1 //===- Parser.cpp - MLIR Parser Implementation ----------------------------===// 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 file implements the parser for the MLIR textual form. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "mlir/Parser.h" 14 #include "Lexer.h" 15 #include "mlir/Analysis/Verifier.h" 16 #include "mlir/IR/AffineExpr.h" 17 #include "mlir/IR/AffineMap.h" 18 #include "mlir/IR/Attributes.h" 19 #include "mlir/IR/Builders.h" 20 #include "mlir/IR/Dialect.h" 21 #include "mlir/IR/DialectImplementation.h" 22 #include "mlir/IR/IntegerSet.h" 23 #include "mlir/IR/Location.h" 24 #include "mlir/IR/MLIRContext.h" 25 #include "mlir/IR/Module.h" 26 #include "mlir/IR/OpImplementation.h" 27 #include "mlir/IR/StandardTypes.h" 28 #include "llvm/ADT/APInt.h" 29 #include "llvm/ADT/DenseMap.h" 30 #include "llvm/ADT/StringExtras.h" 31 #include "llvm/ADT/StringSet.h" 32 #include "llvm/ADT/bit.h" 33 #include "llvm/Support/PrettyStackTrace.h" 34 #include "llvm/Support/SMLoc.h" 35 #include "llvm/Support/SourceMgr.h" 36 #include <algorithm> 37 using namespace mlir; 38 using llvm::MemoryBuffer; 39 using llvm::SMLoc; 40 using llvm::SourceMgr; 41 42 namespace { 43 class Parser; 44 45 //===----------------------------------------------------------------------===// 46 // SymbolState 47 //===----------------------------------------------------------------------===// 48 49 /// This class contains record of any parsed top-level symbols. 50 struct SymbolState { 51 // A map from attribute alias identifier to Attribute. 52 llvm::StringMap<Attribute> attributeAliasDefinitions; 53 54 // A map from type alias identifier to Type. 55 llvm::StringMap<Type> typeAliasDefinitions; 56 57 /// A set of locations into the main parser memory buffer for each of the 58 /// active nested parsers. Given that some nested parsers, i.e. custom dialect 59 /// parsers, operate on a temporary memory buffer, this provides an anchor 60 /// point for emitting diagnostics. 61 SmallVector<llvm::SMLoc, 1> nestedParserLocs; 62 63 /// The top-level lexer that contains the original memory buffer provided by 64 /// the user. This is used by nested parsers to get a properly encoded source 65 /// location. 66 Lexer *topLevelLexer = nullptr; 67 }; 68 69 //===----------------------------------------------------------------------===// 70 // ParserState 71 //===----------------------------------------------------------------------===// 72 73 /// This class refers to all of the state maintained globally by the parser, 74 /// such as the current lexer position etc. 75 struct ParserState { 76 ParserState(const llvm::SourceMgr &sourceMgr, MLIRContext *ctx, 77 SymbolState &symbols) 78 : context(ctx), lex(sourceMgr, ctx), curToken(lex.lexToken()), 79 symbols(symbols), parserDepth(symbols.nestedParserLocs.size()) { 80 // Set the top level lexer for the symbol state if one doesn't exist. 81 if (!symbols.topLevelLexer) 82 symbols.topLevelLexer = &lex; 83 } 84 ~ParserState() { 85 // Reset the top level lexer if it refers the lexer in our state. 86 if (symbols.topLevelLexer == &lex) 87 symbols.topLevelLexer = nullptr; 88 } 89 ParserState(const ParserState &) = delete; 90 void operator=(const ParserState &) = delete; 91 92 /// The context we're parsing into. 93 MLIRContext *const context; 94 95 /// The lexer for the source file we're parsing. 96 Lexer lex; 97 98 /// This is the next token that hasn't been consumed yet. 99 Token curToken; 100 101 /// The current state for symbol parsing. 102 SymbolState &symbols; 103 104 /// The depth of this parser in the nested parsing stack. 105 size_t parserDepth; 106 }; 107 108 //===----------------------------------------------------------------------===// 109 // Parser 110 //===----------------------------------------------------------------------===// 111 112 /// This class implement support for parsing global entities like types and 113 /// shared entities like SSA names. It is intended to be subclassed by 114 /// specialized subparsers that include state, e.g. when a local symbol table. 115 class Parser { 116 public: 117 Builder builder; 118 119 Parser(ParserState &state) : builder(state.context), state(state) {} 120 121 // Helper methods to get stuff from the parser-global state. 122 ParserState &getState() const { return state; } 123 MLIRContext *getContext() const { return state.context; } 124 const llvm::SourceMgr &getSourceMgr() { return state.lex.getSourceMgr(); } 125 126 /// Parse a comma-separated list of elements up until the specified end token. 127 ParseResult 128 parseCommaSeparatedListUntil(Token::Kind rightToken, 129 const std::function<ParseResult()> &parseElement, 130 bool allowEmptyList = true); 131 132 /// Parse a comma separated list of elements that must have at least one entry 133 /// in it. 134 ParseResult 135 parseCommaSeparatedList(const std::function<ParseResult()> &parseElement); 136 137 ParseResult parsePrettyDialectSymbolName(StringRef &prettyName); 138 139 // We have two forms of parsing methods - those that return a non-null 140 // pointer on success, and those that return a ParseResult to indicate whether 141 // they returned a failure. The second class fills in by-reference arguments 142 // as the results of their action. 143 144 //===--------------------------------------------------------------------===// 145 // Error Handling 146 //===--------------------------------------------------------------------===// 147 148 /// Emit an error and return failure. 149 InFlightDiagnostic emitError(const Twine &message = {}) { 150 return emitError(state.curToken.getLoc(), message); 151 } 152 InFlightDiagnostic emitError(SMLoc loc, const Twine &message = {}); 153 154 /// Encode the specified source location information into an attribute for 155 /// attachment to the IR. 156 Location getEncodedSourceLocation(llvm::SMLoc loc) { 157 // If there are no active nested parsers, we can get the encoded source 158 // location directly. 159 if (state.parserDepth == 0) 160 return state.lex.getEncodedSourceLocation(loc); 161 // Otherwise, we need to re-encode it to point to the top level buffer. 162 return state.symbols.topLevelLexer->getEncodedSourceLocation( 163 remapLocationToTopLevelBuffer(loc)); 164 } 165 166 /// Remaps the given SMLoc to the top level lexer of the parser. This is used 167 /// to adjust locations of potentially nested parsers to ensure that they can 168 /// be emitted properly as diagnostics. 169 llvm::SMLoc remapLocationToTopLevelBuffer(llvm::SMLoc loc) { 170 // If there are no active nested parsers, we can return location directly. 171 SymbolState &symbols = state.symbols; 172 if (state.parserDepth == 0) 173 return loc; 174 assert(symbols.topLevelLexer && "expected valid top-level lexer"); 175 176 // Otherwise, we need to remap the location to the main parser. This is 177 // simply offseting the location onto the location of the last nested 178 // parser. 179 size_t offset = loc.getPointer() - state.lex.getBufferBegin(); 180 auto *rawLoc = 181 symbols.nestedParserLocs[state.parserDepth - 1].getPointer() + offset; 182 return llvm::SMLoc::getFromPointer(rawLoc); 183 } 184 185 //===--------------------------------------------------------------------===// 186 // Token Parsing 187 //===--------------------------------------------------------------------===// 188 189 /// Return the current token the parser is inspecting. 190 const Token &getToken() const { return state.curToken; } 191 StringRef getTokenSpelling() const { return state.curToken.getSpelling(); } 192 193 /// If the current token has the specified kind, consume it and return true. 194 /// If not, return false. 195 bool consumeIf(Token::Kind kind) { 196 if (state.curToken.isNot(kind)) 197 return false; 198 consumeToken(kind); 199 return true; 200 } 201 202 /// Advance the current lexer onto the next token. 203 void consumeToken() { 204 assert(state.curToken.isNot(Token::eof, Token::error) && 205 "shouldn't advance past EOF or errors"); 206 state.curToken = state.lex.lexToken(); 207 } 208 209 /// Advance the current lexer onto the next token, asserting what the expected 210 /// current token is. This is preferred to the above method because it leads 211 /// to more self-documenting code with better checking. 212 void consumeToken(Token::Kind kind) { 213 assert(state.curToken.is(kind) && "consumed an unexpected token"); 214 consumeToken(); 215 } 216 217 /// Consume the specified token if present and return success. On failure, 218 /// output a diagnostic and return failure. 219 ParseResult parseToken(Token::Kind expectedToken, const Twine &message); 220 221 //===--------------------------------------------------------------------===// 222 // Type Parsing 223 //===--------------------------------------------------------------------===// 224 225 ParseResult parseFunctionResultTypes(SmallVectorImpl<Type> &elements); 226 ParseResult parseTypeListNoParens(SmallVectorImpl<Type> &elements); 227 ParseResult parseTypeListParens(SmallVectorImpl<Type> &elements); 228 229 /// Optionally parse a type. 230 OptionalParseResult parseOptionalType(Type &type); 231 232 /// Parse an arbitrary type. 233 Type parseType(); 234 235 /// Parse a complex type. 236 Type parseComplexType(); 237 238 /// Parse an extended type. 239 Type parseExtendedType(); 240 241 /// Parse a function type. 242 Type parseFunctionType(); 243 244 /// Parse a memref type. 245 Type parseMemRefType(); 246 247 /// Parse a non function type. 248 Type parseNonFunctionType(); 249 250 /// Parse a tensor type. 251 Type parseTensorType(); 252 253 /// Parse a tuple type. 254 Type parseTupleType(); 255 256 /// Parse a vector type. 257 VectorType parseVectorType(); 258 ParseResult parseDimensionListRanked(SmallVectorImpl<int64_t> &dimensions, 259 bool allowDynamic = true); 260 ParseResult parseXInDimensionList(); 261 262 /// Parse strided layout specification. 263 ParseResult parseStridedLayout(int64_t &offset, 264 SmallVectorImpl<int64_t> &strides); 265 266 // Parse a brace-delimiter list of comma-separated integers with `?` as an 267 // unknown marker. 268 ParseResult parseStrideList(SmallVectorImpl<int64_t> &dimensions); 269 270 //===--------------------------------------------------------------------===// 271 // Attribute Parsing 272 //===--------------------------------------------------------------------===// 273 274 /// Parse an arbitrary attribute with an optional type. 275 Attribute parseAttribute(Type type = {}); 276 277 /// Parse an attribute dictionary. 278 ParseResult parseAttributeDict(SmallVectorImpl<NamedAttribute> &attributes); 279 280 /// Parse an extended attribute. 281 Attribute parseExtendedAttr(Type type); 282 283 /// Parse a float attribute. 284 Attribute parseFloatAttr(Type type, bool isNegative); 285 286 /// Parse a decimal or a hexadecimal literal, which can be either an integer 287 /// or a float attribute. 288 Attribute parseDecOrHexAttr(Type type, bool isNegative); 289 290 /// Parse an opaque elements attribute. 291 Attribute parseOpaqueElementsAttr(Type attrType); 292 293 /// Parse a dense elements attribute. 294 Attribute parseDenseElementsAttr(Type attrType); 295 ShapedType parseElementsLiteralType(Type type); 296 297 /// Parse a sparse elements attribute. 298 Attribute parseSparseElementsAttr(Type attrType); 299 300 //===--------------------------------------------------------------------===// 301 // Location Parsing 302 //===--------------------------------------------------------------------===// 303 304 /// Parse an inline location. 305 ParseResult parseLocation(LocationAttr &loc); 306 307 /// Parse a raw location instance. 308 ParseResult parseLocationInstance(LocationAttr &loc); 309 310 /// Parse a callsite location instance. 311 ParseResult parseCallSiteLocation(LocationAttr &loc); 312 313 /// Parse a fused location instance. 314 ParseResult parseFusedLocation(LocationAttr &loc); 315 316 /// Parse a name or FileLineCol location instance. 317 ParseResult parseNameOrFileLineColLocation(LocationAttr &loc); 318 319 /// Parse an optional trailing location. 320 /// 321 /// trailing-location ::= (`loc` `(` location `)`)? 322 /// 323 ParseResult parseOptionalTrailingLocation(Location &loc) { 324 // If there is a 'loc' we parse a trailing location. 325 if (!getToken().is(Token::kw_loc)) 326 return success(); 327 328 // Parse the location. 329 LocationAttr directLoc; 330 if (parseLocation(directLoc)) 331 return failure(); 332 loc = directLoc; 333 return success(); 334 } 335 336 //===--------------------------------------------------------------------===// 337 // Affine Parsing 338 //===--------------------------------------------------------------------===// 339 340 /// Parse a reference to either an affine map, or an integer set. 341 ParseResult parseAffineMapOrIntegerSetReference(AffineMap &map, 342 IntegerSet &set); 343 ParseResult parseAffineMapReference(AffineMap &map); 344 ParseResult parseIntegerSetReference(IntegerSet &set); 345 346 /// Parse an AffineMap where the dim and symbol identifiers are SSA ids. 347 ParseResult 348 parseAffineMapOfSSAIds(AffineMap &map, 349 function_ref<ParseResult(bool)> parseElement, 350 OpAsmParser::Delimiter delimiter); 351 352 private: 353 /// The Parser is subclassed and reinstantiated. Do not add additional 354 /// non-trivial state here, add it to the ParserState class. 355 ParserState &state; 356 }; 357 } // end anonymous namespace 358 359 //===----------------------------------------------------------------------===// 360 // Helper methods. 361 //===----------------------------------------------------------------------===// 362 363 /// Parse a comma separated list of elements that must have at least one entry 364 /// in it. 365 ParseResult Parser::parseCommaSeparatedList( 366 const std::function<ParseResult()> &parseElement) { 367 // Non-empty case starts with an element. 368 if (parseElement()) 369 return failure(); 370 371 // Otherwise we have a list of comma separated elements. 372 while (consumeIf(Token::comma)) { 373 if (parseElement()) 374 return failure(); 375 } 376 return success(); 377 } 378 379 /// Parse a comma-separated list of elements, terminated with an arbitrary 380 /// token. This allows empty lists if allowEmptyList is true. 381 /// 382 /// abstract-list ::= rightToken // if allowEmptyList == true 383 /// abstract-list ::= element (',' element)* rightToken 384 /// 385 ParseResult Parser::parseCommaSeparatedListUntil( 386 Token::Kind rightToken, const std::function<ParseResult()> &parseElement, 387 bool allowEmptyList) { 388 // Handle the empty case. 389 if (getToken().is(rightToken)) { 390 if (!allowEmptyList) 391 return emitError("expected list element"); 392 consumeToken(rightToken); 393 return success(); 394 } 395 396 if (parseCommaSeparatedList(parseElement) || 397 parseToken(rightToken, "expected ',' or '" + 398 Token::getTokenSpelling(rightToken) + "'")) 399 return failure(); 400 401 return success(); 402 } 403 404 //===----------------------------------------------------------------------===// 405 // DialectAsmParser 406 //===----------------------------------------------------------------------===// 407 408 namespace { 409 /// This class provides the main implementation of the DialectAsmParser that 410 /// allows for dialects to parse attributes and types. This allows for dialect 411 /// hooking into the main MLIR parsing logic. 412 class CustomDialectAsmParser : public DialectAsmParser { 413 public: 414 CustomDialectAsmParser(StringRef fullSpec, Parser &parser) 415 : fullSpec(fullSpec), nameLoc(parser.getToken().getLoc()), 416 parser(parser) {} 417 ~CustomDialectAsmParser() override {} 418 419 /// Emit a diagnostic at the specified location and return failure. 420 InFlightDiagnostic emitError(llvm::SMLoc loc, const Twine &message) override { 421 return parser.emitError(loc, message); 422 } 423 424 /// Return a builder which provides useful access to MLIRContext, global 425 /// objects like types and attributes. 426 Builder &getBuilder() const override { return parser.builder; } 427 428 /// Get the location of the next token and store it into the argument. This 429 /// always succeeds. 