1 | //===- AsmMatcherEmitter.cpp - Generate an assembly matcher ---------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This tablegen backend emits a target specifier matcher for converting parsed |
10 | // assembly operands in the MCInst structures. It also emits a matcher for |
11 | // custom operand parsing. |
12 | // |
13 | // Converting assembly operands into MCInst structures |
14 | // --------------------------------------------------- |
15 | // |
16 | // The input to the target specific matcher is a list of literal tokens and |
17 | // operands. The target specific parser should generally eliminate any syntax |
18 | // which is not relevant for matching; for example, comma tokens should have |
19 | // already been consumed and eliminated by the parser. Most instructions will |
20 | // end up with a single literal token (the instruction name) and some number of |
21 | // operands. |
22 | // |
23 | // Some example inputs, for X86: |
24 | // 'addl' (immediate ...) (register ...) |
25 | // 'add' (immediate ...) (memory ...) |
26 | // 'call' '*' %epc |
27 | // |
28 | // The assembly matcher is responsible for converting this input into a precise |
29 | // machine instruction (i.e., an instruction with a well defined encoding). This |
30 | // mapping has several properties which complicate matching: |
31 | // |
32 | // - It may be ambiguous; many architectures can legally encode particular |
33 | // variants of an instruction in different ways (for example, using a smaller |
34 | // encoding for small immediates). Such ambiguities should never be |
35 | // arbitrarily resolved by the assembler, the assembler is always responsible |
36 | // for choosing the "best" available instruction. |
37 | // |
38 | // - It may depend on the subtarget or the assembler context. Instructions |
39 | // which are invalid for the current mode, but otherwise unambiguous (e.g., |
40 | // an SSE instruction in a file being assembled for i486) should be accepted |
41 | // and rejected by the assembler front end. However, if the proper encoding |
42 | // for an instruction is dependent on the assembler context then the matcher |
43 | // is responsible for selecting the correct machine instruction for the |
44 | // current mode. |
45 | // |
46 | // The core matching algorithm attempts to exploit the regularity in most |
47 | // instruction sets to quickly determine the set of possibly matching |
48 | // instructions, and the simplify the generated code. Additionally, this helps |
49 | // to ensure that the ambiguities are intentionally resolved by the user. |
50 | // |
51 | // The matching is divided into two distinct phases: |
52 | // |
53 | // 1. Classification: Each operand is mapped to the unique set which (a) |
54 | // contains it, and (b) is the largest such subset for which a single |
55 | // instruction could match all members. |
56 | // |
57 | // For register classes, we can generate these subgroups automatically. For |
58 | // arbitrary operands, we expect the user to define the classes and their |
59 | // relations to one another (for example, 8-bit signed immediates as a |
60 | // subset of 32-bit immediates). |
61 | // |
62 | // By partitioning the operands in this way, we guarantee that for any |
63 | // tuple of classes, any single instruction must match either all or none |
64 | // of the sets of operands which could classify to that tuple. |
65 | // |
66 | // In addition, the subset relation amongst classes induces a partial order |
67 | // on such tuples, which we use to resolve ambiguities. |
68 | // |
69 | // 2. The input can now be treated as a tuple of classes (static tokens are |
70 | // simple singleton sets). Each such tuple should generally map to a single |
71 | // instruction (we currently ignore cases where this isn't true, whee!!!), |
72 | // which we can emit a simple matcher for. |
73 | // |
74 | // Custom Operand Parsing |
75 | // ---------------------- |
76 | // |
77 | // Some targets need a custom way to parse operands, some specific instructions |
78 | // can contain arguments that can represent processor flags and other kinds of |
79 | // identifiers that need to be mapped to specific values in the final encoded |
80 | // instructions. The target specific custom operand parsing works in the |
81 | // following way: |
82 | // |
83 | // 1. A operand match table is built, each entry contains a mnemonic, an |
84 | // operand class, a mask for all operand positions for that same |
85 | // class/mnemonic and target features to be checked while trying to match. |
86 | // |
87 | // 2. The operand matcher will try every possible entry with the same |
88 | // mnemonic and will check if the target feature for this mnemonic also |
89 | // matches. After that, if the operand to be matched has its index |
90 | // present in the mask, a successful match occurs. Otherwise, fallback |
91 | // to the regular operand parsing. |
92 | // |
93 | // 3. For a match success, each operand class that has a 'ParserMethod' |
94 | // becomes part of a switch from where the custom method is called. |
95 | // |
96 | //===----------------------------------------------------------------------===// |
97 | |
98 | #include "Common/CodeGenInstAlias.h" |
99 | #include "Common/CodeGenInstruction.h" |
100 | #include "Common/CodeGenRegisters.h" |
101 | #include "Common/CodeGenTarget.h" |
102 | #include "Common/SubtargetFeatureInfo.h" |
103 | #include "Common/Types.h" |
104 | #include "llvm/ADT/CachedHashString.h" |
105 | #include "llvm/ADT/PointerUnion.h" |
106 | #include "llvm/ADT/STLExtras.h" |
107 | #include "llvm/ADT/SmallPtrSet.h" |
108 | #include "llvm/ADT/SmallVector.h" |
109 | #include "llvm/ADT/StringExtras.h" |
110 | #include "llvm/Support/CommandLine.h" |
111 | #include "llvm/Support/Debug.h" |
112 | #include "llvm/Support/ErrorHandling.h" |
113 | #include "llvm/Support/FormatVariadic.h" |
114 | #include "llvm/TableGen/Error.h" |
115 | #include "llvm/TableGen/Record.h" |
116 | #include "llvm/TableGen/StringMatcher.h" |
117 | #include "llvm/TableGen/StringToOffsetTable.h" |
118 | #include "llvm/TableGen/TableGenBackend.h" |
119 | #include <cassert> |
120 | #include <cctype> |
121 | #include <forward_list> |
122 | #include <map> |
123 | #include <set> |
124 | |
125 | using namespace llvm; |
126 | |
127 | #define DEBUG_TYPE "asm-matcher-emitter" |
128 | |
129 | static cl::OptionCategory AsmMatcherEmitterCat("Options for -gen-asm-matcher" ); |
130 | |
131 | static cl::opt<std::string> |
132 | MatchPrefix("match-prefix" , cl::init(Val: "" ), |
133 | cl::desc("Only match instructions with the given prefix" ), |
134 | cl::cat(AsmMatcherEmitterCat)); |
135 | |
136 | namespace { |
137 | class AsmMatcherInfo; |
138 | |
139 | // Register sets are used as keys in some second-order sets TableGen creates |
140 | // when generating its data structures. This means that the order of two |
141 | // RegisterSets can be seen in the outputted AsmMatcher tables occasionally, and |
142 | // can even affect compiler output (at least seen in diagnostics produced when |
143 | // all matches fail). So we use a type that sorts them consistently. |
144 | typedef std::set<const Record *, LessRecordByID> RegisterSet; |
145 | |
146 | class AsmMatcherEmitter { |
147 | const RecordKeeper &Records; |
148 | |
149 | public: |
150 | AsmMatcherEmitter(const RecordKeeper &R) : Records(R) {} |
151 | |
152 | void run(raw_ostream &o); |
153 | }; |
154 | |
155 | /// ClassInfo - Helper class for storing the information about a particular |
156 | /// class of operands which can be matched. |
157 | struct ClassInfo { |
158 | enum ClassInfoKind { |
159 | /// Invalid kind, for use as a sentinel value. |
160 | Invalid = 0, |
161 | |
162 | /// The class for a particular token. |
163 | Token, |
164 | |
165 | /// The (first) register class, subsequent register classes are |
166 | /// RegisterClass0+1, and so on. |
167 | RegisterClass0, |
168 | |
169 | /// The (first) user defined class, subsequent user defined classes are |
170 | /// UserClass0+1, and so on. |
171 | UserClass0 = 1 << 16 |
172 | }; |
173 | |
174 | /// Kind - The class kind, which is either a predefined kind, or (UserClass0 + |
175 | /// N) for the Nth user defined class. |
176 | unsigned Kind; |
177 | |
178 | /// SuperClasses - The super classes of this class. Note that for simplicities |
179 | /// sake user operands only record their immediate super class, while register |
180 | /// operands include all superclasses. |
181 | std::vector<ClassInfo *> SuperClasses; |
182 | |
183 | /// Name - The full class name, suitable for use in an enum. |
184 | std::string Name; |
185 | |
186 | /// ClassName - The unadorned generic name for this class (e.g., Token). |
187 | std::string ClassName; |
188 | |
189 | /// ValueName - The name of the value this class represents; for a token this |
190 | /// is the literal token string, for an operand it is the TableGen class (or |
191 | /// empty if this is a derived class). |
192 | std::string ValueName; |
193 | |
194 | /// PredicateMethod - The name of the operand method to test whether the |
195 | /// operand matches this class; this is not valid for Token or register kinds. |
196 | std::string PredicateMethod; |
197 | |
198 | /// RenderMethod - The name of the operand method to add this operand to an |
199 | /// MCInst; this is not valid for Token or register kinds. |
200 | std::string RenderMethod; |
201 | |
202 | /// ParserMethod - The name of the operand method to do a target specific |
203 | /// parsing on the operand. |
204 | std::string ParserMethod; |
205 | |
206 | /// For register classes: the records for all the registers in this class. |
207 | RegisterSet Registers; |
208 | |
209 | /// For custom match classes: the diagnostic kind for when the predicate |
210 | /// fails. |
211 | std::string DiagnosticType; |
212 | |
213 | /// For custom match classes: the diagnostic string for when the predicate |
214 | /// fails. |
215 | std::string DiagnosticString; |
216 | |
217 | /// Is this operand optional and not always required. |
218 | bool IsOptional; |
219 | |
220 | /// DefaultMethod - The name of the method that returns the default operand |
221 | /// for optional operand |
222 | std::string DefaultMethod; |
223 | |
224 | public: |
225 | /// isRegisterClass() - Check if this is a register class. |
226 | bool isRegisterClass() const { |
227 | return Kind >= RegisterClass0 && Kind < UserClass0; |
228 | } |
229 | |
230 | /// isUserClass() - Check if this is a user defined class. |
231 | bool isUserClass() const { return Kind >= UserClass0; } |
232 | |
233 | /// isRelatedTo - Check whether this class is "related" to \p RHS. Classes |
234 | /// are related if they are in the same class hierarchy. |
235 | bool isRelatedTo(const ClassInfo &RHS) const { |
236 | // Tokens are only related to tokens. |
237 | if (Kind == Token || RHS.Kind == Token) |
238 | return Kind == Token && RHS.Kind == Token; |
239 | |
240 | // Registers classes are only related to registers classes, and only if |
241 | // their intersection is non-empty. |
242 | if (isRegisterClass() || RHS.isRegisterClass()) { |
243 | if (!isRegisterClass() || !RHS.isRegisterClass()) |
244 | return false; |
245 | |
246 | std::vector<const Record *> Tmp; |
247 | std::set_intersection(first1: Registers.begin(), last1: Registers.end(), |
248 | first2: RHS.Registers.begin(), last2: RHS.Registers.end(), |
249 | result: std::back_inserter(x&: Tmp), comp: LessRecordByID()); |
250 | |
251 | return !Tmp.empty(); |
252 | } |
253 | |
254 | // Otherwise we have two users operands; they are related if they are in the |
255 | // same class hierarchy. |
256 | // |
257 | // FIXME: This is an oversimplification, they should only be related if they |
258 | // intersect, however we don't have that information. |
259 | assert(isUserClass() && RHS.isUserClass() && "Unexpected class!" ); |
260 | const ClassInfo *Root = this; |
261 | while (!Root->SuperClasses.empty()) |
262 | Root = Root->SuperClasses.front(); |
263 | |
264 | const ClassInfo *RHSRoot = &RHS; |
265 | while (!RHSRoot->SuperClasses.empty()) |
266 | RHSRoot = RHSRoot->SuperClasses.front(); |
267 | |
268 | return Root == RHSRoot; |
269 | } |
270 | |
271 | /// isSubsetOf - Test whether this class is a subset of \p RHS. |
272 | bool isSubsetOf(const ClassInfo &RHS) const { |
273 | // This is a subset of RHS if it is the same class... |
274 | if (this == &RHS) |
275 | return true; |
276 | |
277 | // ... or if any of its super classes are a subset of RHS. |
278 | SmallVector<const ClassInfo *, 16> Worklist(SuperClasses.begin(), |
279 | SuperClasses.end()); |
280 | SmallPtrSet<const ClassInfo *, 16> Visited; |
281 | while (!Worklist.empty()) { |
282 | auto *CI = Worklist.pop_back_val(); |
283 | if (CI == &RHS) |
284 | return true; |
285 | for (auto *Super : CI->SuperClasses) |
286 | if (Visited.insert(Ptr: Super).second) |
287 | Worklist.push_back(Elt: Super); |
288 | } |
289 | |
290 | return false; |
291 | } |
292 | |
293 | int getTreeDepth() const { |
294 | int Depth = 0; |
295 | const ClassInfo *Root = this; |
296 | while (!Root->SuperClasses.empty()) { |
297 | Depth++; |
298 | Root = Root->SuperClasses.front(); |
299 | } |
300 | return Depth; |
301 | } |
302 | |
303 | const ClassInfo *findRoot() const { |
304 | const ClassInfo *Root = this; |
305 | while (!Root->SuperClasses.empty()) |
306 | Root = Root->SuperClasses.front(); |
307 | return Root; |
308 | } |
309 | |
310 | /// Compare two classes. This does not produce a total ordering, but does |
311 | /// guarantee that subclasses are sorted before their parents, and that the |
312 | /// ordering is transitive. |
313 | bool operator<(const ClassInfo &RHS) const { |
314 | if (this == &RHS) |
315 | return false; |
316 | |
317 | // First, enforce the ordering between the three different types of class. |
318 | // Tokens sort before registers, which sort before user classes. |
319 | if (Kind == Token) { |
320 | if (RHS.Kind != Token) |
321 | return true; |
322 | assert(RHS.Kind == Token); |
323 | } else if (isRegisterClass()) { |
324 | if (RHS.Kind == Token) |
325 | return false; |
326 | else if (RHS.isUserClass()) |
327 | return true; |
328 | assert(RHS.isRegisterClass()); |
329 | } else if (isUserClass()) { |
330 | if (!RHS.isUserClass()) |
331 | return false; |
332 | assert(RHS.isUserClass()); |
333 | } else { |
334 | llvm_unreachable("Unknown ClassInfoKind" ); |
335 | } |
336 | |
337 | if (Kind == Token || isUserClass()) { |
338 | // Related tokens and user classes get sorted by depth in the inheritence |
339 | // tree (so that subclasses are before their parents). |
340 | if (isRelatedTo(RHS)) { |
341 | if (getTreeDepth() > RHS.getTreeDepth()) |
342 | return true; |
343 | if (getTreeDepth() < RHS.getTreeDepth()) |
344 | return false; |
345 | } else { |
346 | // Unrelated tokens and user classes are ordered by the name of their |
347 | // root nodes, so that there is a consistent ordering between |
348 | // unconnected trees. |
349 | return findRoot()->ValueName < RHS.findRoot()->ValueName; |
350 | } |
351 | } else if (isRegisterClass()) { |
352 | // For register sets, sort by number of registers. This guarantees that |
353 | // a set will always sort before all of it's strict supersets. |
354 | if (Registers.size() != RHS.Registers.size()) |
355 | return Registers.size() < RHS.Registers.size(); |
356 | } else { |
357 | llvm_unreachable("Unknown ClassInfoKind" ); |
358 | } |
359 | |
360 | // FIXME: We should be able to just return false here, as we only need a |
361 | // partial order (we use stable sorts, so this is deterministic) and the |
362 | // name of a class shouldn't be significant. However, some of the backends |
363 | // accidentally rely on this behaviour, so it will have to stay like this |
364 | // until they are fixed. |
365 | return ValueName < RHS.ValueName; |
366 | } |
367 | }; |
368 | |
369 | class AsmVariantInfo { |
370 | public: |
371 | StringRef RegisterPrefix; |
372 | StringRef TokenizingCharacters; |
373 | StringRef SeparatorCharacters; |
374 | StringRef BreakCharacters; |
375 | StringRef Name; |
376 | int AsmVariantNo; |
377 | }; |
378 | |
379 | bool getPreferSmallerInstructions(CodeGenTarget const &Target) { |
380 | return Target.getAsmParser()->getValueAsBit(FieldName: "PreferSmallerInstructions" ); |
381 | } |
382 | |
383 | /// MatchableInfo - Helper class for storing the necessary information for an |
384 | /// instruction or alias which is capable of being matched. |
385 | struct MatchableInfo { |
386 | struct AsmOperand { |
387 | /// Token - This is the token that the operand came from. |
388 | StringRef Token; |
389 | |
390 | /// The unique class instance this operand should match. |
391 | ClassInfo *Class = nullptr; |
392 | |
393 | /// The operand name this is, if anything. |
394 | StringRef SrcOpName; |
395 | |
396 | /// The operand name this is, before renaming for tied operands. |
397 | StringRef OrigSrcOpName; |
398 | |
399 | /// The suboperand index within SrcOpName, or -1 for the entire operand. |
400 | int SubOpIdx = -1; |
401 | |
402 | /// Whether the token is "isolated", i.e., it is preceded and followed |
403 | /// by separators. |
404 | bool IsIsolatedToken; |
405 | |
406 | /// Register record if this token is singleton register. |
407 | const Record *SingletonReg = nullptr; |
408 | |
409 | explicit AsmOperand(bool IsIsolatedToken, StringRef T) |
410 | : Token(T), IsIsolatedToken(IsIsolatedToken) {} |
411 | }; |
412 | |
413 | /// ResOperand - This represents a single operand in the result instruction |
414 | /// generated by the match. In cases (like addressing modes) where a single |
415 | /// assembler operand expands to multiple MCOperands, this represents the |
416 | /// single assembler operand, not the MCOperand. |
417 | struct ResOperand { |
418 | enum { |
419 | /// RenderAsmOperand - This represents an operand result that is |
420 | /// generated by calling the render method on the assembly operand. The |
421 | /// corresponding AsmOperand is specified by AsmOperandNum. |
422 | RenderAsmOperand, |
423 | |
424 | /// TiedOperand - This represents a result operand that is a duplicate of |
425 | /// a previous result operand. |
426 | TiedOperand, |
427 | |
428 | /// ImmOperand - This represents an immediate value that is dumped into |
429 | /// the operand. |
430 | ImmOperand, |
431 | |
432 | /// RegOperand - This represents a fixed register that is dumped in. |
433 | RegOperand |
434 | } Kind; |
435 | |
436 | /// Tuple containing the index of the (earlier) result operand that should |
437 | /// be copied from, as well as the indices of the corresponding (parsed) |
438 | /// operands in the asm string. |
439 | struct TiedOperandsTuple { |
440 | unsigned ResOpnd; |
441 | unsigned SrcOpnd1Idx; |
442 | unsigned SrcOpnd2Idx; |
443 | }; |
444 | |
445 | union { |
446 | /// This is the operand # in the AsmOperands list that this should be |
447 | /// copied from. |
448 | unsigned AsmOperandNum; |
449 | |
450 | /// Description of tied operands. |
451 | TiedOperandsTuple TiedOperands; |
452 | |
453 | /// ImmVal - This is the immediate value added to the instruction. |
454 | int64_t ImmVal; |
455 | |
456 | /// Register - This is the register record. |
457 | const Record *Register; |
458 | }; |
459 | |
460 | /// MINumOperands - The number of MCInst operands populated by this |
461 | /// operand. |
462 | unsigned MINumOperands; |
463 | |
464 | static ResOperand getRenderedOp(unsigned AsmOpNum, unsigned NumOperands) { |
465 | ResOperand X; |
466 | X.Kind = RenderAsmOperand; |
467 | X.AsmOperandNum = AsmOpNum; |
468 | X.MINumOperands = NumOperands; |
469 | return X; |
470 | } |
471 | |
472 | static ResOperand getTiedOp(unsigned TiedOperandNum, unsigned SrcOperand1, |
473 | unsigned SrcOperand2) { |
474 | ResOperand X; |
475 | X.Kind = TiedOperand; |
476 | X.TiedOperands = {.ResOpnd: TiedOperandNum, .SrcOpnd1Idx: SrcOperand1, .SrcOpnd2Idx: SrcOperand2}; |
477 | X.MINumOperands = 1; |
478 | return X; |
479 | } |
480 | |
481 | static ResOperand getImmOp(int64_t Val) { |
482 | ResOperand X; |
483 | X.Kind = ImmOperand; |
484 | X.ImmVal = Val; |
485 | X.MINumOperands = 1; |
486 | return X; |
487 | } |
488 | |
489 | static ResOperand getRegOp(const Record *Reg) { |
490 | ResOperand X; |
491 | X.Kind = RegOperand; |
492 | X.Register = Reg; |
493 | X.MINumOperands = 1; |
494 | return X; |
495 | } |
496 | }; |
497 | |
498 | /// AsmVariantID - Target's assembly syntax variant no. |
499 | int AsmVariantID; |
500 | |
501 | /// AsmString - The assembly string for this instruction (with variants |
502 | /// removed), e.g. "movsx $src, $dst". |
503 | std::string AsmString; |
504 | |
505 | /// TheDef - This is the definition of the instruction or InstAlias that this |
506 | /// matchable came from. |
507 | const Record *const TheDef; |
508 | |
509 | // ResInstSize - The size of the resulting instruction for this matchable. |
510 | unsigned ResInstSize; |
511 | |
512 | /// DefRec - This is the definition that it came from. |
513 | PointerUnion<const CodeGenInstruction *, const CodeGenInstAlias *> DefRec; |
514 | |
515 | const CodeGenInstruction *getResultInst() const { |
516 | if (isa<const CodeGenInstruction *>(Val: DefRec)) |
517 | return cast<const CodeGenInstruction *>(Val: DefRec); |
518 | return cast<const CodeGenInstAlias *>(Val: DefRec)->ResultInst; |
519 | } |
520 | |
521 | /// ResOperands - This is the operand list that should be built for the result |
522 | /// MCInst. |
523 | SmallVector<ResOperand, 8> ResOperands; |
524 | |
525 | /// Mnemonic - This is the first token of the matched instruction, its |
526 | /// mnemonic. |
527 | StringRef Mnemonic; |
528 | |
529 | /// AsmOperands - The textual operands that this instruction matches, |
530 | /// annotated with a class and where in the OperandList they were defined. |
531 | /// This directly corresponds to the tokenized AsmString after the mnemonic is |
532 | /// removed. |
533 | SmallVector<AsmOperand, 8> AsmOperands; |
534 | |
535 | /// Predicates - The required subtarget features to match this instruction. |
536 | SmallVector<const SubtargetFeatureInfo *, 4> RequiredFeatures; |
537 | |
538 | /// ConversionFnKind - The enum value which is passed to the generated |
539 | /// convertToMCInst to convert parsed operands into an MCInst for this |
540 | /// function. |
541 | std::string ConversionFnKind; |
542 | |
543 | /// If this instruction is deprecated in some form. |
