " " << SFI.getEnumName() << " = (1ULL << " << SFI.Index << "),\n";
2246 }
2247 OS << " Feature_None = 0\n";
2248 OS << "};\n\n";
2249 }
2250
2251 /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types.
emitOperandDiagnosticTypes(AsmMatcherInfo & Info,raw_ostream & OS)2252 static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) {
2253 // Get the set of diagnostic types from all of the operand classes.
2254 std::set<StringRef> Types;
2255 for (std::map<Record*, ClassInfo*>::const_iterator
2256 I = Info.AsmOperandClasses.begin(),
2257 E = Info.AsmOperandClasses.end(); I != E; ++I) {
2258 if (!I->second->DiagnosticType.empty())
2259 Types.insert(I->second->DiagnosticType);
2260 }
2261
2262 if (Types.empty()) return;
2263
2264 // Now emit the enum entries.
2265 for (std::set<StringRef>::const_iterator I = Types.begin(), E = Types.end();
2266 I != E; ++I)
2267 OS << " Match_" << *I << ",\n";
2268 OS << " END_OPERAND_DIAGNOSTIC_TYPES\n";
2269 }
2270
2271 /// emitGetSubtargetFeatureName - Emit the helper function to get the
2272 /// user-level name for a subtarget feature.
emitGetSubtargetFeatureName(AsmMatcherInfo & Info,raw_ostream & OS)2273 static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) {
2274 OS << "// User-level names for subtarget features that participate in\n"
2275 << "// instruction matching.\n"
2276 << "static const char *getSubtargetFeatureName(uint64_t Val) {\n";
2277 if (!Info.SubtargetFeatures.empty()) {
2278 OS << " switch(Val) {\n";
2279 for (const auto &SF : Info.SubtargetFeatures) {
2280 const SubtargetFeatureInfo &SFI = SF.second;
2281 // FIXME: Totally just a placeholder name to get the algorithm working.
2282 OS << " case " << SFI.getEnumName() << ": return \""
2283 << SFI.TheDef->getValueAsString("PredicateName") << "\";\n";
2284 }
2285 OS << " default: return \"(unknown)\";\n";
2286 OS << " }\n";
2287 } else {
2288 // Nothing to emit, so skip the switch
2289 OS << " return \"(unknown)\";\n";
2290 }
2291 OS << "}\n\n";
2292 }
2293
2294 /// emitComputeAvailableFeatures - Emit the function to compute the list of
2295 /// available features given a subtarget.
emitComputeAvailableFeatures(AsmMatcherInfo & Info,raw_ostream & OS)2296 static void emitComputeAvailableFeatures(AsmMatcherInfo &Info,
2297 raw_ostream &OS) {
2298 std::string ClassName =
2299 Info.AsmParser->getValueAsString("AsmParserClassName");
2300
2301 OS << "uint64_t " << Info.Target.getName() << ClassName << "::\n"
2302 << "ComputeAvailableFeatures(const FeatureBitset& FB) const {\n";
2303 OS << " uint64_t Features = 0;\n";
2304 for (const auto &SF : Info.SubtargetFeatures) {
2305 const SubtargetFeatureInfo &SFI = SF.second;
2306
2307 OS << " if (";
2308 std::string CondStorage =
2309 SFI.TheDef->getValueAsString("AssemblerCondString");
2310 StringRef Conds = CondStorage;
2311 std::pair<StringRef,StringRef> Comma = Conds.split(',');
2312 bool First = true;
2313 do {
2314 if (!First)
2315 OS << " && ";
2316
2317 bool Neg = false;
2318 StringRef Cond = Comma.first;
2319 if (Cond[0] == '!') {
2320 Neg = true;
2321 Cond = Cond.substr(1);
2322 }
2323
2324 OS << "(";
2325 if (Neg)
2326 OS << "!";
2327 OS << "FB[" << Info.Target.getName() << "::" << Cond << "])";
2328
2329 if (Comma.second.empty())
2330 break;
2331
2332 First = false;
2333 Comma = Comma.second.split(',');
2334 } while (true);
2335
2336 OS << ")\n";
2337 OS << " Features |= " << SFI.getEnumName() << ";\n";
2338 }
2339 OS << " return Features;\n";
2340 OS << "}\n\n";
2341 }
2342
GetAliasRequiredFeatures(Record * R,const AsmMatcherInfo & Info)2343 static std::string GetAliasRequiredFeatures(Record *R,
2344 const AsmMatcherInfo &Info) {
2345 std::vector<Record*> ReqFeatures = R->getValueAsListOfDefs("Predicates");
2346 std::string Result;
2347 unsigned NumFeatures = 0;
2348 for (unsigned i = 0, e = ReqFeatures.size(); i != e; ++i) {
2349 const SubtargetFeatureInfo *F = Info.getSubtargetFeature(ReqFeatures[i]);
2350
2351 if (!F)
2352 PrintFatalError(R->getLoc(), "Predicate '" + ReqFeatures[i]->getName() +
2353 "' is not marked as an AssemblerPredicate!");
2354
2355 if (NumFeatures)
2356 Result += '|';
2357
2358 Result += F->getEnumName();
2359 ++NumFeatures;
2360 }
2361
2362 if (NumFeatures > 1)
2363 Result = '(' + Result + ')';
2364 return Result;
2365 }
2366
emitMnemonicAliasVariant(raw_ostream & OS,const AsmMatcherInfo & Info,std::vector<Record * > & Aliases,unsigned Indent=0,StringRef AsmParserVariantName=StringRef ())2367 static void emitMnemonicAliasVariant(raw_ostream &OS,const AsmMatcherInfo &Info,
2368 std::vector<Record*> &Aliases,
2369 unsigned Indent = 0,
2370 StringRef AsmParserVariantName = StringRef()){
2371 // Keep track of all the aliases from a mnemonic. Use an std::map so that the
2372 // iteration order of the map is stable.