430 llvm::SMLoc getCurrentLocation() override { 431 return parser.getToken().getLoc(); 432 } 433 434 /// Return the location of the original name token. 435 llvm::SMLoc getNameLoc() const override { return nameLoc; } 436 437 /// Re-encode the given source location as an MLIR location and return it. 438 Location getEncodedSourceLoc(llvm::SMLoc loc) override { 439 return parser.getEncodedSourceLocation(loc); 440 } 441 442 /// Returns the full specification of the symbol being parsed. This allows 443 /// for using a separate parser if necessary. 444 StringRef getFullSymbolSpec() const override { return fullSpec; } 445 446 /// Parse a floating point value from the stream. 447 ParseResult parseFloat(double &result) override { 448 bool negative = parser.consumeIf(Token::minus); 449 Token curTok = parser.getToken(); 450 451 // Check for a floating point value. 452 if (curTok.is(Token::floatliteral)) { 453 auto val = curTok.getFloatingPointValue(); 454 if (!val.hasValue()) 455 return emitError(curTok.getLoc(), "floating point value too large"); 456 parser.consumeToken(Token::floatliteral); 457 result = negative ? -*val : *val; 458 return success(); 459 } 460 461 // TODO(riverriddle) support hex floating point values. 462 return emitError(getCurrentLocation(), "expected floating point literal"); 463 } 464 465 /// Parse an optional integer value from the stream. 466 OptionalParseResult parseOptionalInteger(uint64_t &result) override { 467 Token curToken = parser.getToken(); 468 if (curToken.isNot(Token::integer, Token::minus)) 469 return llvm::None; 470 471 bool negative = parser.consumeIf(Token::minus); 472 Token curTok = parser.getToken(); 473 if (parser.parseToken(Token::integer, "expected integer value")) 474 return failure(); 475 476 auto val = curTok.getUInt64IntegerValue(); 477 if (!val) 478 return emitError(curTok.getLoc(), "integer value too large"); 479 result = negative ? -*val : *val; 480 return success(); 481 } 482 483 //===--------------------------------------------------------------------===// 484 // Token Parsing 485 //===--------------------------------------------------------------------===// 486 487 /// Parse a `->` token. 488 ParseResult parseArrow() override { 489 return parser.parseToken(Token::arrow, "expected '->'"); 490 } 491 492 /// Parses a `->` if present. 493 ParseResult parseOptionalArrow() override { 494 return success(parser.consumeIf(Token::arrow)); 495 } 496 497 /// Parse a '{' token. 498 ParseResult parseLBrace() override { 499 return parser.parseToken(Token::l_brace, "expected '{'"); 500 } 501 502 /// Parse a '{' token if present 503 ParseResult parseOptionalLBrace() override { 504 return success(parser.consumeIf(Token::l_brace)); 505 } 506 507 /// Parse a `}` token. 508 ParseResult parseRBrace() override { 509 return parser.parseToken(Token::r_brace, "expected '}'"); 510 } 511 512 /// Parse a `}` token if present 513 ParseResult parseOptionalRBrace() override { 514 return success(parser.consumeIf(Token::r_brace)); 515 } 516 517 /// Parse a `:` token. 518 ParseResult parseColon() override { 519 return parser.parseToken(Token::colon, "expected ':'"); 520 } 521 522 /// Parse a `:` token if present. 523 ParseResult parseOptionalColon() override { 524 return success(parser.consumeIf(Token::colon)); 525 } 526 527 /// Parse a `,` token. 528 ParseResult parseComma() override { 529 return parser.parseToken(Token::comma, "expected ','"); 530 } 531 532 /// Parse a `,` token if present. 533 ParseResult parseOptionalComma() override { 534 return success(parser.consumeIf(Token::comma)); 535 } 536 537 /// Parses a `...` if present. 538 ParseResult parseOptionalEllipsis() override { 539 return success(parser.consumeIf(Token::ellipsis)); 540 } 541 542 /// Parse a `=` token. 543 ParseResult parseEqual() override { 544 return parser.parseToken(Token::equal, "expected '='"); 545 } 546 547 /// Parse a '<' token. 548 ParseResult parseLess() override { 549 return parser.parseToken(Token::less, "expected '<'"); 550 } 551 552 /// Parse a `<` token if present. 553 ParseResult parseOptionalLess() override { 554 return success(parser.consumeIf(Token::less)); 555 } 556 557 /// Parse a '>' token. 558 ParseResult parseGreater() override { 559 return parser.parseToken(Token::greater, "expected '>'"); 560 } 561 562 /// Parse a `>` token if present. 563 ParseResult parseOptionalGreater() override { 564 return success(parser.consumeIf(Token::greater)); 565 } 566 567 /// Parse a `(` token. 568 ParseResult parseLParen() override { 569 return parser.parseToken(Token::l_paren, "expected '('"); 570 } 571 572 /// Parses a '(' if present. 573 ParseResult parseOptionalLParen() override { 574 return success(parser.consumeIf(Token::l_paren)); 575 } 576 577 /// Parse a `)` token. 578 ParseResult parseRParen() override { 579 return parser.parseToken(Token::r_paren, "expected ')'"); 580 } 581 582 /// Parses a ')' if present. 583 ParseResult parseOptionalRParen() override { 584 return success(parser.consumeIf(Token::r_paren)); 585 } 586 587 /// Parse a `[` token. 588 ParseResult parseLSquare() override { 589 return parser.parseToken(Token::l_square, "expected '['"); 590 } 591 592 /// Parses a '[' if present. 593 ParseResult parseOptionalLSquare() override { 594 return success(parser.consumeIf(Token::l_square)); 595 } 596 597 /// Parse a `]` token. 598 ParseResult parseRSquare() override { 599 return parser.parseToken(Token::r_square, "expected ']'"); 600 } 601 602 /// Parses a ']' if present. 603 ParseResult parseOptionalRSquare() override { 604 return success(parser.consumeIf(Token::r_square)); 605 } 606 607 /// Parses a '?' if present. 608 ParseResult parseOptionalQuestion() override { 609 return success(parser.consumeIf(Token::question)); 610 } 611 612 /// Parses a '*' if present. 613 ParseResult parseOptionalStar() override { 614 return success(parser.consumeIf(Token::star)); 615 } 616 617 /// Returns if the current token corresponds to a keyword. 618 bool isCurrentTokenAKeyword() const { 619 return parser.getToken().is(Token::bare_identifier) || 620 parser.getToken().isKeyword(); 621 } 622 623 /// Parse the given keyword if present. 624 ParseResult parseOptionalKeyword(StringRef keyword) override { 625 // Check that the current token has the same spelling. 626 if (!isCurrentTokenAKeyword() || parser.getTokenSpelling() != keyword) 627 return failure(); 628 parser.consumeToken(); 629 return success(); 630 } 631 632 /// Parse a keyword, if present, into 'keyword'. 633 ParseResult parseOptionalKeyword(StringRef *keyword) override { 634 // Check that the current token is a keyword. 635 if (!isCurrentTokenAKeyword()) 636 return failure(); 637 638 *keyword = parser.getTokenSpelling(); 639 parser.consumeToken(); 640 return success(); 641 } 642 643 //===--------------------------------------------------------------------===// 644 // Attribute Parsing 645 //===--------------------------------------------------------------------===// 646 647 /// Parse an arbitrary attribute and return it in result. 648 ParseResult parseAttribute(Attribute &result, Type type) override { 649 result = parser.parseAttribute(type); 650 return success(static_cast<bool>(result)); 651 } 652 653 /// Parse an affine map instance into 'map'. 654 ParseResult parseAffineMap(AffineMap &map) override { 655 return parser.parseAffineMapReference(map); 656 } 657 658 /// Parse an integer set instance into 'set'. 659 ParseResult printIntegerSet(IntegerSet &set) override { 660 return parser.parseIntegerSetReference(set); 661 } 662 663 //===--------------------------------------------------------------------===// 664 // Type Parsing 665 //===--------------------------------------------------------------------===// 666 667 ParseResult parseType(Type &result) override { 668 result = parser.parseType(); 669 return success(static_cast<bool>(result)); 670 } 671 672 ParseResult parseDimensionList(SmallVectorImpl<int64_t> &dimensions, 673 bool allowDynamic) override { 674 return parser.parseDimensionListRanked(dimensions, allowDynamic); 675 } 676 677 private: 678 /// The full symbol specification. 679 StringRef fullSpec; 680 681 /// The source location of the dialect symbol. 682 SMLoc nameLoc; 683 684 /// The main parser. 685 Parser &parser; 686 }; 687 } // namespace 688 689 /// Parse the body of a pretty dialect symbol, which starts and ends with <>'s, 690 /// and may be recursive. Return with the 'prettyName' StringRef encompassing 691 /// the entire pretty name. 692 /// 693 /// pretty-dialect-sym-body ::= '<' pretty-dialect-sym-contents+ '>' 694 /// pretty-dialect-sym-contents ::= pretty-dialect-sym-body 695 /// | '(' pretty-dialect-sym-contents+ ')' 696 /// | '[' pretty-dialect-sym-contents+ ']' 697 /// | '{' pretty-dialect-sym-contents+ '}' 698 /// | '[^[<({>\])}\0]+' 699 /// 700 ParseResult Parser::parsePrettyDialectSymbolName(StringRef &prettyName) { 701 // Pretty symbol names are a relatively unstructured format that contains a 702 // series of properly nested punctuation, with anything else in the middle. 703 // Scan ahead to find it and consume it if successful, otherwise emit an 704 // error. 705 auto *curPtr = getTokenSpelling().data(); 706 707 SmallVector<char, 8> nestedPunctuation; 708 709 // Scan over the nested punctuation, bailing out on error and consuming until 710 // we find the end. We know that we're currently looking at the '<', so we 711 // can go until we find the matching '>' character. 712 assert(*curPtr == '<'); 713 do { 714 char c = *curPtr++; 715 switch (c) { 716 case '\0': 717 // This also handles the EOF case. 718 return emitError("unexpected nul or EOF in pretty dialect name"); 719 case '<': 720 case '[': 721 case '(': 722 case '{': 723 nestedPunctuation.push_back(c); 724 continue; 725 726 case '-': 727 // The sequence `->` is treated as special token. 728 if (*curPtr == '>') 729 ++curPtr; 730 continue; 731 732 case '>': 733 if (nestedPunctuation.pop_back_val() != '<') 734 return emitError("unbalanced '>' character in pretty dialect name"); 735 break; 736 case ']': 737 if (nestedPunctuation.pop_back_val() != '[') 738 return emitError("unbalanced ']' character in pretty dialect name"); 739 break; 740 case ')': 741 if (nestedPunctuation.pop_back_val() != '(') 742 return emitError("unbalanced ')' character in pretty dialect name"); 743 break; 744 case '}': 745 if (nestedPunctuation.pop_back_val() != '{') 746 return emitError("unbalanced '}' character in pretty dialect name"); 747 break; 748 749 default: 750 continue; 751 } 752 } while (!nestedPunctuation.empty()); 753 754 // Ok, we succeeded, remember where we stopped, reset the lexer to know it is 755 // consuming all this stuff, and return. 756 state.lex.resetPointer(curPtr); 757 758 unsigned length = curPtr - prettyName.begin(); 759 prettyName = StringRef(prettyName.begin(), length); 760 consumeToken(); 761 return success(); 762 } 763 764 /// Parse an extended dialect symbol. 765 template <typename Symbol, typename SymbolAliasMap, typename CreateFn> 766 static Symbol parseExtendedSymbol(Parser &p, Token::Kind identifierTok, 767 SymbolAliasMap &aliases, 768 CreateFn &&createSymbol) { 769 // Parse the dialect namespace. 770 StringRef identifier = p.getTokenSpelling().drop_front(); 771 auto loc = p.getToken().getLoc(); 772 p.consumeToken(identifierTok); 773 774 // If there is no '<' token following this, and if the typename contains no 775 // dot, then we are parsing a symbol alias. 776 if (p.getToken().isNot(Token::less) && !identifier.contains('.')) { 777 // Check for an alias for this type. 778 auto aliasIt = aliases.find(identifier); 779 if (aliasIt == aliases.end()) 780 return (p.emitError("undefined symbol alias id '" + identifier + "'"), 781 nullptr); 782 return aliasIt->second; 783 } 784 785 // Otherwise, we are parsing a dialect-specific symbol. If the name contains 786 // a dot, then this is the "pretty" form. If not, it is the verbose form that 787 // looks like <"...">. 788 std::string symbolData; 789 auto dialectName = identifier; 790 791 // Handle the verbose form, where "identifier" is a simple dialect name. 792 if (!identifier.contains('.')) { 793 // Consume the '<'. 794 if (p.parseToken(Token::less, "expected '<' in dialect type")) 795 return nullptr; 796 797 // Parse the symbol specific data. 798 if (p.getToken().isNot(Token::string)) 799 return (p.emitError("expected string literal data in dialect symbol"), 800 nullptr); 801 symbolData = p.getToken().getStringValue(); 802 loc = llvm::SMLoc::getFromPointer(p.getToken().getLoc().getPointer() + 1); 803 p.consumeToken(Token::string); 804 805 // Consume the '>'. 806 if (p.parseToken(Token::greater, "expected '>' in dialect symbol")) 807 return nullptr; 808 } else { 809 // Ok, the dialect name is the part of the identifier before the dot, the 810 // part after the dot is the dialect's symbol, or the start thereof. 811 auto dotHalves = identifier.split('.'); 812 dialectName = dotHalves.first; 813 auto prettyName = dotHalves.second; 814 loc = llvm::SMLoc::getFromPointer(prettyName.data()); 815 816 // If the dialect's symbol is followed immediately by a <, then lex the body 817 // of it into prettyName. 818 if (p.getToken().is(Token::less) && 819 prettyName.bytes_end() == p.getTokenSpelling().bytes_begin()) { 820 if (p.parsePrettyDialectSymbolName(prettyName)) 821 return nullptr; 822 } 823 824 symbolData = prettyName.str(); 825 } 826 827 // Record the name location of the type remapped to the top level buffer. 828 llvm::SMLoc locInTopLevelBuffer = p.remapLocationToTopLevelBuffer(loc); 829 p.getState().symbols.nestedParserLocs.push_back(locInTopLevelBuffer); 830 831 // Call into the provided symbol construction function. 832 Symbol sym = createSymbol(dialectName, symbolData, loc); 833 834 // Pop the last parser location. 835 p.getState().symbols.nestedParserLocs.pop_back(); 836 return sym; 837 } 838 839 /// Parses a symbol, of type 'T', and returns it if parsing was successful. If 840 /// parsing failed, nullptr is returned. The number of bytes read from the input 841 /// string is returned in 'numRead'. 842 template <typename T, typename ParserFn> 843 static T parseSymbol(StringRef inputStr, MLIRContext *context, 844 SymbolState &symbolState, ParserFn &&parserFn, 845 size_t *numRead = nullptr) { 846 SourceMgr sourceMgr; 847 auto memBuffer = MemoryBuffer::getMemBuffer( 848 inputStr, /*BufferName=*/"<mlir_parser_buffer>", 849 /*RequiresNullTerminator=*/false); 850 sourceMgr.AddNewSourceBuffer(std::move(memBuffer), SMLoc()); 851 ParserState state(sourceMgr, context, symbolState); 852 Parser parser(state); 853 854 Token startTok = parser.getToken(); 855 T symbol = parserFn(parser); 856 if (!symbol) 857 return T(); 858 859 // If 'numRead' is valid, then provide the number of bytes that were read. 860 Token endTok = parser.getToken(); 861 if (numRead) { 862 *numRead = static_cast<size_t>(endTok.getLoc().getPointer() - 863 startTok.getLoc().getPointer()); 864 865 // Otherwise, ensure that all of the tokens were parsed. 866 } else if (startTok.getLoc() != endTok.getLoc() && endTok.isNot(Token::eof)) { 867 parser.emitError(endTok.getLoc(), "encountered unexpected token"); 868 return T(); 869 } 870 return symbol; 871 } 872 873 //===----------------------------------------------------------------------===// 874 // Error Handling 875 //===----------------------------------------------------------------------===// 876 877 InFlightDiagnostic Parser::emitError(SMLoc loc, const Twine &message) { 878 auto diag = mlir::emitError(getEncodedSourceLocation(loc), message); 879 880 // If we hit a parse error in response to a lexer error, then the lexer 881 // already reported the error. 882 if (getToken().is(Token::error)) 883 diag.abandon(); 884 return diag; 885 } 886 887 //===----------------------------------------------------------------------===// 888 // Token Parsing 889 //===----------------------------------------------------------------------===// 890 891 /// Consume the specified token if present and return success. On failure, 892 /// output a diagnostic and return failure. 