544 | bool HasDeprecation = false; |
545 | |
546 | /// If this is an alias, this is use to determine whether or not to using |
547 | /// the conversion function defined by the instruction's AsmMatchConverter |
548 | /// or to use the function generated by the alias. |
549 | bool UseInstAsmMatchConverter; |
550 | |
551 | MatchableInfo(const CodeGenInstruction &CGI) |
552 | : AsmVariantID(0), AsmString(CGI.AsmString), TheDef(CGI.TheDef), |
553 | ResInstSize(TheDef->getValueAsInt(FieldName: "Size" )), DefRec(&CGI), |
554 | UseInstAsmMatchConverter(true) {} |
555 | |
556 | MatchableInfo(std::unique_ptr<const CodeGenInstAlias> Alias) |
557 | : AsmVariantID(0), AsmString(Alias->AsmString), TheDef(Alias->TheDef), |
558 | ResInstSize(Alias->ResultInst->TheDef->getValueAsInt(FieldName: "Size" )), |
559 | DefRec(Alias.release()), UseInstAsmMatchConverter(TheDef->getValueAsBit( |
560 | FieldName: "UseInstAsmMatchConverter" )) {} |
561 | |
562 | // Could remove this and the dtor if PointerUnion supported unique_ptr |
563 | // elements with a dynamic failure/assertion (like the one below) in the case |
564 | // where it was copied while being in an owning state. |
565 | MatchableInfo(const MatchableInfo &RHS) |
566 | : AsmVariantID(RHS.AsmVariantID), AsmString(RHS.AsmString), |
567 | TheDef(RHS.TheDef), ResInstSize(RHS.ResInstSize), DefRec(RHS.DefRec), |
568 | ResOperands(RHS.ResOperands), Mnemonic(RHS.Mnemonic), |
569 | AsmOperands(RHS.AsmOperands), RequiredFeatures(RHS.RequiredFeatures), |
570 | ConversionFnKind(RHS.ConversionFnKind), |
571 | HasDeprecation(RHS.HasDeprecation), |
572 | UseInstAsmMatchConverter(RHS.UseInstAsmMatchConverter) { |
573 | assert(!isa<const CodeGenInstAlias *>(DefRec)); |
574 | } |
575 | |
576 | ~MatchableInfo() { |
577 | delete dyn_cast_if_present<const CodeGenInstAlias *>(Val&: DefRec); |
578 | } |
579 | |
580 | // Two-operand aliases clone from the main matchable, but mark the second |
581 | // operand as a tied operand of the first for purposes of the assembler. |
582 | void formTwoOperandAlias(StringRef Constraint); |
583 | |
584 | void initialize(const AsmMatcherInfo &Info, |
585 | SmallPtrSetImpl<const Record *> &SingletonRegisters, |
586 | AsmVariantInfo const &Variant, bool HasMnemonicFirst); |
587 | |
588 | /// validate - Return true if this matchable is a valid thing to match against |
589 | /// and perform a bunch of validity checking. |
590 | bool validate(StringRef , bool IsAlias) const; |
591 | |
592 | /// findAsmOperand - Find the AsmOperand with the specified name and |
593 | /// suboperand index. |
594 | int findAsmOperand(StringRef N, int SubOpIdx) const { |
595 | auto I = find_if(Range: AsmOperands, P: [&](const AsmOperand &Op) { |
596 | return Op.SrcOpName == N && Op.SubOpIdx == SubOpIdx; |
597 | }); |
598 | return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; |
599 | } |
600 | |
601 | /// findAsmOperandNamed - Find the first AsmOperand with the specified name. |
602 | /// This does not check the suboperand index. |
603 | int findAsmOperandNamed(StringRef N, int LastIdx = -1) const { |
604 | auto I = |
605 | llvm::find_if(Range: llvm::drop_begin(RangeOrContainer: AsmOperands, N: LastIdx + 1), |
606 | P: [&](const AsmOperand &Op) { return Op.SrcOpName == N; }); |
607 | return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; |
608 | } |
609 | |
610 | int findAsmOperandOriginallyNamed(StringRef N) const { |
611 | auto I = find_if(Range: AsmOperands, P: [&](const AsmOperand &Op) { |
612 | return Op.OrigSrcOpName == N; |
613 | }); |
614 | return (I != AsmOperands.end()) ? I - AsmOperands.begin() : -1; |
615 | } |
616 | |
617 | void buildInstructionResultOperands(); |
618 | void buildAliasResultOperands(bool AliasConstraintsAreChecked); |
619 | |
620 | /// shouldBeMatchedBefore - Compare two matchables for ordering. |
621 | bool shouldBeMatchedBefore(const MatchableInfo &RHS, |
622 | bool PreferSmallerInstructions) const { |
623 | // The primary comparator is the instruction mnemonic. |
624 | if (int Cmp = Mnemonic.compare_insensitive(RHS: RHS.Mnemonic)) |
625 | return Cmp == -1; |
626 | |
627 | // (Optionally) Order by the resultant instuctions size. |
628 | // eg. for ARM thumb instructions smaller encodings should be preferred. |
629 | if (PreferSmallerInstructions && ResInstSize != RHS.ResInstSize) |
630 | return ResInstSize < RHS.ResInstSize; |
631 | |
632 | if (AsmOperands.size() != RHS.AsmOperands.size()) |
633 | return AsmOperands.size() < RHS.AsmOperands.size(); |
634 | |
635 | // Compare lexicographically by operand. The matcher validates that other |
636 | // orderings wouldn't be ambiguous using \see couldMatchAmbiguouslyWith(). |
637 | for (const auto &[LHSOp, RHSOp] : zip_equal(t: AsmOperands, u: RHS.AsmOperands)) { |
638 | if (*LHSOp.Class < *RHSOp.Class) |
639 | return true; |
640 | if (*RHSOp.Class < *LHSOp.Class) |
641 | return false; |
642 | } |
643 | |
644 | // For X86 AVX/AVX512 instructions, we prefer vex encoding because the |
645 | // vex encoding size is smaller. Since X86InstrSSE.td is included ahead |
646 | // of X86InstrAVX512.td, the AVX instruction ID is less than AVX512 ID. |
647 | // We use the ID to sort AVX instruction before AVX512 instruction in |
648 | // matching table. As well as InstAlias. |
649 | if (getResultInst()->TheDef->isSubClassOf(Name: "Instruction" ) && |
650 | getResultInst()->TheDef->getValueAsBit(FieldName: "HasPositionOrder" ) && |
651 | RHS.getResultInst()->TheDef->isSubClassOf(Name: "Instruction" ) && |
652 | RHS.getResultInst()->TheDef->getValueAsBit(FieldName: "HasPositionOrder" )) |
653 | return getResultInst()->TheDef->getID() < |
654 | RHS.getResultInst()->TheDef->getID(); |
655 | |
656 | // Give matches that require more features higher precedence. This is useful |
657 | // because we cannot define AssemblerPredicates with the negation of |
658 | // processor features. For example, ARM v6 "nop" may be either a HINT or |
659 | // MOV. With v6, we want to match HINT. The assembler has no way to |
660 | // predicate MOV under "NoV6", but HINT will always match first because it |
661 | // requires V6 while MOV does not. |
662 | if (RequiredFeatures.size() != RHS.RequiredFeatures.size()) |
663 | return RequiredFeatures.size() > RHS.RequiredFeatures.size(); |
664 | |
665 | return false; |
666 | } |
667 | |
668 | /// couldMatchAmbiguouslyWith - Check whether this matchable could |
669 | /// ambiguously match the same set of operands as \p RHS (without being a |
670 | /// strictly superior match). |
671 | bool couldMatchAmbiguouslyWith(const MatchableInfo &RHS, |
672 | bool PreferSmallerInstructions) const { |
673 | // The primary comparator is the instruction mnemonic. |
674 | if (Mnemonic != RHS.Mnemonic) |
675 | return false; |
676 | |
677 | // Different variants can't conflict. |
678 | if (AsmVariantID != RHS.AsmVariantID) |
679 | return false; |
680 | |
681 | // The size of instruction is unambiguous. |
682 | if (PreferSmallerInstructions && ResInstSize != RHS.ResInstSize) |
683 | return false; |
684 | |
685 | // The number of operands is unambiguous. |
686 | if (AsmOperands.size() != RHS.AsmOperands.size()) |
687 | return false; |
688 | |
689 | // Otherwise, make sure the ordering of the two instructions is unambiguous |
690 | // by checking that either (a) a token or operand kind discriminates them, |
691 | // or (b) the ordering among equivalent kinds is consistent. |
692 | |
693 | // Tokens and operand kinds are unambiguous (assuming a correct target |
694 | // specific parser). |
695 | for (const auto &[LHSOp, RHSOp] : zip_equal(t: AsmOperands, u: RHS.AsmOperands)) { |
696 | if (LHSOp.Class->Kind != RHSOp.Class->Kind || |
697 | LHSOp.Class->Kind == ClassInfo::Token) |
698 | if (*LHSOp.Class < *RHSOp.Class || *RHSOp.Class < *LHSOp.Class) |
699 | return false; |
700 | } |
701 | |
702 | // Otherwise, this operand could commute if all operands are equivalent, or |
703 | // there is a pair of operands that compare less than and a pair that |
704 | // compare greater than. |
705 | bool HasLT = false, HasGT = false; |
706 | for (const auto &[LHSOp, RHSOp] : zip_equal(t: AsmOperands, u: RHS.AsmOperands)) { |
707 | if (*LHSOp.Class < *RHSOp.Class) |
708 | HasLT = true; |
709 | if (*RHSOp.Class < *LHSOp.Class) |
710 | HasGT = true; |
711 | } |
712 | |
713 | return HasLT == HasGT; |
714 | } |
715 | |
716 | void dump() const; |
717 | |
718 | private: |
719 | void tokenizeAsmString(AsmMatcherInfo const &Info, |
720 | AsmVariantInfo const &Variant); |
721 | void addAsmOperand(StringRef Token, bool IsIsolatedToken = false); |
722 | }; |
723 | |
724 | struct OperandMatchEntry { |
725 | unsigned OperandMask; |
726 | const MatchableInfo *MI; |
727 | ClassInfo *CI; |
728 | |
729 | static OperandMatchEntry create(const MatchableInfo *mi, ClassInfo *ci, |
730 | unsigned opMask) { |
731 | OperandMatchEntry X; |
732 | X.OperandMask = opMask; |
733 | X.CI = ci; |
734 | X.MI = mi; |
735 | return X; |
736 | } |
737 | }; |
738 | |
739 | class AsmMatcherInfo { |
740 | public: |
741 | /// Tracked Records |
742 | const RecordKeeper &Records; |
743 | |
744 | /// The tablegen AsmParser record. |
745 | const Record *AsmParser; |
746 | |
747 | /// Target - The target information. |
748 | const CodeGenTarget &Target; |
749 | |
750 | /// The classes which are needed for matching. |
751 | std::forward_list<ClassInfo> Classes; |
752 | |
753 | /// The information on the matchables to match. |
754 | std::vector<std::unique_ptr<MatchableInfo>> Matchables; |
755 | |
756 | /// Info for custom matching operands by user defined methods. |
757 | std::vector<OperandMatchEntry> OperandMatchInfo; |
758 | |
759 | /// Map of Register records to their class information. |
760 | typedef std::map<const Record *, ClassInfo *, LessRecordByID> |
761 | RegisterClassesTy; |
762 | RegisterClassesTy RegisterClasses; |
763 | |
764 | /// Map of Predicate records to their subtarget information. |
765 | SubtargetFeatureInfoMap SubtargetFeatures; |
766 | |
767 | /// Map of AsmOperandClass records to their class information. |
768 | std::map<const Record *, ClassInfo *> AsmOperandClasses; |
769 | |
770 | /// Map of RegisterClass records to their class information. |
771 | std::map<const Record *, ClassInfo *> RegisterClassClasses; |
772 | |
773 | private: |
774 | /// Map of token to class information which has already been constructed. |
775 | std::map<std::string, ClassInfo *> TokenClasses; |
776 | |
777 | private: |
778 | /// getTokenClass - Lookup or create the class for the given token. |
779 | ClassInfo *getTokenClass(StringRef Token); |
780 | |
781 | /// getOperandClass - Lookup or create the class for the given operand. |
782 | ClassInfo *getOperandClass(const CGIOperandList::OperandInfo &OI, |
783 | int SubOpIdx); |
784 | ClassInfo *getOperandClass(const Record *Rec, int SubOpIdx); |
785 | |
786 | /// buildRegisterClasses - Build the ClassInfo* instances for register |
787 | /// classes. |
788 | void |
789 | buildRegisterClasses(SmallPtrSetImpl<const Record *> &SingletonRegisters); |
790 | |
791 | /// buildOperandClasses - Build the ClassInfo* instances for user defined |
792 | /// operand classes. |
793 | void buildOperandClasses(); |
794 | |
795 | void buildInstructionOperandReference(MatchableInfo *II, StringRef OpName, |
796 | unsigned AsmOpIdx); |
797 | void buildAliasOperandReference(MatchableInfo *II, StringRef OpName, |
798 | MatchableInfo::AsmOperand &Op); |
799 | |
800 | public: |
801 | AsmMatcherInfo(const Record *AsmParser, const CodeGenTarget &Target, |
802 | const RecordKeeper &Records); |
803 | |
804 | /// Construct the various tables used during matching. |
805 | void buildInfo(); |
806 | |
807 | /// buildOperandMatchInfo - Build the necessary information to handle user |
808 | /// defined operand parsing methods. |
809 | void buildOperandMatchInfo(); |
810 | |
811 | /// getSubtargetFeature - Lookup or create the subtarget feature info for the |
812 | /// given operand. |
813 | const SubtargetFeatureInfo *getSubtargetFeature(const Record *Def) const { |
814 | assert(Def->isSubClassOf("Predicate" ) && "Invalid predicate type!" ); |
815 | const auto &I = SubtargetFeatures.find(x: Def); |
816 | return I == SubtargetFeatures.end() ? nullptr : &I->second; |
817 | } |
818 | |
819 | const RecordKeeper &getRecords() const { return Records; } |
820 | |
821 | bool hasOptionalOperands() const { |
822 | return any_of(Range: Classes, |
823 | P: [](const ClassInfo &Class) { return Class.IsOptional; }); |
824 | } |
825 | }; |
826 | |
827 | } // end anonymous namespace |
828 | |
829 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
830 | LLVM_DUMP_METHOD void MatchableInfo::dump() const { |
831 | errs() << TheDef->getName() << " -- " |
832 | << "flattened:\"" << AsmString << "\"\n" ; |
833 | |
834 | errs() << " variant: " << AsmVariantID << "\n" ; |
835 | |
836 | for (const auto &[Idx, Op] : enumerate(AsmOperands)) { |
837 | errs() << " op[" << Idx << "] = " << Op.Class->ClassName << " - " ; |
838 | errs() << '\"' << Op.Token << "\"\n" ; |
839 | } |
840 | } |
841 | #endif |
842 | |
843 | static std::pair<StringRef, StringRef> |
844 | parseTwoOperandConstraint(StringRef S, ArrayRef<SMLoc> Loc) { |
845 | // Split via the '='. |
846 | std::pair<StringRef, StringRef> Ops = S.split(Separator: '='); |
847 | if (Ops.second == "" ) |
848 | PrintFatalError(ErrorLoc: Loc, Msg: "missing '=' in two-operand alias constraint" ); |
849 | // Trim whitespace and the leading '$' on the operand names. |
850 | size_t start = Ops.first.find_first_of(C: '$'); |
851 | if (start == std::string::npos) |
852 | PrintFatalError(ErrorLoc: Loc, Msg: "expected '$' prefix on asm operand name" ); |
853 | Ops.first = Ops.first.substr(Start: start + 1); |
854 | size_t end = Ops.first.find_last_of(Chars: " \t" ); |
855 | Ops.first = Ops.first.slice(Start: 0, End: end); |
856 | // Now the second operand. |
857 | start = Ops.second.find_first_of(C: '$'); |
858 | if (start == std::string::npos) |
859 | PrintFatalError(ErrorLoc: Loc, Msg: "expected '$' prefix on asm operand name" ); |
860 | Ops.second = Ops.second.substr(Start: start + 1); |
861 | end = Ops.second.find_last_of(Chars: " \t" ); |
862 | Ops.first = Ops.first.slice(Start: 0, End: end); |
863 | return Ops; |
864 | } |
865 | |
866 | void MatchableInfo::formTwoOperandAlias(StringRef Constraint) { |
867 | // Figure out which operands are aliased and mark them as tied. |
868 | std::pair<StringRef, StringRef> Ops = |
869 | parseTwoOperandConstraint(S: Constraint, Loc: TheDef->getLoc()); |
870 | |
871 | // Find the AsmOperands that refer to the operands we're aliasing. |
872 | int SrcAsmOperand = findAsmOperandNamed(N: Ops.first); |
873 | int DstAsmOperand = findAsmOperandNamed(N: Ops.second); |
874 | if (SrcAsmOperand == -1) |
875 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
876 | Msg: "unknown source two-operand alias operand '" + Ops.first + |
877 | "'." ); |
878 | if (DstAsmOperand == -1) |
879 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
880 | Msg: "unknown destination two-operand alias operand '" + |
881 | Ops.second + "'." ); |
882 | |
883 | // Find the ResOperand that refers to the operand we're aliasing away |
884 | // and update it to refer to the combined operand instead. |
885 | for (ResOperand &Op : ResOperands) { |
886 | if (Op.Kind == ResOperand::RenderAsmOperand && |
887 | Op.AsmOperandNum == (unsigned)SrcAsmOperand) { |
888 | Op.AsmOperandNum = DstAsmOperand; |
889 | break; |
890 | } |
891 | } |
892 | // Remove the AsmOperand for the alias operand. |
893 | AsmOperands.erase(CI: AsmOperands.begin() + SrcAsmOperand); |
894 | // Adjust the ResOperand references to any AsmOperands that followed |
895 | // the one we just deleted. |
896 | for (ResOperand &Op : ResOperands) { |
897 | switch (Op.Kind) { |
898 | default: |
899 | // Nothing to do for operands that don't reference AsmOperands. |
900 | break; |
901 | case ResOperand::RenderAsmOperand: |
902 | if (Op.AsmOperandNum > (unsigned)SrcAsmOperand) |
903 | --Op.AsmOperandNum; |
904 | break; |
905 | } |
906 | } |
907 | } |
908 | |
909 | /// extractSingletonRegisterForAsmOperand - Extract singleton register, |
910 | /// if present, from specified token. |
911 | static void extractSingletonRegisterForAsmOperand(MatchableInfo::AsmOperand &Op, |
912 | const AsmMatcherInfo &Info, |
913 | StringRef RegisterPrefix) { |
914 | StringRef Tok = Op.Token; |
915 | |
916 | // If this token is not an isolated token, i.e., it isn't separated from |
917 | // other tokens (e.g. with whitespace), don't interpret it as a register name. |
918 | if (!Op.IsIsolatedToken) |
919 | return; |
920 | |
921 | if (RegisterPrefix.empty()) { |
922 | std::string LoweredTok = Tok.lower(); |
923 | if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(Name: LoweredTok)) |
924 | Op.SingletonReg = Reg->TheDef; |
925 | return; |
926 | } |
927 | |
928 | if (!Tok.starts_with(Prefix: RegisterPrefix)) |
929 | return; |
930 | |
931 | StringRef RegName = Tok.substr(Start: RegisterPrefix.size()); |
932 | if (const CodeGenRegister *Reg = Info.Target.getRegisterByName(Name: RegName)) |
933 | Op.SingletonReg = Reg->TheDef; |
934 | |
935 | // If there is no register prefix (i.e. "%" in "%eax"), then this may |
936 | // be some random non-register token, just ignore it. |
937 | } |
938 | |
939 | void MatchableInfo::initialize( |
940 | const AsmMatcherInfo &Info, |
941 | SmallPtrSetImpl<const Record *> &SingletonRegisters, |
942 | AsmVariantInfo const &Variant, bool HasMnemonicFirst) { |
943 | AsmVariantID = Variant.AsmVariantNo; |
944 | AsmString = CodeGenInstruction::FlattenAsmStringVariants( |
945 | AsmString, Variant: Variant.AsmVariantNo); |
946 | |
947 | tokenizeAsmString(Info, Variant); |
948 | |
949 | // The first token of the instruction is the mnemonic, which must be a |
950 | // simple string, not a $foo variable or a singleton register. |
951 | if (AsmOperands.empty()) |
952 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
953 | Msg: "Instruction '" + TheDef->getName() + "' has no tokens" ); |
954 | |
955 | assert(!AsmOperands[0].Token.empty()); |
956 | if (HasMnemonicFirst) { |
957 | Mnemonic = AsmOperands[0].Token; |
958 | if (Mnemonic[0] == '$') |
959 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
960 | Msg: "Invalid instruction mnemonic '" + Mnemonic + "'!" ); |
961 | |
962 | // Remove the first operand, it is tracked in the mnemonic field. |
963 | AsmOperands.erase(CI: AsmOperands.begin()); |
964 | } else if (AsmOperands[0].Token[0] != '$') |
965 | Mnemonic = AsmOperands[0].Token; |
966 | |
967 | // Compute the require features. |
968 | for (const Record *Predicate : TheDef->getValueAsListOfDefs(FieldName: "Predicates" )) |
969 | if (const SubtargetFeatureInfo *Feature = |
970 | Info.getSubtargetFeature(Def: Predicate)) |
971 | RequiredFeatures.push_back(Elt: Feature); |
972 | |
973 | // Collect singleton registers, if used. |
974 | for (MatchableInfo::AsmOperand &Op : AsmOperands) { |
975 | extractSingletonRegisterForAsmOperand(Op, Info, RegisterPrefix: Variant.RegisterPrefix); |
976 | if (Op.SingletonReg) |
977 | SingletonRegisters.insert(Ptr: Op.SingletonReg); |
978 | } |
979 | |
980 | const RecordVal *DepMask = TheDef->getValue(Name: "DeprecatedFeatureMask" ); |
981 | if (!DepMask) |
982 | DepMask = TheDef->getValue(Name: "ComplexDeprecationPredicate" ); |
983 | |
984 | HasDeprecation = |
985 | DepMask ? !DepMask->getValue()->getAsUnquotedString().empty() : false; |
986 | } |
987 | |
988 | /// Append an AsmOperand for the given substring of AsmString. |
989 | void MatchableInfo::addAsmOperand(StringRef Token, bool IsIsolatedToken) { |
990 | AsmOperands.push_back(Elt: AsmOperand(IsIsolatedToken, Token)); |
991 | } |
992 | |
993 | /// tokenizeAsmString - Tokenize a simplified assembly string. |
994 | void MatchableInfo::tokenizeAsmString(const AsmMatcherInfo &Info, |
995 | AsmVariantInfo const &Variant) { |
996 | StringRef String = AsmString; |
997 | size_t Prev = 0; |
998 | bool InTok = false; |
999 | bool IsIsolatedToken = true; |
1000 | for (size_t i = 0, e = String.size(); i != e; ++i) { |
1001 | char Char = String[i]; |
1002 | if (Variant.BreakCharacters.contains(C: Char)) { |
1003 | if (InTok) { |
1004 | addAsmOperand(Token: String.substr(Start: Prev, N: i - Prev), IsIsolatedToken: false); |
1005 | Prev = i; |
1006 | IsIsolatedToken = false; |
1007 | } |
1008 | InTok = true; |
1009 | continue; |
1010 | } |
1011 | if (Variant.TokenizingCharacters.contains(C: Char)) { |
1012 | if (InTok) { |
1013 | addAsmOperand(Token: String.substr(Start: Prev, N: i - Prev), IsIsolatedToken); |
1014 | InTok = false; |
1015 | IsIsolatedToken = false; |
1016 | } |
1017 | addAsmOperand(Token: String.substr(Start: i, N: 1), IsIsolatedToken); |
1018 | Prev = i + 1; |
1019 | IsIsolatedToken = true; |
1020 | continue; |
1021 | } |
1022 | if (Variant.SeparatorCharacters.contains(C: Char)) { |
1023 | if (InTok) { |
1024 | addAsmOperand(Token: String.substr(Start: Prev, N: i - Prev), IsIsolatedToken); |
1025 | InTok = false; |
1026 | } |
1027 | Prev = i + 1; |
1028 | IsIsolatedToken = true; |
1029 | continue; |
1030 | } |
1031 | |
1032 | switch (Char) { |
1033 | case '\\': |
1034 | if (InTok) { |
1035 | addAsmOperand(Token: String.substr(Start: Prev, N: i - Prev), IsIsolatedToken: false); |
1036 | InTok = false; |
1037 | IsIsolatedToken = false; |
1038 | } |
1039 | ++i; |
1040 | assert(i != String.size() && "Invalid quoted character" ); |
1041 | addAsmOperand(Token: String.substr(Start: i, N: 1), IsIsolatedToken); |
1042 | Prev = i + 1; |
1043 | IsIsolatedToken = false; |
1044 | break; |
1045 | |
1046 | case '$': { |
1047 | if (InTok) { |
1048 | addAsmOperand(Token: String.substr(Start: Prev, N: i - Prev), IsIsolatedToken); |
1049 | InTok = false; |
1050 | IsIsolatedToken = false; |
1051 | } |
1052 | |
1053 | // If this isn't "${", start new identifier looking like "$xxx" |
1054 | if (i + 1 == String.size() || String[i + 1] != '{') { |
1055 | Prev = i; |
1056 | break; |
1057 | } |
1058 | |
1059 | size_t EndPos = String.find(C: '}', From: i); |
1060 | assert(EndPos != StringRef::npos && |
1061 | "Missing brace in operand reference!" ); |
1062 | addAsmOperand(Token: String.substr(Start: i, N: EndPos + 1 - i), IsIsolatedToken); |
1063 | Prev = EndPos + 1; |
1064 | i = EndPos; |
1065 | IsIsolatedToken = false; |
1066 | break; |
1067 | } |
1068 | |
1069 | default: |
1070 | InTok = true; |
1071 | break; |
1072 | } |
1073 | } |