2373 std::map<std::string, std::vector<Record*> > AliasesFromMnemonic;
2374
2375 for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
2376 Record *R = Aliases[i];
2377 // FIXME: Allow AssemblerVariantName to be a comma separated list.
2378 std::string AsmVariantName = R->getValueAsString("AsmVariantName");
2379 if (AsmVariantName != AsmParserVariantName)
2380 continue;
2381 AliasesFromMnemonic[R->getValueAsString("FromMnemonic")].push_back(R);
2382 }
2383 if (AliasesFromMnemonic.empty())
2384 return;
2385
2386 // Process each alias a "from" mnemonic at a time, building the code executed
2387 // by the string remapper.
2388 std::vector<StringMatcher::StringPair> Cases;
2389 for (std::map<std::string, std::vector<Record*> >::iterator
2390 I = AliasesFromMnemonic.begin(), E = AliasesFromMnemonic.end();
2391 I != E; ++I) {
2392 const std::vector<Record*> &ToVec = I->second;
2393
2394 // Loop through each alias and emit code that handles each case. If there
2395 // are two instructions without predicates, emit an error. If there is one,
2396 // emit it last.
2397 std::string MatchCode;
2398 int AliasWithNoPredicate = -1;
2399
2400 for (unsigned i = 0, e = ToVec.size(); i != e; ++i) {
2401 Record *R = ToVec[i];
2402 std::string FeatureMask = GetAliasRequiredFeatures(R, Info);
2403
2404 // If this unconditionally matches, remember it for later and diagnose
2405 // duplicates.
2406 if (FeatureMask.empty()) {
2407 if (AliasWithNoPredicate != -1) {
2408 // We can't have two aliases from the same mnemonic with no predicate.
2409 PrintError(ToVec[AliasWithNoPredicate]->getLoc(),
2410 "two MnemonicAliases with the same 'from' mnemonic!");
2411 PrintFatalError(R->getLoc(), "this is the other MnemonicAlias.");
2412 }
2413
2414 AliasWithNoPredicate = i;
2415 continue;
2416 }
2417 if (R->getValueAsString("ToMnemonic") == I->first)
2418 PrintFatalError(R->getLoc(), "MnemonicAlias to the same string");
2419
2420 if (!MatchCode.empty())
2421 MatchCode += "else ";
2422 MatchCode += "if ((Features & " + FeatureMask + ") == "+FeatureMask+")\n";
2423 MatchCode += " Mnemonic = \"" +R->getValueAsString("ToMnemonic")+"\";\n";
2424 }
2425
2426 if (AliasWithNoPredicate != -1) {
2427 Record *R = ToVec[AliasWithNoPredicate];
2428 if (!MatchCode.empty())
2429 MatchCode += "else\n ";
2430 MatchCode += "Mnemonic = \"" + R->getValueAsString("ToMnemonic")+"\";\n";
2431 }
2432
2433 MatchCode += "return;";
2434
2435 Cases.push_back(std::make_pair(I->first, MatchCode));
2436 }
2437 StringMatcher("Mnemonic", Cases, OS).Emit(Indent);
2438 }
2439
2440 /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions,
2441 /// emit a function for them and return true, otherwise return false.
emitMnemonicAliases(raw_ostream & OS,const AsmMatcherInfo & Info,CodeGenTarget & Target)2442 static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info,
2443 CodeGenTarget &Target) {
2444 // Ignore aliases when match-prefix is set.