893 ParseResult Parser::parseToken(Token::Kind expectedToken, 894 const Twine &message) { 895 if (consumeIf(expectedToken)) 896 return success(); 897 return emitError(message); 898 } 899 900 //===----------------------------------------------------------------------===// 901 // Type Parsing 902 //===----------------------------------------------------------------------===// 903 904 /// Optionally parse a type. 905 OptionalParseResult Parser::parseOptionalType(Type &type) { 906 // There are many different starting tokens for a type, check them here. 907 switch (getToken().getKind()) { 908 case Token::l_paren: 909 case Token::kw_memref: 910 case Token::kw_tensor: 911 case Token::kw_complex: 912 case Token::kw_tuple: 913 case Token::kw_vector: 914 case Token::inttype: 915 case Token::kw_bf16: 916 case Token::kw_f16: 917 case Token::kw_f32: 918 case Token::kw_f64: 919 case Token::kw_index: 920 case Token::kw_none: 921 case Token::exclamation_identifier: 922 return failure(!(type = parseType())); 923 924 default: 925 return llvm::None; 926 } 927 } 928 929 /// Parse an arbitrary type. 930 /// 931 /// type ::= function-type 932 /// | non-function-type 933 /// 934 Type Parser::parseType() { 935 if (getToken().is(Token::l_paren)) 936 return parseFunctionType(); 937 return parseNonFunctionType(); 938 } 939 940 /// Parse a function result type. 941 /// 942 /// function-result-type ::= type-list-parens 943 /// | non-function-type 944 /// 945 ParseResult Parser::parseFunctionResultTypes(SmallVectorImpl<Type> &elements) { 946 if (getToken().is(Token::l_paren)) 947 return parseTypeListParens(elements); 948 949 Type t = parseNonFunctionType(); 950 if (!t) 951 return failure(); 952 elements.push_back(t); 953 return success(); 954 } 955 956 /// Parse a list of types without an enclosing parenthesis. The list must have 957 /// at least one member. 958 /// 959 /// type-list-no-parens ::= type (`,` type)* 960 /// 961 ParseResult Parser::parseTypeListNoParens(SmallVectorImpl<Type> &elements) { 962 auto parseElt = [&]() -> ParseResult { 963 auto elt = parseType(); 964 elements.push_back(elt); 965 return elt ? success() : failure(); 966 }; 967 968 return parseCommaSeparatedList(parseElt); 969 } 970 971 /// Parse a parenthesized list of types. 972 /// 973 /// type-list-parens ::= `(` `)` 974 /// | `(` type-list-no-parens `)` 975 /// 976 ParseResult Parser::parseTypeListParens(SmallVectorImpl<Type> &elements) { 977 if (parseToken(Token::l_paren, "expected '('")) 978 return failure(); 979 980 // Handle empty lists. 981 if (getToken().is(Token::r_paren)) 982 return consumeToken(), success(); 983 984 if (parseTypeListNoParens(elements) || 985 parseToken(Token::r_paren, "expected ')'")) 986 return failure(); 987 return success(); 988 } 989 990 /// Parse a complex type. 991 /// 992 /// complex-type ::= `complex` `<` type `>` 993 /// 994 Type Parser::parseComplexType() { 995 consumeToken(Token::kw_complex); 996 997 // Parse the '<'. 998 if (parseToken(Token::less, "expected '<' in complex type")) 999 return nullptr; 1000 1001 llvm::SMLoc elementTypeLoc = getToken().getLoc(); 1002 auto elementType = parseType(); 1003 if (!elementType || 1004 parseToken(Token::greater, "expected '>' in complex type")) 1005 return nullptr; 1006 if (!elementType.isa<FloatType>() && !elementType.isa<IntegerType>()) 1007 return emitError(elementTypeLoc, "invalid element type for complex"), 1008 nullptr; 1009 1010 return ComplexType::get(elementType); 1011 } 1012 1013 /// Parse an extended type. 1014 /// 1015 /// extended-type ::= (dialect-type | type-alias) 1016 /// dialect-type ::= `!` dialect-namespace `<` `"` type-data `"` `>` 1017 /// dialect-type ::= `!` alias-name pretty-dialect-attribute-body? 1018 /// type-alias ::= `!` alias-name 1019 /// 1020 Type Parser::parseExtendedType() { 1021 return parseExtendedSymbol<Type>( 1022 *this, Token::exclamation_identifier, state.symbols.typeAliasDefinitions, 1023 [&](StringRef dialectName, StringRef symbolData, 1024 llvm::SMLoc loc) -> Type { 1025 // If we found a registered dialect, then ask it to parse the type. 1026 if (auto *dialect = state.context->getRegisteredDialect(dialectName)) { 1027 return parseSymbol<Type>( 1028 symbolData, state.context, state.symbols, [&](Parser &parser) { 1029 CustomDialectAsmParser customParser(symbolData, parser); 1030 return dialect->parseType(customParser); 1031 }); 1032 } 1033 1034 // Otherwise, form a new opaque type. 1035 return OpaqueType::getChecked( 1036 Identifier::get(dialectName, state.context), symbolData, 1037 state.context, getEncodedSourceLocation(loc)); 1038 }); 1039 } 1040 1041 /// Parse a function type. 1042 /// 1043 /// function-type ::= type-list-parens `->` function-result-type 1044 /// 1045 Type Parser::parseFunctionType() { 1046 assert(getToken().is(Token::l_paren)); 1047 1048 SmallVector<Type, 4> arguments, results; 1049 if (parseTypeListParens(arguments) || 1050 parseToken(Token::arrow, "expected '->' in function type") || 1051 parseFunctionResultTypes(results)) 1052 return nullptr; 1053 1054 return builder.getFunctionType(arguments, results); 1055 } 1056 1057 /// Parse the offset and strides from a strided layout specification. 1058 /// 1059 /// strided-layout ::= `offset:` dimension `,` `strides: ` stride-list 1060 /// 1061 ParseResult Parser::parseStridedLayout(int64_t &offset, 1062 SmallVectorImpl<int64_t> &strides) { 1063 // Parse offset. 1064 consumeToken(Token::kw_offset); 1065 if (!consumeIf(Token::colon)) 1066 return emitError("expected colon after `offset` keyword"); 1067 auto maybeOffset = getToken().getUnsignedIntegerValue(); 1068 bool question = getToken().is(Token::question); 1069 if (!maybeOffset && !question) 1070 return emitError("invalid offset"); 1071 offset = maybeOffset ? static_cast<int64_t>(maybeOffset.getValue()) 1072 : MemRefType::getDynamicStrideOrOffset(); 1073 consumeToken(); 1074 1075 if (!consumeIf(Token::comma)) 1076 return emitError("expected comma after offset value"); 1077 1078 // Parse stride list. 1079 if (!consumeIf(Token::kw_strides)) 1080 return emitError("expected `strides` keyword after offset specification"); 1081 if (!consumeIf(Token::colon)) 1082 return emitError("expected colon after `strides` keyword"); 1083 if (failed(parseStrideList(strides))) 1084 return emitError("invalid braces-enclosed stride list"); 1085 if (llvm::any_of(strides, [](int64_t st) { return st == 0; })) 1086 return emitError("invalid memref stride"); 1087 1088 return success(); 1089 } 1090 1091 /// Parse a memref type. 1092 /// 1093 /// memref-type ::= ranked-memref-type | unranked-memref-type 1094 /// 1095 /// ranked-memref-type ::= `memref` `<` dimension-list-ranked type 1096 /// (`,` semi-affine-map-composition)? (`,` 1097 /// memory-space)? `>` 1098 /// 1099 /// unranked-memref-type ::= `memref` `<*x` type (`,` memory-space)? `>` 1100 /// 1101 /// semi-affine-map-composition ::= (semi-affine-map `,` )* semi-affine-map 1102 /// memory-space ::= integer-literal /* | TODO: address-space-id */ 1103 /// 1104 Type Parser::parseMemRefType() { 1105 consumeToken(Token::kw_memref); 1106 1107 if (parseToken(Token::less, "expected '<' in memref type")) 1108 return nullptr; 1109 1110 bool isUnranked; 1111 SmallVector<int64_t, 4> dimensions; 1112 1113 if (consumeIf(Token::star)) { 1114 // This is an unranked memref type. 1115 isUnranked = true; 1116 if (parseXInDimensionList()) 1117 return nullptr; 1118 1119 } else { 1120 isUnranked = false; 1121 if (parseDimensionListRanked(dimensions)) 1122 return nullptr; 1123 } 1124 1125 // Parse the element type. 1126 auto typeLoc = getToken().getLoc(); 1127 auto elementType = parseType(); 1128 if (!elementType) 1129 return nullptr; 1130 1131 // Check that memref is formed from allowed types. 1132 if (!elementType.isIntOrFloat() && !elementType.isa<VectorType>() && 1133 !elementType.isa<ComplexType>()) 1134 return emitError(typeLoc, "invalid memref element type"), nullptr; 1135 1136 // Parse semi-affine-map-composition. 1137 SmallVector<AffineMap, 2> affineMapComposition; 1138 Optional<unsigned> memorySpace; 1139 unsigned numDims = dimensions.size(); 1140 1141 auto parseElt = [&]() -> ParseResult { 1142 // Check for the memory space. 1143 if (getToken().is(Token::integer)) { 1144 if (memorySpace) 1145 return emitError("multiple memory spaces specified in memref type"); 1146 memorySpace = getToken().getUnsignedIntegerValue(); 1147 if (!memorySpace.hasValue()) 1148 return emitError("invalid memory space in memref type"); 1149 consumeToken(Token::integer); 1150 return success(); 1151 } 1152 if (isUnranked) 1153 return emitError("cannot have affine map for unranked memref type"); 1154 if (memorySpace) 1155 return emitError("expected memory space to be last in memref type"); 1156 1157 AffineMap map; 1158 llvm::SMLoc mapLoc = getToken().getLoc(); 1159 if (getToken().is(Token::kw_offset)) { 1160 int64_t offset; 1161 SmallVector<int64_t, 4> strides; 1162 if (failed(parseStridedLayout(offset, strides))) 1163 return failure(); 1164 // Construct strided affine map. 1165 map = makeStridedLinearLayoutMap(strides, offset, state.context); 1166 } else { 1167 // Parse an affine map attribute. 1168 auto affineMap = parseAttribute(); 1169 if (!affineMap) 1170 return failure(); 1171 auto affineMapAttr = affineMap.dyn_cast<AffineMapAttr>(); 1172 if (!affineMapAttr) 1173 return emitError("expected affine map in memref type"); 1174 map = affineMapAttr.getValue(); 1175 } 1176 1177 if (map.getNumDims() != numDims) { 1178 size_t i = affineMapComposition.size(); 1179 return emitError(mapLoc, "memref affine map dimension mismatch between ") 1180 << (i == 0 ? Twine("memref rank") : "affine map " + Twine(i)) 1181 << " and affine map" << i + 1 << ": " << numDims 1182 << " != " << map.getNumDims(); 1183 } 1184 numDims = map.getNumResults(); 1185 affineMapComposition.push_back(map); 1186 return success(); 1187 }; 1188 1189 // Parse a list of mappings and address space if present. 1190 if (!consumeIf(Token::greater)) { 1191 // Parse comma separated list of affine maps, followed by memory space. 1192 if (parseToken(Token::comma, "expected ',' or '>' in memref type") || 1193 parseCommaSeparatedListUntil(Token::greater, parseElt, 1194 /*allowEmptyList=*/false)) { 1195 return nullptr; 1196 } 1197 } 1198 1199 if (isUnranked) 1200 return UnrankedMemRefType::get(elementType, memorySpace.getValueOr(0)); 1201 1202 return MemRefType::get(dimensions, elementType, affineMapComposition, 1203 memorySpace.getValueOr(0)); 1204 } 1205 1206 /// Parse any type except the function type. 1207 /// 1208 /// non-function-type ::= integer-type 1209 /// | index-type 1210 /// | float-type 1211 /// | extended-type 1212 /// | vector-type 1213 /// | tensor-type 1214 /// | memref-type 1215 /// | complex-type 1216 /// | tuple-type 1217 /// | none-type 1218 /// 1219 /// index-type ::= `index` 1220 /// float-type ::= `f16` | `bf16` | `f32` | `f64` 1221 /// none-type ::= `none` 1222 /// 1223 Type Parser::parseNonFunctionType() { 1224 switch (getToken().getKind()) { 1225 default: 1226 return (emitError("expected non-function type"), nullptr); 1227 case Token::kw_memref: 1228 return parseMemRefType(); 1229 case Token::kw_tensor: 1230 return parseTensorType(); 1231 case Token::kw_complex: 1232 return parseComplexType(); 1233 case Token::kw_tuple: 1234 return parseTupleType(); 1235 case Token::kw_vector: 1236 return parseVectorType(); 1237 // integer-type 1238 case Token::inttype: { 1239 auto width = getToken().getIntTypeBitwidth(); 1240 if (!width.hasValue()) 1241 return (emitError("invalid integer width"), nullptr); 1242 if (width.getValue() > IntegerType::kMaxWidth) { 1243 emitError(getToken().getLoc(), "integer bitwidth is limited to ") 1244 << IntegerType::kMaxWidth << " bits"; 1245 return nullptr; 1246 } 1247 1248 IntegerType::SignednessSemantics signSemantics = IntegerType::Signless; 1249 if (Optional<bool> signedness = getToken().getIntTypeSignedness()) 1250 signSemantics = *signedness ? IntegerType::Signed : IntegerType::Unsigned; 1251 1252 auto loc = getEncodedSourceLocation(getToken().getLoc()); 1253 consumeToken(Token::inttype); 1254 return IntegerType::getChecked(width.getValue(), signSemantics, loc); 1255 } 1256 1257 // float-type 1258 case Token::kw_bf16: 1259 consumeToken(Token::kw_bf16); 1260 return builder.getBF16Type(); 1261 case Token::kw_f16: 1262 consumeToken(Token::kw_f16); 1263 return builder.getF16Type(); 1264 case Token::kw_f32: 1265 consumeToken(Token::kw_f32); 1266 return builder.getF32Type(); 1267 case Token::kw_f64: 1268 consumeToken(Token::kw_f64); 1269 return builder.getF64Type(); 1270 1271 // index-type 1272 case Token::kw_index: 1273 consumeToken(Token::kw_index); 1274 return builder.getIndexType(); 1275 1276 // none-type 1277 case Token::kw_none: 1278 consumeToken(Token::kw_none); 1279 return builder.getNoneType(); 1280 1281 // extended type 1282 case Token::exclamation_identifier: 1283 return parseExtendedType(); 1284 } 1285 } 1286 1287 /// Parse a tensor type. 1288 /// 1289 /// tensor-type ::= `tensor` `<` dimension-list type `>` 1290 /// dimension-list ::= dimension-list-ranked | `*x` 1291 /// 1292 Type Parser::parseTensorType() { 1293 consumeToken(Token::kw_tensor); 1294 1295 if (parseToken(Token::less, "expected '<' in tensor type")) 1296 return nullptr; 1297 1298 bool isUnranked; 1299 SmallVector<int64_t, 4> dimensions; 1300 1301 if (consumeIf(Token::star)) { 1302 // This is an unranked tensor type. 1303 isUnranked = true; 1304 1305 if (parseXInDimensionList()) 1306 return nullptr; 1307 1308 } else { 1309 isUnranked = false; 1310 if (parseDimensionListRanked(dimensions)) 1311 return nullptr; 1312 } 1313 1314 // Parse the element type. 1315 auto elementTypeLoc = getToken().getLoc(); 1316 auto elementType = parseType(); 1317 if (!elementType || parseToken(Token::greater, "expected '>' in tensor type")) 1318 return nullptr; 1319 if (!TensorType::isValidElementType(elementType)) 1320 return emitError(elementTypeLoc, "invalid tensor element type"), nullptr; 1321 1322 if (isUnranked) 1323 return UnrankedTensorType::get(elementType); 1324 return RankedTensorType::get(dimensions, elementType); 1325 } 1326 1327 /// Parse a tuple type. 1328 /// 1329 /// tuple-type ::= `tuple` `<` (type (`,` type)*)? `>` 1330 /// 1331 Type Parser::parseTupleType() { 1332 consumeToken(Token::kw_tuple); 1333 1334 // Parse the '<'. 1335 if (parseToken(Token::less, "expected '<' in tuple type")) 1336 return nullptr; 1337 1338 // Check for an empty tuple by directly parsing '>'. 1339 if (consumeIf(Token::greater)) 1340 return TupleType::get(getContext()); 1341 1342 // Parse the element types and the '>'. 1343 SmallVector<Type, 4> types; 1344 if (parseTypeListNoParens(types) || 1345 parseToken(Token::greater, "expected '>' in tuple type")) 1346 return nullptr; 1347 1348 return TupleType::get(types, getContext()); 1349 } 1350 1351 /// Parse a vector type. 1352 /// 1353 /// vector-type ::= `vector` `<` non-empty-static-dimension-list type `>` 1354 /// non-empty-static-dimension-list ::= decimal-literal `x` 1355 /// static-dimension-list 1356 /// static-dimension-list ::= (decimal-literal `x`)* 1357 /// 1358 VectorType Parser::parseVectorType() { 1359 consumeToken(Token::kw_vector); 1360 1361 if (parseToken(Token::less, "expected '<' in vector type")) 1362 return nullptr; 1363 1364 SmallVector<int64_t, 4> dimensions; 1365 if (parseDimensionListRanked(dimensions, /*allowDynamic=*/false)) 1366 return nullptr; 1367 if (dimensions.empty()) 1368 return (emitError("expected dimension size in vector type"), nullptr); 1369 if (any_of(dimensions, [](int64_t i) { return i <= 0; })) 1370 return emitError(getToken().getLoc(), 1371 "vector types must have positive constant sizes"), 1372 nullptr; 1373 1374 // Parse the element type. 1375 auto typeLoc = getToken().getLoc(); 1376 auto elementType = parseType(); 1377 if (!elementType || parseToken(Token::greater, "expected '>' in vector type")) 1378 return nullptr; 1379 if (!VectorType::isValidElementType(elementType)) 1380 return emitError(typeLoc, "vector elements must be int or float type"), 1381 nullptr; 1382 1383 return VectorType::get(dimensions, elementType); 1384 } 1385 1386 /// Parse a dimension list of a tensor or memref type. This populates the 1387 /// dimension list, using -1 for the `?