1074 | if (InTok && Prev != String.size()) |
1075 | addAsmOperand(Token: String.substr(Start: Prev), IsIsolatedToken); |
1076 | } |
1077 | |
1078 | bool MatchableInfo::validate(StringRef , bool IsAlias) const { |
1079 | // Reject matchables with no .s string. |
1080 | if (AsmString.empty()) |
1081 | PrintFatalError(ErrorLoc: TheDef->getLoc(), Msg: "instruction with empty asm string" ); |
1082 | |
1083 | // Reject any matchables with a newline in them, they should be marked |
1084 | // isCodeGenOnly if they are pseudo instructions. |
1085 | if (AsmString.find(c: '\n') != std::string::npos) |
1086 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
1087 | Msg: "multiline instruction is not valid for the asmparser, " |
1088 | "mark it isCodeGenOnly" ); |
1089 | |
1090 | // Remove comments from the asm string. We know that the asmstring only |
1091 | // has one line. |
1092 | if (!CommentDelimiter.empty() && |
1093 | StringRef(AsmString).contains(Other: CommentDelimiter)) |
1094 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
1095 | Msg: "asmstring for instruction has comment character in it, " |
1096 | "mark it isCodeGenOnly" ); |
1097 | |
1098 | // Reject matchables with operand modifiers, these aren't something we can |
1099 | // handle, the target should be refactored to use operands instead of |
1100 | // modifiers. |
1101 | // |
1102 | // Also, check for instructions which reference the operand multiple times, |
1103 | // if they don't define a custom AsmMatcher: this implies a constraint that |
1104 | // the built-in matching code would not honor. |
1105 | std::set<std::string> OperandNames; |
1106 | for (const AsmOperand &Op : AsmOperands) { |
1107 | StringRef Tok = Op.Token; |
1108 | if (Tok[0] == '$' && Tok.contains(C: ':')) |
1109 | PrintFatalError( |
1110 | ErrorLoc: TheDef->getLoc(), |
1111 | Msg: "matchable with operand modifier '" + Tok + |
1112 | "' not supported by asm matcher. Mark isCodeGenOnly!" ); |
1113 | // Verify that any operand is only mentioned once. |
1114 | // We reject aliases and ignore instructions for now. |
1115 | if (!IsAlias && TheDef->getValueAsString(FieldName: "AsmMatchConverter" ).empty() && |
1116 | Tok[0] == '$' && !OperandNames.insert(x: Tok.str()).second) { |
1117 | LLVM_DEBUG({ |
1118 | errs() << "warning: '" << TheDef->getName() << "': " |
1119 | << "ignoring instruction with tied operand '" << Tok << "'\n" ; |
1120 | }); |
1121 | return false; |
1122 | } |
1123 | } |
1124 | |
1125 | return true; |
1126 | } |
1127 | |
1128 | static std::string getEnumNameForToken(StringRef Str) { |
1129 | std::string Res; |
1130 | |
1131 | for (char C : Str) { |
1132 | switch (C) { |
1133 | case '*': |
1134 | Res += "_STAR_" ; |
1135 | break; |
1136 | case '%': |
1137 | Res += "_PCT_" ; |
1138 | break; |
1139 | case ':': |
1140 | Res += "_COLON_" ; |
1141 | break; |
1142 | case '!': |
1143 | Res += "_EXCLAIM_" ; |
1144 | break; |
1145 | case '.': |
1146 | Res += "_DOT_" ; |
1147 | break; |
1148 | case '<': |
1149 | Res += "_LT_" ; |
1150 | break; |
1151 | case '>': |
1152 | Res += "_GT_" ; |
1153 | break; |
1154 | case '-': |
1155 | Res += "_MINUS_" ; |
1156 | break; |
1157 | case '#': |
1158 | Res += "_HASH_" ; |
1159 | break; |
1160 | default: |
1161 | if (isAlnum(C)) |
1162 | Res += C; |
1163 | else |
1164 | Res += "_" + utostr(X: (unsigned)C) + "_" ; |
1165 | } |
1166 | } |
1167 | |
1168 | return Res; |
1169 | } |
1170 | |
1171 | ClassInfo *AsmMatcherInfo::getTokenClass(StringRef Token) { |
1172 | ClassInfo *&Entry = TokenClasses[Token.str()]; |
1173 | |
1174 | if (!Entry) { |
1175 | Classes.emplace_front(); |
1176 | Entry = &Classes.front(); |
1177 | Entry->Kind = ClassInfo::Token; |
1178 | Entry->ClassName = "Token" ; |
1179 | Entry->Name = "MCK_" + getEnumNameForToken(Str: Token); |
1180 | Entry->ValueName = Token.str(); |
1181 | Entry->PredicateMethod = "<invalid>" ; |
1182 | Entry->RenderMethod = "<invalid>" ; |
1183 | Entry->ParserMethod = "" ; |
1184 | Entry->DiagnosticType = "" ; |
1185 | Entry->IsOptional = false; |
1186 | Entry->DefaultMethod = "<invalid>" ; |
1187 | } |
1188 | |
1189 | return Entry; |
1190 | } |
1191 | |
1192 | ClassInfo * |
1193 | AsmMatcherInfo::getOperandClass(const CGIOperandList::OperandInfo &OI, |
1194 | int SubOpIdx) { |
1195 | const Record *Rec = OI.Rec; |
1196 | if (SubOpIdx != -1) |
1197 | Rec = cast<DefInit>(Val: OI.MIOperandInfo->getArg(Num: SubOpIdx))->getDef(); |
1198 | return getOperandClass(Rec, SubOpIdx); |
1199 | } |
1200 | |
1201 | ClassInfo *AsmMatcherInfo::getOperandClass(const Record *Rec, int SubOpIdx) { |
1202 | if (Rec->isSubClassOf(Name: "RegisterOperand" )) { |
1203 | // RegisterOperand may have an associated ParserMatchClass. If it does, |
1204 | // use it, else just fall back to the underlying register class. |
1205 | const RecordVal *R = Rec->getValue(Name: "ParserMatchClass" ); |
1206 | if (!R || !R->getValue()) |
1207 | PrintFatalError(ErrorLoc: Rec->getLoc(), |
1208 | Msg: "Record `" + Rec->getName() + |
1209 | "' does not have a ParserMatchClass!\n" ); |
1210 | |
1211 | if (const DefInit *DI = dyn_cast<DefInit>(Val: R->getValue())) { |
1212 | const Record *MatchClass = DI->getDef(); |
1213 | if (ClassInfo *CI = AsmOperandClasses[MatchClass]) |
1214 | return CI; |
1215 | } |
1216 | |
1217 | // No custom match class. Just use the register class. |
1218 | const Record *ClassRec = Rec->getValueAsDef(FieldName: "RegClass" ); |
1219 | if (!ClassRec) |
1220 | PrintFatalError(ErrorLoc: Rec->getLoc(), |
1221 | Msg: "RegisterOperand `" + Rec->getName() + |
1222 | "' has no associated register class!\n" ); |
1223 | if (ClassInfo *CI = RegisterClassClasses[ClassRec]) |
1224 | return CI; |
1225 | PrintFatalError(ErrorLoc: Rec->getLoc(), Msg: "register class has no class info!" ); |
1226 | } |
1227 | |
1228 | if (Rec->isSubClassOf(Name: "RegisterClass" )) { |
1229 | if (ClassInfo *CI = RegisterClassClasses[Rec]) |
1230 | return CI; |
1231 | PrintFatalError(ErrorLoc: Rec->getLoc(), Msg: "register class has no class info!" ); |
1232 | } |
1233 | |
1234 | if (!Rec->isSubClassOf(Name: "Operand" )) |
1235 | PrintFatalError(ErrorLoc: Rec->getLoc(), |
1236 | Msg: "Operand `" + Rec->getName() + |
1237 | "' does not derive from class Operand!\n" ); |
1238 | const Record *MatchClass = Rec->getValueAsDef(FieldName: "ParserMatchClass" ); |
1239 | if (ClassInfo *CI = AsmOperandClasses[MatchClass]) |
1240 | return CI; |
1241 | |
1242 | PrintFatalError(ErrorLoc: Rec->getLoc(), Msg: "operand has no match class!" ); |
1243 | } |
1244 | |
1245 | struct LessRegisterSet { |
1246 | bool operator()(const RegisterSet &LHS, const RegisterSet &RHS) const { |
1247 | // std::set<T> defines its own compariso "operator<", but it |
1248 | // performs a lexicographical comparison by T's innate comparison |
1249 | // for some reason. We don't want non-deterministic pointer |
1250 | // comparisons so use this instead. |
1251 | return std::lexicographical_compare(first1: LHS.begin(), last1: LHS.end(), first2: RHS.begin(), |
1252 | last2: RHS.end(), comp: LessRecordByID()); |
1253 | } |
1254 | }; |
1255 | |
1256 | void AsmMatcherInfo::buildRegisterClasses( |
1257 | SmallPtrSetImpl<const Record *> &SingletonRegisters) { |
1258 | const auto &Registers = Target.getRegBank().getRegisters(); |
1259 | auto &RegClassList = Target.getRegBank().getRegClasses(); |
1260 | |
1261 | typedef std::set<RegisterSet, LessRegisterSet> RegisterSetSet; |
1262 | |
1263 | // The register sets used for matching. |
1264 | RegisterSetSet RegisterSets; |
1265 | |
1266 | // Gather the defined sets. |
1267 | for (const CodeGenRegisterClass &RC : RegClassList) |
1268 | RegisterSets.insert( |
1269 | x: RegisterSet(RC.getOrder().begin(), RC.getOrder().end())); |
1270 | |
1271 | // Add any required singleton sets. |
1272 | for (const Record *Rec : SingletonRegisters) { |
1273 | RegisterSets.insert(x: RegisterSet(&Rec, &Rec + 1)); |
1274 | } |
1275 | |
1276 | // Introduce derived sets where necessary (when a register does not determine |
1277 | // a unique register set class), and build the mapping of registers to the set |
1278 | // they should classify to. |
1279 | std::map<const Record *, RegisterSet> RegisterMap; |
1280 | for (const CodeGenRegister &CGR : Registers) { |
1281 | // Compute the intersection of all sets containing this register. |
1282 | RegisterSet ContainingSet; |
1283 | |
1284 | for (const RegisterSet &RS : RegisterSets) { |
1285 | if (!RS.count(x: CGR.TheDef)) |
1286 | continue; |
1287 | |
1288 | if (ContainingSet.empty()) { |
1289 | ContainingSet = RS; |
1290 | continue; |
1291 | } |
1292 | |
1293 | RegisterSet Tmp; |
1294 | std::set_intersection(first1: ContainingSet.begin(), last1: ContainingSet.end(), |
1295 | first2: RS.begin(), last2: RS.end(), |
1296 | result: std::inserter(x&: Tmp, i: Tmp.begin()), comp: LessRecordByID()); |
1297 | ContainingSet = std::move(Tmp); |
1298 | } |
1299 | |
1300 | if (!ContainingSet.empty()) { |
1301 | RegisterSets.insert(x: ContainingSet); |
1302 | RegisterMap.try_emplace(k: CGR.TheDef, args&: ContainingSet); |
1303 | } |
1304 | } |
1305 | |
1306 | // Construct the register classes. |
1307 | std::map<RegisterSet, ClassInfo *, LessRegisterSet> RegisterSetClasses; |
1308 | unsigned Index = 0; |
1309 | for (const RegisterSet &RS : RegisterSets) { |
1310 | Classes.emplace_front(); |
1311 | ClassInfo *CI = &Classes.front(); |
1312 | CI->Kind = ClassInfo::RegisterClass0 + Index; |
1313 | CI->ClassName = "Reg" + utostr(X: Index); |
1314 | CI->Name = "MCK_Reg" + utostr(X: Index); |
1315 | CI->ValueName = "" ; |
1316 | CI->PredicateMethod = "" ; // unused |
1317 | CI->RenderMethod = "addRegOperands" ; |
1318 | CI->Registers = RS; |
1319 | // FIXME: diagnostic type. |
1320 | CI->DiagnosticType = "" ; |
1321 | CI->IsOptional = false; |
1322 | CI->DefaultMethod = "" ; // unused |
1323 | RegisterSetClasses.try_emplace(k: RS, args&: CI); |
1324 | ++Index; |
1325 | } |
1326 | |
1327 | // Find the superclasses; we could compute only the subgroup lattice edges, |
1328 | // but there isn't really a point. |
1329 | for (const RegisterSet &RS : RegisterSets) { |
1330 | ClassInfo *CI = RegisterSetClasses[RS]; |
1331 | for (const RegisterSet &RS2 : RegisterSets) |
1332 | if (RS != RS2 && llvm::includes(Range1: RS2, Range2: RS, C: LessRecordByID())) |
1333 | CI->SuperClasses.push_back(x: RegisterSetClasses[RS2]); |
1334 | } |
1335 | |
1336 | // Name the register classes which correspond to a user defined RegisterClass. |
1337 | for (const CodeGenRegisterClass &RC : RegClassList) { |
1338 | // Def will be NULL for non-user defined register classes. |
1339 | const Record *Def = RC.getDef(); |
1340 | if (!Def) |
1341 | continue; |
1342 | ClassInfo *CI = RegisterSetClasses[RegisterSet(RC.getOrder().begin(), |
1343 | RC.getOrder().end())]; |
1344 | if (CI->ValueName.empty()) { |
1345 | CI->ClassName = RC.getName(); |
1346 | CI->Name = "MCK_" + RC.getName(); |
1347 | CI->ValueName = RC.getName(); |
1348 | } else { |
1349 | CI->ValueName = CI->ValueName + "," + RC.getName(); |
1350 | } |
1351 | |
1352 | const Init *DiagnosticType = Def->getValueInit(FieldName: "DiagnosticType" ); |
1353 | if (const StringInit *SI = dyn_cast<StringInit>(Val: DiagnosticType)) |
1354 | CI->DiagnosticType = SI->getValue().str(); |
1355 | |
1356 | const Init *DiagnosticString = Def->getValueInit(FieldName: "DiagnosticString" ); |
1357 | if (const StringInit *SI = dyn_cast<StringInit>(Val: DiagnosticString)) |
1358 | CI->DiagnosticString = SI->getValue().str(); |
1359 | |
1360 | // If we have a diagnostic string but the diagnostic type is not specified |
1361 | // explicitly, create an anonymous diagnostic type. |
1362 | if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty()) |
1363 | CI->DiagnosticType = RC.getName(); |
1364 | |
1365 | RegisterClassClasses.try_emplace(k: Def, args&: CI); |
1366 | } |
1367 | |
1368 | // Populate the map for individual registers. |
1369 | for (auto &It : RegisterMap) |
1370 | RegisterClasses[It.first] = RegisterSetClasses[It.second]; |
1371 | |
1372 | // Name the register classes which correspond to singleton registers. |
1373 | for (const Record *Rec : SingletonRegisters) { |
1374 | ClassInfo *CI = RegisterClasses[Rec]; |
1375 | assert(CI && "Missing singleton register class info!" ); |
1376 | |
1377 | if (CI->ValueName.empty()) { |
1378 | CI->ClassName = Rec->getName().str(); |
1379 | CI->Name = "MCK_" + Rec->getName().str(); |
1380 | CI->ValueName = Rec->getName().str(); |
1381 | } else { |
1382 | CI->ValueName = CI->ValueName + "," + Rec->getName().str(); |
1383 | } |
1384 | } |
1385 | } |
1386 | |
1387 | void AsmMatcherInfo::buildOperandClasses() { |
1388 | ArrayRef<const Record *> AsmOperands = |
1389 | Records.getAllDerivedDefinitions(ClassName: "AsmOperandClass" ); |
1390 | |
1391 | // Pre-populate AsmOperandClasses map. |
1392 | for (const Record *Rec : AsmOperands) { |
1393 | Classes.emplace_front(); |
1394 | AsmOperandClasses[Rec] = &Classes.front(); |
1395 | } |
1396 | |
1397 | unsigned Index = 0; |
1398 | for (const Record *Rec : AsmOperands) { |
1399 | ClassInfo *CI = AsmOperandClasses[Rec]; |
1400 | CI->Kind = ClassInfo::UserClass0 + Index; |
1401 | |
1402 | const ListInit *Supers = Rec->getValueAsListInit(FieldName: "SuperClasses" ); |
1403 | for (const Init *I : Supers->getElements()) { |
1404 | const DefInit *DI = dyn_cast<DefInit>(Val: I); |
1405 | if (!DI) { |
1406 | PrintError(ErrorLoc: Rec->getLoc(), Msg: "Invalid super class reference!" ); |
1407 | continue; |
1408 | } |
1409 | |
1410 | ClassInfo *SC = AsmOperandClasses[DI->getDef()]; |
1411 | if (!SC) |
1412 | PrintError(ErrorLoc: Rec->getLoc(), Msg: "Invalid super class reference!" ); |
1413 | else |
1414 | CI->SuperClasses.push_back(x: SC); |
1415 | } |
1416 | CI->ClassName = Rec->getValueAsString(FieldName: "Name" ).str(); |
1417 | CI->Name = "MCK_" + CI->ClassName; |
1418 | CI->ValueName = Rec->getName().str(); |
1419 | |
1420 | // Get or construct the predicate method name. |
1421 | const Init *PMName = Rec->getValueInit(FieldName: "PredicateMethod" ); |
1422 | if (const StringInit *SI = dyn_cast<StringInit>(Val: PMName)) { |
1423 | CI->PredicateMethod = SI->getValue().str(); |
1424 | } else { |
1425 | assert(isa<UnsetInit>(PMName) && "Unexpected PredicateMethod field!" ); |
1426 | CI->PredicateMethod = "is" + CI->ClassName; |
1427 | } |
1428 | |
1429 | // Get or construct the render method name. |
1430 | const Init *RMName = Rec->getValueInit(FieldName: "RenderMethod" ); |
1431 | if (const StringInit *SI = dyn_cast<StringInit>(Val: RMName)) { |
1432 | CI->RenderMethod = SI->getValue().str(); |
1433 | } else { |
1434 | assert(isa<UnsetInit>(RMName) && "Unexpected RenderMethod field!" ); |
1435 | CI->RenderMethod = "add" + CI->ClassName + "Operands" ; |
1436 | } |
1437 | |
1438 | // Get the parse method name or leave it as empty. |
1439 | const Init *PRMName = Rec->getValueInit(FieldName: "ParserMethod" ); |
1440 | if (const StringInit *SI = dyn_cast<StringInit>(Val: PRMName)) |
1441 | CI->ParserMethod = SI->getValue().str(); |
1442 | |
1443 | // Get the diagnostic type and string or leave them as empty. |
1444 | const Init *DiagnosticType = Rec->getValueInit(FieldName: "DiagnosticType" ); |
1445 | if (const StringInit *SI = dyn_cast<StringInit>(Val: DiagnosticType)) |
1446 | CI->DiagnosticType = SI->getValue().str(); |
1447 | const Init *DiagnosticString = Rec->getValueInit(FieldName: "DiagnosticString" ); |
1448 | if (const StringInit *SI = dyn_cast<StringInit>(Val: DiagnosticString)) |
1449 | CI->DiagnosticString = SI->getValue().str(); |
1450 | // If we have a DiagnosticString, we need a DiagnosticType for use within |
1451 | // the matcher. |
1452 | if (!CI->DiagnosticString.empty() && CI->DiagnosticType.empty()) |
1453 | CI->DiagnosticType = CI->ClassName; |
1454 | |
1455 | const Init *IsOptional = Rec->getValueInit(FieldName: "IsOptional" ); |
1456 | if (const BitInit *BI = dyn_cast<BitInit>(Val: IsOptional)) |
1457 | CI->IsOptional = BI->getValue(); |
1458 | |
1459 | // Get or construct the default method name. |
1460 | const Init *DMName = Rec->getValueInit(FieldName: "DefaultMethod" ); |
1461 | if (const StringInit *SI = dyn_cast<StringInit>(Val: DMName)) { |
1462 | CI->DefaultMethod = SI->getValue().str(); |
1463 | } else { |
1464 | assert(isa<UnsetInit>(DMName) && "Unexpected DefaultMethod field!" ); |
1465 | CI->DefaultMethod = "default" + CI->ClassName + "Operands" ; |
1466 | } |
1467 | |
1468 | ++Index; |
1469 | } |
1470 | } |
1471 | |
1472 | AsmMatcherInfo::AsmMatcherInfo(const Record *asmParser, |
1473 | const CodeGenTarget &target, |
1474 | const RecordKeeper &records) |
1475 | : Records(records), AsmParser(asmParser), Target(target) {} |
1476 | |
1477 | /// buildOperandMatchInfo - Build the necessary information to handle user |
1478 | /// defined operand parsing methods. |
1479 | void AsmMatcherInfo::buildOperandMatchInfo() { |
1480 | /// Map containing a mask with all operands indices that can be found for |
1481 | /// that class inside a instruction. |
1482 | typedef std::map<ClassInfo *, unsigned, deref<std::less<>>> OpClassMaskTy; |
1483 | OpClassMaskTy OpClassMask; |
1484 | |
1485 | bool CallCustomParserForAllOperands = |
1486 | AsmParser->getValueAsBit(FieldName: "CallCustomParserForAllOperands" ); |
1487 | for (const auto &MI : Matchables) { |
1488 | OpClassMask.clear(); |
1489 | |
1490 | // Keep track of all operands of this instructions which belong to the |
1491 | // same class. |
1492 | unsigned NumOptionalOps = 0; |
1493 | for (const auto &[Idx, Op] : enumerate(First&: MI->AsmOperands)) { |
1494 | if (CallCustomParserForAllOperands || !Op.Class->ParserMethod.empty()) { |
1495 | unsigned &OperandMask = OpClassMask[Op.Class]; |
1496 | OperandMask |= maskTrailingOnes<unsigned>(N: NumOptionalOps + 1) |
1497 | << (Idx - NumOptionalOps); |
1498 | } |
1499 | if (Op.Class->IsOptional) |
1500 | ++NumOptionalOps; |
1501 | } |
1502 | |
1503 | // Generate operand match info for each mnemonic/operand class pair. |
1504 | for (const auto [CI, OpMask] : OpClassMask) { |
1505 | OperandMatchInfo.push_back( |
1506 | x: OperandMatchEntry::create(mi: MI.get(), ci: CI, opMask: OpMask)); |
1507 | } |
1508 | } |
1509 | } |
1510 | |
1511 | void AsmMatcherInfo::buildInfo() { |
1512 | // Build information about all of the AssemblerPredicates. |
1513 | SubtargetFeaturesInfoVec SubtargetFeaturePairs = |
1514 | SubtargetFeatureInfo::getAll(Records); |
1515 | SubtargetFeatures.insert(first: SubtargetFeaturePairs.begin(), |
1516 | last: SubtargetFeaturePairs.end()); |
1517 | #ifndef NDEBUG |
1518 | for (const auto &Pair : SubtargetFeatures) |
1519 | LLVM_DEBUG(Pair.second.dump()); |
1520 | #endif // NDEBUG |
1521 | |
1522 | bool HasMnemonicFirst = AsmParser->getValueAsBit(FieldName: "HasMnemonicFirst" ); |
1523 | bool ReportMultipleNearMisses = |
1524 | AsmParser->getValueAsBit(FieldName: "ReportMultipleNearMisses" ); |
1525 | |
1526 | // Parse the instructions; we need to do this first so that we can gather the |
1527 | // singleton register classes. |
1528 | SmallPtrSet<const Record *, 16> SingletonRegisters; |
1529 | unsigned VariantCount = Target.getAsmParserVariantCount(); |
1530 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
1531 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
1532 | StringRef = |
1533 | AsmVariant->getValueAsString(FieldName: "CommentDelimiter" ); |
1534 | AsmVariantInfo Variant; |
1535 | Variant.RegisterPrefix = AsmVariant->getValueAsString(FieldName: "RegisterPrefix" ); |
1536 | Variant.TokenizingCharacters = |
1537 | AsmVariant->getValueAsString(FieldName: "TokenizingCharacters" ); |
1538 | Variant.SeparatorCharacters = |
1539 | AsmVariant->getValueAsString(FieldName: "SeparatorCharacters" ); |
1540 | Variant.BreakCharacters = AsmVariant->getValueAsString(FieldName: "BreakCharacters" ); |
1541 | Variant.Name = AsmVariant->getValueAsString(FieldName: "Name" ); |
1542 | Variant.AsmVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
1543 | |
1544 | for (const CodeGenInstruction *CGI : Target.getInstructionsByEnumValue()) { |
1545 | |
1546 | // If the tblgen -match-prefix option is specified (for tblgen hackers), |
1547 | // filter the set of instructions we consider. |
1548 | if (!StringRef(CGI->TheDef->getName()).starts_with(Prefix: MatchPrefix)) |
1549 | continue; |
1550 | |
1551 | // Ignore "codegen only" instructions. |
1552 | if (CGI->TheDef->getValueAsBit(FieldName: "isCodeGenOnly" )) |
1553 | continue; |
1554 | |
1555 | // Ignore instructions for different instructions |
1556 | StringRef V = CGI->TheDef->getValueAsString(FieldName: "AsmVariantName" ); |
1557 | if (!V.empty() && V != Variant.Name) |
1558 | continue; |
1559 | |
1560 | auto II = std::make_unique<MatchableInfo>(args: *CGI); |
1561 | |
1562 | II->initialize(Info: *this, SingletonRegisters, Variant, HasMnemonicFirst); |
1563 | |
1564 | // Ignore instructions which shouldn't be matched and diagnose invalid |
1565 | // instruction definitions with an error. |
1566 | if (!II->validate(CommentDelimiter, IsAlias: false)) |
1567 | continue; |
1568 | |
1569 | Matchables.push_back(x: std::move(II)); |
1570 | } |
1571 | |
1572 | // Parse all of the InstAlias definitions and stick them in the list of |
1573 | // matchables. |
1574 | for (const Record *InstAlias : |
1575 | Records.getAllDerivedDefinitions(ClassName: "InstAlias" )) { |
1576 | auto Alias = std::make_unique<CodeGenInstAlias>(args&: InstAlias, args: Target); |
1577 | |
1578 | // If the tblgen -match-prefix option is specified (for tblgen hackers), |
1579 | // filter the set of instruction aliases we consider, based on the target |
1580 | // instruction. |
1581 | if (!StringRef(Alias->ResultInst->TheDef->getName()) |
1582 | .starts_with(Prefix: MatchPrefix)) |
1583 | continue; |
1584 | |
1585 | StringRef V = Alias->TheDef->getValueAsString(FieldName: "AsmVariantName" ); |
1586 | if (!V.empty() && V != Variant.Name) |
1587 | continue; |
1588 | |
1589 | auto II = std::make_unique<MatchableInfo>(args: std::move(Alias)); |
1590 | |
1591 | II->initialize(Info: *this, SingletonRegisters, Variant, HasMnemonicFirst); |
1592 | |
1593 | // Validate the alias definitions. |
1594 | II->validate(CommentDelimiter, IsAlias: true); |
1595 | |
1596 | Matchables.push_back(x: std::move(II)); |
1597 | } |
1598 | } |
1599 | |
1600 | // Build info for the register classes. |
1601 | buildRegisterClasses(SingletonRegisters); |
1602 | |
1603 | // Build info for the user defined assembly operand classes. |