2445 if (!MatchPrefix.empty())
2446 return false;
2447
2448 std::vector<Record*> Aliases =
2449 Info.getRecords().getAllDerivedDefinitions("MnemonicAlias");
2450 if (Aliases.empty()) return false;
2451
2452 OS << "static void applyMnemonicAliases(StringRef &Mnemonic, "
2453 "uint64_t Features, unsigned VariantID) {\n";
2454 OS << " switch (VariantID) {\n";
2455 unsigned VariantCount = Target.getAsmParserVariantCount();
2456 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2457 Record *AsmVariant = Target.getAsmParserVariant(VC);
2458 int AsmParserVariantNo = AsmVariant->getValueAsInt("Variant");
2459 std::string AsmParserVariantName = AsmVariant->getValueAsString("Name");
2460 OS << " case " << AsmParserVariantNo << ":\n";
2461 emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2,
2462 AsmParserVariantName);
2463 OS << " break;\n";
2464 }
2465 OS << " }\n";
2466
2467 // Emit aliases that apply to all variants.
2468 emitMnemonicAliasVariant(OS, Info, Aliases);
2469
2470 OS << "}\n\n";
2471
2472 return true;
2473 }
2474
emitCustomOperandParsing(raw_ostream & OS,CodeGenTarget & Target,const AsmMatcherInfo & Info,StringRef ClassName,StringToOffsetTable & StringTable,unsigned MaxMnemonicIndex)2475 static void emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target,
2476 const AsmMatcherInfo &Info, StringRef ClassName,
2477 StringToOffsetTable &StringTable,
2478 unsigned MaxMnemonicIndex) {
2479 unsigned MaxMask = 0;
2480 for (std::vector<OperandMatchEntry>::const_iterator it =
2481 Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2482 it != ie; ++it) {
2483 MaxMask |= it->OperandMask;
2484 }
2485
2486 // Emit the static custom operand parsing table;
2487 OS << "namespace {\n";
2488 OS << " struct OperandMatchEntry {\n";
2489 OS << " " << getMinimalRequiredFeaturesType(Info)
2490 << " RequiredFeatures;\n";
2491 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
2492 << " Mnemonic;\n";
2493 OS << " " << getMinimalTypeForRange(std::distance(
2494 Info.Classes.begin(), Info.Classes.end())) << " Class;\n";
2495 OS << " " << getMinimalTypeForRange(MaxMask)
2496 << " OperandMask;\n\n";
2497 OS << " StringRef getMnemonic() const {\n";
2498 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
2499 OS << " MnemonicTable[Mnemonic]);\n";
2500 OS << " }\n";
2501 OS << " };\n\n";
2502
2503 OS << " // Predicate for searching for an opcode.\n";
2504 OS << " struct LessOpcodeOperand {\n";
2505 OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n";
2506 OS << " return LHS.getMnemonic() < RHS;\n";
2507 OS << " }\n";
2508 OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n";
2509 OS << " return LHS < RHS.getMnemonic();\n";
2510 OS << " }\n";
2511 OS << " bool operator()(const OperandMatchEntry &LHS,";
2512 OS << " const OperandMatchEntry &RHS) {\n";
2513 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
2514 OS << " }\n";
2515 OS << " };\n";
2516
2517 OS << "} // end anonymous namespace.\n\n";
2518
2519 OS << "static const OperandMatchEntry OperandMatchTable["
2520 << Info.OperandMatchInfo.size() << "] = {\n";
2521
2522 OS << " /* Operand List Mask, Mnemonic, Operand Class, Features */\n";
2523 for (std::vector<OperandMatchEntry>::const_iterator it =
2524 Info.OperandMatchInfo.begin(), ie = Info.OperandMatchInfo.end();
2525 it != ie; ++it) {
2526 const OperandMatchEntry &OMI = *it;
2527 const MatchableInfo &II = *OMI.MI;
2528
2529 OS << " { ";
2530
2531 // Write the required features mask.
2532 if (!II.RequiredFeatures.empty()) {
2533 for (unsigned i = 0, e = II.RequiredFeatures.size(); i != e; ++i) {
2534 if (i) OS << "|";
2535 OS << II.RequiredFeatures[i]->getEnumName();
2536 }
2537 } else
2538 OS << "0";
2539
2540 // Store a pascal-style length byte in the mnemonic.
2541 std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.str();
2542 OS << ", " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2543 << " /* " << II.Mnemonic << " */, ";
2544
2545 OS << OMI.CI->Name;
2546
2547 OS << ", " << OMI.OperandMask;
2548 OS << " /* ";
2549 bool printComma = false;
2550 for (int i = 0, e = 31; i !=e; ++i)
2551 if (OMI.OperandMask & (1 << i)) {
2552 if (printComma)
2553 OS << ", ";
2554 OS << i;
2555 printComma = true;
2556 }
2557 OS << " */";
2558
2559 OS << " },\n";
2560 }
2561 OS << "};\n\n";
2562
2563 // Emit the operand class switch to call the correct custom parser for
2564 // the found operand class.