` dimensions if `allowDynamic` is set and 1388 /// errors out on `?` otherwise. 1389 /// 1390 /// dimension-list-ranked ::= (dimension `x`)* 1391 /// dimension ::= `?` | decimal-literal 1392 /// 1393 /// When `allowDynamic` is not set, this is used to parse: 1394 /// 1395 /// static-dimension-list ::= (decimal-literal `x`)* 1396 ParseResult 1397 Parser::parseDimensionListRanked(SmallVectorImpl<int64_t> &dimensions, 1398 bool allowDynamic) { 1399 while (getToken().isAny(Token::integer, Token::question)) { 1400 if (consumeIf(Token::question)) { 1401 if (!allowDynamic) 1402 return emitError("expected static shape"); 1403 dimensions.push_back(-1); 1404 } else { 1405 // Hexadecimal integer literals (starting with `0x`) are not allowed in 1406 // aggregate type declarations. Therefore, `0xf32` should be processed as 1407 // a sequence of separate elements `0`, `x`, `f32`. 1408 if (getTokenSpelling().size() > 1 && getTokenSpelling()[1] == 'x') { 1409 // We can get here only if the token is an integer literal. Hexadecimal 1410 // integer literals can only start with `0x` (`1x` wouldn't lex as a 1411 // literal, just `1` would, at which point we don't get into this 1412 // branch). 1413 assert(getTokenSpelling()[0] == '0' && "invalid integer literal"); 1414 dimensions.push_back(0); 1415 state.lex.resetPointer(getTokenSpelling().data() + 1); 1416 consumeToken(); 1417 } else { 1418 // Make sure this integer value is in bound and valid. 1419 auto dimension = getToken().getUnsignedIntegerValue(); 1420 if (!dimension.hasValue()) 1421 return emitError("invalid dimension"); 1422 dimensions.push_back((int64_t)dimension.getValue()); 1423 consumeToken(Token::integer); 1424 } 1425 } 1426 1427 // Make sure we have an 'x' or something like 'xbf32'. 1428 if (parseXInDimensionList()) 1429 return failure(); 1430 } 1431 1432 return success(); 1433 } 1434 1435 /// Parse an 'x' token in a dimension list, handling the case where the x is 1436 /// juxtaposed with an element type, as in "xf32", leaving the "f32" as the next 1437 /// token. 1438 ParseResult Parser::parseXInDimensionList() { 1439 if (getToken().isNot(Token::bare_identifier) || getTokenSpelling()[0] != 'x') 1440 return emitError("expected 'x' in dimension list"); 1441 1442 // If we had a prefix of 'x', lex the next token immediately after the 'x'. 1443 if (getTokenSpelling().size() != 1) 1444 state.lex.resetPointer(getTokenSpelling().data() + 1); 1445 1446 // Consume the 'x'. 1447 consumeToken(Token::bare_identifier); 1448 1449 return success(); 1450 } 1451 1452 // Parse a comma-separated list of dimensions, possibly empty: 1453 // stride-list ::= `[` (dimension (`,` dimension)*)? `]` 1454 ParseResult Parser::parseStrideList(SmallVectorImpl<int64_t> &dimensions) { 1455 if (!consumeIf(Token::l_square)) 1456 return failure(); 1457 // Empty list early exit. 1458 if (consumeIf(Token::r_square)) 1459 return success(); 1460 while (true) { 1461 if (consumeIf(Token::question)) { 1462 dimensions.push_back(MemRefType::getDynamicStrideOrOffset()); 1463 } else { 1464 // This must be an integer value. 1465 int64_t val; 1466 if (getToken().getSpelling().getAsInteger(10, val)) 1467 return emitError("invalid integer value: ") << getToken().getSpelling(); 1468 // Make sure it is not the one value for `?`. 1469 if (ShapedType::isDynamic(val)) 1470 return emitError("invalid integer value: ") 1471 << getToken().getSpelling() 1472 << ", use `?` to specify a dynamic dimension"; 1473 dimensions.push_back(val); 1474 consumeToken(Token::integer); 1475 } 1476 if (!consumeIf(Token::comma)) 1477 break; 1478 } 1479 if (!consumeIf(Token::r_square)) 1480 return failure(); 1481 return success(); 1482 } 1483 1484 //===----------------------------------------------------------------------===// 1485 // Attribute parsing. 1486 //===----------------------------------------------------------------------===// 1487 1488 /// Return the symbol reference referred to by the given token, that is known to 1489 /// be an @-identifier. 1490 static std::string extractSymbolReference(Token tok) { 1491 assert(tok.is(Token::at_identifier) && "expected valid @-identifier"); 1492 StringRef nameStr = tok.getSpelling().drop_front(); 1493 1494 // Check to see if the reference is a string literal, or a bare identifier. 1495 if (nameStr.front() == '"') 1496 return tok.getStringValue(); 1497 return std::string(nameStr); 1498 } 1499 1500 /// Parse an arbitrary attribute. 1501 /// 1502 /// attribute-value ::= `unit` 1503 /// | bool-literal 1504 /// | integer-literal (`:` (index-type | integer-type))? 1505 /// | float-literal (`:` float-type)? 1506 /// | string-literal (`:` type)? 1507 /// | type 1508 /// | `[` (attribute-value (`,` attribute-value)*)? `]` 1509 /// | `{` (attribute-entry (`,` attribute-entry)*)? `}` 1510 /// | symbol-ref-id (`::` symbol-ref-id)* 1511 /// | `dense` `<` attribute-value `>` `:` 1512 /// (tensor-type | vector-type) 1513 /// | `sparse` `<` attribute-value `,` attribute-value `>` 1514 /// `:` (tensor-type | vector-type) 1515 /// | `opaque` `<` dialect-namespace `,` hex-string-literal 1516 /// `>` `:` (tensor-type | vector-type) 1517 /// | extended-attribute 1518 /// 1519 Attribute Parser::parseAttribute(Type type) { 1520 switch (getToken().getKind()) { 1521 // Parse an AffineMap or IntegerSet attribute. 1522 case Token::kw_affine_map: { 1523 consumeToken(Token::kw_affine_map); 1524 1525 AffineMap map; 1526 if (parseToken(Token::less, "expected '<' in affine map") || 1527 parseAffineMapReference(map) || 1528 parseToken(Token::greater, "expected '>' in affine map")) 1529 return Attribute(); 1530 return AffineMapAttr::get(map); 1531 } 1532 case Token::kw_affine_set: { 1533 consumeToken(Token::kw_affine_set); 1534 1535 IntegerSet set; 1536 if (parseToken(Token::less, "expected '<' in integer set") || 1537 parseIntegerSetReference(set) || 1538 parseToken(Token::greater, "expected '>' in integer set")) 1539 return Attribute(); 1540 return IntegerSetAttr::get(set); 1541 } 1542 1543 // Parse an array attribute. 1544 case Token::l_square: { 1545 consumeToken(Token::l_square); 1546 1547 SmallVector<Attribute, 4> elements; 1548 auto parseElt = [&]() -> ParseResult { 1549 elements.push_back(parseAttribute()); 1550 return elements.back() ? success() : failure(); 1551 }; 1552 1553 if (parseCommaSeparatedListUntil(Token::r_square, parseElt)) 1554 return nullptr; 1555 return builder.getArrayAttr(elements); 1556 } 1557 1558 // Parse a boolean attribute. 1559 case Token::kw_false: 1560 consumeToken(Token::kw_false); 1561 return builder.getBoolAttr(false); 1562 case Token::kw_true: 1563 consumeToken(Token::kw_true); 1564 return builder.getBoolAttr(true); 1565 1566 // Parse a dense elements attribute. 1567 case Token::kw_dense: 1568 return parseDenseElementsAttr(type); 1569 1570 // Parse a dictionary attribute. 1571 case Token::l_brace: { 1572 SmallVector<NamedAttribute, 4> elements; 1573 if (parseAttributeDict(elements)) 1574 return nullptr; 1575 return builder.getDictionaryAttr(elements); 1576 } 1577 1578 // Parse an extended attribute, i.e. alias or dialect attribute. 1579 case Token::hash_identifier: 1580 return parseExtendedAttr(type); 1581 1582 // Parse floating point and integer attributes. 1583 case Token::floatliteral: 1584 return parseFloatAttr(type, /*isNegative=*/false); 1585 case Token::integer: 1586 return parseDecOrHexAttr(type, /*isNegative=*/false); 1587 case Token::minus: { 1588 consumeToken(Token::minus); 1589 if (getToken().is(Token::integer)) 1590 return parseDecOrHexAttr(type, /*isNegative=*/true); 1591 if (getToken().is(Token::floatliteral)) 1592 return parseFloatAttr(type, /*isNegative=*/true); 1593 1594 return (emitError("expected constant integer or floating point value"), 1595 nullptr); 1596 } 1597 1598 // Parse a location attribute. 1599 case Token::kw_loc: { 1600 LocationAttr attr; 1601 return failed(parseLocation(attr)) ? Attribute() : attr; 1602 } 1603 1604 // Parse an opaque elements attribute. 1605 case Token::kw_opaque: 1606 return parseOpaqueElementsAttr(type); 1607 1608 // Parse a sparse elements attribute. 1609 case Token::kw_sparse: 1610 return parseSparseElementsAttr(type); 1611 1612 // Parse a string attribute. 1613 case Token::string: { 1614 auto val = getToken().getStringValue(); 1615 consumeToken(Token::string); 1616 // Parse the optional trailing colon type if one wasn't explicitly provided. 1617 if (!type && consumeIf(Token::colon) && !(type = parseType())) 1618 return Attribute(); 1619 1620 return type ? StringAttr::get(val, type) 1621 : StringAttr::get(val, getContext()); 1622 } 1623 1624 // Parse a symbol reference attribute. 1625 case Token::at_identifier: { 1626 std::string nameStr = extractSymbolReference(getToken()); 1627 consumeToken(Token::at_identifier); 1628 1629 // Parse any nested references. 1630 std::vector<FlatSymbolRefAttr> nestedRefs; 1631 while (getToken().is(Token::colon)) { 1632 // Check for the '::' prefix. 1633 const char *curPointer = getToken().getLoc().getPointer(); 1634 consumeToken(Token::colon); 1635 if (!consumeIf(Token::colon)) { 1636 state.lex.resetPointer(curPointer); 1637 consumeToken(); 1638 break; 1639 } 1640 // Parse the reference itself. 1641 auto curLoc = getToken().getLoc(); 1642 if (getToken().isNot(Token::at_identifier)) { 1643 emitError(curLoc, "expected nested symbol reference identifier"); 1644 return Attribute(); 1645 } 1646 1647 std::string nameStr = extractSymbolReference(getToken()); 1648 consumeToken(Token::at_identifier); 1649 nestedRefs.push_back(SymbolRefAttr::get(nameStr, getContext())); 1650 } 1651 1652 return builder.getSymbolRefAttr(nameStr, nestedRefs); 1653 } 1654 1655 // Parse a 'unit' attribute. 1656 case Token::kw_unit: 1657 consumeToken(Token::kw_unit); 1658 return builder.getUnitAttr(); 1659 1660 default: 1661 // Parse a type attribute. 1662 if (Type type = parseType()) 1663 return TypeAttr::get(type); 1664 return nullptr; 1665 } 1666 } 1667 1668 /// Attribute dictionary. 1669 /// 1670 /// attribute-dict ::= `{` `}` 1671 /// | `{` attribute-entry (`,` attribute-entry)* `}` 1672 /// attribute-entry ::= (bare-id | string-literal) `=` attribute-value 1673 /// 1674 ParseResult 1675 Parser::parseAttributeDict(SmallVectorImpl<NamedAttribute> &attributes) { 1676 if (parseToken(Token::l_brace, "expected '{' in attribute dictionary")) 1677 return failure(); 1678 1679 auto parseElt = [&]() -> ParseResult { 1680 // The name of an attribute can either be a bare identifier, or a string. 1681 Optional<Identifier> nameId; 1682 if (getToken().is(Token::string)) 1683 nameId = builder.getIdentifier(getToken().getStringValue()); 1684 else if (getToken().isAny(Token::bare_identifier, Token::inttype) || 1685 getToken().isKeyword()) 1686 nameId = builder.getIdentifier(getTokenSpelling()); 1687 else 1688 return emitError("expected attribute name"); 1689 consumeToken(); 1690 1691 // Try to parse the '=' for the attribute value. 1692 if (!consumeIf(Token::equal)) { 1693 // If there is no '=', we treat this as a unit attribute. 1694 attributes.push_back({*nameId, builder.getUnitAttr()}); 1695 return success(); 1696 } 1697 1698 auto attr = parseAttribute(); 1699 if (!attr) 1700 return failure(); 1701 1702 attributes.push_back({*nameId, attr}); 1703 return success(); 1704 }; 1705 1706 if (parseCommaSeparatedListUntil(Token::r_brace, parseElt)) 1707 return failure(); 1708 1709 return success(); 1710 } 1711 1712 /// Parse an extended attribute. 1713 /// 1714 /// extended-attribute ::= (dialect-attribute | attribute-alias) 1715 /// dialect-attribute ::= `#` dialect-namespace `<` `"` attr-data `"` `>` 1716 /// dialect-attribute ::= `#` alias-name pretty-dialect-sym-body? 1717 /// attribute-alias ::= `#` alias-name 1718 /// 1719 Attribute Parser::parseExtendedAttr(Type type) { 1720 Attribute attr = parseExtendedSymbol<Attribute>( 1721 *this, Token::hash_identifier, state.symbols.attributeAliasDefinitions, 1722 [&](StringRef dialectName, StringRef symbolData, 1723 llvm::SMLoc loc) -> Attribute { 1724 // Parse an optional trailing colon type. 1725 Type attrType = type; 1726 if (consumeIf(Token::colon) && !(attrType = parseType())) 1727 return Attribute(); 1728 1729 // If we found a registered dialect, then ask it to parse the attribute. 1730 if (auto *dialect = state.context->getRegisteredDialect(dialectName)) { 1731 return parseSymbol<Attribute>( 1732 symbolData, state.context, state.symbols, [&](Parser &parser) { 1733 CustomDialectAsmParser customParser(symbolData, parser); 1734 return dialect->parseAttribute(customParser, attrType); 1735 }); 1736 } 1737 1738 // Otherwise, form a new opaque attribute. 1739 return OpaqueAttr::getChecked( 1740 Identifier::get(dialectName, state.context), symbolData, 1741 attrType ? attrType : NoneType::get(state.context), 1742 getEncodedSourceLocation(loc)); 1743 }); 1744 1745 // Ensure that the attribute has the same type as requested. 1746 if (attr && type && attr.getType() != type) { 1747 emitError("attribute type different than expected: expected ") 1748 << type << ", but got " << attr.getType(); 1749 return nullptr; 1750 } 1751 return attr; 1752 } 1753 1754 /// Parse a float attribute. 1755 Attribute Parser::parseFloatAttr(Type type, bool isNegative) { 1756 auto val = getToken().getFloatingPointValue(); 1757 if (!val.hasValue()) 1758 return (emitError("floating point value too large for attribute"), nullptr); 1759 consumeToken(Token::floatliteral); 1760 if (!type) { 1761 // Default to F64 when no type is specified. 1762 if (!consumeIf(Token::colon)) 1763 type = builder.getF64Type(); 1764 else if (!(type = parseType())) 1765 return nullptr; 1766 } 1767 if (!type.isa<FloatType>()) 1768 return (emitError("floating point value not valid for specified type"), 1769 nullptr); 1770 return FloatAttr::get(type, isNegative ? -val.getValue() : val.getValue()); 1771 } 1772 1773 /// Construct a float attribute bitwise equivalent to the integer literal. 1774 static Optional<APFloat> buildHexadecimalFloatLiteral(Parser *p, FloatType type, 1775 uint64_t value) { 1776 // FIXME: bfloat is currently stored as a double internally because it doesn't 1777 // have valid APFloat semantics. 1778 if (type.isF64() || type.isBF16()) 1779 return APFloat(type.getFloatSemantics(), APInt(/*numBits=*/64, value)); 1780 1781 APInt apInt(type.getWidth(), value); 1782 if (apInt != value) { 1783 p->emitError("hexadecimal float constant out of range for type"); 1784 return llvm::None; 1785 } 1786 return APFloat(type.getFloatSemantics(), apInt); 1787 } 1788 1789 /// Construct an APint from a parsed value, a known attribute type and 1790 /// sign. 1791 static Optional<APInt> buildAttributeAPInt(Type type, bool isNegative, 1792 StringRef spelling) { 1793 // Parse the integer value into an APInt that is big enough to hold the value. 1794 APInt result; 1795 bool isHex = spelling.size() > 1 && spelling[1] == 'x'; 1796 if (spelling.getAsInteger(isHex ? 0 : 10, result)) 1797 return llvm::None; 1798 1799 // Extend or truncate the bitwidth to the right size. 1800 unsigned width = type.isIndex() ? IndexType::kInternalStorageBitWidth 1801 : type.getIntOrFloatBitWidth(); 1802 if (width > result.getBitWidth()) { 1803 result = result.zext(width); 1804 } else if (width < result.getBitWidth()) { 1805 // The parser can return an unnecessarily wide result with leading zeros. 1806 // This isn't a problem, but truncating off bits is bad. 1807 if (result.countLeadingZeros() < result.getBitWidth() - width) 1808 return llvm::None; 1809 1810 result = result.trunc(width); 1811 } 1812 1813 if (isNegative) { 1814 // The value is negative, we have an overflow if the sign bit is not set 1815 // in the negated apInt. 