1604 | buildOperandClasses(); |
1605 | |
1606 | // Build the information about matchables, now that we have fully formed |
1607 | // classes. |
1608 | std::vector<std::unique_ptr<MatchableInfo>> NewMatchables; |
1609 | for (auto &II : Matchables) { |
1610 | // Parse the tokens after the mnemonic. |
1611 | // Note: buildInstructionOperandReference may insert new AsmOperands, so |
1612 | // don't precompute the loop bound, i.e., cannot use range based for loop |
1613 | // here. |
1614 | for (size_t Idx = 0; Idx < II->AsmOperands.size(); ++Idx) { |
1615 | MatchableInfo::AsmOperand &Op = II->AsmOperands[Idx]; |
1616 | StringRef Token = Op.Token; |
1617 | // Check for singleton registers. |
1618 | if (const Record *RegRecord = Op.SingletonReg) { |
1619 | Op.Class = RegisterClasses[RegRecord]; |
1620 | assert(Op.Class && Op.Class->Registers.size() == 1 && |
1621 | "Unexpected class for singleton register" ); |
1622 | continue; |
1623 | } |
1624 | |
1625 | // Check for simple tokens. |
1626 | if (Token[0] != '$') { |
1627 | Op.Class = getTokenClass(Token); |
1628 | continue; |
1629 | } |
1630 | |
1631 | if (Token.size() > 1 && isdigit(Token[1])) { |
1632 | Op.Class = getTokenClass(Token); |
1633 | continue; |
1634 | } |
1635 | |
1636 | // Otherwise this is an operand reference. |
1637 | StringRef OperandName; |
1638 | if (Token[1] == '{') |
1639 | OperandName = Token.substr(Start: 2, N: Token.size() - 3); |
1640 | else |
1641 | OperandName = Token.substr(Start: 1); |
1642 | |
1643 | if (isa<const CodeGenInstruction *>(Val: II->DefRec)) |
1644 | buildInstructionOperandReference(II: II.get(), OpName: OperandName, AsmOpIdx: Idx); |
1645 | else |
1646 | buildAliasOperandReference(II: II.get(), OpName: OperandName, Op); |
1647 | } |
1648 | |
1649 | if (isa<const CodeGenInstruction *>(Val: II->DefRec)) { |
1650 | II->buildInstructionResultOperands(); |
1651 | // If the instruction has a two-operand alias, build up the |
1652 | // matchable here. We'll add them in bulk at the end to avoid |
1653 | // confusing this loop. |
1654 | StringRef Constraint = |
1655 | II->TheDef->getValueAsString(FieldName: "TwoOperandAliasConstraint" ); |
1656 | if (Constraint != "" ) { |
1657 | // Start by making a copy of the original matchable. |
1658 | auto AliasII = std::make_unique<MatchableInfo>(args&: *II); |
1659 | |
1660 | // Adjust it to be a two-operand alias. |
1661 | AliasII->formTwoOperandAlias(Constraint); |
1662 | |
1663 | // Add the alias to the matchables list. |
1664 | NewMatchables.push_back(x: std::move(AliasII)); |
1665 | } |
1666 | } else { |
1667 | // FIXME: The tied operands checking is not yet integrated with the |
1668 | // framework for reporting multiple near misses. To prevent invalid |
1669 | // formats from being matched with an alias if a tied-operands check |
1670 | // would otherwise have disallowed it, we just disallow such constructs |
1671 | // in TableGen completely. |
1672 | II->buildAliasResultOperands(AliasConstraintsAreChecked: !ReportMultipleNearMisses); |
1673 | } |
1674 | } |
1675 | if (!NewMatchables.empty()) |
1676 | Matchables.insert(position: Matchables.end(), |
1677 | first: std::make_move_iterator(i: NewMatchables.begin()), |
1678 | last: std::make_move_iterator(i: NewMatchables.end())); |
1679 | |
1680 | // Process token alias definitions and set up the associated superclass |
1681 | // information. |
1682 | for (const Record *Rec : Records.getAllDerivedDefinitions(ClassName: "TokenAlias" )) { |
1683 | ClassInfo *FromClass = getTokenClass(Token: Rec->getValueAsString(FieldName: "FromToken" )); |
1684 | ClassInfo *ToClass = getTokenClass(Token: Rec->getValueAsString(FieldName: "ToToken" )); |
1685 | if (FromClass == ToClass) |
1686 | PrintFatalError(ErrorLoc: Rec->getLoc(), |
1687 | Msg: "error: Destination value identical to source value." ); |
1688 | FromClass->SuperClasses.push_back(x: ToClass); |
1689 | } |
1690 | |
1691 | // Reorder classes so that classes precede super classes. |
1692 | Classes.sort(); |
1693 | |
1694 | #ifdef EXPENSIVE_CHECKS |
1695 | // Verify that the table is sorted and operator < works transitively. |
1696 | for (auto I = Classes.begin(), E = Classes.end(); I != E; ++I) { |
1697 | for (auto J = I; J != E; ++J) { |
1698 | assert(!(*J < *I)); |
1699 | assert(I == J || !J->isSubsetOf(*I)); |
1700 | } |
1701 | } |
1702 | #endif |
1703 | } |
1704 | |
1705 | /// buildInstructionOperandReference - The specified operand is a reference to a |
1706 | /// named operand such as $src. Resolve the Class and OperandInfo pointers. |
1707 | void AsmMatcherInfo::buildInstructionOperandReference(MatchableInfo *II, |
1708 | StringRef OperandName, |
1709 | unsigned AsmOpIdx) { |
1710 | const CodeGenInstruction &CGI = *cast<const CodeGenInstruction *>(Val&: II->DefRec); |
1711 | const CGIOperandList &Operands = CGI.Operands; |
1712 | MatchableInfo::AsmOperand *Op = &II->AsmOperands[AsmOpIdx]; |
1713 | |
1714 | // Map this token to an operand. |
1715 | std::optional<unsigned> Idx = Operands.findOperandNamed(Name: OperandName); |
1716 | if (!Idx) |
1717 | PrintFatalError(ErrorLoc: II->TheDef->getLoc(), |
1718 | Msg: "error: unable to find operand: '" + OperandName + "'" ); |
1719 | // If the instruction operand has multiple suboperands, but the parser |
1720 | // match class for the asm operand is still the default "ImmAsmOperand", |
1721 | // then handle each suboperand separately. |
1722 | if (Op->SubOpIdx == -1 && Operands[*Idx].MINumOperands > 1) { |
1723 | const Record *Rec = Operands[*Idx].Rec; |
1724 | assert(Rec->isSubClassOf("Operand" ) && "Unexpected operand!" ); |
1725 | const Record *MatchClass = Rec->getValueAsDef(FieldName: "ParserMatchClass" ); |
1726 | if (MatchClass && MatchClass->getValueAsString(FieldName: "Name" ) == "Imm" ) { |
1727 | // Insert remaining suboperands after AsmOpIdx in II->AsmOperands. |
1728 | StringRef Token = Op->Token; // save this in case Op gets moved |
1729 | for (unsigned SI = 1, SE = Operands[*Idx].MINumOperands; SI != SE; ++SI) { |
1730 | MatchableInfo::AsmOperand NewAsmOp(/*IsIsolatedToken=*/true, Token); |
1731 | NewAsmOp.SubOpIdx = SI; |
1732 | II->AsmOperands.insert(I: II->AsmOperands.begin() + AsmOpIdx + SI, |
1733 | Elt: NewAsmOp); |
1734 | } |
1735 | // Replace Op with first suboperand. |
1736 | Op = &II->AsmOperands[AsmOpIdx]; // update the pointer in case it moved |
1737 | Op->SubOpIdx = 0; |
1738 | } |
1739 | } |
1740 | |
1741 | // Set up the operand class. |
1742 | Op->Class = getOperandClass(OI: Operands[*Idx], SubOpIdx: Op->SubOpIdx); |
1743 | Op->OrigSrcOpName = OperandName; |
1744 | |
1745 | // If the named operand is tied, canonicalize it to the untied operand. |
1746 | // For example, something like: |
1747 | // (outs GPR:$dst), (ins GPR:$src) |
1748 | // with an asmstring of |
1749 | // "inc $src" |
1750 | // we want to canonicalize to: |
1751 | // "inc $dst" |
1752 | // so that we know how to provide the $dst operand when filling in the result. |
1753 | int OITied = -1; |
1754 | if (Operands[*Idx].MINumOperands == 1) |
1755 | OITied = Operands[*Idx].getTiedRegister(); |
1756 | if (OITied != -1) { |
1757 | // The tied operand index is an MIOperand index, find the operand that |
1758 | // contains it. |
1759 | auto [OpIdx, SubopIdx] = Operands.getSubOperandNumber(Op: OITied); |
1760 | OperandName = Operands[OpIdx].Name; |
1761 | Op->SubOpIdx = SubopIdx; |
1762 | } |
1763 | |
1764 | Op->SrcOpName = OperandName; |
1765 | } |
1766 | |
1767 | /// buildAliasOperandReference - When parsing an operand reference out of the |
1768 | /// matching string (e.g. "movsx $src, $dst"), determine what the class of the |
1769 | /// operand reference is by looking it up in the result pattern definition. |
1770 | void AsmMatcherInfo::buildAliasOperandReference(MatchableInfo *II, |
1771 | StringRef OperandName, |
1772 | MatchableInfo::AsmOperand &Op) { |
1773 | const CodeGenInstAlias &CGA = *cast<const CodeGenInstAlias *>(Val&: II->DefRec); |
1774 | |
1775 | // Set up the operand class. |
1776 | for (const auto &[ResultOp, SubOpIdx] : |
1777 | zip_equal(t: CGA.ResultOperands, u: CGA.ResultInstOperandIndex)) { |
1778 | if (ResultOp.isRecord() && ResultOp.getName() == OperandName) { |
1779 | // It's safe to go with the first one we find, because CodeGenInstAlias |
1780 | // validates that all operands with the same name have the same record. |
1781 | Op.SubOpIdx = SubOpIdx.second; |
1782 | // Use the match class from the Alias definition, not the |
1783 | // destination instruction, as we may have an immediate that's |
1784 | // being munged by the match class. |
1785 | Op.Class = getOperandClass(Rec: ResultOp.getRecord(), SubOpIdx: Op.SubOpIdx); |
1786 | Op.SrcOpName = OperandName; |
1787 | Op.OrigSrcOpName = OperandName; |
1788 | return; |
1789 | } |
1790 | } |
1791 | |
1792 | PrintFatalError(ErrorLoc: II->TheDef->getLoc(), |
1793 | Msg: "error: unable to find operand: '" + OperandName + "'" ); |
1794 | } |
1795 | |
1796 | void MatchableInfo::buildInstructionResultOperands() { |
1797 | const CodeGenInstruction *ResultInst = getResultInst(); |
1798 | |
1799 | // Loop over all operands of the result instruction, determining how to |
1800 | // populate them. |
1801 | for (const CGIOperandList::OperandInfo &OpInfo : ResultInst->Operands) { |
1802 | // If this is a tied operand, just copy from the previously handled operand. |
1803 | int TiedOp = -1; |
1804 | if (OpInfo.MINumOperands == 1) |
1805 | TiedOp = OpInfo.getTiedRegister(); |
1806 | if (TiedOp != -1) { |
1807 | int TiedSrcOperand = findAsmOperandOriginallyNamed(N: OpInfo.Name); |
1808 | if (TiedSrcOperand != -1 && |
1809 | ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand) |
1810 | ResOperands.push_back(Elt: ResOperand::getTiedOp( |
1811 | TiedOperandNum: TiedOp, SrcOperand1: ResOperands[TiedOp].AsmOperandNum, SrcOperand2: TiedSrcOperand)); |
1812 | else |
1813 | ResOperands.push_back(Elt: ResOperand::getTiedOp(TiedOperandNum: TiedOp, SrcOperand1: 0, SrcOperand2: 0)); |
1814 | continue; |
1815 | } |
1816 | |
1817 | int SrcOperand = findAsmOperandNamed(N: OpInfo.Name); |
1818 | if (OpInfo.Name.empty() || SrcOperand == -1) { |
1819 | // This may happen for operands that are tied to a suboperand of a |
1820 | // complex operand. Simply use a dummy value here; nobody should |
1821 | // use this operand slot. |
1822 | // FIXME: The long term goal is for the MCOperand list to not contain |
1823 | // tied operands at all. |
1824 | ResOperands.push_back(Elt: ResOperand::getImmOp(Val: 0)); |
1825 | continue; |
1826 | } |
1827 | |
1828 | // Check if the one AsmOperand populates the entire operand. |
1829 | unsigned NumOperands = OpInfo.MINumOperands; |
1830 | if (AsmOperands[SrcOperand].SubOpIdx == -1) { |
1831 | ResOperands.push_back(Elt: ResOperand::getRenderedOp(AsmOpNum: SrcOperand, NumOperands)); |
1832 | continue; |
1833 | } |
1834 | |
1835 | // Add a separate ResOperand for each suboperand. |
1836 | for (unsigned AI = 0; AI < NumOperands; ++AI) { |
1837 | assert(AsmOperands[SrcOperand + AI].SubOpIdx == (int)AI && |
1838 | AsmOperands[SrcOperand + AI].SrcOpName == OpInfo.Name && |
1839 | "unexpected AsmOperands for suboperands" ); |
1840 | ResOperands.push_back(Elt: ResOperand::getRenderedOp(AsmOpNum: SrcOperand + AI, NumOperands: 1)); |
1841 | } |
1842 | } |
1843 | } |
1844 | |
1845 | void MatchableInfo::buildAliasResultOperands(bool AliasConstraintsAreChecked) { |
1846 | const CodeGenInstAlias &CGA = *cast<const CodeGenInstAlias *>(Val&: DefRec); |
1847 | const CodeGenInstruction *ResultInst = getResultInst(); |
1848 | |
1849 | // Map of: $reg -> #lastref |
1850 | // where $reg is the name of the operand in the asm string |
1851 | // where #lastref is the last processed index where $reg was referenced in |
1852 | // the asm string. |
1853 | SmallDenseMap<StringRef, int> OperandRefs; |
1854 | |
1855 | // Loop over all operands of the result instruction, determining how to |
1856 | // populate them. |
1857 | unsigned AliasOpNo = 0; |
1858 | unsigned LastOpNo = CGA.ResultInstOperandIndex.size(); |
1859 | for (const auto &[Idx, OpInfo] : enumerate(First: ResultInst->Operands)) { |
1860 | // If this is a tied operand, just copy from the previously handled operand. |
1861 | int TiedOp = -1; |
1862 | if (OpInfo.MINumOperands == 1) |
1863 | TiedOp = OpInfo.getTiedRegister(); |
1864 | if (TiedOp != -1) { |
1865 | unsigned SrcOp1 = 0; |
1866 | unsigned SrcOp2 = 0; |
1867 | |
1868 | // If an operand has been specified twice in the asm string, |
1869 | // add the two source operand's indices to the TiedOp so that |
1870 | // at runtime the 'tied' constraint is checked. |
1871 | if (ResOperands[TiedOp].Kind == ResOperand::RenderAsmOperand) { |
1872 | SrcOp1 = ResOperands[TiedOp].AsmOperandNum; |
1873 | |
1874 | // Find the next operand (similarly named operand) in the string. |
1875 | StringRef Name = AsmOperands[SrcOp1].SrcOpName; |
1876 | auto Insert = OperandRefs.try_emplace(Key: Name, Args&: SrcOp1); |
1877 | SrcOp2 = findAsmOperandNamed(N: Name, LastIdx: Insert.first->second); |
1878 | |
1879 | // Not updating the record in OperandRefs will cause TableGen |
1880 | // to fail with an error at the end of this function. |
1881 | if (AliasConstraintsAreChecked) |
1882 | Insert.first->second = SrcOp2; |
1883 | |
1884 | // In case it only has one reference in the asm string, |
1885 | // it doesn't need to be checked for tied constraints. |
1886 | SrcOp2 = (SrcOp2 == (unsigned)-1) ? SrcOp1 : SrcOp2; |
1887 | } |
1888 | |
1889 | // If the alias operand is of a different operand class, we only want |
1890 | // to benefit from the tied-operands check and just match the operand |
1891 | // as a normal, but not copy the original (TiedOp) to the result |
1892 | // instruction. We do this by passing -1 as the tied operand to copy. |
1893 | if (OpInfo.Rec->getName() != |
1894 | ResultInst->Operands[TiedOp].Rec->getName()) { |
1895 | SrcOp1 = ResOperands[TiedOp].AsmOperandNum; |
1896 | int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second; |
1897 | StringRef Name = CGA.ResultOperands[AliasOpNo].getName(); |
1898 | SrcOp2 = findAsmOperand(N: Name, SubOpIdx: SubIdx); |
1899 | ResOperands.push_back( |
1900 | Elt: ResOperand::getTiedOp(TiedOperandNum: (unsigned)-1, SrcOperand1: SrcOp1, SrcOperand2: SrcOp2)); |
1901 | } else { |
1902 | ResOperands.push_back(Elt: ResOperand::getTiedOp(TiedOperandNum: TiedOp, SrcOperand1: SrcOp1, SrcOperand2: SrcOp2)); |
1903 | continue; |
1904 | } |
1905 | } |
1906 | |
1907 | // Handle all the suboperands for this operand. |
1908 | StringRef OpName = OpInfo.Name; |
1909 | for (; AliasOpNo < LastOpNo && |
1910 | CGA.ResultInstOperandIndex[AliasOpNo].first == Idx; |
1911 | ++AliasOpNo) { |
1912 | int SubIdx = CGA.ResultInstOperandIndex[AliasOpNo].second; |
1913 | |
1914 | // Find out what operand from the asmparser that this MCInst operand |
1915 | // comes from. |
1916 | switch (CGA.ResultOperands[AliasOpNo].Kind) { |
1917 | case CodeGenInstAlias::ResultOperand::K_Record: { |
1918 | StringRef Name = CGA.ResultOperands[AliasOpNo].getName(); |
1919 | int SrcOperand = findAsmOperand(N: Name, SubOpIdx: SubIdx); |
1920 | if (SrcOperand == -1) |
1921 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
1922 | Msg: "Instruction '" + TheDef->getName() + |
1923 | "' has operand '" + OpName + |
1924 | "' that doesn't appear in asm string!" ); |
1925 | |
1926 | // Add it to the operand references. If it is added a second time, the |
1927 | // record won't be updated and it will fail later on. |
1928 | OperandRefs.try_emplace(Key: Name, Args&: SrcOperand); |
1929 | |
1930 | unsigned NumOperands = (SubIdx == -1 ? OpInfo.MINumOperands : 1); |
1931 | ResOperands.push_back( |
1932 | Elt: ResOperand::getRenderedOp(AsmOpNum: SrcOperand, NumOperands)); |
1933 | break; |
1934 | } |
1935 | case CodeGenInstAlias::ResultOperand::K_Imm: { |
1936 | int64_t ImmVal = CGA.ResultOperands[AliasOpNo].getImm(); |
1937 | ResOperands.push_back(Elt: ResOperand::getImmOp(Val: ImmVal)); |
1938 | break; |
1939 | } |
1940 | case CodeGenInstAlias::ResultOperand::K_Reg: { |
1941 | const Record *Reg = CGA.ResultOperands[AliasOpNo].getRegister(); |
1942 | ResOperands.push_back(Elt: ResOperand::getRegOp(Reg)); |
1943 | break; |
1944 | } |
1945 | } |
1946 | } |
1947 | } |
1948 | |
1949 | // Check that operands are not repeated more times than is supported. |
1950 | for (auto &T : OperandRefs) { |
1951 | if (T.second != -1 && findAsmOperandNamed(N: T.first, LastIdx: T.second) != -1) |
1952 | PrintFatalError(ErrorLoc: TheDef->getLoc(), |
1953 | Msg: "Operand '" + T.first + "' can never be matched" ); |
1954 | } |
1955 | } |
1956 | |
1957 | static unsigned |
1958 | getConverterOperandID(const std::string &Name, |
1959 | SmallSetVector<CachedHashString, 16> &Table, |
1960 | bool &IsNew) { |
1961 | IsNew = Table.insert(X: CachedHashString(Name)); |
1962 | |
1963 | unsigned ID = IsNew ? Table.size() - 1 : find(Range&: Table, Val: Name) - Table.begin(); |
1964 | |
1965 | assert(ID < Table.size()); |
1966 | |
1967 | return ID; |
1968 | } |
1969 | |
1970 | static unsigned |
1971 | emitConvertFuncs(CodeGenTarget &Target, StringRef ClassName, |
1972 | std::vector<std::unique_ptr<MatchableInfo>> &Infos, |
1973 | bool HasMnemonicFirst, bool HasOptionalOperands, |
1974 | raw_ostream &OS) { |
1975 | SmallSetVector<CachedHashString, 16> OperandConversionKinds; |
1976 | SmallSetVector<CachedHashString, 16> InstructionConversionKinds; |
1977 | std::vector<std::vector<uint8_t>> ConversionTable; |
1978 | size_t MaxRowLength = 2; // minimum is custom converter plus terminator. |
1979 | |
1980 | // TargetOperandClass - This is the target's operand class, like X86Operand. |
1981 | std::string TargetOperandClass = Target.getName().str() + "Operand" ; |
1982 | |
1983 | // Write the convert function to a separate stream, so we can drop it after |
1984 | // the enum. We'll build up the conversion handlers for the individual |
1985 | // operand types opportunistically as we encounter them. |
1986 | std::string ConvertFnBody; |
1987 | raw_string_ostream CvtOS(ConvertFnBody); |
1988 | // Start the unified conversion function. |
1989 | if (HasOptionalOperands) { |
1990 | CvtOS << "void " << Target.getName() << ClassName << "::\n" |
1991 | << "convertToMCInst(unsigned Kind, MCInst &Inst, " |
1992 | << "unsigned Opcode,\n" |
1993 | << " const OperandVector &Operands,\n" |
1994 | << " const SmallBitVector &OptionalOperandsMask,\n" |
1995 | << " ArrayRef<unsigned> DefaultsOffset) {\n" ; |
1996 | } else { |
1997 | CvtOS << "void " << Target.getName() << ClassName << "::\n" |
1998 | << "convertToMCInst(unsigned Kind, MCInst &Inst, " |
1999 | << "unsigned Opcode,\n" |
2000 | << " const OperandVector &Operands) {\n" ; |
2001 | } |
2002 | CvtOS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n" ; |
2003 | CvtOS << " const uint8_t *Converter = ConversionTable[Kind];\n" ; |
2004 | CvtOS << " Inst.setOpcode(Opcode);\n" ; |
2005 | CvtOS << " for (const uint8_t *p = Converter; *p; p += 2) {\n" ; |
2006 | if (HasOptionalOperands) { |
2007 | // When optional operands are involved, formal and actual operand indices |
2008 | // may differ. Map the former to the latter by subtracting the number of |
2009 | // absent optional operands. |
2010 | // FIXME: This is not an operand index in the CVT_Tied case |
2011 | CvtOS << " unsigned OpIdx = *(p + 1) - DefaultsOffset[*(p + 1)];\n" ; |
2012 | } else { |
2013 | CvtOS << " unsigned OpIdx = *(p + 1);\n" ; |
2014 | } |
2015 | CvtOS << " switch (*p) {\n" ; |
2016 | CvtOS << " default: llvm_unreachable(\"invalid conversion entry!\");\n" ; |
2017 | CvtOS << " case CVT_Reg:\n" ; |
2018 | CvtOS << " static_cast<" << TargetOperandClass |
2019 | << " &>(*Operands[OpIdx]).addRegOperands(Inst, 1);\n" ; |
2020 | CvtOS << " break;\n" ; |
2021 | CvtOS << " case CVT_Tied: {\n" ; |
2022 | CvtOS << " assert(*(p + 1) < (size_t)(std::end(TiedAsmOperandTable) -\n" ; |
2023 | CvtOS |
2024 | << " std::begin(TiedAsmOperandTable)) &&\n" ; |
2025 | CvtOS << " \"Tied operand not found\");\n" ; |
2026 | CvtOS << " unsigned TiedResOpnd = TiedAsmOperandTable[*(p + 1)][0];\n" ; |
2027 | CvtOS << " if (TiedResOpnd != (uint8_t)-1)\n" ; |
2028 | CvtOS << " Inst.addOperand(Inst.getOperand(TiedResOpnd));\n" ; |
2029 | CvtOS << " break;\n" ; |
2030 | CvtOS << " }\n" ; |
2031 | |
2032 | std::string OperandFnBody; |
2033 | raw_string_ostream OpOS(OperandFnBody); |
2034 | // Start the operand number lookup function. |
2035 | OpOS << "void " << Target.getName() << ClassName << "::\n" |
2036 | << "convertToMapAndConstraints(unsigned Kind,\n" ; |
2037 | OpOS.indent(NumSpaces: 27); |
2038 | OpOS << "const OperandVector &Operands) {\n" |
2039 | << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n" |
2040 | << " unsigned NumMCOperands = 0;\n" |
2041 | << " const uint8_t *Converter = ConversionTable[Kind];\n" |
2042 | << " for (const uint8_t *p = Converter; *p; p += 2) {\n" |
2043 | << " switch (*p) {\n" |
2044 | << " default: llvm_unreachable(\"invalid conversion entry!\");\n" |
2045 | << " case CVT_Reg:\n" |
2046 | << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" |
2047 | << " Operands[*(p + 1)]->setConstraint(\"r\");\n" |
2048 | << " ++NumMCOperands;\n" |
2049 | << " break;\n" |
2050 | << " case CVT_Tied:\n" |
2051 | << " ++NumMCOperands;\n" |
2052 | << " break;\n" ; |
2053 | |
2054 | // Pre-populate the operand conversion kinds with the standard always |
2055 | // available entries. |
2056 | OperandConversionKinds.insert(X: CachedHashString("CVT_Done" )); |
2057 | OperandConversionKinds.insert(X: CachedHashString("CVT_Reg" )); |
2058 | OperandConversionKinds.insert(X: CachedHashString("CVT_Tied" )); |
2059 | enum { CVT_Done, CVT_Reg, CVT_Tied }; |
2060 | |
2061 | // Map of e.g. <0, 2, 3> -> "Tie_0_2_3" enum label. |
2062 | std::map<std::tuple<uint8_t, uint8_t, uint8_t>, std::string> |
2063 | TiedOperandsEnumMap; |
2064 | |
2065 | for (auto &II : Infos) { |
2066 | // Check if we have a custom match function. |
2067 | StringRef AsmMatchConverter = |
2068 | II->getResultInst()->TheDef->getValueAsString(FieldName: "AsmMatchConverter" ); |
2069 | if (!AsmMatchConverter.empty() && II->UseInstAsmMatchConverter) { |
2070 | std::string Signature = ("ConvertCustom_" + AsmMatchConverter).str(); |
2071 | II->ConversionFnKind = Signature; |
2072 | |
2073 | // Check if we have already generated this signature. |
2074 | if (!InstructionConversionKinds.insert(X: CachedHashString(Signature))) |