2565 OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2566 << Target.getName() << ClassName << "::\n"
2567 << "tryCustomParseOperand(OperandVector"
2568 << " &Operands,\n unsigned MCK) {\n\n"
2569 << " switch(MCK) {\n";
2570
2571 for (const auto &CI : Info.Classes) {
2572 if (CI.ParserMethod.empty())
2573 continue;
2574 OS << " case " << CI.Name << ":\n"
2575 << " return " << CI.ParserMethod << "(Operands);\n";
2576 }
2577
2578 OS << " default:\n";
2579 OS << " return MatchOperand_NoMatch;\n";
2580 OS << " }\n";
2581 OS << " return MatchOperand_NoMatch;\n";
2582 OS << "}\n\n";
2583
2584 // Emit the static custom operand parser. This code is very similar with
2585 // the other matcher. Also use MatchResultTy here just in case we go for
2586 // a better error handling.
2587 OS << Target.getName() << ClassName << "::OperandMatchResultTy "
2588 << Target.getName() << ClassName << "::\n"
2589 << "MatchOperandParserImpl(OperandVector"
2590 << " &Operands,\n StringRef Mnemonic) {\n";
2591
2592 // Emit code to get the available features.
2593 OS << " // Get the current feature set.\n";
2594 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2595
2596 OS << " // Get the next operand index.\n";
2597 OS << " unsigned NextOpNum = Operands.size();\n";
2598
2599 // Emit code to search the table.
2600 OS << " // Search the table.\n";
2601 OS << " std::pair<const OperandMatchEntry*, const OperandMatchEntry*>";
2602 OS << " MnemonicRange\n";
2603 OS << " (OperandMatchTable, OperandMatchTable+";
2604 OS << Info.OperandMatchInfo.size() << ");\n";
2605 OS << " if(!Mnemonic.empty())\n";
2606 OS << " MnemonicRange = std::equal_range(OperandMatchTable,";
2607 OS << " OperandMatchTable+"
2608 << Info.OperandMatchInfo.size() << ", Mnemonic,\n"
2609 << " LessOpcodeOperand());\n\n";
2610
2611 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
2612 OS << " return MatchOperand_NoMatch;\n\n";
2613
2614 OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n"
2615 << " *ie = MnemonicRange.second; it != ie; ++it) {\n";
2616
2617 OS << " // equal_range guarantees that instruction mnemonic matches.\n";
2618 OS << " assert(Mnemonic == it->getMnemonic());\n\n";
2619
2620 // Emit check that the required features are available.
2621 OS << " // check if the available features match\n";
2622 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
2623 << "!= it->RequiredFeatures) {\n";
2624 OS << " continue;\n";
2625 OS << " }\n\n";
2626
2627 // Emit check to ensure the operand number matches.
2628 OS << " // check if the operand in question has a custom parser.\n";
2629 OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n";
2630 OS << " continue;\n\n";
2631
2632 // Emit call to the custom parser method
2633 OS << " // call custom parse method to handle the operand\n";
2634 OS << " OperandMatchResultTy Result = ";
2635 OS << "tryCustomParseOperand(Operands, it->Class);\n";
2636 OS << " if (Result != MatchOperand_NoMatch)\n";
2637 OS << " return Result;\n";
2638 OS << " }\n\n";
2639
2640 OS << " // Okay, we had no match.\n";
2641 OS << " return MatchOperand_NoMatch;\n";
2642 OS << "}\n\n";
2643 }
2644
run(raw_ostream & OS)2645 void AsmMatcherEmitter::run(raw_ostream &OS) {
2646 CodeGenTarget Target(Records);
2647 Record *AsmParser = Target.getAsmParser();
2648 std::string ClassName = AsmParser->getValueAsString("AsmParserClassName");
2649
2650 // Compute the information on the instructions to match.
2651 AsmMatcherInfo Info(AsmParser, Target, Records);
2652 Info.buildInfo();
2653
2654 // Sort the instruction table using the partial order on classes. We use
2655 // stable_sort to ensure that ambiguous instructions are still
2656 // deterministically ordered.
2657 std::stable_sort(Info.Matchables.begin(), Info.Matchables.end(),
2658 [](const std::unique_ptr<MatchableInfo> &a,
2659 const std::unique_ptr<MatchableInfo> &b){
2660 return *a < *b;});
2661
2662 DEBUG_WITH_TYPE("instruction_info", {
2663 for (const auto &MI : Info.Matchables)
2664 MI->dump();
2665 });
2666
2667 // Check for ambiguous matchables.