1816 result.negate(); 1817 if (!result.isSignBitSet()) 1818 return llvm::None; 1819 } else if ((type.isSignedInteger() || type.isIndex()) && 1820 result.isSignBitSet()) { 1821 // The value is a positive signed integer or index, 1822 // we have an overflow if the sign bit is set. 1823 return llvm::None; 1824 } 1825 1826 return result; 1827 } 1828 1829 /// Parse a decimal or a hexadecimal literal, which can be either an integer 1830 /// or a float attribute. 1831 Attribute Parser::parseDecOrHexAttr(Type type, bool isNegative) { 1832 // Remember if the literal is hexadecimal. 1833 StringRef spelling = getToken().getSpelling(); 1834 auto loc = state.curToken.getLoc(); 1835 bool isHex = spelling.size() > 1 && spelling[1] == 'x'; 1836 1837 consumeToken(Token::integer); 1838 if (!type) { 1839 // Default to i64 if not type is specified. 1840 if (!consumeIf(Token::colon)) 1841 type = builder.getIntegerType(64); 1842 else if (!(type = parseType())) 1843 return nullptr; 1844 } 1845 1846 if (auto floatType = type.dyn_cast<FloatType>()) { 1847 if (isNegative) 1848 return emitError( 1849 loc, 1850 "hexadecimal float literal should not have a leading minus"), 1851 nullptr; 1852 if (!isHex) { 1853 emitError(loc, "unexpected decimal integer literal for a float attribute") 1854 .attachNote() 1855 << "add a trailing dot to make the literal a float"; 1856 return nullptr; 1857 } 1858 1859 auto val = Token::getUInt64IntegerValue(spelling); 1860 if (!val.hasValue()) 1861 return emitError("integer constant out of range for attribute"), nullptr; 1862 1863 // Construct a float attribute bitwise equivalent to the integer literal. 1864 Optional<APFloat> apVal = 1865 buildHexadecimalFloatLiteral(this, floatType, *val); 1866 return apVal ? FloatAttr::get(floatType, *apVal) : Attribute(); 1867 } 1868 1869 if (!type.isa<IntegerType>() && !type.isa<IndexType>()) 1870 return emitError(loc, "integer literal not valid for specified type"), 1871 nullptr; 1872 1873 if (isNegative && type.isUnsignedInteger()) { 1874 emitError(loc, 1875 "negative integer literal not valid for unsigned integer type"); 1876 return nullptr; 1877 } 1878 1879 Optional<APInt> apInt = buildAttributeAPInt(type, isNegative, spelling); 1880 if (!apInt) 1881 return emitError(loc, "integer constant out of range for attribute"), 1882 nullptr; 1883 return builder.getIntegerAttr(type, *apInt); 1884 } 1885 1886 /// Parse elements values stored within a hex etring. On success, the values are 1887 /// stored into 'result'. 1888 static ParseResult parseElementAttrHexValues(Parser &parser, Token tok, 1889 std::string &result) { 1890 std::string val = tok.getStringValue(); 1891 if (val.size() < 2 || val[0] != '0' || val[1] != 'x') 1892 return parser.emitError(tok.getLoc(), 1893 "elements hex string should start with '0x'"); 1894 1895 StringRef hexValues = StringRef(val).drop_front(2); 1896 if (!llvm::all_of(hexValues, llvm::isHexDigit)) 1897 return parser.emitError(tok.getLoc(), 1898 "elements hex string only contains hex digits"); 1899 1900 result = llvm::fromHex(hexValues); 1901 return success(); 1902 } 1903 1904 /// Parse an opaque elements attribute. 1905 Attribute Parser::parseOpaqueElementsAttr(Type attrType) { 1906 consumeToken(Token::kw_opaque); 1907 if (parseToken(Token::less, "expected '<' after 'opaque'")) 1908 return nullptr; 1909 1910 if (getToken().isNot(Token::string)) 1911 return (emitError("expected dialect namespace"), nullptr); 1912 1913 auto name = getToken().getStringValue(); 1914 auto *dialect = builder.getContext()->getRegisteredDialect(name); 1915 // TODO(shpeisman): Allow for having an unknown dialect on an opaque 1916 // attribute. Otherwise, it can't be roundtripped without having the dialect 1917 // registered. 1918 if (!dialect) 1919 return (emitError("no registered dialect with namespace '" + name + "'"), 1920 nullptr); 1921 consumeToken(Token::string); 1922 1923 if (parseToken(Token::comma, "expected ','")) 1924 return nullptr; 1925 1926 Token hexTok = getToken(); 1927 if (parseToken(Token::string, "elements hex string should start with '0x'") || 1928 parseToken(Token::greater, "expected '>'")) 1929 return nullptr; 1930 auto type = parseElementsLiteralType(attrType); 1931 if (!type) 1932 return nullptr; 1933 1934 std::string data; 1935 if (parseElementAttrHexValues(*this, hexTok, data)) 1936 return nullptr; 1937 return OpaqueElementsAttr::get(dialect, type, data); 1938 } 1939 1940 namespace { 1941 class TensorLiteralParser { 1942 public: 1943 TensorLiteralParser(Parser &p) : p(p) {} 1944 1945 /// Parse the elements of a tensor literal. If 'allowHex' is true, the parser 1946 /// may also parse a tensor literal that is store as a hex string. 1947 ParseResult parse(bool allowHex); 1948 1949 /// Build a dense attribute instance with the parsed elements and the given 1950 /// shaped type. 1951 DenseElementsAttr getAttr(llvm::SMLoc loc, ShapedType type); 1952 1953 ArrayRef<int64_t> getShape() const { return shape; } 1954 1955 private: 1956 enum class ElementKind { Boolean, Integer, Float, String }; 1957 1958 /// Return a string to represent the given element kind. 1959 const char *getElementKindStr(ElementKind kind) { 1960 switch (kind) { 1961 case ElementKind::Boolean: 1962 return "'boolean'"; 1963 case ElementKind::Integer: 1964 return "'integer'"; 1965 case ElementKind::Float: 1966 return "'float'"; 1967 case ElementKind::String: 1968 return "'string'"; 1969 } 1970 llvm_unreachable("unknown element kind"); 1971 } 1972 1973 /// Build a Dense Integer attribute for the given type. 1974 DenseElementsAttr getIntAttr(llvm::SMLoc loc, ShapedType type, Type eltTy); 1975 1976 /// Build a Dense Float attribute for the given type. 1977 DenseElementsAttr getFloatAttr(llvm::SMLoc loc, ShapedType type, 1978 FloatType eltTy); 1979 1980 /// Build a Dense String attribute for the given type. 1981 DenseElementsAttr getStringAttr(llvm::SMLoc loc, ShapedType type, Type eltTy); 1982 1983 /// Build a Dense attribute with hex data for the given type. 1984 DenseElementsAttr getHexAttr(llvm::SMLoc loc, ShapedType type); 1985 1986 /// Parse a single element, returning failure if it isn't a valid element 1987 /// literal. For example: 1988 /// parseElement(1) -> Success, 1 1989 /// parseElement([1]) -> Failure 1990 ParseResult parseElement(); 1991 1992 /// Parse a list of either lists or elements, returning the dimensions of the 1993 /// parsed sub-tensors in dims. For example: 1994 /// parseList([1, 2, 3]) -> Success, [3] 1995 /// parseList([[1, 2], [3, 4]]) -> Success, [2, 2] 1996 /// parseList([[1, 2], 3]) -> Failure 1997 /// parseList([[1, [2, 3]], [4, [5]]]) -> Failure 1998 ParseResult parseList(SmallVectorImpl<int64_t> &dims); 1999 2000 /// Parse a literal that was printed as a hex string. 2001 ParseResult parseHexElements(); 2002 2003 Parser &p; 2004 2005 /// The shape inferred from the parsed elements. 2006 SmallVector<int64_t, 4> shape; 2007 2008 /// Storage used when parsing elements, this is a pair of <is_negated, token>. 2009 std::vector<std::pair<bool, Token>> storage; 2010 2011 /// A flag that indicates the type of elements that have been parsed. 2012 Optional<ElementKind> knownEltKind; 2013 2014 /// Storage used when parsing elements that were stored as hex values. 2015 Optional<Token> hexStorage; 2016 }; 2017 } // namespace 2018 2019 /// Parse the elements of a tensor literal. If 'allowHex' is true, the parser 2020 /// may also parse a tensor literal that is store as a hex string. 2021 ParseResult TensorLiteralParser::parse(bool allowHex) { 2022 // If hex is allowed, check for a string literal. 2023 if (allowHex && p.getToken().is(Token::string)) { 2024 hexStorage = p.getToken(); 2025 p.consumeToken(Token::string); 2026 return success(); 2027 } 2028 // Otherwise, parse a list or an individual element. 2029 if (p.getToken().is(Token::l_square)) 2030 return parseList(shape); 2031 return parseElement(); 2032 } 2033 2034 /// Build a dense attribute instance with the parsed elements and the given 2035 /// shaped type. 2036 DenseElementsAttr TensorLiteralParser::getAttr(llvm::SMLoc loc, 2037 ShapedType type) { 2038 Type eltType = type.getElementType(); 2039 2040 // Check to see if we parse the literal from a hex string. 2041 if (hexStorage.hasValue() && eltType.isIntOrFloat()) 2042 return getHexAttr(loc, type); 2043 2044 // Check that the parsed storage size has the same number of elements to the 2045 // type, or is a known splat. 2046 if (!shape.empty() && getShape() != type.getShape()) { 2047 p.emitError(loc) << "inferred shape of elements literal ([" << getShape() 2048 << "]) does not match type ([" << type.getShape() << "])"; 2049 return nullptr; 2050 } 2051 2052 // If the type is an integer, build a set of APInt values from the storage 2053 // with the correct bitwidth. 2054 if (auto intTy = eltType.dyn_cast<IntegerType>()) 2055 return getIntAttr(loc, type, intTy); 2056 if (auto indexTy = eltType.dyn_cast<IndexType>()) 2057 return getIntAttr(loc, type, indexTy); 2058 2059 // If parsing a floating point type. 2060 if (auto floatTy = eltType.dyn_cast<FloatType>()) 2061 return getFloatAttr(loc, type, floatTy); 2062 2063 // Other types are assumed to be string representations. 2064 return getStringAttr(loc, type, type.getElementType()); 2065 } 2066 2067 /// Build a Dense Integer attribute for the given type. 2068 DenseElementsAttr TensorLiteralParser::getIntAttr(llvm::SMLoc loc, 2069 ShapedType type, Type eltTy) { 2070 std::vector<APInt> intElements; 2071 intElements.reserve(storage.size()); 2072 auto isUintType = type.getElementType().isUnsignedInteger(); 2073 for (const auto &signAndToken : storage) { 2074 bool isNegative = signAndToken.first; 2075 const Token &token = signAndToken.second; 2076 auto tokenLoc = token.getLoc(); 2077 2078 if (isNegative && isUintType) { 2079 p.emitError(tokenLoc) 2080 << "expected unsigned integer elements, but parsed negative value"; 2081 return nullptr; 2082 } 2083 2084 // Check to see if floating point values were parsed. 2085 if (token.is(Token::floatliteral)) { 2086 p.emitError(tokenLoc) 2087 << "expected integer elements, but parsed floating-point"; 2088 return nullptr; 2089 } 2090 2091 assert(token.isAny(Token::integer, Token::kw_true, Token::kw_false) && 2092 "unexpected token type"); 2093 if (token.isAny(Token::kw_true, Token::kw_false)) { 2094 if (!eltTy.isInteger(1)) { 2095 p.emitError(tokenLoc) 2096 << "expected i1 type for 'true' or 'false' values"; 2097 return nullptr; 2098 } 2099 APInt apInt(1, token.is(Token::kw_true), /*isSigned=*/false); 2100 intElements.push_back(apInt); 2101 continue; 2102 } 2103 2104 // Create APInt values for each element with the correct bitwidth. 2105 Optional<APInt> apInt = 2106 buildAttributeAPInt(eltTy, isNegative, token.getSpelling()); 2107 if (!apInt) 2108 return (p.emitError(tokenLoc, "integer constant out of range for type"), 2109 nullptr); 2110 intElements.push_back(*apInt); 2111 } 2112 2113 return DenseElementsAttr::get(type, intElements); 2114 } 2115 2116 /// Build a Dense Float attribute for the given type. 2117 DenseElementsAttr TensorLiteralParser::getFloatAttr(llvm::SMLoc loc, 2118 ShapedType type, 2119 FloatType eltTy) { 2120 std::vector<APFloat> floatValues; 2121 floatValues.reserve(storage.size()); 2122 for (const auto &signAndToken : storage) { 2123 bool isNegative = signAndToken.first; 2124 const Token &token = signAndToken.second; 2125 2126 // Handle hexadecimal float literals. 2127 if (token.is(Token::integer) && token.getSpelling().startswith("0x")) { 2128 if (isNegative) { 2129 p.emitError(token.getLoc()) 2130 << "hexadecimal float literal should not have a leading minus"; 2131 return nullptr; 2132 } 2133 auto val = token.getUInt64IntegerValue(); 2134 if (!val.hasValue()) { 2135 p.emitError("hexadecimal float constant out of range for attribute"); 2136 return nullptr; 2137 } 2138 Optional<APFloat> apVal = buildHexadecimalFloatLiteral(&p, eltTy, *val); 2139 if (!apVal) 2140 return nullptr; 2141 floatValues.push_back(*apVal); 2142 continue; 2143 } 2144 2145 // Check to see if any decimal integers or booleans were parsed. 2146 if (!token.is(Token::floatliteral)) { 2147 p.emitError() << "expected floating-point elements, but parsed integer"; 2148 return nullptr; 2149 } 2150 2151 // Build the float values from tokens. 2152 auto val = token.getFloatingPointValue(); 2153 if (!val.hasValue()) { 2154 p.emitError("floating point value too large for attribute"); 2155 return nullptr; 2156 } 2157 // Treat BF16 as double because it is not supported in LLVM's APFloat. 2158 APFloat apVal(isNegative ? -*val : *val); 2159 if (!eltTy.isBF16() && !eltTy.isF64()) { 2160 bool unused; 2161 apVal.convert(eltTy.getFloatSemantics(), APFloat::rmNearestTiesToEven, 2162 &unused); 2163 } 2164 floatValues.push_back(apVal); 2165 } 2166 2167 return DenseElementsAttr::get(type, floatValues); 2168 } 2169 2170 /// Build a Dense String attribute for the given type. 2171 DenseElementsAttr TensorLiteralParser::getStringAttr(llvm::SMLoc loc, 2172 ShapedType type, 2173 Type eltTy) { 2174 if (hexStorage.hasValue()) { 2175 auto stringValue = hexStorage.getValue().getStringValue(); 2176 return DenseStringElementsAttr::get(type, {stringValue}); 2177 } 2178 2179 std::vector<std::string> stringValues; 2180 std::vector<StringRef> stringRefValues; 2181 stringValues.reserve(storage.size()); 2182 stringRefValues.reserve(storage.size()); 2183 2184 for (auto val : storage) { 2185 stringValues.push_back(val.second.getStringValue()); 2186 stringRefValues.push_back(stringValues.back()); 2187 } 2188 2189 return DenseStringElementsAttr::get(type, stringRefValues); 2190 } 2191 2192 /// Build a Dense attribute with hex data for the given type. 2193 DenseElementsAttr TensorLiteralParser::getHexAttr(llvm::SMLoc loc, 2194 ShapedType type) { 2195 Type elementType = type.getElementType(); 2196 if (!elementType.isa<FloatType>() && !elementType.isa<IntegerType>()) { 2197 p.emitError(loc) << "expected floating-point or integer element type, got " 2198 << elementType; 2199 return nullptr; 2200 } 2201 2202 std::string data; 2203 if (parseElementAttrHexValues(p, hexStorage.getValue(), data)) 2204 return nullptr; 2205 2206 // Check that the size of the hex data corresponds to the size of the type, or 2207 // a splat of the type. 2208 // TODO: bf16 is currently stored as a double, this should be removed when 2209 // APFloat properly supports it. 2210 int64_t elementWidth = 2211 elementType.isBF16() ? 64 : elementType.getIntOrFloatBitWidth(); 2212 if (static_cast<int64_t>(data.size() * CHAR_BIT) != 2213 (type.getNumElements() * elementWidth)) { 2214 p.emitError(loc) << "elements hex data size is invalid for provided type: " 2215 << type; 2216 return nullptr; 2217 } 2218 2219 return DenseElementsAttr::getFromRawBuffer( 2220 type, ArrayRef<char>(data.data(), data.size()), /*isSplatBuffer=*/false); 2221 } 2222 2223 ParseResult TensorLiteralParser::parseElement() { 2224 switch (p.getToken().getKind()) { 2225 // Parse a boolean element. 2226 case Token::kw_true: 2227 case Token::kw_false: 2228 case Token::floatliteral: 2229 case Token::integer: 2230 storage.emplace_back(/*isNegative=*/false, p.getToken()); 2231 p.consumeToken(); 2232 break; 2233 2234 // Parse a signed integer or a negative floating-point element. 