2075 | continue; |
2076 | |
2077 | // Remember this converter for the kind enum. |
2078 | unsigned KindID = OperandConversionKinds.size(); |
2079 | OperandConversionKinds.insert( |
2080 | X: CachedHashString("CVT_" + getEnumNameForToken(Str: AsmMatchConverter))); |
2081 | |
2082 | // Add the converter row for this instruction. |
2083 | ConversionTable.emplace_back(); |
2084 | ConversionTable.back().push_back(x: KindID); |
2085 | ConversionTable.back().push_back(x: CVT_Done); |
2086 | |
2087 | // Add the handler to the conversion driver function. |
2088 | CvtOS << " case CVT_" << getEnumNameForToken(Str: AsmMatchConverter) |
2089 | << ":\n" |
2090 | << " " << AsmMatchConverter << "(Inst, Operands);\n" |
2091 | << " break;\n" ; |
2092 | |
2093 | // FIXME: Handle the operand number lookup for custom match functions. |
2094 | continue; |
2095 | } |
2096 | |
2097 | // Build the conversion function signature. |
2098 | std::string Signature = "Convert" ; |
2099 | |
2100 | std::vector<uint8_t> ConversionRow; |
2101 | |
2102 | // Compute the convert enum and the case body. |
2103 | MaxRowLength = std::max(a: MaxRowLength, b: II->ResOperands.size() * 2 + 1); |
2104 | |
2105 | for (const auto &[Idx, OpInfo] : enumerate(First&: II->ResOperands)) { |
2106 | // Generate code to populate each result operand. |
2107 | switch (OpInfo.Kind) { |
2108 | case MatchableInfo::ResOperand::RenderAsmOperand: { |
2109 | // This comes from something we parsed. |
2110 | const MatchableInfo::AsmOperand &Op = |
2111 | II->AsmOperands[OpInfo.AsmOperandNum]; |
2112 | |
2113 | // Registers are always converted the same, don't duplicate the |
2114 | // conversion function based on them. |
2115 | Signature += "__" ; |
2116 | std::string Class; |
2117 | Class = Op.Class->isRegisterClass() ? "Reg" : Op.Class->ClassName; |
2118 | Signature += Class; |
2119 | Signature += utostr(X: OpInfo.MINumOperands); |
2120 | Signature += "_" + itostr(X: OpInfo.AsmOperandNum); |
2121 | |
2122 | // Add the conversion kind, if necessary, and get the associated ID |
2123 | // the index of its entry in the vector). |
2124 | std::string Name = |
2125 | "CVT_" + |
2126 | (Op.Class->isRegisterClass() ? "Reg" : Op.Class->RenderMethod); |
2127 | if (Op.Class->IsOptional) { |
2128 | // For optional operands we must also care about DefaultMethod |
2129 | assert(HasOptionalOperands); |
2130 | Name += "_" + Op.Class->DefaultMethod; |
2131 | } |
2132 | Name = getEnumNameForToken(Str: Name); |
2133 | |
2134 | bool IsNewConverter = false; |
2135 | unsigned ID = |
2136 | getConverterOperandID(Name, Table&: OperandConversionKinds, IsNew&: IsNewConverter); |
2137 | |
2138 | // Add the operand entry to the instruction kind conversion row. |
2139 | ConversionRow.push_back(x: ID); |
2140 | ConversionRow.push_back(x: OpInfo.AsmOperandNum + HasMnemonicFirst); |
2141 | |
2142 | if (!IsNewConverter) |
2143 | break; |
2144 | |
2145 | // This is a new operand kind. Add a handler for it to the |
2146 | // converter driver. |
2147 | CvtOS << " case " << Name << ":\n" ; |
2148 | if (Op.Class->IsOptional) { |
2149 | // If optional operand is not present in actual instruction then we |
2150 | // should call its DefaultMethod before RenderMethod |
2151 | assert(HasOptionalOperands); |
2152 | CvtOS << " if (OptionalOperandsMask[*(p + 1) - 1]) {\n" |
2153 | << " " << Op.Class->DefaultMethod << "()" |
2154 | << "->" << Op.Class->RenderMethod << "(Inst, " |
2155 | << OpInfo.MINumOperands << ");\n" |
2156 | << " } else {\n" |
2157 | << " static_cast<" << TargetOperandClass |
2158 | << " &>(*Operands[OpIdx])." << Op.Class->RenderMethod |
2159 | << "(Inst, " << OpInfo.MINumOperands << ");\n" |
2160 | << " }\n" ; |
2161 | } else { |
2162 | CvtOS << " static_cast<" << TargetOperandClass |
2163 | << " &>(*Operands[OpIdx])." << Op.Class->RenderMethod |
2164 | << "(Inst, " << OpInfo.MINumOperands << ");\n" ; |
2165 | } |
2166 | CvtOS << " break;\n" ; |
2167 | |
2168 | // Add a handler for the operand number lookup. |
2169 | OpOS << " case " << Name << ":\n" |
2170 | << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" ; |
2171 | |
2172 | if (Op.Class->isRegisterClass()) |
2173 | OpOS << " Operands[*(p + 1)]->setConstraint(\"r\");\n" ; |
2174 | else |
2175 | OpOS << " Operands[*(p + 1)]->setConstraint(\"m\");\n" ; |
2176 | OpOS << " NumMCOperands += " << OpInfo.MINumOperands << ";\n" |
2177 | << " break;\n" ; |
2178 | break; |
2179 | } |
2180 | case MatchableInfo::ResOperand::TiedOperand: { |
2181 | // If this operand is tied to a previous one, just copy the MCInst |
2182 | // operand from the earlier one.We can only tie single MCOperand values. |
2183 | assert(OpInfo.MINumOperands == 1 && "Not a singular MCOperand" ); |
2184 | uint8_t TiedOp = OpInfo.TiedOperands.ResOpnd; |
2185 | uint8_t SrcOp1 = OpInfo.TiedOperands.SrcOpnd1Idx + HasMnemonicFirst; |
2186 | uint8_t SrcOp2 = OpInfo.TiedOperands.SrcOpnd2Idx + HasMnemonicFirst; |
2187 | assert((Idx > TiedOp || TiedOp == (uint8_t)-1) && |
2188 | "Tied operand precedes its target!" ); |
2189 | auto TiedTupleName = std::string("Tie" ) + utostr(X: TiedOp) + '_' + |
2190 | utostr(X: SrcOp1) + '_' + utostr(X: SrcOp2); |
2191 | Signature += "__" + TiedTupleName; |
2192 | ConversionRow.push_back(x: CVT_Tied); |
2193 | ConversionRow.push_back(x: TiedOp); |
2194 | ConversionRow.push_back(x: SrcOp1); |
2195 | ConversionRow.push_back(x: SrcOp2); |
2196 | |
2197 | // Also create an 'enum' for this combination of tied operands. |
2198 | auto Key = std::tuple(TiedOp, SrcOp1, SrcOp2); |
2199 | TiedOperandsEnumMap.emplace(args&: Key, args&: TiedTupleName); |
2200 | break; |
2201 | } |
2202 | case MatchableInfo::ResOperand::ImmOperand: { |
2203 | int64_t Val = OpInfo.ImmVal; |
2204 | std::string Ty = "imm_" + itostr(X: Val); |
2205 | Ty = getEnumNameForToken(Str: Ty); |
2206 | Signature += "__" + Ty; |
2207 | |
2208 | std::string Name = "CVT_" + Ty; |
2209 | bool IsNewConverter = false; |
2210 | unsigned ID = |
2211 | getConverterOperandID(Name, Table&: OperandConversionKinds, IsNew&: IsNewConverter); |
2212 | // Add the operand entry to the instruction kind conversion row. |
2213 | ConversionRow.push_back(x: ID); |
2214 | ConversionRow.push_back(x: 0); |
2215 | |
2216 | if (!IsNewConverter) |
2217 | break; |
2218 | |
2219 | CvtOS << " case " << Name << ":\n" |
2220 | << " Inst.addOperand(MCOperand::createImm(" << Val << "));\n" |
2221 | << " break;\n" ; |
2222 | |
2223 | OpOS << " case " << Name << ":\n" |
2224 | << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" |
2225 | << " Operands[*(p + 1)]->setConstraint(\"\");\n" |
2226 | << " ++NumMCOperands;\n" |
2227 | << " break;\n" ; |
2228 | break; |
2229 | } |
2230 | case MatchableInfo::ResOperand::RegOperand: { |
2231 | std::string Reg, Name; |
2232 | if (!OpInfo.Register) { |
2233 | Name = "reg0" ; |
2234 | Reg = "0" ; |
2235 | } else { |
2236 | Reg = getQualifiedName(R: OpInfo.Register); |
2237 | Name = "reg" + OpInfo.Register->getName().str(); |
2238 | } |
2239 | Signature += "__" + Name; |
2240 | Name = "CVT_" + Name; |
2241 | bool IsNewConverter = false; |
2242 | unsigned ID = |
2243 | getConverterOperandID(Name, Table&: OperandConversionKinds, IsNew&: IsNewConverter); |
2244 | // Add the operand entry to the instruction kind conversion row. |
2245 | ConversionRow.push_back(x: ID); |
2246 | ConversionRow.push_back(x: 0); |
2247 | |
2248 | if (!IsNewConverter) |
2249 | break; |
2250 | CvtOS << " case " << Name << ":\n" |
2251 | << " Inst.addOperand(MCOperand::createReg(" << Reg << "));\n" |
2252 | << " break;\n" ; |
2253 | |
2254 | OpOS << " case " << Name << ":\n" |
2255 | << " Operands[*(p + 1)]->setMCOperandNum(NumMCOperands);\n" |
2256 | << " Operands[*(p + 1)]->setConstraint(\"m\");\n" |
2257 | << " ++NumMCOperands;\n" |
2258 | << " break;\n" ; |
2259 | } |
2260 | } |
2261 | } |
2262 | |
2263 | // If there were no operands, add to the signature to that effect |
2264 | if (Signature == "Convert" ) |
2265 | Signature += "_NoOperands" ; |
2266 | |
2267 | II->ConversionFnKind = Signature; |
2268 | |
2269 | // Save the signature. If we already have it, don't add a new row |
2270 | // to the table. |
2271 | if (!InstructionConversionKinds.insert(X: CachedHashString(Signature))) |
2272 | continue; |
2273 | |
2274 | // Add the row to the table. |
2275 | ConversionTable.push_back(x: std::move(ConversionRow)); |
2276 | } |
2277 | |
2278 | // Finish up the converter driver function. |
2279 | CvtOS << " }\n }\n}\n\n" ; |
2280 | |
2281 | // Finish up the operand number lookup function. |
2282 | OpOS << " }\n }\n}\n\n" ; |
2283 | |
2284 | // Output a static table for tied operands. |
2285 | if (TiedOperandsEnumMap.size()) { |
2286 | // The number of tied operand combinations will be small in practice, |
2287 | // but just add the assert to be sure. |
2288 | assert(TiedOperandsEnumMap.size() <= 254 && |
2289 | "Too many tied-operand combinations to reference with " |
2290 | "an 8bit offset from the conversion table, where index " |
2291 | "'255' is reserved as operand not to be copied." ); |
2292 | |
2293 | OS << "enum {\n" ; |
2294 | for (auto &KV : TiedOperandsEnumMap) { |
2295 | OS << " " << KV.second << ",\n" ; |
2296 | } |
2297 | OS << "};\n\n" ; |
2298 | |
2299 | OS << "static const uint8_t TiedAsmOperandTable[][3] = {\n" ; |
2300 | for (auto &KV : TiedOperandsEnumMap) { |
2301 | OS << " /* " << KV.second << " */ { " << utostr(X: std::get<0>(t: KV.first)) |
2302 | << ", " << utostr(X: std::get<1>(t: KV.first)) << ", " |
2303 | << utostr(X: std::get<2>(t: KV.first)) << " },\n" ; |
2304 | } |
2305 | OS << "};\n\n" ; |
2306 | } else { |
2307 | OS << "static const uint8_t TiedAsmOperandTable[][3] = " |
2308 | "{ /* empty */ {0, 0, 0} };\n\n" ; |
2309 | } |
2310 | |
2311 | OS << "namespace {\n" ; |
2312 | |
2313 | // Output the operand conversion kind enum. |
2314 | OS << "enum OperatorConversionKind {\n" ; |
2315 | for (const auto &Converter : OperandConversionKinds) |
2316 | OS << " " << Converter << ",\n" ; |
2317 | OS << " CVT_NUM_CONVERTERS\n" ; |
2318 | OS << "};\n\n" ; |
2319 | |
2320 | // Output the instruction conversion kind enum. |
2321 | OS << "enum InstructionConversionKind {\n" ; |
2322 | for (const auto &Signature : InstructionConversionKinds) |
2323 | OS << " " << Signature << ",\n" ; |
2324 | OS << " CVT_NUM_SIGNATURES\n" ; |
2325 | OS << "};\n\n" ; |
2326 | |
2327 | OS << "} // end anonymous namespace\n\n" ; |
2328 | |
2329 | // Output the conversion table. |
2330 | OS << "static const uint8_t ConversionTable[CVT_NUM_SIGNATURES][" |
2331 | << MaxRowLength << "] = {\n" ; |
2332 | |
2333 | for (unsigned Row = 0, ERow = ConversionTable.size(); Row != ERow; ++Row) { |
2334 | assert(ConversionTable[Row].size() % 2 == 0 && "bad conversion row!" ); |
2335 | OS << " // " << InstructionConversionKinds[Row] << "\n" ; |
2336 | OS << " { " ; |
2337 | for (unsigned i = 0, e = ConversionTable[Row].size(); i != e; i += 2) { |
2338 | const auto &OCK = OperandConversionKinds[ConversionTable[Row][i]]; |
2339 | OS << OCK << ", " ; |
2340 | if (OCK != CachedHashString("CVT_Tied" )) { |
2341 | OS << (unsigned)(ConversionTable[Row][i + 1]) << ", " ; |
2342 | continue; |
2343 | } |
2344 | |
2345 | // For a tied operand, emit a reference to the TiedAsmOperandTable |
2346 | // that contains the operand to copy, and the parsed operands to |
2347 | // check for their tied constraints. |
2348 | auto Key = std::tuple((uint8_t)ConversionTable[Row][i + 1], |
2349 | (uint8_t)ConversionTable[Row][i + 2], |
2350 | (uint8_t)ConversionTable[Row][i + 3]); |
2351 | auto TiedOpndEnum = TiedOperandsEnumMap.find(x: Key); |
2352 | assert(TiedOpndEnum != TiedOperandsEnumMap.end() && |
2353 | "No record for tied operand pair" ); |
2354 | OS << TiedOpndEnum->second << ", " ; |
2355 | i += 2; |
2356 | } |
2357 | OS << "CVT_Done },\n" ; |
2358 | } |
2359 | |
2360 | OS << "};\n\n" ; |
2361 | |
2362 | // Spit out the conversion driver function. |
2363 | OS << ConvertFnBody; |
2364 | |
2365 | // Spit out the operand number lookup function. |
2366 | OS << OperandFnBody; |
2367 | |
2368 | return ConversionTable.size(); |
2369 | } |
2370 | |
2371 | /// emitMatchClassEnumeration - Emit the enumeration for match class kinds. |
2372 | static void emitMatchClassEnumeration(CodeGenTarget &Target, |
2373 | std::forward_list<ClassInfo> &Infos, |
2374 | raw_ostream &OS) { |
2375 | OS << "namespace {\n\n" ; |
2376 | |
2377 | OS << "/// MatchClassKind - The kinds of classes which participate in\n" |
2378 | << "/// instruction matching.\n" ; |
2379 | OS << "enum MatchClassKind {\n" ; |
2380 | OS << " InvalidMatchClass = 0,\n" ; |
2381 | OS << " OptionalMatchClass = 1,\n" ; |
2382 | ClassInfo::ClassInfoKind LastKind = ClassInfo::Token; |
2383 | StringRef LastName = "OptionalMatchClass" ; |
2384 | for (const auto &CI : Infos) { |
2385 | if (LastKind == ClassInfo::Token && CI.Kind != ClassInfo::Token) { |
2386 | OS << " MCK_LAST_TOKEN = " << LastName << ",\n" ; |
2387 | } else if (LastKind < ClassInfo::UserClass0 && |
2388 | CI.Kind >= ClassInfo::UserClass0) { |
2389 | OS << " MCK_LAST_REGISTER = " << LastName << ",\n" ; |
2390 | } |
2391 | LastKind = (ClassInfo::ClassInfoKind)CI.Kind; |
2392 | LastName = CI.Name; |
2393 | |
2394 | OS << " " << CI.Name << ", // " ; |
2395 | if (CI.Kind == ClassInfo::Token) { |
2396 | OS << "'" << CI.ValueName << "'\n" ; |
2397 | } else if (CI.isRegisterClass()) { |
2398 | if (!CI.ValueName.empty()) |
2399 | OS << "register class '" << CI.ValueName << "'\n" ; |
2400 | else |
2401 | OS << "derived register class\n" ; |
2402 | } else { |
2403 | OS << "user defined class '" << CI.ValueName << "'\n" ; |
2404 | } |
2405 | } |
2406 | OS << " NumMatchClassKinds\n" ; |
2407 | OS << "};\n\n" ; |
2408 | |
2409 | OS << "} // end anonymous namespace\n\n" ; |
2410 | } |
2411 | |
2412 | /// emitMatchClassDiagStrings - Emit a function to get the diagnostic text to be |
2413 | /// used when an assembly operand does not match the expected operand class. |
2414 | static void emitOperandMatchErrorDiagStrings(AsmMatcherInfo &Info, |
2415 | raw_ostream &OS) { |
2416 | // If the target does not use DiagnosticString for any operands, don't emit |
2417 | // an unused function. |
2418 | if (llvm::all_of(Range&: Info.Classes, P: [](const ClassInfo &CI) { |
2419 | return CI.DiagnosticString.empty(); |
2420 | })) |
2421 | return; |
2422 | |
2423 | OS << "static const char *getMatchKindDiag(" << Info.Target.getName() |
2424 | << "AsmParser::" << Info.Target.getName() |
2425 | << "MatchResultTy MatchResult) {\n" ; |
2426 | OS << " switch (MatchResult) {\n" ; |
2427 | |
2428 | for (const auto &CI : Info.Classes) { |
2429 | if (!CI.DiagnosticString.empty()) { |
2430 | assert(!CI.DiagnosticType.empty() && |
2431 | "DiagnosticString set without DiagnosticType" ); |
2432 | OS << " case " << Info.Target.getName() << "AsmParser::Match_" |
2433 | << CI.DiagnosticType << ":\n" ; |
2434 | OS << " return \"" << CI.DiagnosticString << "\";\n" ; |
2435 | } |
2436 | } |
2437 | |
2438 | OS << " default:\n" ; |
2439 | OS << " return nullptr;\n" ; |
2440 | |
2441 | OS << " }\n" ; |
2442 | OS << "}\n\n" ; |
2443 | } |
2444 | |
2445 | static void emitRegisterMatchErrorFunc(AsmMatcherInfo &Info, raw_ostream &OS) { |
2446 | OS << "static unsigned getDiagKindFromRegisterClass(MatchClassKind " |
2447 | "RegisterClass) {\n" ; |
2448 | if (none_of(Range&: Info.Classes, P: [](const ClassInfo &CI) { |
2449 | return CI.isRegisterClass() && !CI.DiagnosticType.empty(); |
2450 | })) { |
2451 | OS << " return MCTargetAsmParser::Match_InvalidOperand;\n" ; |
2452 | } else { |
2453 | OS << " switch (RegisterClass) {\n" ; |
2454 | for (const auto &CI : Info.Classes) { |
2455 | if (CI.isRegisterClass() && !CI.DiagnosticType.empty()) { |
2456 | OS << " case " << CI.Name << ":\n" ; |
2457 | OS << " return " << Info.Target.getName() << "AsmParser::Match_" |
2458 | << CI.DiagnosticType << ";\n" ; |
2459 | } |
2460 | } |
2461 | |
2462 | OS << " default:\n" ; |
2463 | OS << " return MCTargetAsmParser::Match_InvalidOperand;\n" ; |
2464 | |
2465 | OS << " }\n" ; |
2466 | } |
2467 | OS << "}\n\n" ; |
2468 | } |
2469 | |
2470 | /// emitValidateOperandClass - Emit the function to validate an operand class. |
2471 | static void emitValidateOperandClass(const CodeGenTarget &Target, |
2472 | AsmMatcherInfo &Info, raw_ostream &OS) { |
2473 | OS << "static unsigned validateOperandClass(MCParsedAsmOperand &GOp, " |
2474 | << "MatchClassKind Kind) {\n" ; |
2475 | OS << " " << Info.Target.getName() << "Operand &Operand = (" |
2476 | << Info.Target.getName() << "Operand &)GOp;\n" ; |
2477 | |
2478 | // The InvalidMatchClass is not to match any operand. |
2479 | OS << " if (Kind == InvalidMatchClass)\n" ; |
2480 | OS << " return MCTargetAsmParser::Match_InvalidOperand;\n\n" ; |
2481 | |
2482 | // Check for Token operands first. |
2483 | // FIXME: Use a more specific diagnostic type. |
2484 | OS << " if (Operand.isToken() && Kind <= MCK_LAST_TOKEN)\n" ; |
2485 | OS << " return isSubclass(matchTokenString(Operand.getToken()), Kind) ?\n" |
2486 | << " MCTargetAsmParser::Match_Success :\n" |
2487 | << " MCTargetAsmParser::Match_InvalidOperand;\n\n" ; |
2488 | |
2489 | // Check the user classes. We don't care what order since we're only |
2490 | // actually matching against one of them. |
2491 | OS << " switch (Kind) {\n" |
2492 | " default: break;\n" ; |
2493 | for (const auto &CI : Info.Classes) { |
2494 | if (!CI.isUserClass()) |
2495 | continue; |
2496 | |
2497 | OS << " case " << CI.Name << ": {\n" ; |
2498 | OS << " DiagnosticPredicate DP(Operand." << CI.PredicateMethod |
2499 | << "());\n" ; |
2500 | OS << " if (DP.isMatch())\n" ; |
2501 | OS << " return MCTargetAsmParser::Match_Success;\n" ; |
2502 | if (!CI.DiagnosticType.empty()) { |
2503 | OS << " if (DP.isNearMatch())\n" ; |
2504 | OS << " return " << Info.Target.getName() << "AsmParser::Match_" |
2505 | << CI.DiagnosticType << ";\n" ; |
2506 | OS << " break;\n" ; |
2507 | } else { |
2508 | OS << " break;\n" ; |
2509 | } |
2510 | OS << " }\n" ; |
2511 | } |
2512 | OS << " } // end switch (Kind)\n\n" ; |
2513 | |
2514 | // Check for register operands, including sub-classes. |
2515 | const auto &Regs = Target.getRegBank().getRegisters(); |
2516 | StringRef Namespace = Regs.front().TheDef->getValueAsString(FieldName: "Namespace" ); |
2517 | SmallVector<StringRef> Table(1 + Regs.size(), "InvalidMatchClass" ); |
2518 | for (const auto &RC : Info.RegisterClasses) { |
2519 | const auto &Reg = Target.getRegBank().getReg(RC.first); |
2520 | Table[Reg->EnumValue] = RC.second->Name; |
2521 | } |
2522 | OS << " if (Operand.isReg()) {\n" ; |
2523 | OS << " static constexpr uint16_t Table[" << Namespace |
2524 | << "::NUM_TARGET_REGS] = {\n" ; |
2525 | for (auto &MatchClassName : Table) |
2526 | OS << " " << MatchClassName << ",\n" ; |
2527 | OS << " };\n\n" ; |
2528 | OS << " MCRegister Reg = Operand.getReg();\n" ; |
2529 | OS << " MatchClassKind OpKind = Reg.isPhysical() ? " |
2530 | "(MatchClassKind)Table[Reg.id()] : InvalidMatchClass;\n" ; |
2531 | OS << " return isSubclass(OpKind, Kind) ? " |
2532 | << "(unsigned)MCTargetAsmParser::Match_Success :\n " |
2533 | << " getDiagKindFromRegisterClass(Kind);\n }\n\n" ; |
2534 | |
2535 | // Expected operand is a register, but actual is not. |
2536 | OS << " if (Kind > MCK_LAST_TOKEN && Kind <= MCK_LAST_REGISTER)\n" ; |
2537 | OS << " return getDiagKindFromRegisterClass(Kind);\n\n" ; |
2538 | |
2539 | // Generic fallthrough match failure case for operands that don't have |
2540 | // specialized diagnostic types. |
2541 | OS << " return MCTargetAsmParser::Match_InvalidOperand;\n" ; |
2542 | OS << "}\n\n" ; |
2543 | } |
2544 | |
2545 | /// emitIsSubclass - Emit the subclass predicate function. |
2546 | static void emitIsSubclass(CodeGenTarget &Target, |
2547 | std::forward_list<ClassInfo> &Infos, |
2548 | raw_ostream &OS) { |
2549 | OS << "/// isSubclass - Compute whether \\p A is a subclass of \\p B.\n" ; |
2550 | OS << "static bool isSubclass(MatchClassKind A, MatchClassKind B) {\n" ; |
2551 | OS << " if (A == B)\n" ; |
2552 | OS << " return true;\n\n" ; |
2553 | |
2554 | // TODO: Use something like SequenceToOffsetTable to allow sequences to |
2555 | // overlap in this table. |
2556 | SmallVector<bool> SuperClassData; |
2557 | |
2558 | OS << " [[maybe_unused]] static constexpr struct {\n" ; |
2559 | OS << " uint32_t Offset;\n" ; |
2560 | OS << " uint16_t Start;\n" ; |
2561 | OS << " uint16_t Length;\n" ; |
2562 | OS << " } Table[] = {\n" ; |
2563 | OS << " {0, 0, 0},\n" ; // InvalidMatchClass |
2564 | OS << " {0, 0, 0},\n" ; // OptionalMatchClass |
2565 | for (const auto &A : Infos) { |
2566 | SmallVector<bool> SuperClasses; |
2567 | SuperClasses.push_back(Elt: false); // InvalidMatchClass |
2568 | SuperClasses.push_back(Elt: A.IsOptional); // OptionalMatchClass |
2569 | for (const auto &B : Infos) |
2570 | SuperClasses.push_back(Elt: &A != &B && A.isSubsetOf(RHS: B)); |
2571 | |
2572 | // Trim leading and trailing zeros. |
2573 | auto End = find_if(Range: reverse(C&: SuperClasses), P: [](bool B) { return B; }).base(); |
2574 | auto Start = |
2575 | std::find_if(first: SuperClasses.begin(), last: End, pred: [](bool B) { return B; }); |
2576 | |
2577 | unsigned Offset = SuperClassData.size(); |
2578 | SuperClassData.append(in_start: Start, in_end: End); |
2579 | |
2580 | OS << " {" << Offset << ", " << (Start - SuperClasses.begin()) << ", " |
2581 | << (End - Start) << "},\n" ; |
2582 | } |
2583 | OS << " };\n\n" ; |
2584 | |
2585 | if (SuperClassData.empty()) { |
2586 | OS << " return false;\n" ; |
2587 | } else { |
2588 | // Dump the boolean data packed into bytes. |
2589 | SuperClassData.append(NumInputs: -SuperClassData.size() % 8, Elt: false); |
2590 | OS << " static constexpr uint8_t Data[] = {\n" ; |
2591 | for (unsigned I = 0, E = SuperClassData.size(); I < E; I += 8) { |
2592 | unsigned Byte = 0; |
2593 | for (unsigned J = 0; J < 8; ++J) |
2594 | Byte |= (unsigned)SuperClassData[I + J] << J; |
2595 | OS << formatv(Fmt: " {:X2},\n" , Vals&: Byte); |
2596 | } |
2597 | OS << " };\n\n" ; |
2598 | |
2599 | OS << " auto &Entry = Table[A];\n" ; |
2600 | OS << " unsigned Idx = B - Entry.Start;\n" ; |
2601 | OS << " if (Idx >= Entry.Length)\n" ; |
2602 | OS << " return false;\n" ; |
2603 | OS << " Idx += Entry.Offset;\n" ; |
2604 | OS << " return (Data[Idx / 8] >> (Idx % 8)) & 1;\n" ; |
2605 | } |
2606 | OS << "}\n\n" ; |