2668 DEBUG_WITH_TYPE("ambiguous_instrs", {
2669 unsigned NumAmbiguous = 0;
2670 for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E;
2671 ++I) {
2672 for (auto J = std::next(I); J != E; ++J) {
2673 const MatchableInfo &A = **I;
2674 const MatchableInfo &B = **J;
2675
2676 if (A.couldMatchAmbiguouslyWith(B)) {
2677 errs() << "warning: ambiguous matchables:\n";
2678 A.dump();
2679 errs() << "\nis incomparable with:\n";
2680 B.dump();
2681 errs() << "\n\n";
2682 ++NumAmbiguous;
2683 }
2684 }
2685 }
2686 if (NumAmbiguous)
2687 errs() << "warning: " << NumAmbiguous
2688 << " ambiguous matchables!\n";
2689 });
2690
2691 // Compute the information on the custom operand parsing.
2692 Info.buildOperandMatchInfo();
2693
2694 // Write the output.
2695
2696 // Information for the class declaration.
2697 OS << "\n#ifdef GET_ASSEMBLER_HEADER\n";
2698 OS << "#undef GET_ASSEMBLER_HEADER\n";
2699 OS << " // This should be included into the middle of the declaration of\n";
2700 OS << " // your subclasses implementation of MCTargetAsmParser.\n";
2701 OS << " uint64_t ComputeAvailableFeatures(const FeatureBitset& FB) const;\n";
2702 OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, "
2703 << "unsigned Opcode,\n"
2704 << " const OperandVector "
2705 << "&Operands);\n";
2706 OS << " void convertToMapAndConstraints(unsigned Kind,\n ";
2707 OS << " const OperandVector &Operands) override;\n";
2708 OS << " bool mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) override;\n";
2709 OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n"
2710 << " MCInst &Inst,\n"
2711 << " uint64_t &ErrorInfo,"
2712 << " bool matchingInlineAsm,\n"
2713 << " unsigned VariantID = 0);\n";
2714
2715 if (!Info.OperandMatchInfo.empty()) {
2716 OS << "\n enum OperandMatchResultTy {\n";
2717 OS << " MatchOperand_Success, // operand matched successfully\n";
2718 OS << " MatchOperand_NoMatch, // operand did not match\n";
2719 OS << " MatchOperand_ParseFail // operand matched but had errors\n";
2720 OS << " };\n";
2721 OS << " OperandMatchResultTy MatchOperandParserImpl(\n";
2722 OS << " OperandVector &Operands,\n";
2723 OS << " StringRef Mnemonic);\n";
2724
2725 OS << " OperandMatchResultTy tryCustomParseOperand(\n";
2726 OS << " OperandVector &Operands,\n";
2727 OS << " unsigned MCK);\n\n";
2728 }
2729
2730 OS << "#endif // GET_ASSEMBLER_HEADER_INFO\n\n";
2731
2732 // Emit the operand match diagnostic enum names.
2733 OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n";
2734 OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2735 emitOperandDiagnosticTypes(Info, OS);
2736 OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n";
2737
2738
2739 OS << "\n#ifdef GET_REGISTER_MATCHER\n";
2740 OS << "#undef GET_REGISTER_MATCHER\n\n";
2741
2742 // Emit the subtarget feature enumeration.
2743 emitSubtargetFeatureFlagEnumeration(Info, OS);
2744
2745 // Emit the function to match a register name to number.
2746 // This should be omitted for Mips target
2747 if (AsmParser->getValueAsBit("ShouldEmitMatchRegisterName"))
2748 emitMatchRegisterName(Target, AsmParser, OS);
2749
2750 OS << "#endif // GET_REGISTER_MATCHER\n\n";
2751
2752 OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n";
2753 OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n";
2754
2755 // Generate the helper function to get the names for subtarget features.
2756 emitGetSubtargetFeatureName(Info, OS);
2757
2758 OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n";
2759
2760 OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n";
2761 OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n";
2762
2763 // Generate the function that remaps for mnemonic aliases.
2764 bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target);
2765
2766 // Generate the convertToMCInst function to convert operands into an MCInst.
2767 // Also, generate the convertToMapAndConstraints function for MS-style inline
2768 // assembly. The latter doesn't actually generate a MCInst.
2769 emitConvertFuncs(Target, ClassName, Info.Matchables, OS);
2770
2771 // Emit the enumeration for classes which participate in matching.
2772 emitMatchClassEnumeration(Target, Info.Classes, OS);
2773
2774 // Emit the routine to match token strings to their match class.
2775 emitMatchTokenString(Target, Info.Classes, OS);
2776
2777 // Emit the subclass predicate routine.
2778 emitIsSubclass(Target, Info.Classes, OS);
2779
2780 // Emit the routine to validate an operand against a match class.
2781 emitValidateOperandClass(Info, OS);
2782
2783 // Emit the available features compute function.
2784 emitComputeAvailableFeatures(Info, OS);
2785
2786
2787 StringToOffsetTable StringTable;
2788
2789 size_t MaxNumOperands = 0;
2790 unsigned MaxMnemonicIndex = 0;
2791 bool HasDeprecation = false;
2792 for (const auto &MI : Info.Matchables) {
2793 MaxNumOperands = std::max(MaxNumOperands, MI->AsmOperands.size());
2794 HasDeprecation |= MI->HasDeprecation;
2795
2796 // Store a pascal-style length byte in the mnemonic.