2235 case Token::minus: 2236 p.consumeToken(Token::minus); 2237 if (!p.getToken().isAny(Token::floatliteral, Token::integer)) 2238 return p.emitError("expected integer or floating point literal"); 2239 storage.emplace_back(/*isNegative=*/true, p.getToken()); 2240 p.consumeToken(); 2241 break; 2242 2243 case Token::string: 2244 storage.emplace_back(/*isNegative=*/ false, p.getToken()); 2245 p.consumeToken(); 2246 break; 2247 default: 2248 return p.emitError("expected element literal of primitive type"); 2249 } 2250 2251 return success(); 2252 } 2253 2254 /// Parse a list of either lists or elements, returning the dimensions of the 2255 /// parsed sub-tensors in dims. For example: 2256 /// parseList([1, 2, 3]) -> Success, [3] 2257 /// parseList([[1, 2], [3, 4]]) -> Success, [2, 2] 2258 /// parseList([[1, 2], 3]) -> Failure 2259 /// parseList([[1, [2, 3]], [4, [5]]]) -> Failure 2260 ParseResult TensorLiteralParser::parseList(SmallVectorImpl<int64_t> &dims) { 2261 p.consumeToken(Token::l_square); 2262 2263 auto checkDims = [&](const SmallVectorImpl<int64_t> &prevDims, 2264 const SmallVectorImpl<int64_t> &newDims) -> ParseResult { 2265 if (prevDims == newDims) 2266 return success(); 2267 return p.emitError("tensor literal is invalid; ranks are not consistent " 2268 "between elements"); 2269 }; 2270 2271 bool first = true; 2272 SmallVector<int64_t, 4> newDims; 2273 unsigned size = 0; 2274 auto parseCommaSeparatedList = [&]() -> ParseResult { 2275 SmallVector<int64_t, 4> thisDims; 2276 if (p.getToken().getKind() == Token::l_square) { 2277 if (parseList(thisDims)) 2278 return failure(); 2279 } else if (parseElement()) { 2280 return failure(); 2281 } 2282 ++size; 2283 if (!first) 2284 return checkDims(newDims, thisDims); 2285 newDims = thisDims; 2286 first = false; 2287 return success(); 2288 }; 2289 if (p.parseCommaSeparatedListUntil(Token::r_square, parseCommaSeparatedList)) 2290 return failure(); 2291 2292 // Return the sublists' dimensions with 'size' prepended. 2293 dims.clear(); 2294 dims.push_back(size); 2295 dims.append(newDims.begin(), newDims.end()); 2296 return success(); 2297 } 2298 2299 /// Parse a dense elements attribute. 2300 Attribute Parser::parseDenseElementsAttr(Type attrType) { 2301 consumeToken(Token::kw_dense); 2302 if (parseToken(Token::less, "expected '<' after 'dense'")) 2303 return nullptr; 2304 2305 // Parse the literal data. 2306 TensorLiteralParser literalParser(*this); 2307 if (literalParser.parse(/*allowHex=*/true)) 2308 return nullptr; 2309 2310 if (parseToken(Token::greater, "expected '>'")) 2311 return nullptr; 2312 2313 auto typeLoc = getToken().getLoc(); 2314 auto type = parseElementsLiteralType(attrType); 2315 if (!type) 2316 return nullptr; 2317 return literalParser.getAttr(typeLoc, type); 2318 } 2319 2320 /// Shaped type for elements attribute. 2321 /// 2322 /// elements-literal-type ::= vector-type | ranked-tensor-type 2323 /// 2324 /// This method also checks the type has static shape. 2325 ShapedType Parser::parseElementsLiteralType(Type type) { 2326 // If the user didn't provide a type, parse the colon type for the literal. 2327 if (!type) { 2328 if (parseToken(Token::colon, "expected ':'")) 2329 return nullptr; 2330 if (!(type = parseType())) 2331 return nullptr; 2332 } 2333 2334 if (!type.isa<RankedTensorType>() && !type.isa<VectorType>()) { 2335 emitError("elements literal must be a ranked tensor or vector type"); 2336 return nullptr; 2337 } 2338 2339 auto sType = type.cast<ShapedType>(); 2340 if (!sType.hasStaticShape()) 2341 return (emitError("elements literal type must have static shape"), nullptr); 2342 2343 return sType; 2344 } 2345 2346 /// Parse a sparse elements attribute. 2347 Attribute Parser::parseSparseElementsAttr(Type attrType) { 2348 consumeToken(Token::kw_sparse); 2349 if (parseToken(Token::less, "Expected '<' after 'sparse'")) 2350 return nullptr; 2351 2352 /// Parse the indices. We don't allow hex values here as we may need to use 2353 /// the inferred shape. 2354 auto indicesLoc = getToken().getLoc(); 2355 TensorLiteralParser indiceParser(*this); 2356 if (indiceParser.parse(/*allowHex=*/false)) 2357 return nullptr; 2358 2359 if (parseToken(Token::comma, "expected ','")) 2360 return nullptr; 2361 2362 /// Parse the values. 2363 auto valuesLoc = getToken().getLoc(); 2364 TensorLiteralParser valuesParser(*this); 2365 if (valuesParser.parse(/*allowHex=*/true)) 2366 return nullptr; 2367 2368 if (parseToken(Token::greater, "expected '>'")) 2369 return nullptr; 2370 2371 auto type = parseElementsLiteralType(attrType); 2372 if (!type) 2373 return nullptr; 2374 2375 // If the indices are a splat, i.e. the literal parser parsed an element and 2376 // not a list, we set the shape explicitly. The indices are represented by a 2377 // 2-dimensional shape where the second dimension is the rank of the type. 2378 // Given that the parsed indices is a splat, we know that we only have one 2379 // indice and thus one for the first dimension. 2380 auto indiceEltType = builder.getIntegerType(64); 2381 ShapedType indicesType; 2382 if (indiceParser.getShape().empty()) { 2383 indicesType = RankedTensorType::get({1, type.getRank()}, indiceEltType); 2384 } else { 2385 // Otherwise, set the shape to the one parsed by the literal parser. 2386 indicesType = RankedTensorType::get(indiceParser.getShape(), indiceEltType); 2387 } 2388 auto indices = indiceParser.getAttr(indicesLoc, indicesType); 2389 2390 // If the values are a splat, set the shape explicitly based on the number of 2391 // indices. The number of indices is encoded in the first dimension of the 2392 // indice shape type. 2393 auto valuesEltType = type.getElementType(); 2394 ShapedType valuesType = 2395 valuesParser.getShape().empty() 2396 ? RankedTensorType::get({indicesType.getDimSize(0)}, valuesEltType) 2397 : RankedTensorType::get(valuesParser.getShape(), valuesEltType); 2398 auto values = valuesParser.getAttr(valuesLoc, valuesType); 2399 2400 /// Sanity check. 2401 if (valuesType.getRank() != 1) 2402 return (emitError("expected 1-d tensor for values"), nullptr); 2403 2404 auto sameShape = (indicesType.getRank() == 1) || 2405 (type.getRank() == indicesType.getDimSize(1)); 2406 auto sameElementNum = indicesType.getDimSize(0) == valuesType.getDimSize(0); 2407 if (!sameShape || !sameElementNum) { 2408 emitError() << "expected shape ([" << type.getShape() 2409 << "]); inferred shape of indices literal ([" 2410 << indicesType.getShape() 2411 << "]); inferred shape of values literal ([" 2412 << valuesType.getShape() << "])"; 2413 return nullptr; 2414 } 2415 2416 // Build the sparse elements attribute by the indices and values. 2417 return SparseElementsAttr::get(type, indices, values); 2418 } 2419 2420 //===----------------------------------------------------------------------===// 2421 // Location parsing. 2422 //===----------------------------------------------------------------------===// 2423 2424 /// Parse a location. 2425 /// 2426 /// location ::= `loc` inline-location 2427 /// inline-location ::= '(' location-inst ')' 2428 /// 2429 ParseResult Parser::parseLocation(LocationAttr &loc) { 2430 // Check for 'loc' identifier. 2431 if (parseToken(Token::kw_loc, "expected 'loc' keyword")) 2432 return emitError(); 2433 2434 // Parse the inline-location. 2435 if (parseToken(Token::l_paren, "expected '(' in inline location") || 2436 parseLocationInstance(loc) || 2437 parseToken(Token::r_paren, "expected ')' in inline location")) 2438 return failure(); 2439 return success(); 2440 } 2441 2442 /// Specific location instances. 2443 /// 2444 /// location-inst ::= filelinecol-location | 2445 /// name-location | 2446 /// callsite-location | 2447 /// fused-location | 2448 /// unknown-location 2449 /// filelinecol-location ::= string-literal ':' integer-literal 2450 /// ':' integer-literal 2451 /// name-location ::= string-literal 2452 /// callsite-location ::= 'callsite' '(' location-inst 'at' location-inst ')' 2453 /// fused-location ::= fused ('<' attribute-value '>')? 2454 /// '[' location-inst (location-inst ',')* ']' 2455 /// unknown-location ::= 'unknown' 2456 /// 2457 ParseResult Parser::parseCallSiteLocation(LocationAttr &loc) { 2458 consumeToken(Token::bare_identifier); 2459 2460 // Parse the '('. 2461 if (parseToken(Token::l_paren, "expected '(' in callsite location")) 2462 return failure(); 2463 2464 // Parse the callee location. 2465 LocationAttr calleeLoc; 2466 if (parseLocationInstance(calleeLoc)) 2467 return failure(); 2468 2469 // Parse the 'at'. 2470 if (getToken().isNot(Token::bare_identifier) || 2471 getToken().getSpelling() != "at") 2472 return emitError("expected 'at' in callsite location"); 2473 consumeToken(Token::bare_identifier); 2474 2475 // Parse the caller location. 2476 LocationAttr callerLoc; 2477 if (parseLocationInstance(callerLoc)) 2478 return failure(); 2479 2480 // Parse the ')'. 2481 if (parseToken(Token::r_paren, "expected ')' in callsite location")) 2482 return failure(); 2483 2484 // Return the callsite location. 2485 loc = CallSiteLoc::get(calleeLoc, callerLoc); 2486 return success(); 2487 } 2488 2489 ParseResult Parser::parseFusedLocation(LocationAttr &loc) { 2490 consumeToken(Token::bare_identifier); 2491 2492 // Try to parse the optional metadata. 2493 Attribute metadata; 2494 if (consumeIf(Token::less)) { 2495 metadata = parseAttribute(); 2496 if (!metadata) 2497 return emitError("expected valid attribute metadata"); 2498 // Parse the '>' token. 2499 if (parseToken(Token::greater, 2500 "expected '>' after fused location metadata")) 2501 return failure(); 2502 } 2503 2504 SmallVector<Location, 4> locations; 2505 auto parseElt = [&] { 2506 LocationAttr newLoc; 2507 if (parseLocationInstance(newLoc)) 2508 return failure(); 2509 locations.push_back(newLoc); 2510 return success(); 2511 }; 2512 2513 if (parseToken(Token::l_square, "expected '[' in fused location") || 2514 parseCommaSeparatedList(parseElt) || 2515 parseToken(Token::r_square, "expected ']' in fused location")) 2516 return failure(); 2517 2518 // Return the fused location. 2519 loc = FusedLoc::get(locations, metadata, getContext()); 2520 return success(); 2521 } 2522 2523 ParseResult Parser::parseNameOrFileLineColLocation(LocationAttr &loc) { 2524 auto *ctx = getContext(); 2525 auto str = getToken().getStringValue(); 2526 consumeToken(Token::string); 2527 2528 // If the next token is ':' this is a filelinecol location. 2529 if (consumeIf(Token::colon)) { 2530 // Parse the line number. 2531 if (getToken().isNot(Token::integer)) 2532 return emitError("expected integer line number in FileLineColLoc"); 2533 auto line = getToken().getUnsignedIntegerValue(); 2534 if (!line.hasValue()) 2535 return emitError("expected integer line number in FileLineColLoc"); 2536 consumeToken(Token::integer); 2537 2538 // Parse the ':'. 2539 if (parseToken(Token::colon, "expected ':' in FileLineColLoc")) 2540 return failure(); 2541 2542 // Parse the column number. 2543 if (getToken().isNot(Token::integer)) 2544 return emitError("expected integer column number in FileLineColLoc"); 2545 auto column = getToken().getUnsignedIntegerValue(); 2546 if (!column.hasValue()) 2547 return emitError("expected integer column number in FileLineColLoc"); 2548 consumeToken(Token::integer); 2549 2550 loc = FileLineColLoc::get(str, line.getValue(), column.getValue(), ctx); 2551 return success(); 2552 } 2553 2554 // Otherwise, this is a NameLoc. 2555 2556 // Check for a child location. 2557 if (consumeIf(Token::l_paren)) { 2558 auto childSourceLoc = getToken().getLoc(); 2559 2560 // Parse the child location. 2561 LocationAttr childLoc; 2562 if (parseLocationInstance(childLoc)) 2563 return failure(); 2564 2565 // The child must not be another NameLoc. 2566 if (childLoc.isa<NameLoc>()) 2567 return emitError(childSourceLoc, 2568 "child of NameLoc cannot be another NameLoc"); 2569 loc = NameLoc::get(Identifier::get(str, ctx), childLoc); 2570 2571 // Parse the closing ')'. 2572 if (parseToken(Token::r_paren, 2573 "expected ')' after child location of NameLoc")) 2574 return failure(); 2575 } else { 2576 loc = NameLoc::get(Identifier::get(str, ctx), ctx); 2577 } 2578 2579 return success(); 2580 } 2581 2582 ParseResult Parser::parseLocationInstance(LocationAttr &loc) { 2583 // Handle either name or filelinecol locations. 2584 if (getToken().is(Token::string)) 2585 return parseNameOrFileLineColLocation(loc); 2586 2587 // Bare tokens required for other cases. 2588 if (!getToken().is(Token::bare_identifier)) 2589 return emitError("expected location instance"); 2590 2591 // Check for the 'callsite' signifying a callsite location. 2592 if (getToken().getSpelling() == "callsite") 2593 return parseCallSiteLocation(loc); 2594 2595 // If the token is 'fused', then this is a fused location. 2596 if (getToken().getSpelling() == "fused") 2597 return parseFusedLocation(loc); 2598 2599 // Check for a 'unknown' for an unknown location. 2600 if (getToken().getSpelling() == "unknown") { 2601 consumeToken(Token::bare_identifier); 2602 loc = UnknownLoc::get(getContext()); 2603 return success(); 2604 } 2605 2606 return emitError("expected location instance"); 2607 } 2608 2609 //===----------------------------------------------------------------------===// 2610 // Affine parsing. 2611 //===----------------------------------------------------------------------===// 2612 2613 /// Lower precedence ops (all at the same precedence level). LNoOp is false in 2614 /// the boolean sense. 2615 enum AffineLowPrecOp { 2616 /// Null value. 2617 LNoOp, 2618 Add, 2619 Sub 2620 }; 2621 2622 /// Higher precedence ops - all at the same precedence level. HNoOp is false 2623 /// in the boolean sense. 2624 enum AffineHighPrecOp { 2625 /// Null value. 2626 HNoOp, 2627 Mul, 2628 FloorDiv, 2629 CeilDiv, 2630 Mod 2631 }; 2632 2633 namespace { 2634 /// This is a specialized parser for affine structures (affine maps, affine 2635 /// expressions, and integer sets), maintaining the state transient to their 2636 /// bodies. 2637 class AffineParser : public Parser { 2638 public: 2639 AffineParser(ParserState &state, bool allowParsingSSAIds = false, 2640 function_ref<ParseResult(bool)> parseElement = nullptr) 2641 : Parser(state), allowParsingSSAIds(allowParsingSSAIds), 2642 parseElement(parseElement), numDimOperands(0), numSymbolOperands(0) {} 2643 2644 AffineMap parseAffineMapRange(unsigned numDims, unsigned numSymbols); 2645 ParseResult parseAffineMapOrIntegerSetInline(AffineMap &map, IntegerSet &set); 2646 IntegerSet parseIntegerSetConstraints(unsigned numDims, unsigned numSymbols); 2647 ParseResult parseAffineMapOfSSAIds(AffineMap &map, 2648 OpAsmParser::Delimiter delimiter); 2649 void getDimsAndSymbolSSAIds(SmallVectorImpl<StringRef> &dimAndSymbolSSAIds, 2650 unsigned &numDims); 2651 2652 private: 2653 // Binary affine op parsing. 2654 AffineLowPrecOp consumeIfLowPrecOp(); 2655 AffineHighPrecOp consumeIfHighPrecOp(); 2656 2657 // Identifier lists for polyhedral structures. 2658 ParseResult parseDimIdList(unsigned &numDims); 2659 ParseResult parseSymbolIdList(unsigned &numSymbols); 2660 ParseResult parseDimAndOptionalSymbolIdList(unsigned &numDims, 2661 unsigned &numSymbols); 2662 ParseResult parseIdentifierDefinition(AffineExpr idExpr); 2663 2664 AffineExpr parseAffineExpr(); 2665 AffineExpr parseParentheticalExpr(); 2666 AffineExpr parseNegateExpression(AffineExpr lhs); 2667 AffineExpr parseIntegerExpr(); 2668 AffineExpr parseBareIdExpr(); 2669 AffineExpr parseSSAIdExpr(bool isSymbol); 2670 AffineExpr parseSymbolSSAIdExpr(); 2671 2672 AffineExpr getAffineBinaryOpExpr(AffineHighPrecOp op, AffineExpr lhs, 2673 AffineExpr rhs, SMLoc opLoc); 2674 AffineExpr getAffineBinaryOpExpr(AffineLowPrecOp op, AffineExpr lhs, 2675 AffineExpr rhs); 2676 AffineExpr parseAffineOperandExpr(AffineExpr lhs); 2677 AffineExpr parseAffineLowPrecOpExpr(AffineExpr llhs, AffineLowPrecOp llhsOp); 2678 AffineExpr parseAffineHighPrecOpExpr(AffineExpr llhs, AffineHighPrecOp llhsOp, 2679 SMLoc llhsOpLoc); 2680 AffineExpr parseAffineConstraint(bool *isEq); 2681 2682 private: 2683 bool allowParsingSSAIds; 2684 function_ref<ParseResult(bool)> parseElement; 2685 unsigned numDimOperands; 2686 unsigned numSymbolOperands; 2687 SmallVector<std::pair<StringRef, AffineExpr>, 4> dimsAndSymbols; 2688 }; 2689 } // end anonymous namespace 2690 2691 /// Create an affine binary high precedence op expression (mul's, div's, mod). 2692 /// opLoc is the location of the op token to be used to report errors 2693 /// for non-conforming expressions. 2694 AffineExpr AffineParser::getAffineBinaryOpExpr(AffineHighPrecOp op, 2695 AffineExpr lhs, AffineExpr rhs, 2696 SMLoc opLoc) { 2697 // TODO: make the error location info accurate. 