2607 | } |
2608 | |
2609 | /// emitMatchTokenString - Emit the function to match a token string to the |
2610 | /// appropriate match class value. |
2611 | static void emitMatchTokenString(CodeGenTarget &Target, |
2612 | std::forward_list<ClassInfo> &Infos, |
2613 | raw_ostream &OS) { |
2614 | // Construct the match list. |
2615 | std::vector<StringMatcher::StringPair> Matches; |
2616 | for (const auto &CI : Infos) { |
2617 | if (CI.Kind == ClassInfo::Token) |
2618 | Matches.emplace_back(args: CI.ValueName, args: "return " + CI.Name + ";" ); |
2619 | } |
2620 | |
2621 | OS << "static MatchClassKind matchTokenString(StringRef Name) {\n" ; |
2622 | |
2623 | StringMatcher("Name" , Matches, OS).Emit(); |
2624 | |
2625 | OS << " return InvalidMatchClass;\n" ; |
2626 | OS << "}\n\n" ; |
2627 | } |
2628 | |
2629 | /// emitMatchRegisterName - Emit the function to match a string to the target |
2630 | /// specific register enum. |
2631 | static void emitMatchRegisterName(const CodeGenTarget &Target, |
2632 | const Record *AsmParser, raw_ostream &OS) { |
2633 | // Construct the match list. |
2634 | std::vector<StringMatcher::StringPair> Matches; |
2635 | const auto &Regs = Target.getRegBank().getRegisters(); |
2636 | std::string Namespace = |
2637 | Regs.front().TheDef->getValueAsString(FieldName: "Namespace" ).str(); |
2638 | for (const CodeGenRegister &Reg : Regs) { |
2639 | StringRef AsmName = Reg.TheDef->getValueAsString(FieldName: "AsmName" ); |
2640 | if (AsmName.empty()) |
2641 | continue; |
2642 | |
2643 | Matches.emplace_back(args: AsmName.str(), args: "return " + Namespace + |
2644 | "::" + Reg.getName().str() + ';'); |
2645 | } |
2646 | |
2647 | OS << "static MCRegister MatchRegisterName(StringRef Name) {\n" ; |
2648 | |
2649 | bool IgnoreDuplicates = |
2650 | AsmParser->getValueAsBit(FieldName: "AllowDuplicateRegisterNames" ); |
2651 | StringMatcher("Name" , Matches, OS).Emit(Indent: 0, IgnoreDuplicates); |
2652 | |
2653 | OS << " return " << Namespace << "::NoRegister;\n" ; |
2654 | OS << "}\n\n" ; |
2655 | } |
2656 | |
2657 | /// Emit the function to match a string to the target |
2658 | /// specific register enum. |
2659 | static void emitMatchRegisterAltName(const CodeGenTarget &Target, |
2660 | const Record *AsmParser, raw_ostream &OS) { |
2661 | // Construct the match list. |
2662 | std::vector<StringMatcher::StringPair> Matches; |
2663 | const auto &Regs = Target.getRegBank().getRegisters(); |
2664 | std::string Namespace = |
2665 | Regs.front().TheDef->getValueAsString(FieldName: "Namespace" ).str(); |
2666 | for (const CodeGenRegister &Reg : Regs) { |
2667 | |
2668 | auto AltNames = Reg.TheDef->getValueAsListOfStrings(FieldName: "AltNames" ); |
2669 | |
2670 | for (auto AltName : AltNames) { |
2671 | AltName = StringRef(AltName).trim(); |
2672 | |
2673 | // don't handle empty alternative names |
2674 | if (AltName.empty()) |
2675 | continue; |
2676 | |
2677 | Matches.emplace_back(args: AltName.str(), args: "return " + Namespace + |
2678 | "::" + Reg.getName().str() + ';'); |
2679 | } |
2680 | } |
2681 | |
2682 | OS << "static MCRegister MatchRegisterAltName(StringRef Name) {\n" ; |
2683 | |
2684 | bool IgnoreDuplicates = |
2685 | AsmParser->getValueAsBit(FieldName: "AllowDuplicateRegisterNames" ); |
2686 | StringMatcher("Name" , Matches, OS).Emit(Indent: 0, IgnoreDuplicates); |
2687 | |
2688 | OS << " return " << Namespace << "::NoRegister;\n" ; |
2689 | OS << "}\n\n" ; |
2690 | } |
2691 | |
2692 | /// emitOperandDiagnosticTypes - Emit the operand matching diagnostic types. |
2693 | static void emitOperandDiagnosticTypes(AsmMatcherInfo &Info, raw_ostream &OS) { |
2694 | // Get the set of diagnostic types from all of the operand classes. |
2695 | std::set<StringRef> Types; |
2696 | for (const auto &OpClassEntry : Info.AsmOperandClasses) { |
2697 | if (!OpClassEntry.second->DiagnosticType.empty()) |
2698 | Types.insert(x: OpClassEntry.second->DiagnosticType); |
2699 | } |
2700 | for (const auto &OpClassEntry : Info.RegisterClassClasses) { |
2701 | if (!OpClassEntry.second->DiagnosticType.empty()) |
2702 | Types.insert(x: OpClassEntry.second->DiagnosticType); |
2703 | } |
2704 | |
2705 | if (Types.empty()) |
2706 | return; |
2707 | |
2708 | // Now emit the enum entries. |
2709 | for (StringRef Type : Types) |
2710 | OS << " Match_" << Type << ",\n" ; |
2711 | OS << " END_OPERAND_DIAGNOSTIC_TYPES\n" ; |
2712 | } |
2713 | |
2714 | /// emitGetSubtargetFeatureName - Emit the helper function to get the |
2715 | /// user-level name for a subtarget feature. |
2716 | static void emitGetSubtargetFeatureName(AsmMatcherInfo &Info, raw_ostream &OS) { |
2717 | OS << "// User-level names for subtarget features that participate in\n" |
2718 | << "// instruction matching.\n" |
2719 | << "static const char *getSubtargetFeatureName(uint64_t Val) {\n" ; |
2720 | if (!Info.SubtargetFeatures.empty()) { |
2721 | OS << " switch(Val) {\n" ; |
2722 | for (const auto &SF : Info.SubtargetFeatures) { |
2723 | const SubtargetFeatureInfo &SFI = SF.second; |
2724 | // FIXME: Totally just a placeholder name to get the algorithm working. |
2725 | OS << " case " << SFI.getEnumBitName() << ": return \"" |
2726 | << SFI.TheDef->getValueAsString(FieldName: "PredicateName" ) << "\";\n" ; |
2727 | } |
2728 | OS << " default: return \"(unknown)\";\n" ; |
2729 | OS << " }\n" ; |
2730 | } else { |
2731 | // Nothing to emit, so skip the switch |
2732 | OS << " return \"(unknown)\";\n" ; |
2733 | } |
2734 | OS << "}\n\n" ; |
2735 | } |
2736 | |
2737 | static std::string GetAliasRequiredFeatures(const Record *R, |
2738 | const AsmMatcherInfo &Info) { |
2739 | std::string Result; |
2740 | |
2741 | bool First = true; |
2742 | for (const Record *RF : R->getValueAsListOfDefs(FieldName: "Predicates" )) { |
2743 | const SubtargetFeatureInfo *F = Info.getSubtargetFeature(Def: RF); |
2744 | if (!F) |
2745 | PrintFatalError(ErrorLoc: R->getLoc(), |
2746 | Msg: "Predicate '" + RF->getName() + |
2747 | "' is not marked as an AssemblerPredicate!" ); |
2748 | if (!First) |
2749 | Result += " && " ; |
2750 | Result += "Features.test(" + F->getEnumBitName() + ')'; |
2751 | First = false; |
2752 | } |
2753 | |
2754 | return Result; |
2755 | } |
2756 | |
2757 | static void |
2758 | emitMnemonicAliasVariant(raw_ostream &OS, const AsmMatcherInfo &Info, |
2759 | ArrayRef<const Record *> Aliases, unsigned Indent = 0, |
2760 | StringRef AsmParserVariantName = StringRef()) { |
2761 | // Keep track of all the aliases from a mnemonic. Use an std::map so that the |
2762 | // iteration order of the map is stable. |
2763 | std::map<std::string, std::vector<const Record *>> AliasesFromMnemonic; |
2764 | |
2765 | for (const Record *R : Aliases) { |
2766 | // FIXME: Allow AssemblerVariantName to be a comma separated list. |
2767 | StringRef AsmVariantName = R->getValueAsString(FieldName: "AsmVariantName" ); |
2768 | if (AsmVariantName != AsmParserVariantName) |
2769 | continue; |
2770 | AliasesFromMnemonic[R->getValueAsString(FieldName: "FromMnemonic" ).lower()].push_back( |
2771 | x: R); |
2772 | } |
2773 | if (AliasesFromMnemonic.empty()) |
2774 | return; |
2775 | |
2776 | // Process each alias a "from" mnemonic at a time, building the code executed |
2777 | // by the string remapper. |
2778 | std::vector<StringMatcher::StringPair> Cases; |
2779 | for (const auto &AliasEntry : AliasesFromMnemonic) { |
2780 | // Loop through each alias and emit code that handles each case. If there |
2781 | // are two instructions without predicates, emit an error. If there is one, |
2782 | // emit it last. |
2783 | std::string MatchCode; |
2784 | int AliasWithNoPredicate = -1; |
2785 | |
2786 | ArrayRef<const Record *> ToVec = AliasEntry.second; |
2787 | for (const auto &[Idx, R] : enumerate(First&: ToVec)) { |
2788 | std::string FeatureMask = GetAliasRequiredFeatures(R, Info); |
2789 | |
2790 | // If this unconditionally matches, remember it for later and diagnose |
2791 | // duplicates. |
2792 | if (FeatureMask.empty()) { |
2793 | if (AliasWithNoPredicate != -1 && |
2794 | R->getValueAsString(FieldName: "ToMnemonic" ) != |
2795 | ToVec[AliasWithNoPredicate]->getValueAsString(FieldName: "ToMnemonic" )) { |
2796 | // We can't have two different aliases from the same mnemonic with no |
2797 | // predicate. |
2798 | PrintError( |
2799 | ErrorLoc: ToVec[AliasWithNoPredicate]->getLoc(), |
2800 | Msg: "two different MnemonicAliases with the same 'from' mnemonic!" ); |
2801 | PrintFatalError(ErrorLoc: R->getLoc(), Msg: "this is the other MnemonicAlias." ); |
2802 | } |
2803 | |
2804 | AliasWithNoPredicate = Idx; |
2805 | continue; |
2806 | } |
2807 | if (R->getValueAsString(FieldName: "ToMnemonic" ) == AliasEntry.first) |
2808 | PrintFatalError(ErrorLoc: R->getLoc(), Msg: "MnemonicAlias to the same string" ); |
2809 | |
2810 | if (!MatchCode.empty()) |
2811 | MatchCode += "else " ; |
2812 | MatchCode += "if (" + FeatureMask + ")\n" ; |
2813 | MatchCode += " Mnemonic = \"" ; |
2814 | MatchCode += R->getValueAsString(FieldName: "ToMnemonic" ).lower(); |
2815 | MatchCode += "\";\n" ; |
2816 | } |
2817 | |
2818 | if (AliasWithNoPredicate != -1) { |
2819 | const Record *R = ToVec[AliasWithNoPredicate]; |
2820 | if (!MatchCode.empty()) |
2821 | MatchCode += "else\n " ; |
2822 | MatchCode += "Mnemonic = \"" ; |
2823 | MatchCode += R->getValueAsString(FieldName: "ToMnemonic" ).lower(); |
2824 | MatchCode += "\";\n" ; |
2825 | } |
2826 | |
2827 | MatchCode += "return;" ; |
2828 | |
2829 | Cases.emplace_back(args: AliasEntry.first, args&: MatchCode); |
2830 | } |
2831 | StringMatcher("Mnemonic" , Cases, OS).Emit(Indent); |
2832 | } |
2833 | |
2834 | /// emitMnemonicAliases - If the target has any MnemonicAlias<> definitions, |
2835 | /// emit a function for them and return true, otherwise return false. |
2836 | static bool emitMnemonicAliases(raw_ostream &OS, const AsmMatcherInfo &Info, |
2837 | CodeGenTarget &Target) { |
2838 | // Ignore aliases when match-prefix is set. |
2839 | if (!MatchPrefix.empty()) |
2840 | return false; |
2841 | |
2842 | ArrayRef<const Record *> Aliases = |
2843 | Info.getRecords().getAllDerivedDefinitions(ClassName: "MnemonicAlias" ); |
2844 | if (Aliases.empty()) |
2845 | return false; |
2846 | |
2847 | OS << "static void applyMnemonicAliases(StringRef &Mnemonic, " |
2848 | "const FeatureBitset &Features, unsigned VariantID) {\n" ; |
2849 | unsigned VariantCount = Target.getAsmParserVariantCount(); |
2850 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
2851 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
2852 | int AsmParserVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
2853 | StringRef AsmParserVariantName = AsmVariant->getValueAsString(FieldName: "Name" ); |
2854 | |
2855 | // If the variant doesn't have a name, defer to the emitMnemonicAliasVariant |
2856 | // call after the loop. |
2857 | if (AsmParserVariantName.empty()) { |
2858 | assert(VariantCount == 1 && "Multiple variants should each be named" ); |
2859 | continue; |
2860 | } |
2861 | |
2862 | if (VC == 0) |
2863 | OS << " switch (VariantID) {\n" ; |
2864 | OS << " case " << AsmParserVariantNo << ":\n" ; |
2865 | emitMnemonicAliasVariant(OS, Info, Aliases, /*Indent=*/2, |
2866 | AsmParserVariantName); |
2867 | OS << " break;\n" ; |
2868 | |
2869 | if (VC == VariantCount - 1) |
2870 | OS << " }\n" ; |
2871 | } |
2872 | |
2873 | // Emit aliases that apply to all variants. |
2874 | emitMnemonicAliasVariant(OS, Info, Aliases); |
2875 | |
2876 | OS << "}\n\n" ; |
2877 | |
2878 | return true; |
2879 | } |
2880 | |
2881 | static void |
2882 | emitCustomOperandParsing(raw_ostream &OS, CodeGenTarget &Target, |
2883 | const AsmMatcherInfo &Info, StringRef ClassName, |
2884 | const StringToOffsetTable &StringTable, |
2885 | unsigned MaxMnemonicIndex, unsigned MaxFeaturesIndex, |
2886 | bool HasMnemonicFirst, const Record &AsmParser) { |
2887 | unsigned MaxMask = 0; |
2888 | for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { |
2889 | MaxMask |= OMI.OperandMask; |
2890 | } |
2891 | |
2892 | // Emit the static custom operand parsing table; |
2893 | OS << "namespace {\n" ; |
2894 | OS << " struct OperandMatchEntry {\n" ; |
2895 | OS << " " << getMinimalTypeForRange(Range: MaxMnemonicIndex) << " Mnemonic;\n" ; |
2896 | OS << " " << getMinimalTypeForRange(Range: MaxMask) << " OperandMask;\n" ; |
2897 | OS << " " |
2898 | << getMinimalTypeForRange( |
2899 | Range: std::distance(first: Info.Classes.begin(), last: Info.Classes.end()) + |
2900 | 2 /* Include 'InvalidMatchClass' and 'OptionalMatchClass' */) |
2901 | << " Class;\n" ; |
2902 | OS << " " << getMinimalTypeForRange(Range: MaxFeaturesIndex) |
2903 | << " RequiredFeaturesIdx;\n\n" ; |
2904 | OS << " StringRef getMnemonic() const {\n" ; |
2905 | OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n" ; |
2906 | OS << " MnemonicTable[Mnemonic]);\n" ; |
2907 | OS << " }\n" ; |
2908 | OS << " };\n\n" ; |
2909 | |
2910 | OS << " // Predicate for searching for an opcode.\n" ; |
2911 | OS << " struct LessOpcodeOperand {\n" ; |
2912 | OS << " bool operator()(const OperandMatchEntry &LHS, StringRef RHS) {\n" ; |
2913 | OS << " return LHS.getMnemonic() < RHS;\n" ; |
2914 | OS << " }\n" ; |
2915 | OS << " bool operator()(StringRef LHS, const OperandMatchEntry &RHS) {\n" ; |
2916 | OS << " return LHS < RHS.getMnemonic();\n" ; |
2917 | OS << " }\n" ; |
2918 | OS << " bool operator()(const OperandMatchEntry &LHS," ; |
2919 | OS << " const OperandMatchEntry &RHS) {\n" ; |
2920 | OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n" ; |
2921 | OS << " }\n" ; |
2922 | OS << " };\n" ; |
2923 | |
2924 | OS << "} // end anonymous namespace\n\n" ; |
2925 | |
2926 | OS << "static const OperandMatchEntry OperandMatchTable[" |
2927 | << Info.OperandMatchInfo.size() << "] = {\n" ; |
2928 | |
2929 | OS << " /* Operand List Mnemonic, Mask, Operand Class, Features */\n" ; |
2930 | for (const OperandMatchEntry &OMI : Info.OperandMatchInfo) { |
2931 | const MatchableInfo &II = *OMI.MI; |
2932 | |
2933 | OS << " { " ; |
2934 | |
2935 | // Store a pascal-style length byte in the mnemonic. |
2936 | std::string LenMnemonic = char(II.Mnemonic.size()) + II.Mnemonic.lower(); |
2937 | OS << *StringTable.GetStringOffset(Str: LenMnemonic) << " /* " << II.Mnemonic |
2938 | << " */, " ; |
2939 | |
2940 | OS << OMI.OperandMask; |
2941 | OS << " /* " ; |
2942 | ListSeparator LS; |
2943 | for (int i = 0, e = 31; i != e; ++i) |
2944 | if (OMI.OperandMask & (1 << i)) |
2945 | OS << LS << i; |
2946 | OS << " */, " ; |
2947 | |
2948 | OS << OMI.CI->Name; |
2949 | |
2950 | // Write the required features mask. |
2951 | OS << ", AMFBS" ; |
2952 | if (II.RequiredFeatures.empty()) |
2953 | OS << "_None" ; |
2954 | else |
2955 | for (const auto &F : II.RequiredFeatures) |
2956 | OS << '_' << F->TheDef->getName(); |
2957 | |
2958 | OS << " },\n" ; |
2959 | } |
2960 | OS << "};\n\n" ; |
2961 | |
2962 | // Emit the operand class switch to call the correct custom parser for |
2963 | // the found operand class. |
2964 | OS << "ParseStatus " << Target.getName() << ClassName << "::\n" |
2965 | << "tryCustomParseOperand(OperandVector" |
2966 | << " &Operands,\n unsigned MCK) {\n\n" |
2967 | << " switch(MCK) {\n" ; |
2968 | |
2969 | for (const auto &CI : Info.Classes) { |
2970 | if (CI.ParserMethod.empty()) |
2971 | continue; |
2972 | OS << " case " << CI.Name << ":\n" |
2973 | << " return " << CI.ParserMethod << "(Operands);\n" ; |
2974 | } |
2975 | |
2976 | OS << " default:\n" ; |
2977 | OS << " return ParseStatus::NoMatch;\n" ; |
2978 | OS << " }\n" ; |
2979 | OS << " return ParseStatus::NoMatch;\n" ; |
2980 | OS << "}\n\n" ; |
2981 | |
2982 | // Emit the static custom operand parser. This code is very similar with |
2983 | // the other matcher. Also use MatchResultTy here just in case we go for |
2984 | // a better error handling. |
2985 | OS << "ParseStatus " << Target.getName() << ClassName << "::\n" |
2986 | << "MatchOperandParserImpl(OperandVector" |
2987 | << " &Operands,\n StringRef Mnemonic,\n" |
2988 | << " bool ParseForAllFeatures) {\n" ; |
2989 | |
2990 | // Emit code to get the available features. |
2991 | OS << " // Get the current feature set.\n" ; |
2992 | OS << " const FeatureBitset &AvailableFeatures = " |
2993 | "getAvailableFeatures();\n\n" ; |
2994 | |
2995 | OS << " // Get the next operand index.\n" ; |
2996 | OS << " unsigned NextOpNum = Operands.size()" |
2997 | << (HasMnemonicFirst ? " - 1" : "" ) << ";\n" ; |
2998 | |
2999 | // Emit code to search the table. |
3000 | OS << " // Search the table.\n" ; |
3001 | if (HasMnemonicFirst) { |
3002 | OS << " auto MnemonicRange =\n" ; |
3003 | OS << " std::equal_range(std::begin(OperandMatchTable), " |
3004 | "std::end(OperandMatchTable),\n" ; |
3005 | OS << " Mnemonic, LessOpcodeOperand());\n\n" ; |
3006 | } else { |
3007 | OS << " auto MnemonicRange = std::pair(std::begin(OperandMatchTable)," |
3008 | " std::end(OperandMatchTable));\n" ; |
3009 | OS << " if (!Mnemonic.empty())\n" ; |
3010 | OS << " MnemonicRange =\n" ; |
3011 | OS << " std::equal_range(std::begin(OperandMatchTable), " |
3012 | "std::end(OperandMatchTable),\n" ; |
3013 | OS << " Mnemonic, LessOpcodeOperand());\n\n" ; |
3014 | } |
3015 | |
3016 | OS << " if (MnemonicRange.first == MnemonicRange.second)\n" ; |
3017 | OS << " return ParseStatus::NoMatch;\n\n" ; |
3018 | |
3019 | OS << " for (const OperandMatchEntry *it = MnemonicRange.first,\n" |
3020 | << " *ie = MnemonicRange.second; it != ie; ++it) {\n" ; |
3021 | |
3022 | OS << " // equal_range guarantees that instruction mnemonic matches.\n" ; |
3023 | OS << " assert(Mnemonic == it->getMnemonic());\n\n" ; |
3024 | |
3025 | // Emit check that the required features are available. |
3026 | OS << " // check if the available features match\n" ; |
3027 | OS << " const FeatureBitset &RequiredFeatures = " |
3028 | "FeatureBitsets[it->RequiredFeaturesIdx];\n" ; |
3029 | OS << " if (!ParseForAllFeatures && (AvailableFeatures & " |
3030 | "RequiredFeatures) != RequiredFeatures)\n" ; |
3031 | OS << " continue;\n\n" ; |
3032 | |
3033 | // Emit check to ensure the operand number matches. |
3034 | OS << " // check if the operand in question has a custom parser.\n" ; |
3035 | OS << " if (!(it->OperandMask & (1 << NextOpNum)))\n" ; |
3036 | OS << " continue;\n\n" ; |
3037 | |
3038 | // Emit call to the custom parser method |
3039 | StringRef ParserName = AsmParser.getValueAsString(FieldName: "OperandParserMethod" ); |
3040 | if (ParserName.empty()) |
3041 | ParserName = "tryCustomParseOperand" ; |
3042 | OS << " // call custom parse method to handle the operand\n" ; |
3043 | OS << " ParseStatus Result = " << ParserName << "(Operands, it->Class);\n" ; |
3044 | OS << " if (!Result.isNoMatch())\n" ; |
3045 | OS << " return Result;\n" ; |
3046 | OS << " }\n\n" ; |
3047 | |
3048 | OS << " // Okay, we had no match.\n" ; |
3049 | OS << " return ParseStatus::NoMatch;\n" ; |
3050 | OS << "}\n\n" ; |
3051 | } |
3052 | |
3053 | static void emitAsmTiedOperandConstraints(CodeGenTarget &Target, |
3054 | AsmMatcherInfo &Info, raw_ostream &OS, |
3055 | bool HasOptionalOperands) { |
3056 | std::string AsmParserName = |
3057 | Info.AsmParser->getValueAsString(FieldName: "AsmParserClassName" ).str(); |
3058 | OS << "static bool " ; |
3059 | OS << "checkAsmTiedOperandConstraints(const " << Target.getName() |
3060 | << AsmParserName << "&AsmParser,\n" ; |
3061 | OS << " unsigned Kind, const OperandVector " |
3062 | "&Operands,\n" ; |
3063 | if (HasOptionalOperands) |
3064 | OS << " ArrayRef<unsigned> DefaultsOffset,\n" ; |
3065 | OS << " uint64_t &ErrorInfo) {\n" ; |
3066 | OS << " assert(Kind < CVT_NUM_SIGNATURES && \"Invalid signature!\");\n" ; |
3067 | OS << " const uint8_t *Converter = ConversionTable[Kind];\n" ; |
3068 | OS << " for (const uint8_t *p = Converter; *p; p += 2) {\n" ; |
3069 | OS << " switch (*p) {\n" ; |
3070 | OS << " case CVT_Tied: {\n" ; |
3071 | OS << " unsigned OpIdx = *(p + 1);\n" ; |
3072 | OS << " assert(OpIdx < (size_t)(std::end(TiedAsmOperandTable) -\n" ; |
3073 | OS << " std::begin(TiedAsmOperandTable)) &&\n" ; |
3074 | OS << " \"Tied operand not found\");\n" ; |
3075 | OS << " unsigned OpndNum1 = TiedAsmOperandTable[OpIdx][1];\n" ; |
3076 | OS << " unsigned OpndNum2 = TiedAsmOperandTable[OpIdx][2];\n" ; |
3077 | if (HasOptionalOperands) { |
3078 | // When optional operands are involved, formal and actual operand indices |
3079 | // may differ. Map the former to the latter by subtracting the number of |
3080 | // absent optional operands. |
3081 | OS << " OpndNum1 = OpndNum1 - DefaultsOffset[OpndNum1];\n" ; |
3082 | OS << " OpndNum2 = OpndNum2 - DefaultsOffset[OpndNum2];\n" ; |
3083 | } |
3084 | OS << " if (OpndNum1 != OpndNum2) {\n" ; |
3085 | OS << " auto &SrcOp1 = Operands[OpndNum1];\n" ; |
3086 | OS << " auto &SrcOp2 = Operands[OpndNum2];\n" ; |
3087 | OS << " if (!AsmParser.areEqualRegs(*SrcOp1, *SrcOp2)) {\n" ; |
3088 | OS << " ErrorInfo = OpndNum2;\n" ; |
3089 | OS << " return false;\n" ; |
3090 | OS << " }\n" ; |
3091 | OS << " }\n" ; |
3092 | OS << " break;\n" ; |
3093 | OS << " }\n" ; |
3094 | OS << " default:\n" ; |
3095 | OS << " break;\n" ; |
3096 | OS << " }\n" ; |
3097 | OS << " }\n" ; |
3098 | OS << " return true;\n" ; |
3099 | OS << "}\n\n" ; |
3100 | } |
3101 | |
3102 | static void emitMnemonicSpellChecker(raw_ostream &OS, CodeGenTarget &Target, |
3103 | unsigned VariantCount) { |
3104 | OS << "static std::string " << Target.getName() |
3105 | << "MnemonicSpellCheck(StringRef S, const FeatureBitset &FBS," |
3106 | << " unsigned VariantID) {\n" ; |
3107 | if (!VariantCount) |
3108 | OS << " return \"\";" ; |
3109 | else { |
3110 | OS << " const unsigned MaxEditDist = 2;\n" ; |
3111 | OS << " std::vector<StringRef> Candidates;\n" ; |
3112 | OS << " StringRef Prev = \"\";\n\n" ; |
3113 | |
3114 | OS << " // Find the appropriate table for this asm variant.\n" ; |
3115 | OS << " const MatchEntry *Start, *End;\n" ; |
3116 | OS << " switch (VariantID) {\n" ; |
3117 | OS << " default: llvm_unreachable(\"invalid variant!\");\n" ; |
3118 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
3119 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
3120 | int AsmVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
3121 | OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC |
3122 | << "); End = std::end(MatchTable" << VC << "); break;\n" ; |
3123 | } |
3124 | OS << " }\n\n" ; |
3125 | OS << " for (auto I = Start; I < End; I++) {\n" ; |