2797 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
2798 MaxMnemonicIndex = std::max(MaxMnemonicIndex,
2799 StringTable.GetOrAddStringOffset(LenMnemonic, false));
2800 }
2801
2802 OS << "static const char *const MnemonicTable =\n";
2803 StringTable.EmitString(OS);
2804 OS << ";\n\n";
2805
2806 // Emit the static match table; unused classes get initalized to 0 which is
2807 // guaranteed to be InvalidMatchClass.
2808 //
2809 // FIXME: We can reduce the size of this table very easily. First, we change
2810 // it so that store the kinds in separate bit-fields for each index, which
2811 // only needs to be the max width used for classes at that index (we also need
2812 // to reject based on this during classification). If we then make sure to
2813 // order the match kinds appropriately (putting mnemonics last), then we
2814 // should only end up using a few bits for each class, especially the ones
2815 // following the mnemonic.
2816 OS << "namespace {\n";
2817 OS << " struct MatchEntry {\n";
2818 OS << " " << getMinimalTypeForRange(MaxMnemonicIndex)
2819 << " Mnemonic;\n";
2820 OS << " uint16_t Opcode;\n";
2821 OS << " " << getMinimalTypeForRange(Info.Matchables.size())
2822 << " ConvertFn;\n";
2823 OS << " " << getMinimalRequiredFeaturesType(Info)
2824 << " RequiredFeatures;\n";
2825 OS << " " << getMinimalTypeForRange(
2826 std::distance(Info.Classes.begin(), Info.Classes.end()))
2827 << " Classes[" << MaxNumOperands << "];\n";
2828 OS << " StringRef getMnemonic() const {\n";
2829 OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n";
2830 OS << " MnemonicTable[Mnemonic]);\n";
2831 OS << " }\n";
2832 OS << " };\n\n";
2833
2834 OS << " // Predicate for searching for an opcode.\n";
2835 OS << " struct LessOpcode {\n";
2836 OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n";
2837 OS << " return LHS.getMnemonic() < RHS;\n";
2838 OS << " }\n";
2839 OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n";
2840 OS << " return LHS < RHS.getMnemonic();\n";
2841 OS << " }\n";
2842 OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n";
2843 OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n";
2844 OS << " }\n";
2845 OS << " };\n";
2846
2847 OS << "} // end anonymous namespace.\n\n";
2848
2849 unsigned VariantCount = Target.getAsmParserVariantCount();
2850 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2851 Record *AsmVariant = Target.getAsmParserVariant(VC);
2852 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2853
2854 OS << "static const MatchEntry MatchTable" << VC << "[] = {\n";
2855
2856 for (const auto &MI : Info.Matchables) {
2857 if (MI->AsmVariantID != AsmVariantNo)
2858 continue;
2859
2860 // Store a pascal-style length byte in the mnemonic.
2861 std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.str();
2862 OS << " { " << StringTable.GetOrAddStringOffset(LenMnemonic, false)
2863 << " /* " << MI->Mnemonic << " */, "
2864 << Target.getName() << "::"
2865 << MI->getResultInst()->TheDef->getName() << ", "
2866 << MI->ConversionFnKind << ", ";
2867
2868 // Write the required features mask.
2869 if (!MI->RequiredFeatures.empty()) {
2870 for (unsigned i = 0, e = MI->RequiredFeatures.size(); i != e; ++i) {
2871 if (i) OS << "|";
2872 OS << MI->RequiredFeatures[i]->getEnumName();
2873 }
2874 } else
2875 OS << "0";
2876
2877 OS << ", { ";
2878 for (unsigned i = 0, e = MI->AsmOperands.size(); i != e; ++i) {
2879 const MatchableInfo::AsmOperand &Op = MI->AsmOperands[i];
2880
2881 if (i) OS << ", ";
2882 OS << Op.Class->Name;
2883 }
2884 OS << " }, },\n";
2885 }
2886
2887 OS << "};\n\n";
2888 }
2889
2890 // A method to determine if a mnemonic is in the list.
2891 OS << "bool " << Target.getName() << ClassName << "::\n"
2892 << "mnemonicIsValid(StringRef Mnemonic, unsigned VariantID) {\n";
2893 OS << " // Find the appropriate table for this asm variant.\n";
2894 OS << " const MatchEntry *Start, *End;\n";
2895 OS << " switch (VariantID) {\n";
2896 OS << " default: llvm_unreachable(\"invalid variant!\");\n";
2897 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2898 Record *AsmVariant = Target.getAsmParserVariant(VC);
2899 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2900 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
2901 << "); End = std::end(MatchTable" << VC << "); break;\n";
2902 }
2903 OS << " }\n";
2904 OS << " // Search the table.\n";
2905 OS << " std::pair<const MatchEntry*, const MatchEntry*> MnemonicRange =\n";
2906 OS << " std::equal_range(Start, End, Mnemonic, LessOpcode());\n";
2907 OS << " return MnemonicRange.first != MnemonicRange.second;\n";
2908 OS << "}\n\n";
2909
2910 // Finally, build the match function.