2698 switch (op) { 2699 case Mul: 2700 if (!lhs.isSymbolicOrConstant() && !rhs.isSymbolicOrConstant()) { 2701 emitError(opLoc, "non-affine expression: at least one of the multiply " 2702 "operands has to be either a constant or symbolic"); 2703 return nullptr; 2704 } 2705 return lhs * rhs; 2706 case FloorDiv: 2707 if (!rhs.isSymbolicOrConstant()) { 2708 emitError(opLoc, "non-affine expression: right operand of floordiv " 2709 "has to be either a constant or symbolic"); 2710 return nullptr; 2711 } 2712 return lhs.floorDiv(rhs); 2713 case CeilDiv: 2714 if (!rhs.isSymbolicOrConstant()) { 2715 emitError(opLoc, "non-affine expression: right operand of ceildiv " 2716 "has to be either a constant or symbolic"); 2717 return nullptr; 2718 } 2719 return lhs.ceilDiv(rhs); 2720 case Mod: 2721 if (!rhs.isSymbolicOrConstant()) { 2722 emitError(opLoc, "non-affine expression: right operand of mod " 2723 "has to be either a constant or symbolic"); 2724 return nullptr; 2725 } 2726 return lhs % rhs; 2727 case HNoOp: 2728 llvm_unreachable("can't create affine expression for null high prec op"); 2729 return nullptr; 2730 } 2731 llvm_unreachable("Unknown AffineHighPrecOp"); 2732 } 2733 2734 /// Create an affine binary low precedence op expression (add, sub). 2735 AffineExpr AffineParser::getAffineBinaryOpExpr(AffineLowPrecOp op, 2736 AffineExpr lhs, AffineExpr rhs) { 2737 switch (op) { 2738 case AffineLowPrecOp::Add: 2739 return lhs + rhs; 2740 case AffineLowPrecOp::Sub: 2741 return lhs - rhs; 2742 case AffineLowPrecOp::LNoOp: 2743 llvm_unreachable("can't create affine expression for null low prec op"); 2744 return nullptr; 2745 } 2746 llvm_unreachable("Unknown AffineLowPrecOp"); 2747 } 2748 2749 /// Consume this token if it is a lower precedence affine op (there are only 2750 /// two precedence levels). 2751 AffineLowPrecOp AffineParser::consumeIfLowPrecOp() { 2752 switch (getToken().getKind()) { 2753 case Token::plus: 2754 consumeToken(Token::plus); 2755 return AffineLowPrecOp::Add; 2756 case Token::minus: 2757 consumeToken(Token::minus); 2758 return AffineLowPrecOp::Sub; 2759 default: 2760 return AffineLowPrecOp::LNoOp; 2761 } 2762 } 2763 2764 /// Consume this token if it is a higher precedence affine op (there are only 2765 /// two precedence levels) 2766 AffineHighPrecOp AffineParser::consumeIfHighPrecOp() { 2767 switch (getToken().getKind()) { 2768 case Token::star: 2769 consumeToken(Token::star); 2770 return Mul; 2771 case Token::kw_floordiv: 2772 consumeToken(Token::kw_floordiv); 2773 return FloorDiv; 2774 case Token::kw_ceildiv: 2775 consumeToken(Token::kw_ceildiv); 2776 return CeilDiv; 2777 case Token::kw_mod: 2778 consumeToken(Token::kw_mod); 2779 return Mod; 2780 default: 2781 return HNoOp; 2782 } 2783 } 2784 2785 /// Parse a high precedence op expression list: mul, div, and mod are high 2786 /// precedence binary ops, i.e., parse a 2787 /// expr_1 op_1 expr_2 op_2 ... expr_n 2788 /// where op_1, op_2 are all a AffineHighPrecOp (mul, div, mod). 2789 /// All affine binary ops are left associative. 2790 /// Given llhs, returns (llhs llhsOp lhs) op rhs, or (lhs op rhs) if llhs is 2791 /// null. If no rhs can be found, returns (llhs llhsOp lhs) or lhs if llhs is 2792 /// null. llhsOpLoc is the location of the llhsOp token that will be used to 2793 /// report an error for non-conforming expressions. 2794 AffineExpr AffineParser::parseAffineHighPrecOpExpr(AffineExpr llhs, 2795 AffineHighPrecOp llhsOp, 2796 SMLoc llhsOpLoc) { 2797 AffineExpr lhs = parseAffineOperandExpr(llhs); 2798 if (!lhs) 2799 return nullptr; 2800 2801 // Found an LHS. Parse the remaining expression. 2802 auto opLoc = getToken().getLoc(); 2803 if (AffineHighPrecOp op = consumeIfHighPrecOp()) { 2804 if (llhs) { 2805 AffineExpr expr = getAffineBinaryOpExpr(llhsOp, llhs, lhs, opLoc); 2806 if (!expr) 2807 return nullptr; 2808 return parseAffineHighPrecOpExpr(expr, op, opLoc); 2809 } 2810 // No LLHS, get RHS 2811 return parseAffineHighPrecOpExpr(lhs, op, opLoc); 2812 } 2813 2814 // This is the last operand in this expression. 2815 if (llhs) 2816 return getAffineBinaryOpExpr(llhsOp, llhs, lhs, llhsOpLoc); 2817 2818 // No llhs, 'lhs' itself is the expression. 2819 return lhs; 2820 } 2821 2822 /// Parse an affine expression inside parentheses. 2823 /// 2824 /// affine-expr ::= `(` affine-expr `)` 2825 AffineExpr AffineParser::parseParentheticalExpr() { 2826 if (parseToken(Token::l_paren, "expected '('")) 2827 return nullptr; 2828 if (getToken().is(Token::r_paren)) 2829 return (emitError("no expression inside parentheses"), nullptr); 2830 2831 auto expr = parseAffineExpr(); 2832 if (!expr) 2833 return nullptr; 2834 if (parseToken(Token::r_paren, "expected ')'")) 2835 return nullptr; 2836 2837 return expr; 2838 } 2839 2840 /// Parse the negation expression. 2841 /// 2842 /// affine-expr ::= `-` affine-expr 2843 AffineExpr AffineParser::parseNegateExpression(AffineExpr lhs) { 2844 if (parseToken(Token::minus, "expected '-'")) 2845 return nullptr; 2846 2847 AffineExpr operand = parseAffineOperandExpr(lhs); 2848 // Since negation has the highest precedence of all ops (including high 2849 // precedence ops) but lower than parentheses, we are only going to use 2850 // parseAffineOperandExpr instead of parseAffineExpr here. 2851 if (!operand) 2852 // Extra error message although parseAffineOperandExpr would have 2853 // complained. Leads to a better diagnostic. 2854 return (emitError("missing operand of negation"), nullptr); 2855 return (-1) * operand; 2856 } 2857 2858 /// Parse a bare id that may appear in an affine expression. 2859 /// 2860 /// affine-expr ::= bare-id 2861 AffineExpr AffineParser::parseBareIdExpr() { 2862 if (getToken().isNot(Token::bare_identifier)) 2863 return (emitError("expected bare identifier"), nullptr); 2864 2865 StringRef sRef = getTokenSpelling(); 2866 for (auto entry : dimsAndSymbols) { 2867 if (entry.first == sRef) { 2868 consumeToken(Token::bare_identifier); 2869 return entry.second; 2870 } 2871 } 2872 2873 return (emitError("use of undeclared identifier"), nullptr); 2874 } 2875 2876 /// Parse an SSA id which may appear in an affine expression. 2877 AffineExpr AffineParser::parseSSAIdExpr(bool isSymbol) { 2878 if (!allowParsingSSAIds) 2879 return (emitError("unexpected ssa identifier"), nullptr); 2880 if (getToken().isNot(Token::percent_identifier)) 2881 return (emitError("expected ssa identifier"), nullptr); 2882 auto name = getTokenSpelling(); 2883 // Check if we already parsed this SSA id. 2884 for (auto entry : dimsAndSymbols) { 2885 if (entry.first == name) { 2886 consumeToken(Token::percent_identifier); 2887 return entry.second; 2888 } 2889 } 2890 // Parse the SSA id and add an AffineDim/SymbolExpr to represent it. 2891 if (parseElement(isSymbol)) 2892 return (emitError("failed to parse ssa identifier"), nullptr); 2893 auto idExpr = isSymbol 2894 ? getAffineSymbolExpr(numSymbolOperands++, getContext()) 2895 : getAffineDimExpr(numDimOperands++, getContext()); 2896 dimsAndSymbols.push_back({name, idExpr}); 2897 return idExpr; 2898 } 2899 2900 AffineExpr AffineParser::parseSymbolSSAIdExpr() { 2901 if (parseToken(Token::kw_symbol, "expected symbol keyword") || 2902 parseToken(Token::l_paren, "expected '(' at start of SSA symbol")) 2903 return nullptr; 2904 AffineExpr symbolExpr = parseSSAIdExpr(/*isSymbol=*/true); 2905 if (!symbolExpr) 2906 return nullptr; 2907 if (parseToken(Token::r_paren, "expected ')' at end of SSA symbol")) 2908 return nullptr; 2909 return symbolExpr; 2910 } 2911 2912 /// Parse a positive integral constant appearing in an affine expression. 2913 /// 2914 /// affine-expr ::= integer-literal 2915 AffineExpr AffineParser::parseIntegerExpr() { 2916 auto val = getToken().getUInt64IntegerValue(); 2917 if (!val.hasValue() || (int64_t)val.getValue() < 0) 2918 return (emitError("constant too large for index"), nullptr); 2919 2920 consumeToken(Token::integer); 2921 return builder.getAffineConstantExpr((int64_t)val.getValue()); 2922 } 2923 2924 /// Parses an expression that can be a valid operand of an affine expression. 2925 /// lhs: if non-null, lhs is an affine expression that is the lhs of a binary 2926 /// operator, the rhs of which is being parsed. This is used to determine 2927 /// whether an error should be emitted for a missing right operand. 2928 // Eg: for an expression without parentheses (like i + j + k + l), each 2929 // of the four identifiers is an operand. For i + j*k + l, j*k is not an 2930 // operand expression, it's an op expression and will be parsed via 2931 // parseAffineHighPrecOpExpression(). However, for i + (j*k) + -l, (j*k) and 2932 // -l are valid operands that will be parsed by this function. 2933 AffineExpr AffineParser::parseAffineOperandExpr(AffineExpr lhs) { 2934 switch (getToken().getKind()) { 2935 case Token::bare_identifier: 2936 return parseBareIdExpr(); 2937 case Token::kw_symbol: 2938 return parseSymbolSSAIdExpr(); 2939 case Token::percent_identifier: 2940 return parseSSAIdExpr(/*isSymbol=*/false); 2941 case Token::integer: 2942 return parseIntegerExpr(); 2943 case Token::l_paren: 2944 return parseParentheticalExpr(); 2945 case Token::minus: 2946 return parseNegateExpression(lhs); 2947 case Token::kw_ceildiv: 2948 case Token::kw_floordiv: 2949 case Token::kw_mod: 2950 case Token::plus: 2951 case Token::star: 2952 if (lhs) 2953 emitError("missing right operand of binary operator"); 2954 else 2955 emitError("missing left operand of binary operator"); 2956 return nullptr; 2957 default: 2958 if (lhs) 2959 emitError("missing right operand of binary operator"); 2960 else 2961 emitError("expected affine expression"); 2962 return nullptr; 2963 } 2964 } 2965 2966 /// Parse affine expressions that are bare-id's, integer constants, 2967 /// parenthetical affine expressions, and affine op expressions that are a 2968 /// composition of those. 2969 /// 2970 /// All binary op's associate from left to right. 2971 /// 2972 /// {add, sub} have lower precedence than {mul, div, and mod}. 2973 /// 2974 /// Add, sub'are themselves at the same precedence level. Mul, floordiv, 2975 /// ceildiv, and mod are at the same higher precedence level. Negation has 2976 /// higher precedence than any binary op. 2977 /// 2978 /// llhs: the affine expression appearing on the left of the one being parsed. 2979 /// This function will return ((llhs llhsOp lhs) op rhs) if llhs is non null, 2980 /// and lhs op rhs otherwise; if there is no rhs, llhs llhsOp lhs is returned 2981 /// if llhs is non-null; otherwise lhs is returned. This is to deal with left 2982 /// associativity. 2983 /// 2984 /// Eg: when the expression is e1 + e2*e3 + e4, with e1 as llhs, this function 2985 /// will return the affine expr equivalent of (e1 + (e2*e3)) + e4, where 2986 /// (e2*e3) will be parsed using parseAffineHighPrecOpExpr(). 2987 AffineExpr AffineParser::parseAffineLowPrecOpExpr(AffineExpr llhs, 2988 AffineLowPrecOp llhsOp) { 2989 AffineExpr lhs; 2990 if (!(lhs = parseAffineOperandExpr(llhs))) 2991 return nullptr; 2992 2993 // Found an LHS. Deal with the ops. 2994 if (AffineLowPrecOp lOp = consumeIfLowPrecOp()) { 2995 if (llhs) { 2996 AffineExpr sum = getAffineBinaryOpExpr(llhsOp, llhs, lhs); 2997 return parseAffineLowPrecOpExpr(sum, lOp); 2998 } 2999 // No LLHS, get RHS and form the expression. 3000 return parseAffineLowPrecOpExpr(lhs, lOp); 3001 } 3002 auto opLoc = getToken().getLoc(); 3003 if (AffineHighPrecOp hOp = consumeIfHighPrecOp()) { 3004 // We have a higher precedence op here. Get the rhs operand for the llhs 3005 // through parseAffineHighPrecOpExpr. 3006 AffineExpr highRes = parseAffineHighPrecOpExpr(lhs, hOp, opLoc); 3007 if (!highRes) 3008 return nullptr; 3009 3010 // If llhs is null, the product forms the first operand of the yet to be 3011 // found expression. If non-null, the op to associate with llhs is llhsOp. 3012 AffineExpr expr = 3013 llhs ? getAffineBinaryOpExpr(llhsOp, llhs, highRes) : highRes; 3014 3015 // Recurse for subsequent low prec op's after the affine high prec op 3016 // expression. 3017 if (AffineLowPrecOp nextOp = consumeIfLowPrecOp()) 3018 return parseAffineLowPrecOpExpr(expr, nextOp); 3019 return expr; 3020 } 3021 // Last operand in the expression list. 3022 if (llhs) 3023 return getAffineBinaryOpExpr(llhsOp, llhs, lhs); 3024 // No llhs, 'lhs' itself is the expression. 3025 return lhs; 3026 } 3027 3028 /// Parse an affine expression. 3029 /// affine-expr ::= `(` affine-expr `)` 3030 /// | `-` affine-expr 3031 /// | affine-expr `+` affine-expr 3032 /// | affine-expr `-` affine-expr 3033 /// | affine-expr `*` affine-expr 3034 /// | affine-expr `floordiv` affine-expr 3035 /// | affine-expr `ceildiv` affine-expr 3036 /// | affine-expr `mod` affine-expr 3037 /// | bare-id 3038 /// | integer-literal 3039 /// 3040 /// Additional conditions are checked depending on the production. For eg., 3041 /// one of the operands for `*` has to be either constant/symbolic; the second 3042 /// operand for floordiv, ceildiv, and mod has to be a positive integer. 3043 AffineExpr AffineParser::parseAffineExpr() { 3044 return parseAffineLowPrecOpExpr(nullptr, AffineLowPrecOp::LNoOp); 3045 } 3046 3047 /// Parse a dim or symbol from the lists appearing before the actual 3048 /// expressions of the affine map. Update our state to store the 3049 /// dimensional/symbolic identifier. 3050 ParseResult AffineParser::parseIdentifierDefinition(AffineExpr idExpr) { 3051 if (getToken().isNot(Token::bare_identifier)) 3052 return emitError("expected bare identifier"); 3053 3054 auto name = getTokenSpelling(); 3055 for (auto entry : dimsAndSymbols) { 3056 if (entry.first == name) 3057 return emitError("redefinition of identifier '" + name + "'"); 3058 } 3059 consumeToken(Token::bare_identifier); 3060 3061 dimsAndSymbols.push_back({name, idExpr}); 3062 return success(); 3063 } 3064 3065 /// Parse the list of dimensional identifiers to an affine map. 3066 ParseResult AffineParser::parseDimIdList(unsigned &numDims) { 3067 if (parseToken(Token::l_paren, 3068 "expected '(' at start of dimensional identifiers list")) { 3069 return failure(); 3070 } 3071 3072 auto parseElt = [&]() -> ParseResult { 3073 auto dimension = getAffineDimExpr(numDims++, getContext()); 3074 return parseIdentifierDefinition(dimension); 3075 }; 3076 return parseCommaSeparatedListUntil(Token::r_paren, parseElt); 3077 } 3078 3079 /// Parse the list of symbolic identifiers to an affine map. 3080 ParseResult AffineParser::parseSymbolIdList(unsigned &numSymbols) { 3081 consumeToken(Token::l_square); 3082 auto parseElt = [&]() -> ParseResult { 3083 auto symbol = getAffineSymbolExpr(numSymbols++, getContext()); 3084 return parseIdentifierDefinition(symbol); 3085 }; 3086 return parseCommaSeparatedListUntil(Token::r_square, parseElt); 3087 } 3088 3089 /// Parse the list of symbolic identifiers to an affine map. 3090 ParseResult 3091 AffineParser::parseDimAndOptionalSymbolIdList(unsigned &numDims, 3092 unsigned &numSymbols) { 3093 if (parseDimIdList(numDims)) { 3094 return failure(); 3095 } 3096 if (!getToken().is(Token::l_square)) { 3097 numSymbols = 0; 3098 return success(); 3099 } 3100 return parseSymbolIdList(numSymbols); 3101 } 3102 3103 /// Parses an ambiguous affine map or integer set definition inline. 3104 ParseResult AffineParser::parseAffineMapOrIntegerSetInline(AffineMap &map, 3105 IntegerSet &set) { 3106 unsigned numDims = 0, numSymbols = 0; 3107 3108 // List of dimensional and optional symbol identifiers. 3109 if (parseDimAndOptionalSymbolIdList(numDims, numSymbols)) { 3110 return failure(); 3111 } 3112 3113 // This is needed for parsing attributes as we wouldn't know whether we would 3114 // be parsing an integer set attribute or an affine map attribute. 