3126 | OS << " // Ignore unsupported instructions.\n" ; |
3127 | OS << " const FeatureBitset &RequiredFeatures = " |
3128 | "FeatureBitsets[I->RequiredFeaturesIdx];\n" ; |
3129 | OS << " if ((FBS & RequiredFeatures) != RequiredFeatures)\n" ; |
3130 | OS << " continue;\n" ; |
3131 | OS << "\n" ; |
3132 | OS << " StringRef T = I->getMnemonic();\n" ; |
3133 | OS << " // Avoid recomputing the edit distance for the same string.\n" ; |
3134 | OS << " if (T == Prev)\n" ; |
3135 | OS << " continue;\n" ; |
3136 | OS << "\n" ; |
3137 | OS << " Prev = T;\n" ; |
3138 | OS << " unsigned Dist = S.edit_distance(T, false, MaxEditDist);\n" ; |
3139 | OS << " if (Dist <= MaxEditDist)\n" ; |
3140 | OS << " Candidates.push_back(T);\n" ; |
3141 | OS << " }\n" ; |
3142 | OS << "\n" ; |
3143 | OS << " if (Candidates.empty())\n" ; |
3144 | OS << " return \"\";\n" ; |
3145 | OS << "\n" ; |
3146 | OS << " std::string Res = \", did you mean: \";\n" ; |
3147 | OS << " unsigned i = 0;\n" ; |
3148 | OS << " for (; i < Candidates.size() - 1; i++)\n" ; |
3149 | OS << " Res += Candidates[i].str() + \", \";\n" ; |
3150 | OS << " return Res + Candidates[i].str() + \"?\";\n" ; |
3151 | } |
3152 | OS << "}\n" ; |
3153 | OS << "\n" ; |
3154 | } |
3155 | |
3156 | static void emitMnemonicChecker(raw_ostream &OS, CodeGenTarget &Target, |
3157 | unsigned VariantCount, bool HasMnemonicFirst, |
3158 | bool HasMnemonicAliases) { |
3159 | OS << "static bool " << Target.getName() |
3160 | << "CheckMnemonic(StringRef Mnemonic,\n" ; |
3161 | OS << " " |
3162 | << "const FeatureBitset &AvailableFeatures,\n" ; |
3163 | OS << " " |
3164 | << "unsigned VariantID) {\n" ; |
3165 | |
3166 | if (!VariantCount) { |
3167 | OS << " return false;\n" ; |
3168 | } else { |
3169 | if (HasMnemonicAliases) { |
3170 | OS << " // Process all MnemonicAliases to remap the mnemonic.\n" ; |
3171 | OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);" ; |
3172 | OS << "\n\n" ; |
3173 | } |
3174 | OS << " // Find the appropriate table for this asm variant.\n" ; |
3175 | OS << " const MatchEntry *Start, *End;\n" ; |
3176 | OS << " switch (VariantID) {\n" ; |
3177 | OS << " default: llvm_unreachable(\"invalid variant!\");\n" ; |
3178 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
3179 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
3180 | int AsmVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
3181 | OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC |
3182 | << "); End = std::end(MatchTable" << VC << "); break;\n" ; |
3183 | } |
3184 | OS << " }\n\n" ; |
3185 | |
3186 | OS << " // Search the table.\n" ; |
3187 | if (HasMnemonicFirst) { |
3188 | OS << " auto MnemonicRange = " |
3189 | "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n" ; |
3190 | } else { |
3191 | OS << " auto MnemonicRange = std::pair(Start, End);\n" ; |
3192 | OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n" ; |
3193 | OS << " if (!Mnemonic.empty())\n" ; |
3194 | OS << " MnemonicRange = " |
3195 | << "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n" ; |
3196 | } |
3197 | |
3198 | OS << " if (MnemonicRange.first == MnemonicRange.second)\n" ; |
3199 | OS << " return false;\n\n" ; |
3200 | |
3201 | OS << " for (const MatchEntry *it = MnemonicRange.first, " |
3202 | << "*ie = MnemonicRange.second;\n" ; |
3203 | OS << " it != ie; ++it) {\n" ; |
3204 | OS << " const FeatureBitset &RequiredFeatures =\n" ; |
3205 | OS << " FeatureBitsets[it->RequiredFeaturesIdx];\n" ; |
3206 | OS << " if ((AvailableFeatures & RequiredFeatures) == " ; |
3207 | OS << "RequiredFeatures)\n" ; |
3208 | OS << " return true;\n" ; |
3209 | OS << " }\n" ; |
3210 | OS << " return false;\n" ; |
3211 | } |
3212 | OS << "}\n" ; |
3213 | OS << "\n" ; |
3214 | } |
3215 | |
3216 | // Emit a function mapping match classes to strings, for debugging. |
3217 | static void emitMatchClassKindNames(std::forward_list<ClassInfo> &Infos, |
3218 | raw_ostream &OS) { |
3219 | OS << "#ifndef NDEBUG\n" ; |
3220 | OS << "const char *getMatchClassName(MatchClassKind Kind) {\n" ; |
3221 | OS << " switch (Kind) {\n" ; |
3222 | |
3223 | OS << " case InvalidMatchClass: return \"InvalidMatchClass\";\n" ; |
3224 | OS << " case OptionalMatchClass: return \"OptionalMatchClass\";\n" ; |
3225 | for (const auto &CI : Infos) { |
3226 | OS << " case " << CI.Name << ": return \"" << CI.Name << "\";\n" ; |
3227 | } |
3228 | OS << " case NumMatchClassKinds: return \"NumMatchClassKinds\";\n" ; |
3229 | |
3230 | OS << " }\n" ; |
3231 | OS << " llvm_unreachable(\"unhandled MatchClassKind!\");\n" ; |
3232 | OS << "}\n\n" ; |
3233 | OS << "#endif // NDEBUG\n" ; |
3234 | } |
3235 | |
3236 | static std::string |
3237 | getNameForFeatureBitset(ArrayRef<const Record *> FeatureBitset) { |
3238 | std::string Name = "AMFBS" ; |
3239 | for (const Record *Feature : FeatureBitset) |
3240 | Name += ("_" + Feature->getName()).str(); |
3241 | return Name; |
3242 | } |
3243 | |
3244 | void AsmMatcherEmitter::run(raw_ostream &OS) { |
3245 | CodeGenTarget Target(Records); |
3246 | const Record *AsmParser = Target.getAsmParser(); |
3247 | StringRef ClassName = AsmParser->getValueAsString(FieldName: "AsmParserClassName" ); |
3248 | |
3249 | emitSourceFileHeader(Desc: "Assembly Matcher Source Fragment" , OS, Record: Records); |
3250 | |
3251 | // Compute the information on the instructions to match. |
3252 | AsmMatcherInfo Info(AsmParser, Target, Records); |
3253 | Info.buildInfo(); |
3254 | |
3255 | bool PreferSmallerInstructions = getPreferSmallerInstructions(Target); |
3256 | // Sort the instruction table using the partial order on classes. We use |
3257 | // stable_sort to ensure that ambiguous instructions are still |
3258 | // deterministically ordered. |
3259 | llvm::stable_sort( |
3260 | Range&: Info.Matchables, |
3261 | C: [PreferSmallerInstructions](const std::unique_ptr<MatchableInfo> &A, |
3262 | const std::unique_ptr<MatchableInfo> &B) { |
3263 | return A->shouldBeMatchedBefore(RHS: *B, PreferSmallerInstructions); |
3264 | }); |
3265 | |
3266 | #ifdef EXPENSIVE_CHECKS |
3267 | // Verify that the table is sorted and operator < works transitively. |
3268 | for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; |
3269 | ++I) { |
3270 | for (auto J = I; J != E; ++J) { |
3271 | assert(!(*J)->shouldBeMatchedBefore(**I, PreferSmallerInstructions)); |
3272 | } |
3273 | } |
3274 | #endif |
3275 | |
3276 | DEBUG_WITH_TYPE("instruction_info" , { |
3277 | for (const auto &MI : Info.Matchables) |
3278 | MI->dump(); |
3279 | }); |
3280 | |
3281 | // Check for ambiguous matchables. |
3282 | DEBUG_WITH_TYPE("ambiguous_instrs" , { |
3283 | unsigned NumAmbiguous = 0; |
3284 | for (auto I = Info.Matchables.begin(), E = Info.Matchables.end(); I != E; |
3285 | ++I) { |
3286 | for (auto J = std::next(I); J != E; ++J) { |
3287 | const MatchableInfo &A = **I; |
3288 | const MatchableInfo &B = **J; |
3289 | |
3290 | if (A.couldMatchAmbiguouslyWith(B, PreferSmallerInstructions)) { |
3291 | errs() << "warning: ambiguous matchables:\n" ; |
3292 | A.dump(); |
3293 | errs() << "\nis incomparable with:\n" ; |
3294 | B.dump(); |
3295 | errs() << "\n\n" ; |
3296 | ++NumAmbiguous; |
3297 | } |
3298 | } |
3299 | } |
3300 | if (NumAmbiguous) |
3301 | errs() << "warning: " << NumAmbiguous << " ambiguous matchables!\n" ; |
3302 | }); |
3303 | |
3304 | // Compute the information on the custom operand parsing. |
3305 | Info.buildOperandMatchInfo(); |
3306 | |
3307 | bool HasMnemonicFirst = AsmParser->getValueAsBit(FieldName: "HasMnemonicFirst" ); |
3308 | bool HasOptionalOperands = Info.hasOptionalOperands(); |
3309 | bool ReportMultipleNearMisses = |
3310 | AsmParser->getValueAsBit(FieldName: "ReportMultipleNearMisses" ); |
3311 | |
3312 | // Write the output. |
3313 | |
3314 | // Information for the class declaration. |
3315 | OS << "\n#ifdef GET_ASSEMBLER_HEADER\n" ; |
3316 | OS << "#undef GET_ASSEMBLER_HEADER\n" ; |
3317 | OS << " // This should be included into the middle of the declaration of\n" ; |
3318 | OS << " // your subclasses implementation of MCTargetAsmParser.\n" ; |
3319 | OS << " FeatureBitset ComputeAvailableFeatures(const FeatureBitset &FB) " |
3320 | "const;\n" ; |
3321 | if (HasOptionalOperands) { |
3322 | OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " |
3323 | << "unsigned Opcode,\n" |
3324 | << " const OperandVector &Operands,\n" |
3325 | << " const SmallBitVector " |
3326 | "&OptionalOperandsMask,\n" |
3327 | << " ArrayRef<unsigned> DefaultsOffset);\n" ; |
3328 | } else { |
3329 | OS << " void convertToMCInst(unsigned Kind, MCInst &Inst, " |
3330 | << "unsigned Opcode,\n" |
3331 | << " const OperandVector &Operands);\n" ; |
3332 | } |
3333 | OS << " void convertToMapAndConstraints(unsigned Kind,\n " ; |
3334 | OS << " const OperandVector &Operands) override;\n" ; |
3335 | OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n" |
3336 | << " MCInst &Inst,\n" ; |
3337 | if (ReportMultipleNearMisses) |
3338 | OS << " SmallVectorImpl<NearMissInfo> " |
3339 | "*NearMisses,\n" ; |
3340 | else |
3341 | OS << " uint64_t &ErrorInfo,\n" |
3342 | << " FeatureBitset &MissingFeatures,\n" ; |
3343 | OS << " bool matchingInlineAsm,\n" |
3344 | << " unsigned VariantID = 0);\n" ; |
3345 | if (!ReportMultipleNearMisses) |
3346 | OS << " unsigned MatchInstructionImpl(const OperandVector &Operands,\n" |
3347 | << " MCInst &Inst,\n" |
3348 | << " uint64_t &ErrorInfo,\n" |
3349 | << " bool matchingInlineAsm,\n" |
3350 | << " unsigned VariantID = 0) {\n" |
3351 | << " FeatureBitset MissingFeatures;\n" |
3352 | << " return MatchInstructionImpl(Operands, Inst, ErrorInfo, " |
3353 | "MissingFeatures,\n" |
3354 | << " matchingInlineAsm, VariantID);\n" |
3355 | << " }\n\n" ; |
3356 | |
3357 | if (!Info.OperandMatchInfo.empty()) { |
3358 | OS << " ParseStatus MatchOperandParserImpl(\n" ; |
3359 | OS << " OperandVector &Operands,\n" ; |
3360 | OS << " StringRef Mnemonic,\n" ; |
3361 | OS << " bool ParseForAllFeatures = false);\n" ; |
3362 | |
3363 | OS << " ParseStatus tryCustomParseOperand(\n" ; |
3364 | OS << " OperandVector &Operands,\n" ; |
3365 | OS << " unsigned MCK);\n\n" ; |
3366 | } |
3367 | |
3368 | OS << "#endif // GET_ASSEMBLER_HEADER\n\n" ; |
3369 | |
3370 | // Emit the operand match diagnostic enum names. |
3371 | OS << "\n#ifdef GET_OPERAND_DIAGNOSTIC_TYPES\n" ; |
3372 | OS << "#undef GET_OPERAND_DIAGNOSTIC_TYPES\n\n" ; |
3373 | emitOperandDiagnosticTypes(Info, OS); |
3374 | OS << "#endif // GET_OPERAND_DIAGNOSTIC_TYPES\n\n" ; |
3375 | |
3376 | OS << "\n#ifdef GET_REGISTER_MATCHER\n" ; |
3377 | OS << "#undef GET_REGISTER_MATCHER\n\n" ; |
3378 | |
3379 | // Emit the subtarget feature enumeration. |
3380 | SubtargetFeatureInfo::emitSubtargetFeatureBitEnumeration( |
3381 | SubtargetFeatures: Info.SubtargetFeatures, OS); |
3382 | |
3383 | // Emit the function to match a register name to number. |
3384 | // This should be omitted for Mips target |
3385 | if (AsmParser->getValueAsBit(FieldName: "ShouldEmitMatchRegisterName" )) |
3386 | emitMatchRegisterName(Target, AsmParser, OS); |
3387 | |
3388 | if (AsmParser->getValueAsBit(FieldName: "ShouldEmitMatchRegisterAltName" )) |
3389 | emitMatchRegisterAltName(Target, AsmParser, OS); |
3390 | |
3391 | OS << "#endif // GET_REGISTER_MATCHER\n\n" ; |
3392 | |
3393 | OS << "\n#ifdef GET_SUBTARGET_FEATURE_NAME\n" ; |
3394 | OS << "#undef GET_SUBTARGET_FEATURE_NAME\n\n" ; |
3395 | |
3396 | // Generate the helper function to get the names for subtarget features. |
3397 | emitGetSubtargetFeatureName(Info, OS); |
3398 | |
3399 | OS << "#endif // GET_SUBTARGET_FEATURE_NAME\n\n" ; |
3400 | |
3401 | OS << "\n#ifdef GET_MATCHER_IMPLEMENTATION\n" ; |
3402 | OS << "#undef GET_MATCHER_IMPLEMENTATION\n\n" ; |
3403 | |
3404 | // Generate the function that remaps for mnemonic aliases. |
3405 | bool HasMnemonicAliases = emitMnemonicAliases(OS, Info, Target); |
3406 | |
3407 | // Generate the convertToMCInst function to convert operands into an MCInst. |
3408 | // Also, generate the convertToMapAndConstraints function for MS-style inline |
3409 | // assembly. The latter doesn't actually generate a MCInst. |
3410 | unsigned NumConverters = |
3411 | emitConvertFuncs(Target, ClassName, Infos&: Info.Matchables, HasMnemonicFirst, |
3412 | HasOptionalOperands, OS); |
3413 | |
3414 | // Emit the enumeration for classes which participate in matching. |
3415 | emitMatchClassEnumeration(Target, Infos&: Info.Classes, OS); |
3416 | |
3417 | // Emit a function to get the user-visible string to describe an operand |
3418 | // match failure in diagnostics. |
3419 | emitOperandMatchErrorDiagStrings(Info, OS); |
3420 | |
3421 | // Emit a function to map register classes to operand match failure codes. |
3422 | emitRegisterMatchErrorFunc(Info, OS); |
3423 | |
3424 | // Emit the routine to match token strings to their match class. |
3425 | emitMatchTokenString(Target, Infos&: Info.Classes, OS); |
3426 | |
3427 | // Emit the subclass predicate routine. |
3428 | emitIsSubclass(Target, Infos&: Info.Classes, OS); |
3429 | |
3430 | // Emit the routine to validate an operand against a match class. |
3431 | emitValidateOperandClass(Target, Info, OS); |
3432 | |
3433 | emitMatchClassKindNames(Infos&: Info.Classes, OS); |
3434 | |
3435 | // Emit the available features compute function. |
3436 | SubtargetFeatureInfo::emitComputeAssemblerAvailableFeatures( |
3437 | TargetName: Info.Target.getName(), ClassName, FuncName: "ComputeAvailableFeatures" , |
3438 | SubtargetFeatures&: Info.SubtargetFeatures, OS); |
3439 | |
3440 | if (!ReportMultipleNearMisses) |
3441 | emitAsmTiedOperandConstraints(Target, Info, OS, HasOptionalOperands); |
3442 | |
3443 | StringToOffsetTable StringTable; |
3444 | |
3445 | size_t MaxNumOperands = 0; |
3446 | unsigned MaxMnemonicIndex = 0; |
3447 | bool HasDeprecation = false; |
3448 | for (const auto &MI : Info.Matchables) { |
3449 | MaxNumOperands = std::max(a: MaxNumOperands, b: MI->AsmOperands.size()); |
3450 | HasDeprecation |= MI->HasDeprecation; |
3451 | |
3452 | // Store a pascal-style length byte in the mnemonic. |
3453 | std::string LenMnemonic = char(MI->Mnemonic.size()) + MI->Mnemonic.lower(); |
3454 | MaxMnemonicIndex = std::max( |
3455 | a: MaxMnemonicIndex, b: StringTable.GetOrAddStringOffset(Str: LenMnemonic, appendZero: false)); |
3456 | } |
3457 | |
3458 | OS << "static const char MnemonicTable[] =\n" ; |
3459 | StringTable.EmitString(O&: OS); |
3460 | OS << ";\n\n" ; |
3461 | |
3462 | std::vector<std::vector<const Record *>> FeatureBitsets; |
3463 | for (const auto &MI : Info.Matchables) { |
3464 | if (MI->RequiredFeatures.empty()) |
3465 | continue; |
3466 | FeatureBitsets.emplace_back(); |
3467 | for (const auto *F : MI->RequiredFeatures) |
3468 | FeatureBitsets.back().push_back(x: F->TheDef); |
3469 | } |
3470 | |
3471 | llvm::sort(C&: FeatureBitsets, |
3472 | Comp: [&](ArrayRef<const Record *> A, ArrayRef<const Record *> B) { |
3473 | if (A.size() != B.size()) |
3474 | return A.size() < B.size(); |
3475 | for (const auto [ARec, BRec] : zip_equal(t&: A, u&: B)) { |
3476 | if (ARec->getName() != BRec->getName()) |
3477 | return ARec->getName() < BRec->getName(); |
3478 | } |
3479 | return false; |
3480 | }); |
3481 | FeatureBitsets.erase(first: llvm::unique(R&: FeatureBitsets), last: FeatureBitsets.end()); |
3482 | OS << "// Feature bitsets.\n" |
3483 | << "enum : " << getMinimalTypeForRange(Range: FeatureBitsets.size()) << " {\n" |
3484 | << " AMFBS_None,\n" ; |
3485 | for (const auto &FeatureBitset : FeatureBitsets) { |
3486 | if (FeatureBitset.empty()) |
3487 | continue; |
3488 | OS << " " << getNameForFeatureBitset(FeatureBitset) << ",\n" ; |
3489 | } |
3490 | OS << "};\n\n" |
3491 | << "static constexpr FeatureBitset FeatureBitsets[] = {\n" |
3492 | << " {}, // AMFBS_None\n" ; |
3493 | for (const auto &FeatureBitset : FeatureBitsets) { |
3494 | if (FeatureBitset.empty()) |
3495 | continue; |
3496 | OS << " {" ; |
3497 | for (const auto &Feature : FeatureBitset) { |
3498 | const auto &I = Info.SubtargetFeatures.find(x: Feature); |
3499 | assert(I != Info.SubtargetFeatures.end() && "Didn't import predicate?" ); |
3500 | OS << I->second.getEnumBitName() << ", " ; |
3501 | } |
3502 | OS << "},\n" ; |
3503 | } |
3504 | OS << "};\n\n" ; |
3505 | |
3506 | // Emit the static match table; unused classes get initialized to 0 which is |
3507 | // guaranteed to be InvalidMatchClass. |
3508 | // |
3509 | // FIXME: We can reduce the size of this table very easily. First, we change |
3510 | // it so that store the kinds in separate bit-fields for each index, which |
3511 | // only needs to be the max width used for classes at that index (we also need |
3512 | // to reject based on this during classification). If we then make sure to |
3513 | // order the match kinds appropriately (putting mnemonics last), then we |
3514 | // should only end up using a few bits for each class, especially the ones |
3515 | // following the mnemonic. |
3516 | OS << "namespace {\n" ; |
3517 | OS << " struct MatchEntry {\n" ; |
3518 | OS << " " << getMinimalTypeForRange(Range: MaxMnemonicIndex) << " Mnemonic;\n" ; |
3519 | OS << " uint16_t Opcode;\n" ; |
3520 | OS << " " << getMinimalTypeForRange(Range: NumConverters) << " ConvertFn;\n" ; |
3521 | OS << " " << getMinimalTypeForRange(Range: FeatureBitsets.size()) |
3522 | << " RequiredFeaturesIdx;\n" ; |
3523 | OS << " " |
3524 | << getMinimalTypeForRange( |
3525 | Range: std::distance(first: Info.Classes.begin(), last: Info.Classes.end()) + |
3526 | 2 /* Include 'InvalidMatchClass' and 'OptionalMatchClass' */) |
3527 | << " Classes[" << MaxNumOperands << "];\n" ; |
3528 | OS << " StringRef getMnemonic() const {\n" ; |
3529 | OS << " return StringRef(MnemonicTable + Mnemonic + 1,\n" ; |
3530 | OS << " MnemonicTable[Mnemonic]);\n" ; |
3531 | OS << " }\n" ; |
3532 | OS << " };\n\n" ; |
3533 | |
3534 | OS << " // Predicate for searching for an opcode.\n" ; |
3535 | OS << " struct LessOpcode {\n" ; |
3536 | OS << " bool operator()(const MatchEntry &LHS, StringRef RHS) {\n" ; |
3537 | OS << " return LHS.getMnemonic() < RHS;\n" ; |
3538 | OS << " }\n" ; |
3539 | OS << " bool operator()(StringRef LHS, const MatchEntry &RHS) {\n" ; |
3540 | OS << " return LHS < RHS.getMnemonic();\n" ; |
3541 | OS << " }\n" ; |
3542 | OS << " bool operator()(const MatchEntry &LHS, const MatchEntry &RHS) {\n" ; |
3543 | OS << " return LHS.getMnemonic() < RHS.getMnemonic();\n" ; |
3544 | OS << " }\n" ; |
3545 | OS << " };\n" ; |
3546 | |
3547 | OS << "} // end anonymous namespace\n\n" ; |
3548 | |
3549 | unsigned VariantCount = Target.getAsmParserVariantCount(); |
3550 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
3551 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
3552 | int AsmVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
3553 | |
3554 | OS << "static const MatchEntry MatchTable" << VC << "[] = {\n" ; |
3555 | |
3556 | for (const auto &MI : Info.Matchables) { |
3557 | if (MI->AsmVariantID != AsmVariantNo) |
3558 | continue; |
3559 | |
3560 | // Store a pascal-style length byte in the mnemonic. |
3561 | std::string LenMnemonic = |
3562 | char(MI->Mnemonic.size()) + MI->Mnemonic.lower(); |
3563 | OS << " { " << *StringTable.GetStringOffset(Str: LenMnemonic) << " /* " |
3564 | << MI->Mnemonic << " */, " << Target.getInstNamespace() |
3565 | << "::" << MI->getResultInst()->TheDef->getName() << ", " |
3566 | << MI->ConversionFnKind << ", " ; |
3567 | |
3568 | // Write the required features mask. |
3569 | OS << "AMFBS" ; |
3570 | if (MI->RequiredFeatures.empty()) |
3571 | OS << "_None" ; |
3572 | else |
3573 | for (const auto &F : MI->RequiredFeatures) |
3574 | OS << '_' << F->TheDef->getName(); |
3575 | |
3576 | OS << ", { " ; |
3577 | ListSeparator LS; |
3578 | for (const MatchableInfo::AsmOperand &Op : MI->AsmOperands) |
3579 | OS << LS << Op.Class->Name; |
3580 | OS << " }, },\n" ; |
3581 | } |
3582 | |
3583 | OS << "};\n\n" ; |
3584 | } |
3585 | |
3586 | OS << "#include \"llvm/Support/Debug.h\"\n" ; |
3587 | OS << "#include \"llvm/Support/Format.h\"\n\n" ; |
3588 | |
3589 | // Finally, build the match function. |
3590 | OS << "unsigned " << Target.getName() << ClassName << "::\n" |
3591 | << "MatchInstructionImpl(const OperandVector &Operands,\n" ; |
3592 | OS << " MCInst &Inst,\n" ; |
3593 | if (ReportMultipleNearMisses) |
3594 | OS << " SmallVectorImpl<NearMissInfo> *NearMisses,\n" ; |
3595 | else |
3596 | OS << " uint64_t &ErrorInfo,\n" |
3597 | << " FeatureBitset &MissingFeatures,\n" ; |
3598 | OS << " bool matchingInlineAsm, unsigned VariantID) {\n" ; |
3599 | |
3600 | if (!ReportMultipleNearMisses) { |
3601 | OS << " // Eliminate obvious mismatches.\n" ; |
3602 | OS << " if (Operands.size() > " << (MaxNumOperands + HasMnemonicFirst) |
3603 | << ") {\n" ; |
3604 | OS << " ErrorInfo = " << (MaxNumOperands + HasMnemonicFirst) << ";\n" ; |
3605 | OS << " return Match_InvalidOperand;\n" ; |
3606 | OS << " }\n\n" ; |
3607 | } |
3608 | |
3609 | // Emit code to get the available features. |
3610 | OS << " // Get the current feature set.\n" ; |
3611 | OS << " const FeatureBitset &AvailableFeatures = " |
3612 | "getAvailableFeatures();\n\n" ; |
3613 | |
3614 | OS << " // Get the instruction mnemonic, which is the first token.\n" ; |
3615 | if (HasMnemonicFirst) { |
3616 | OS << " StringRef Mnemonic = ((" << Target.getName() |
3617 | << "Operand &)*Operands[0]).getToken();\n\n" ; |
3618 | } else { |
3619 | OS << " StringRef Mnemonic;\n" ; |
3620 | OS << " if (Operands[0]->isToken())\n" ; |
3621 | OS << " Mnemonic = ((" << Target.getName() |
3622 | << "Operand &)*Operands[0]).getToken();\n\n" ; |
3623 | } |
3624 | |
3625 | if (HasMnemonicAliases) { |
3626 | OS << " // Process all MnemonicAliases to remap the mnemonic.\n" ; |
3627 | OS << " applyMnemonicAliases(Mnemonic, AvailableFeatures, VariantID);\n\n" ; |
3628 | } |
3629 | |
3630 | // Emit code to compute the class list for this operand vector. |
3631 | if (!ReportMultipleNearMisses) { |
3632 | OS << " // Some state to try to produce better error messages.\n" ; |
3633 | OS << " bool HadMatchOtherThanFeatures = false;\n" ; |
3634 | OS << " bool HadMatchOtherThanPredicate = false;\n" ; |
3635 | OS << " unsigned RetCode = Match_InvalidOperand;\n" ; |
3636 | OS << " MissingFeatures.set();\n" ; |
3637 | OS << " // Set ErrorInfo to the operand that mismatches if it is\n" ; |
3638 | OS << " // wrong for all instances of the instruction.\n" ; |
3639 | OS << " ErrorInfo = ~0ULL;\n" ; |
3640 | } |
3641 | |