2911 OS << "unsigned " << Target.getName() << ClassName << "::\n"
2912 << "MatchInstructionImpl(const OperandVector &Operands,\n";
2913 OS << " MCInst &Inst, uint64_t &ErrorInfo,\n"
2914 << " bool matchingInlineAsm, unsigned VariantID) {\n";
2915
2916 OS << " // Eliminate obvious mismatches.\n";
2917 OS << " if (Operands.size() > " << MaxNumOperands << ") {\n";
2918 OS << " ErrorInfo = " << MaxNumOperands << ";\n";
2919 OS << " return Match_InvalidOperand;\n";
2920 OS << " }\n\n";
2921
2922 // Emit code to get the available features.
2923 OS << " // Get the current feature set.\n";
2924 OS << " uint64_t AvailableFeatures = getAvailableFeatures();\n\n";
2925
2926 OS << " // Get the instruction mnemonic, which is the first token.\n";
2927 OS << " StringRef Mnemonic;\n";
2928 OS << " if (Operands[0]->isToken())\n";
2929 OS << " Mnemonic = ((" << Target.getName()
2930 << "Operand&)*Operands[0]).getToken();\n\n";
2931
2932 if (HasMnemonicAliases) {
2933 OS << " // Process all MnemonicAliases to remap the mnemonic.\n";
2934 OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n";
2935 }
2936
2937 // Emit code to compute the class list for this operand vector.
2938 OS << " // Some state to try to produce better error messages.\n";
2939 OS << " bool HadMatchOtherThanFeatures = false;\n";
2940 OS << " bool HadMatchOtherThanPredicate = false;\n";
2941 OS << " unsigned RetCode = Match_InvalidOperand;\n";
2942 OS << " uint64_t MissingFeatures = ~0ULL;\n";
2943 OS << " // Set ErrorInfo to the operand that mismatches if it is\n";
2944 OS << " // wrong for all instances of the instruction.\n";
2945 OS << " ErrorInfo = ~0ULL;\n";
2946
2947 // Emit code to search the table.
2948 OS << " // Find the appropriate table for this asm variant.\n";
2949 OS << " const MatchEntry *Start, *End;\n";
2950 OS << " switch (VariantID) {\n";
2951 OS << " default: llvm_unreachable(\"invalid variant!\");\n";
2952 for (unsigned VC = 0; VC != VariantCount; ++VC) {
2953 Record *AsmVariant = Target.getAsmParserVariant(VC);
2954 int AsmVariantNo = AsmVariant->getValueAsInt("Variant");
2955 OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC
2956 << "); End = std::end(MatchTable" << VC << "); break;\n";
2957 }
2958 OS << " }\n";
2959 OS << " // Search the table.\n";
2960 OS << " std::pair<const MatchEntry*, const MatchEntry*>"
2961 "MnemonicRange(Start, End);\n";
2962 OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n";
2963 OS << " if (!Mnemonic.empty())\n";
2964 OS << " MnemonicRange = std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n";
2965
2966 OS << " // Return a more specific error code if no mnemonics match.\n";
2967 OS << " if (MnemonicRange.first == MnemonicRange.second)\n";
2968 OS << " return Match_MnemonicFail;\n\n";
2969
2970 OS << " for (const MatchEntry *it = MnemonicRange.first, "
2971 << "*ie = MnemonicRange.second;\n";
2972 OS << " it != ie; ++it) {\n";
2973
2974 // Emit check that the subclasses match.