3642 | if (HasOptionalOperands) { |
3643 | OS << " SmallBitVector OptionalOperandsMask(" << MaxNumOperands << ");\n" ; |
3644 | } |
3645 | |
3646 | // Emit code to search the table. |
3647 | OS << " // Find the appropriate table for this asm variant.\n" ; |
3648 | OS << " const MatchEntry *Start, *End;\n" ; |
3649 | OS << " switch (VariantID) {\n" ; |
3650 | OS << " default: llvm_unreachable(\"invalid variant!\");\n" ; |
3651 | for (unsigned VC = 0; VC != VariantCount; ++VC) { |
3652 | const Record *AsmVariant = Target.getAsmParserVariant(i: VC); |
3653 | int AsmVariantNo = AsmVariant->getValueAsInt(FieldName: "Variant" ); |
3654 | OS << " case " << AsmVariantNo << ": Start = std::begin(MatchTable" << VC |
3655 | << "); End = std::end(MatchTable" << VC << "); break;\n" ; |
3656 | } |
3657 | OS << " }\n" ; |
3658 | |
3659 | OS << " // Search the table.\n" ; |
3660 | if (HasMnemonicFirst) { |
3661 | OS << " auto MnemonicRange = " |
3662 | "std::equal_range(Start, End, Mnemonic, LessOpcode());\n\n" ; |
3663 | } else { |
3664 | OS << " auto MnemonicRange = std::pair(Start, End);\n" ; |
3665 | OS << " unsigned SIndex = Mnemonic.empty() ? 0 : 1;\n" ; |
3666 | OS << " if (!Mnemonic.empty())\n" ; |
3667 | OS << " MnemonicRange = " |
3668 | "std::equal_range(Start, End, Mnemonic.lower(), LessOpcode());\n\n" ; |
3669 | } |
3670 | |
3671 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"AsmMatcher: found \" " |
3672 | "<<\n" |
3673 | << " std::distance(MnemonicRange.first, MnemonicRange.second) <<\n" |
3674 | << " \" encodings with mnemonic '\" << Mnemonic << \"'\\n\");\n\n" ; |
3675 | |
3676 | OS << " // Return a more specific error code if no mnemonics match.\n" ; |
3677 | OS << " if (MnemonicRange.first == MnemonicRange.second)\n" ; |
3678 | OS << " return Match_MnemonicFail;\n\n" ; |
3679 | |
3680 | OS << " for (const MatchEntry *it = MnemonicRange.first, " |
3681 | << "*ie = MnemonicRange.second;\n" ; |
3682 | OS << " it != ie; ++it) {\n" ; |
3683 | OS << " const FeatureBitset &RequiredFeatures = " |
3684 | "FeatureBitsets[it->RequiredFeaturesIdx];\n" ; |
3685 | OS << " bool HasRequiredFeatures =\n" ; |
3686 | OS << " (AvailableFeatures & RequiredFeatures) == RequiredFeatures;\n" ; |
3687 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Trying to match " |
3688 | "opcode \"\n" ; |
3689 | OS << " << MII.getName(it->Opcode) " |
3690 | "<< \"\\n\");\n" ; |
3691 | |
3692 | if (ReportMultipleNearMisses) { |
3693 | OS << " // Some state to record ways in which this instruction did not " |
3694 | "match.\n" ; |
3695 | OS << " NearMissInfo OperandNearMiss = NearMissInfo::getSuccess();\n" ; |
3696 | OS << " NearMissInfo FeaturesNearMiss = NearMissInfo::getSuccess();\n" ; |
3697 | OS << " NearMissInfo EarlyPredicateNearMiss = " |
3698 | "NearMissInfo::getSuccess();\n" ; |
3699 | OS << " NearMissInfo LatePredicateNearMiss = " |
3700 | "NearMissInfo::getSuccess();\n" ; |
3701 | OS << " bool MultipleInvalidOperands = false;\n" ; |
3702 | } |
3703 | |
3704 | if (HasMnemonicFirst) { |
3705 | OS << " // equal_range guarantees that instruction mnemonic matches.\n" ; |
3706 | OS << " assert(Mnemonic == it->getMnemonic());\n" ; |
3707 | } |
3708 | |
3709 | // Emit check that the subclasses match. |
3710 | if (!ReportMultipleNearMisses) |
3711 | OS << " bool OperandsValid = true;\n" ; |
3712 | if (HasOptionalOperands) { |
3713 | OS << " OptionalOperandsMask.reset(0, " << MaxNumOperands << ");\n" ; |
3714 | } |
3715 | OS << " for (unsigned FormalIdx = " << (HasMnemonicFirst ? "0" : "SIndex" ) |
3716 | << ", ActualIdx = " << (HasMnemonicFirst ? "1" : "SIndex" ) |
3717 | << "; FormalIdx != " << MaxNumOperands << "; ++FormalIdx) {\n" ; |
3718 | OS << " auto Formal = " |
3719 | << "static_cast<MatchClassKind>(it->Classes[FormalIdx]);\n" ; |
3720 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n" ; |
3721 | OS << " dbgs() << \" Matching formal operand class \" " |
3722 | "<< getMatchClassName(Formal)\n" ; |
3723 | OS << " << \" against actual operand at index \" " |
3724 | "<< ActualIdx);\n" ; |
3725 | OS << " if (ActualIdx < Operands.size())\n" ; |
3726 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \" (\";\n" ; |
3727 | OS << " Operands[ActualIdx]->print(dbgs(), " |
3728 | "*getContext().getAsmInfo()); dbgs() << " |
3729 | "\"): \");\n" ; |
3730 | OS << " else\n" ; |
3731 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \": \");\n" ; |
3732 | OS << " if (ActualIdx >= Operands.size()) {\n" ; |
3733 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"actual operand " |
3734 | "index out of range\\n\");\n" ; |
3735 | if (ReportMultipleNearMisses) { |
3736 | OS << " bool ThisOperandValid = (Formal == " |
3737 | << "InvalidMatchClass) || " |
3738 | "isSubclass(Formal, OptionalMatchClass);\n" ; |
3739 | OS << " if (!ThisOperandValid) {\n" ; |
3740 | OS << " if (!OperandNearMiss) {\n" ; |
3741 | OS << " // Record info about match failure for later use.\n" ; |
3742 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"recording " |
3743 | "too-few-operands near miss\\n\");\n" ; |
3744 | OS << " OperandNearMiss =\n" ; |
3745 | OS << " NearMissInfo::getTooFewOperands(Formal, " |
3746 | "it->Opcode);\n" ; |
3747 | OS << " } else if (OperandNearMiss.getKind() != " |
3748 | "NearMissInfo::NearMissTooFewOperands) {\n" ; |
3749 | OS << " // If more than one operand is invalid, give up on this " |
3750 | "match entry.\n" ; |
3751 | OS << " DEBUG_WITH_TYPE(\n" ; |
3752 | OS << " \"asm-matcher\",\n" ; |
3753 | OS << " dbgs() << \"second invalid operand, giving up on " |
3754 | "this opcode\\n\");\n" ; |
3755 | OS << " MultipleInvalidOperands = true;\n" ; |
3756 | OS << " break;\n" ; |
3757 | OS << " }\n" ; |
3758 | OS << " } else {\n" ; |
3759 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"but formal " |
3760 | "operand not required\\n\");\n" ; |
3761 | OS << " if (isSubclass(Formal, OptionalMatchClass)) {\n" ; |
3762 | OS << " OptionalOperandsMask.set(FormalIdx);\n" ; |
3763 | OS << " }\n" ; |
3764 | OS << " }\n" ; |
3765 | OS << " continue;\n" ; |
3766 | } else { |
3767 | OS << " if (Formal == InvalidMatchClass) {\n" ; |
3768 | if (HasOptionalOperands) { |
3769 | OS << " OptionalOperandsMask.set(FormalIdx, " << MaxNumOperands |
3770 | << ");\n" ; |
3771 | } |
3772 | OS << " break;\n" ; |
3773 | OS << " }\n" ; |
3774 | OS << " if (isSubclass(Formal, OptionalMatchClass)) {\n" ; |
3775 | if (HasOptionalOperands) { |
3776 | OS << " OptionalOperandsMask.set(FormalIdx);\n" ; |
3777 | } |
3778 | OS << " continue;\n" ; |
3779 | OS << " }\n" ; |
3780 | OS << " OperandsValid = false;\n" ; |
3781 | OS << " ErrorInfo = ActualIdx;\n" ; |
3782 | OS << " break;\n" ; |
3783 | } |
3784 | OS << " }\n" ; |
3785 | OS << " MCParsedAsmOperand &Actual = *Operands[ActualIdx];\n" ; |
3786 | OS << " unsigned Diag = validateOperandClass(Actual, Formal);\n" ; |
3787 | OS << " if (Diag == Match_Success) {\n" ; |
3788 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n" ; |
3789 | OS << " dbgs() << \"match success using generic " |
3790 | "matcher\\n\");\n" ; |
3791 | OS << " ++ActualIdx;\n" ; |
3792 | OS << " continue;\n" ; |
3793 | OS << " }\n" ; |
3794 | OS << " // If the generic handler indicates an invalid operand\n" ; |
3795 | OS << " // failure, check for a special case.\n" ; |
3796 | OS << " if (Diag != Match_Success) {\n" ; |
3797 | OS << " unsigned TargetDiag = validateTargetOperandClass(Actual, " |
3798 | "Formal);\n" ; |
3799 | OS << " if (TargetDiag == Match_Success) {\n" ; |
3800 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\",\n" ; |
3801 | OS << " dbgs() << \"match success using target " |
3802 | "matcher\\n\");\n" ; |
3803 | OS << " ++ActualIdx;\n" ; |
3804 | OS << " continue;\n" ; |
3805 | OS << " }\n" ; |
3806 | OS << " // If the target matcher returned a specific error code use\n" ; |
3807 | OS << " // that, else use the one from the generic matcher.\n" ; |
3808 | OS << " if (TargetDiag != Match_InvalidOperand && " |
3809 | "HasRequiredFeatures)\n" ; |
3810 | OS << " Diag = TargetDiag;\n" ; |
3811 | OS << " }\n" ; |
3812 | OS << " // If current formal operand wasn't matched and it is optional\n" |
3813 | << " // then try to match next formal operand\n" ; |
3814 | OS << " if (Diag == Match_InvalidOperand " |
3815 | << "&& isSubclass(Formal, OptionalMatchClass)) {\n" ; |
3816 | if (HasOptionalOperands) { |
3817 | OS << " OptionalOperandsMask.set(FormalIdx);\n" ; |
3818 | } |
3819 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"ignoring " |
3820 | "optional operand\\n\");\n" ; |
3821 | OS << " continue;\n" ; |
3822 | OS << " }\n" ; |
3823 | |
3824 | if (ReportMultipleNearMisses) { |
3825 | OS << " if (!OperandNearMiss) {\n" ; |
3826 | OS << " // If this is the first invalid operand we have seen, " |
3827 | "record some\n" ; |
3828 | OS << " // information about it.\n" ; |
3829 | OS << " DEBUG_WITH_TYPE(\n" ; |
3830 | OS << " \"asm-matcher\",\n" ; |
3831 | OS << " dbgs()\n" ; |
3832 | OS << " << \"operand match failed, recording near-miss with " |
3833 | "diag code \"\n" ; |
3834 | OS << " << Diag << \"\\n\");\n" ; |
3835 | OS << " OperandNearMiss =\n" ; |
3836 | OS << " NearMissInfo::getMissedOperand(Diag, Formal, " |
3837 | "it->Opcode, ActualIdx);\n" ; |
3838 | OS << " ++ActualIdx;\n" ; |
3839 | OS << " } else {\n" ; |
3840 | OS << " // If more than one operand is invalid, give up on this " |
3841 | "match entry.\n" ; |
3842 | OS << " DEBUG_WITH_TYPE(\n" ; |
3843 | OS << " \"asm-matcher\",\n" ; |
3844 | OS << " dbgs() << \"second operand mismatch, skipping this " |
3845 | "opcode\\n\");\n" ; |
3846 | OS << " MultipleInvalidOperands = true;\n" ; |
3847 | OS << " break;\n" ; |
3848 | OS << " }\n" ; |
3849 | OS << " }\n\n" ; |
3850 | } else { |
3851 | OS << " // If this operand is broken for all of the instances of " |
3852 | "this\n" ; |
3853 | OS << " // mnemonic, keep track of it so we can report loc info.\n" ; |
3854 | OS << " // If we already had a match that only failed due to a\n" ; |
3855 | OS << " // target predicate, that diagnostic is preferred.\n" ; |
3856 | OS << " if (!HadMatchOtherThanPredicate &&\n" ; |
3857 | OS << " (it == MnemonicRange.first || ErrorInfo <= ActualIdx)) " |
3858 | "{\n" ; |
3859 | OS << " if (HasRequiredFeatures && (ErrorInfo != ActualIdx || Diag " |
3860 | "!= Match_InvalidOperand))\n" ; |
3861 | OS << " RetCode = Diag;\n" ; |
3862 | OS << " ErrorInfo = ActualIdx;\n" ; |
3863 | OS << " }\n" ; |
3864 | OS << " // Otherwise, just reject this instance of the mnemonic.\n" ; |
3865 | OS << " OperandsValid = false;\n" ; |
3866 | OS << " break;\n" ; |
3867 | OS << " }\n\n" ; |
3868 | } |
3869 | |
3870 | if (ReportMultipleNearMisses) |
3871 | OS << " if (MultipleInvalidOperands) {\n" ; |
3872 | else |
3873 | OS << " if (!OperandsValid) {\n" ; |
3874 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: " |
3875 | "multiple \"\n" ; |
3876 | OS << " \"operand mismatches, " |
3877 | "ignoring \"\n" ; |
3878 | OS << " \"this opcode\\n\");\n" ; |
3879 | OS << " continue;\n" ; |
3880 | OS << " }\n" ; |
3881 | |
3882 | // Emit check that the required features are available. |
3883 | OS << " if (!HasRequiredFeatures) {\n" ; |
3884 | if (!ReportMultipleNearMisses) |
3885 | OS << " HadMatchOtherThanFeatures = true;\n" ; |
3886 | OS << " FeatureBitset NewMissingFeatures = RequiredFeatures & " |
3887 | "~AvailableFeatures;\n" ; |
3888 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Missing target " |
3889 | "features:\";\n" ; |
3890 | OS << " for (unsigned I = 0, E = " |
3891 | "NewMissingFeatures.size(); I != E; ++I)\n" ; |
3892 | OS << " if (NewMissingFeatures[I])\n" ; |
3893 | OS << " dbgs() << ' ' << I;\n" ; |
3894 | OS << " dbgs() << \"\\n\");\n" ; |
3895 | if (ReportMultipleNearMisses) { |
3896 | OS << " FeaturesNearMiss = " |
3897 | "NearMissInfo::getMissedFeature(NewMissingFeatures);\n" ; |
3898 | } else { |
3899 | OS << " if (NewMissingFeatures.count() <=\n" |
3900 | " MissingFeatures.count())\n" ; |
3901 | OS << " MissingFeatures = NewMissingFeatures;\n" ; |
3902 | OS << " continue;\n" ; |
3903 | } |
3904 | OS << " }\n" ; |
3905 | OS << "\n" ; |
3906 | OS << " Inst.clear();\n\n" ; |
3907 | OS << " Inst.setOpcode(it->Opcode);\n" ; |
3908 | // Verify the instruction with the target-specific match predicate function. |
3909 | OS << " // We have a potential match but have not rendered the operands.\n" |
3910 | << " // Check the target predicate to handle any context sensitive\n" |
3911 | " // constraints.\n" |
3912 | << " // For example, Ties that are referenced multiple times must be\n" |
3913 | " // checked here to ensure the input is the same for each match\n" |
3914 | " // constraints. If we leave it any later the ties will have been\n" |
3915 | " // canonicalized\n" |
3916 | << " unsigned MatchResult;\n" |
3917 | << " if ((MatchResult = checkEarlyTargetMatchPredicate(Inst, " |
3918 | "Operands)) != Match_Success) {\n" |
3919 | << " Inst.clear();\n" ; |
3920 | OS << " DEBUG_WITH_TYPE(\n" ; |
3921 | OS << " \"asm-matcher\",\n" ; |
3922 | OS << " dbgs() << \"Early target match predicate failed with diag " |
3923 | "code \"\n" ; |
3924 | OS << " << MatchResult << \"\\n\");\n" ; |
3925 | if (ReportMultipleNearMisses) { |
3926 | OS << " EarlyPredicateNearMiss = " |
3927 | "NearMissInfo::getMissedPredicate(MatchResult);\n" ; |
3928 | } else { |
3929 | OS << " RetCode = MatchResult;\n" |
3930 | << " HadMatchOtherThanPredicate = true;\n" |
3931 | << " continue;\n" ; |
3932 | } |
3933 | OS << " }\n\n" ; |
3934 | |
3935 | if (ReportMultipleNearMisses) { |
3936 | OS << " // If we did not successfully match the operands, then we can't " |
3937 | "convert to\n" ; |
3938 | OS << " // an MCInst, so bail out on this instruction variant now.\n" ; |
3939 | OS << " if (OperandNearMiss) {\n" ; |
3940 | OS << " // If the operand mismatch was the only problem, report it as " |
3941 | "a near-miss.\n" ; |
3942 | OS << " if (NearMisses && !FeaturesNearMiss && " |
3943 | "!EarlyPredicateNearMiss) {\n" ; |
3944 | OS << " DEBUG_WITH_TYPE(\n" ; |
3945 | OS << " \"asm-matcher\",\n" ; |
3946 | OS << " dbgs()\n" ; |
3947 | OS << " << \"Opcode result: one mismatched operand, adding " |
3948 | "near-miss\\n\");\n" ; |
3949 | OS << " NearMisses->push_back(OperandNearMiss);\n" ; |
3950 | OS << " } else {\n" ; |
3951 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: " |
3952 | "multiple \"\n" ; |
3953 | OS << " \"types of " |
3954 | "mismatch, so not \"\n" ; |
3955 | OS << " \"reporting " |
3956 | "near-miss\\n\");\n" ; |
3957 | OS << " }\n" ; |
3958 | OS << " continue;\n" ; |
3959 | OS << " }\n\n" ; |
3960 | } |
3961 | |
3962 | // When converting parsed operands to MCInst we need to know whether optional |
3963 | // operands were parsed or not so that we can choose the correct converter |
3964 | // function. We also need to know this when checking tied operand constraints. |
3965 | // DefaultsOffset is an array of deltas between the formal (MCInst) and the |
3966 | // actual (parsed operand array) operand indices. When all optional operands |
3967 | // are present, all elements of the array are zeros. If some of the optional |
3968 | // operands are absent, the array might look like '0, 0, 1, 1, 1, 2, 2, 3', |
3969 | // where each increment in value reflects the absence of an optional operand. |
3970 | if (HasOptionalOperands) { |
3971 | OS << " unsigned DefaultsOffset[" << (MaxNumOperands + 1) |
3972 | << "] = { 0 };\n" ; |
3973 | OS << " assert(OptionalOperandsMask.size() == " << (MaxNumOperands) |
3974 | << ");\n" ; |
3975 | OS << " for (unsigned i = 0, NumDefaults = 0; i < " << (MaxNumOperands) |
3976 | << "; ++i) {\n" ; |
3977 | OS << " DefaultsOffset[i + 1] = NumDefaults;\n" ; |
3978 | OS << " NumDefaults += (OptionalOperandsMask[i] ? 1 : 0);\n" ; |
3979 | OS << " }\n\n" ; |
3980 | } |
3981 | |
3982 | OS << " if (matchingInlineAsm) {\n" ; |
3983 | OS << " convertToMapAndConstraints(it->ConvertFn, Operands);\n" ; |
3984 | if (!ReportMultipleNearMisses) { |
3985 | if (HasOptionalOperands) { |
3986 | OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, " |
3987 | "Operands,\n" ; |
3988 | OS << " DefaultsOffset, " |
3989 | "ErrorInfo))\n" ; |
3990 | } else { |
3991 | OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, " |
3992 | "Operands,\n" ; |
3993 | OS << " ErrorInfo))\n" ; |
3994 | } |
3995 | OS << " return Match_InvalidTiedOperand;\n" ; |
3996 | OS << "\n" ; |
3997 | } |
3998 | OS << " return Match_Success;\n" ; |
3999 | OS << " }\n\n" ; |
4000 | OS << " // We have selected a definite instruction, convert the parsed\n" |
4001 | << " // operands into the appropriate MCInst.\n" ; |
4002 | if (HasOptionalOperands) { |
4003 | OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands,\n" |
4004 | << " OptionalOperandsMask, DefaultsOffset);\n" ; |
4005 | } else { |
4006 | OS << " convertToMCInst(it->ConvertFn, Inst, it->Opcode, Operands);\n" ; |
4007 | } |
4008 | OS << "\n" ; |
4009 | |
4010 | // Verify the instruction with the target-specific match predicate function. |
4011 | OS << " // We have a potential match. Check the target predicate to\n" |
4012 | << " // handle any context sensitive constraints.\n" |
4013 | << " if ((MatchResult = checkTargetMatchPredicate(Inst)) !=" |
4014 | << " Match_Success) {\n" |
4015 | << " DEBUG_WITH_TYPE(\"asm-matcher\",\n" |
4016 | << " dbgs() << \"Target match predicate failed with " |
4017 | "diag code \"\n" |
4018 | << " << MatchResult << \"\\n\");\n" |
4019 | << " Inst.clear();\n" ; |
4020 | if (ReportMultipleNearMisses) { |
4021 | OS << " LatePredicateNearMiss = " |
4022 | "NearMissInfo::getMissedPredicate(MatchResult);\n" ; |
4023 | } else { |
4024 | OS << " RetCode = MatchResult;\n" |
4025 | << " HadMatchOtherThanPredicate = true;\n" |
4026 | << " continue;\n" ; |
4027 | } |
4028 | OS << " }\n\n" ; |
4029 | |
4030 | if (ReportMultipleNearMisses) { |
4031 | OS << " int NumNearMisses = ((int)(bool)OperandNearMiss +\n" ; |
4032 | OS << " (int)(bool)FeaturesNearMiss +\n" ; |
4033 | OS << " (int)(bool)EarlyPredicateNearMiss +\n" ; |
4034 | OS << " (int)(bool)LatePredicateNearMiss);\n" ; |
4035 | OS << " if (NumNearMisses == 1) {\n" ; |
4036 | OS << " // We had exactly one type of near-miss, so add that to the " |
4037 | "list.\n" ; |
4038 | OS << " assert(!OperandNearMiss && \"OperandNearMiss was handled " |
4039 | "earlier\");\n" ; |
4040 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: " |
4041 | "found one type of \"\n" ; |
4042 | OS << " \"mismatch, so " |
4043 | "reporting a \"\n" ; |
4044 | OS << " \"near-miss\\n\");\n" ; |
4045 | OS << " if (NearMisses && FeaturesNearMiss)\n" ; |
4046 | OS << " NearMisses->push_back(FeaturesNearMiss);\n" ; |
4047 | OS << " else if (NearMisses && EarlyPredicateNearMiss)\n" ; |
4048 | OS << " NearMisses->push_back(EarlyPredicateNearMiss);\n" ; |
4049 | OS << " else if (NearMisses && LatePredicateNearMiss)\n" ; |
4050 | OS << " NearMisses->push_back(LatePredicateNearMiss);\n" ; |
4051 | OS << "\n" ; |
4052 | OS << " continue;\n" ; |
4053 | OS << " } else if (NumNearMisses > 1) {\n" ; |
4054 | OS << " // This instruction missed in more than one way, so ignore " |
4055 | "it.\n" ; |
4056 | OS << " DEBUG_WITH_TYPE(\"asm-matcher\", dbgs() << \"Opcode result: " |
4057 | "multiple \"\n" ; |
4058 | OS << " \"types of mismatch, " |
4059 | "so not \"\n" ; |
4060 | OS << " \"reporting " |
4061 | "near-miss\\n\");\n" ; |
4062 | OS << " continue;\n" ; |
4063 | OS << " }\n" ; |
4064 | } |
4065 | |
4066 | // Call the post-processing function, if used. |
4067 | StringRef InsnCleanupFn = AsmParser->getValueAsString(FieldName: "AsmParserInstCleanup" ); |
4068 | if (!InsnCleanupFn.empty()) |
4069 | OS << " " << InsnCleanupFn << "(Inst);\n" ; |
4070 | |
4071 | if (HasDeprecation) { |
4072 | OS << " std::string Info;\n" ; |
4073 | OS << " if " |
4074 | "(!getParser().getTargetParser().getTargetOptions()." |
4075 | "MCNoDeprecatedWarn &&\n" ; |
4076 | OS << " MII.getDeprecatedInfo(Inst, getSTI(), Info)) {\n" ; |
4077 | OS << " SMLoc Loc = ((" << Target.getName() |
4078 | << "Operand &)*Operands[0]).getStartLoc();\n" ; |
4079 | OS << " getParser().Warning(Loc, Info, std::nullopt);\n" ; |
4080 | OS << " }\n" ; |
4081 | } |
4082 | |
4083 | if (!ReportMultipleNearMisses) { |
4084 | if (HasOptionalOperands) { |
4085 | OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, " |
4086 | "Operands,\n" ; |
4087 | OS << " DefaultsOffset, " |
4088 | "ErrorInfo))\n" ; |
4089 | } else { |
4090 | OS << " if (!checkAsmTiedOperandConstraints(*this, it->ConvertFn, " |
4091 | "Operands,\n" ; |
4092 | OS << " ErrorInfo))\n" ; |
4093 | } |
4094 | OS << " return Match_InvalidTiedOperand;\n" ; |
4095 | OS << "\n" ; |
4096 | } |
4097 | |
4098 | OS << " DEBUG_WITH_TYPE(\n" ; |
4099 | OS << " \"asm-matcher\",\n" ; |
4100 | OS << " dbgs() << \"Opcode result: complete match, selecting this " |
4101 | "opcode\\n\");\n" ; |
4102 | OS << " return Match_Success;\n" ; |
4103 | OS << " }\n\n" ; |
4104 | |
4105 | if (ReportMultipleNearMisses) { |
4106 | OS << " // No instruction variants matched exactly.\n" ; |
4107 | OS << " return Match_NearMisses;\n" ; |
4108 | } else { |
4109 | OS << " // Okay, we had no match. Try to return a useful error code.\n" ; |
4110 | OS << " if (HadMatchOtherThanPredicate || !HadMatchOtherThanFeatures)\n" ; |
4111 | OS << " return RetCode;\n\n" ; |
4112 | OS << " ErrorInfo = 0;\n" ; |
4113 | OS << " return Match_MissingFeature;\n" ; |
4114 | } |
4115 | OS << "}\n\n" ; |
4116 | |
4117 | if (!Info.OperandMatchInfo.empty()) |
4118 | emitCustomOperandParsing(OS, Target, Info, ClassName, StringTable, |
4119 | MaxMnemonicIndex, MaxFeaturesIndex: FeatureBitsets.size(), |
4120 | HasMnemonicFirst, AsmParser: *AsmParser); |
4121 | |
4122 | OS << "#endif // GET_MATCHER_IMPLEMENTATION\n\n" ; |
4123 | |
4124 | OS << "\n#ifdef GET_MNEMONIC_SPELL_CHECKER\n" ; |
4125 | OS << "#undef GET_MNEMONIC_SPELL_CHECKER\n\n" ; |
4126 | |
4127 | emitMnemonicSpellChecker(OS, Target, VariantCount); |
4128 | |
4129 | OS << "#endif // GET_MNEMONIC_SPELL_CHECKER\n\n" ; |
4130 | |
4131 | OS << "\n#ifdef GET_MNEMONIC_CHECKER\n" ; |
4132 | OS << "#undef GET_MNEMONIC_CHECKER\n\n" ; |
4133 | |
4134 | emitMnemonicChecker(OS, Target, VariantCount, HasMnemonicFirst, |
4135 | HasMnemonicAliases); |
4136 | |
4137 | OS << "#endif // GET_MNEMONIC_CHECKER\n\n" ; |
4138 | } |
4139 | |
4140 | static TableGen::Emitter::OptClass<AsmMatcherEmitter> |
4141 | X("gen-asm-matcher" , "Generate assembly instruction matcher" ); |
4142 | |