2975 OS << " bool OperandsValid = true;\n";
2976 OS << " for (unsigned i = SIndex; i != " << MaxNumOperands << "; ++i) {\n";
2977 OS << " auto Formal = static_cast<MatchClassKind>(it->Classes[i]);\n";
2978 OS << " if (i >= Operands.size()) {\n";
2979 OS << " OperandsValid = (Formal == " <<"InvalidMatchClass);\n";
2980 OS << " if (!OperandsValid) ErrorInfo = i;\n";
2981 OS << " break;\n";
2982 OS << " }\n";
2983 OS << " MCParsedAsmOperand &Actual = *Operands[i];\n";
2984 OS << " unsigned Diag = validateOperandClass(Actual, Formal);\n";
2985 OS << " if (Diag == Match_Success)\n";
2986 OS << " continue;\n";
2987 OS << " // If the generic handler indicates an invalid operand\n";
2988 OS << " // failure, check for a special case.\n";
2989 OS << " if (Diag == Match_InvalidOperand) {\n";
2990 OS << " Diag = validateTargetOperandClass(Actual, Formal);\n";
2991 OS << " if (Diag == Match_Success)\n";
2992 OS << " continue;\n";
2993 OS << " }\n";
2994 OS << " // If this operand is broken for all of the instances of this\n";
2995 OS << " // mnemonic, keep track of it so we can report loc info.\n";
2996 OS << " // If we already had a match that only failed due to a\n";
2997 OS << " // target predicate, that diagnostic is preferred.\n";
2998 OS << " if (!HadMatchOtherThanPredicate &&\n";
2999 OS << " (it == MnemonicRange.first || ErrorInfo <= i)) {\n";
3000 OS << " ErrorInfo = i;\n";
3001 OS << " // InvalidOperand is the default. Prefer specificity.\n";
3002 OS << " if (Diag != Match_InvalidOperand)\n";
3003 OS << " RetCode = Diag;\n";
3004 OS << " }\n";
3005 OS << " // Otherwise, just reject this instance of the mnemonic.\n";
3006 OS << " OperandsValid = false;\n";
3007 OS << " break;\n";
3008 OS << " }\n\n";
3009
3010 OS << " if (!OperandsValid) continue;\n";
3011
3012 // Emit check that the required features are available.
3013 OS << " if ((AvailableFeatures & it->RequiredFeatures) "
3014 << "!= it->RequiredFeatures) {\n";
3015 OS << " HadMatchOtherThanFeatures = true;\n";
3016 OS << " uint64_t NewMissingFeatures = it->RequiredFeatures & "
3017 "~AvailableFeatures;\n";
3018 OS << " if (countPopulation(NewMissingFeatures) <=\n"
3019 " countPopulation(MissingFeatures))\n";
3020 OS << " MissingFeatures = NewMissingFeatures;\n";
3021 OS << " continue;\n";
3022 OS << " }\n";
3023 OS << "\n";
3024 OS << " Inst.clear();\n\n";
3025 OS << " if (matchingInlineAsm) {\n";
3026 OS << " Inst.setOpcode(it->Opcode);\n";
3027 OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n";
3028 OS << " return Match_Success;\n";
3029 OS << " }\n\n";
3030 OS << " // We have selected a definite instruction, convert the parsed\n"
3031 << " // operands into the appropriate MCInst.\n";
3032 OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n";
3033 OS << "\n";
3034
3035 // Verify the instruction with the target-specific match predicate function.
3036 OS << " // We have a potential match. Check the target predicate to\n"
3037 << " // handle any context sensitive constraints.\n"
3038 << " unsigned MatchResult;\n"
3039 << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !="
3040 << " Match_Success) {\n"
3041 << " Inst.clear();\n"
3042 << " RetCode = MatchResult;\n"
3043 << " HadMatchOtherThanPredicate = true;\n"
3044 << " continue;\n"
3045 << " }\n\n";
3046
3047 // Call the post-processing function, if used.
3048 std::string InsnCleanupFn =
3049 AsmParser->getValueAsString("AsmParserInstCleanup");
3050 if (!InsnCleanupFn.empty())
3051 OS << " " << InsnCleanupFn << "(Inst);\n";
3052
3053 if (HasDeprecation) {
3054 OS << " std::string Info;\n";
3055 OS << " if (MII.get(Inst.getOpcode()).getDeprecatedInfo(Inst, getSTI(), Info)) {\n";
3056 OS << " SMLoc Loc = ((" << Target.getName()
3057 << "Operand&)*Operands[0]).getStartLoc();\n";
3058 OS << " getParser().Warning(Loc, Info, None);\n";
3059 OS << " }\n";
3060 }
3061
3062 OS << " return Match_Success;\n";
3063 OS << " }\n\n";
3064
3065 OS << " // Okay, we had no match. Try to return a useful error code.\n";
3066 OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n";
3067 OS << " return RetCode;\n\n";
3068 OS << " // Missing feature matches return which features were missing\n";
3069 OS << " ErrorInfo = MissingFeatures;\n";
3070 OS << " return Match_MissingFeature;\n";
3071 OS << "}\n\n";
3072
3073 if (!Info.OperandMatchInfo.empty())
3074 emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable,
3075 MaxMnemonicIndex);
3076
3077 OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n";
3078 }
3079
3080 namespace llvm {
3081
EmitAsmMatcher(RecordKeeper & RK,raw_ostream & OS)3082 void EmitAsmMatcher(RecordKeeper &RK, raw_ostream &OS) {
3083 emitSourceFileHeader("Assembly Matcher Source Fragment", OS);
3084 AsmMatcherEmitter(RK).run(OS);
3085 }
3086
3087 } // End llvm namespace
3088