3115 bool isArrow = getToken().is(Token::arrow); 3116 bool isColon = getToken().is(Token::colon); 3117 if (!isArrow && !isColon) { 3118 return emitError("expected '->' or ':'"); 3119 } else if (isArrow) { 3120 parseToken(Token::arrow, "expected '->' or '['"); 3121 map = parseAffineMapRange(numDims, numSymbols); 3122 return map ? success() : failure(); 3123 } else if (parseToken(Token::colon, "expected ':' or '['")) { 3124 return failure(); 3125 } 3126 3127 if ((set = parseIntegerSetConstraints(numDims, numSymbols))) 3128 return success(); 3129 3130 return failure(); 3131 } 3132 3133 /// Parse an AffineMap where the dim and symbol identifiers are SSA ids. 3134 ParseResult 3135 AffineParser::parseAffineMapOfSSAIds(AffineMap &map, 3136 OpAsmParser::Delimiter delimiter) { 3137 Token::Kind rightToken; 3138 switch (delimiter) { 3139 case OpAsmParser::Delimiter::Square: 3140 if (parseToken(Token::l_square, "expected '['")) 3141 return failure(); 3142 rightToken = Token::r_square; 3143 break; 3144 case OpAsmParser::Delimiter::Paren: 3145 if (parseToken(Token::l_paren, "expected '('")) 3146 return failure(); 3147 rightToken = Token::r_paren; 3148 break; 3149 default: 3150 return emitError("unexpected delimiter"); 3151 } 3152 3153 SmallVector<AffineExpr, 4> exprs; 3154 auto parseElt = [&]() -> ParseResult { 3155 auto elt = parseAffineExpr(); 3156 exprs.push_back(elt); 3157 return elt ? success() : failure(); 3158 }; 3159 3160 // Parse a multi-dimensional affine expression (a comma-separated list of 3161 // 1-d affine expressions); the list can be empty. Grammar: 3162 // multi-dim-affine-expr ::= `(` `)` 3163 // | `(` affine-expr (`,` affine-expr)* `)` 3164 if (parseCommaSeparatedListUntil(rightToken, parseElt, 3165 /*allowEmptyList=*/true)) 3166 return failure(); 3167 // Parsed a valid affine map. 3168 map = AffineMap::get(numDimOperands, dimsAndSymbols.size() - numDimOperands, 3169 exprs, getContext()); 3170 return success(); 3171 } 3172 3173 /// Parse the range and sizes affine map definition inline. 3174 /// 3175 /// affine-map ::= dim-and-symbol-id-lists `->` multi-dim-affine-expr 3176 /// 3177 /// multi-dim-affine-expr ::= `(` `)` 3178 /// multi-dim-affine-expr ::= `(` affine-expr (`,` affine-expr)* `)` 3179 AffineMap AffineParser::parseAffineMapRange(unsigned numDims, 3180 unsigned numSymbols) { 3181 parseToken(Token::l_paren, "expected '(' at start of affine map range"); 3182 3183 SmallVector<AffineExpr, 4> exprs; 3184 auto parseElt = [&]() -> ParseResult { 3185 auto elt = parseAffineExpr(); 3186 ParseResult res = elt ? success() : failure(); 3187 exprs.push_back(elt); 3188 return res; 3189 }; 3190 3191 // Parse a multi-dimensional affine expression (a comma-separated list of 3192 // 1-d affine expressions). Grammar: 3193 // multi-dim-affine-expr ::= `(` `)` 3194 // | `(` affine-expr (`,` affine-expr)* `)` 3195 if (parseCommaSeparatedListUntil(Token::r_paren, parseElt, true)) 3196 return AffineMap(); 3197 3198 // Parsed a valid affine map. 3199 return AffineMap::get(numDims, numSymbols, exprs, getContext()); 3200 } 3201 3202 /// Parse an affine constraint. 3203 /// affine-constraint ::= affine-expr `>=` `0` 3204 /// | affine-expr `==` `0` 3205 /// 3206 /// isEq is set to true if the parsed constraint is an equality, false if it 3207 /// is an inequality (greater than or equal). 3208 /// 3209 AffineExpr AffineParser::parseAffineConstraint(bool *isEq) { 3210 AffineExpr expr = parseAffineExpr(); 3211 if (!expr) 3212 return nullptr; 3213 3214 if (consumeIf(Token::greater) && consumeIf(Token::equal) && 3215 getToken().is(Token::integer)) { 3216 auto dim = getToken().getUnsignedIntegerValue(); 3217 if (dim.hasValue() && dim.getValue() == 0) { 3218 consumeToken(Token::integer); 3219 *isEq = false; 3220 return expr; 3221 } 3222 return (emitError("expected '0' after '>='"), nullptr); 3223 } 3224 3225 if (consumeIf(Token::equal) && consumeIf(Token::equal) && 3226 getToken().is(Token::integer)) { 3227 auto dim = getToken().getUnsignedIntegerValue(); 3228 if (dim.hasValue() && dim.getValue() == 0) { 3229 consumeToken(Token::integer); 3230 *isEq = true; 3231 return expr; 3232 } 3233 return (emitError("expected '0' after '=='"), nullptr); 3234 } 3235 3236 return (emitError("expected '== 0' or '>= 0' at end of affine constraint"), 3237 nullptr); 3238 } 3239 3240 /// Parse the constraints that are part of an integer set definition. 3241 /// integer-set-inline 3242 /// ::= dim-and-symbol-id-lists `:` 3243 /// '(' affine-constraint-conjunction? ')' 3244 /// affine-constraint-conjunction ::= affine-constraint (`,` 3245 /// affine-constraint)* 3246 /// 3247 IntegerSet AffineParser::parseIntegerSetConstraints(unsigned numDims, 3248 unsigned numSymbols) { 3249 if (parseToken(Token::l_paren, 3250 "expected '(' at start of integer set constraint list")) 3251 return IntegerSet(); 3252 3253 SmallVector<AffineExpr, 4> constraints; 3254 SmallVector<bool, 4> isEqs; 3255 auto parseElt = [&]() -> ParseResult { 3256 bool isEq; 3257 auto elt = parseAffineConstraint(&isEq); 3258 ParseResult res = elt ? success() : failure(); 3259 if (elt) { 3260 constraints.push_back(elt); 3261 isEqs.push_back(isEq); 3262 } 3263 return res; 3264 }; 3265 3266 // Parse a list of affine constraints (comma-separated). 3267 if (parseCommaSeparatedListUntil(Token::r_paren, parseElt, true)) 3268 return IntegerSet(); 3269 3270 // If no constraints were parsed, then treat this as a degenerate 'true' case. 3271 if (constraints.empty()) { 3272 /* 0 == 0 */ 3273 auto zero = getAffineConstantExpr(0, getContext()); 3274 return IntegerSet::get(numDims, numSymbols, zero, true); 3275 } 3276 3277 // Parsed a valid integer set. 3278 return IntegerSet::get(numDims, numSymbols, constraints, isEqs); 3279 } 3280 3281 /// Parse an ambiguous reference to either and affine map or an integer set. 3282 ParseResult Parser::parseAffineMapOrIntegerSetReference(AffineMap &map, 3283 IntegerSet &set) { 3284 return AffineParser(state).parseAffineMapOrIntegerSetInline(map, set); 3285 } 3286 ParseResult Parser::parseAffineMapReference(AffineMap &map) { 3287 llvm::SMLoc curLoc = getToken().getLoc(); 3288 IntegerSet set; 3289 if (parseAffineMapOrIntegerSetReference(map, set)) 3290 return failure(); 3291 if (set) 3292 return emitError(curLoc, "expected AffineMap, but got IntegerSet"); 3293 return success(); 3294 } 3295 ParseResult Parser::parseIntegerSetReference(IntegerSet &set) { 3296 llvm::SMLoc curLoc = getToken().getLoc(); 3297 AffineMap map; 3298 if (parseAffineMapOrIntegerSetReference(map, set)) 3299 return failure(); 3300 if (map) 3301 return emitError(curLoc, "expected IntegerSet, but got AffineMap"); 3302 return success(); 3303 } 3304 3305 /// Parse an AffineMap of SSA ids. The callback 'parseElement' is used to 3306 /// parse SSA value uses encountered while parsing affine expressions. 3307 ParseResult 3308 Parser::parseAffineMapOfSSAIds(AffineMap &map, 3309 function_ref<ParseResult(bool)> parseElement, 3310 OpAsmParser::Delimiter delimiter) { 3311 return AffineParser(state, /*allowParsingSSAIds=*/true, parseElement) 3312 .parseAffineMapOfSSAIds(map, delimiter); 3313 } 3314 3315 //===----------------------------------------------------------------------===// 3316 // OperationParser 3317 //===----------------------------------------------------------------------===// 3318 3319 namespace { 3320 /// This class provides support for parsing operations and regions of 3321 /// operations. 3322 class OperationParser : public Parser { 3323 public: 3324 OperationParser(ParserState &state, ModuleOp moduleOp) 3325 : Parser(state), opBuilder(moduleOp.getBodyRegion()), moduleOp(moduleOp) { 3326 } 3327 3328 ~OperationParser(); 3329 3330 /// After parsing is finished, this function must be called to see if there 3331 /// are any remaining issues. 3332 ParseResult finalize(); 3333 3334 //===--------------------------------------------------------------------===// 3335 // SSA Value Handling 3336 //===--------------------------------------------------------------------===// 3337 3338 /// This represents a use of an SSA value in the program. The first two 3339 /// entries in the tuple are the name and result number of a reference. The 3340 /// third is the location of the reference, which is used in case this ends 3341 /// up being a use of an undefined value. 3342 struct SSAUseInfo { 3343 StringRef name; // Value name, e.g. %42 or %abc 3344 unsigned number; // Number, specified with #12 3345 SMLoc loc; // Location of first definition or use. 3346 }; 3347 3348 /// Push a new SSA name scope to the parser. 3349 void pushSSANameScope(bool isIsolated); 3350 3351 /// Pop the last SSA name scope from the parser. 3352 ParseResult popSSANameScope(); 3353 3354 /// Register a definition of a value with the symbol table. 3355 ParseResult addDefinition(SSAUseInfo useInfo, Value value); 3356 3357 /// Parse an optional list of SSA uses into 'results'. 3358 ParseResult parseOptionalSSAUseList(SmallVectorImpl<SSAUseInfo> &results); 3359 3360 /// Parse a single SSA use into 'result'. 3361 ParseResult parseSSAUse(SSAUseInfo &result); 3362 3363 /// Given a reference to an SSA value and its type, return a reference. This 3364 /// returns null on failure. 3365 Value resolveSSAUse(SSAUseInfo useInfo, Type type); 3366 3367 ParseResult parseSSADefOrUseAndType( 3368 const std::function<ParseResult(SSAUseInfo, Type)> &action); 3369 3370 ParseResult parseOptionalSSAUseAndTypeList(SmallVectorImpl<Value> &results); 3371 3372 /// Return the location of the value identified by its name and number if it 3373 /// has been already reference. 3374 Optional<SMLoc> getReferenceLoc(StringRef name, unsigned number) { 3375 auto &values = isolatedNameScopes.back().values; 3376 if (!values.count(name) || number >= values[name].size()) 3377 return {}; 3378 if (values[name][number].first) 3379 return values[name][number].second; 3380 return {}; 3381 } 3382 3383 //===--------------------------------------------------------------------===// 3384 // Operation Parsing 3385 //===--------------------------------------------------------------------===// 3386 3387 /// Parse an operation instance. 3388 ParseResult parseOperation(); 3389 3390 /// Parse a single operation successor. 3391 ParseResult parseSuccessor(Block *&dest); 3392 3393 /// Parse a comma-separated list of operation successors in brackets. 3394 ParseResult parseSuccessors(SmallVectorImpl<Block *> &destinations); 3395 3396 /// Parse an operation instance that is in the generic form. 3397 Operation *parseGenericOperation(); 3398 3399 /// Parse an operation instance that is in the generic form and insert it at 3400 /// the provided insertion point. 3401 Operation *parseGenericOperation(Block *insertBlock, 3402 Block::iterator insertPt); 3403 3404 /// This is the structure of a result specifier in the assembly syntax, 3405 /// including the name, number of results, and location. 3406 typedef std::tuple<StringRef, unsigned, SMLoc> ResultRecord; 3407 3408 /// Parse an operation instance that is in the op-defined custom form. 3409 /// resultInfo specifies information about the "%name =" specifiers. 3410 Operation *parseCustomOperation(ArrayRef<ResultRecord> resultInfo); 3411 3412 //===--------------------------------------------------------------------===// 3413 // Region Parsing 3414 //===--------------------------------------------------------------------===// 3415 3416 /// Parse a region into 'region' with the provided entry block arguments. 3417 /// 'isIsolatedNameScope' indicates if the naming scope of this region is 3418 /// isolated from those above. 3419 ParseResult parseRegion(Region ®ion, 3420 ArrayRef<std::pair<SSAUseInfo, Type>> entryArguments, 3421 bool isIsolatedNameScope = false); 3422 3423 /// Parse a region body into 'region'. 3424 ParseResult parseRegionBody(Region ®ion); 3425 3426 //===--------------------------------------------------------------------===// 3427 // Block Parsing 3428 //===--------------------------------------------------------------------===// 3429 3430 /// Parse a new block into 'block'. 3431 ParseResult parseBlock(Block *&block); 3432 3433 /// Parse a list of operations into 'block'. 3434 ParseResult parseBlockBody(Block *block); 3435 3436 /// Parse a (possibly empty) list of block arguments. 3437 ParseResult parseOptionalBlockArgList(SmallVectorImpl<BlockArgument> &results, 3438 Block *owner); 3439 3440 /// Get the block with the specified name, creating it if it doesn't 3441 /// already exist. The location specified is the point of use, which allows 3442 /// us to diagnose references to blocks that are not defined precisely. 3443 Block *getBlockNamed(StringRef name, SMLoc loc); 3444 3445 /// Define the block with the specified name. Returns the Block* or nullptr in 3446 /// the case of redefinition. 3447 Block *defineBlockNamed(StringRef name, SMLoc loc, Block *existing); 3448 3449 private: 3450 /// Returns the info for a block at the current scope for the given name. 3451 std::pair<Block *, SMLoc> &getBlockInfoByName(StringRef name) { 3452 return blocksByName.back()[name]; 3453 } 3454 3455 /// Insert a new forward reference to the given block. 3456 void insertForwardRef(Block *block, SMLoc loc) { 3457 forwardRef.back().try_emplace(block, loc); 3458 } 3459 3460 /// Erase any forward reference to the given block. 3461 bool eraseForwardRef(Block *block) { return forwardRef.back().erase(block); } 3462 3463 /// Record that a definition was added at the current scope. 3464 void recordDefinition(StringRef def); 3465 3466 /// Get the value entry for the given SSA name. 3467 SmallVectorImpl<std::pair<Value, SMLoc>> &getSSAValueEntry(StringRef name); 3468 3469 /// Create a forward reference placeholder value with the given location and 3470 /// result type. 3471 Value createForwardRefPlaceholder(SMLoc loc, Type type); 3472 3473 /// Return true if this is a forward reference. 3474 bool isForwardRefPlaceholder(Value value) { 3475 return forwardRefPlaceholders.count(value); 3476 } 3477 3478 /// This struct represents an isolated SSA name scope. This scope may contain 3479 /// other nested non-isolated scopes. These scopes are used for operations 3480 /// that are known to be isolated to allow for reusing names within their 3481 /// regions, even if those names are used above. 3482 struct IsolatedSSANameScope { 3483 /// Record that a definition was added at the current scope. 3484 void recordDefinition(StringRef def) { 3485 definitionsPerScope.back().insert(def); 3486 } 3487 3488 /// Push a nested name scope. 3489 void pushSSANameScope() { definitionsPerScope.push_back({}); } 3490 3491 /// Pop a nested name scope. 3492 void popSSANameScope() { 3493 for (auto &def : definitionsPerScope.pop_back_val()) 3494 values.erase(def.getKey()); 3495 } 3496 3497 /// This keeps track of all of the SSA values we are tracking for each name 3498 /// scope, indexed by their name. This has one entry per result number. 3499 llvm::StringMap<SmallVector<std::pair<Value, SMLoc>, 1>> values; 3500 3501 /// This keeps track of all of the values defined by a specific name scope. 3502 SmallVector<llvm::StringSet<>, 2> definitionsPerScope; 3503 }; 3504 3505 /// A list of isolated name scopes. 3506 SmallVector<IsolatedSSANameScope, 2> isolatedNameScopes; 3507 3508 /// This keeps track of the block names as well as the location of the first 3509 /// reference for each nested name scope. This is used to diagnose invalid 3510 /// block references and memorize them. 3511 SmallVector<DenseMap<StringRef, std::pair<Block *, SMLoc>>, 2> blocksByName; 3512 SmallVector<DenseMap<Block *, SMLoc>, 2> forwardRef; 3513 3514 /// These are all of the placeholders we've made along with the location of 3515 /// their first reference, to allow checking for use of undefined values. 3516 DenseMap<Value, SMLoc> forwardRefPlaceholders; 3517 3518 /// The builder used when creating parsed operation instances. 3519 OpBuilder opBuilder; 3520 3521 /// The top level module operation. 3522 ModuleOp moduleOp; 3523 }; 3524 } // end anonymous namespace 3525 3526 OperationParser::~OperationParser() { 3527 for (auto &fwd : forwardRefPlaceholders) { 3528 // Drop all uses of undefined forward declared reference and destroy 3529 // defining operation. 3530 fwd.first.dropAllUses(); 3531 fwd.first.getDefiningOp()->destroy(); 3532 } 3533 } 3534 3535 /// After parsing is finished, this function must be called to see if there are 3536 /// any remaining issues. 3537 ParseResult OperationParser::finalize() { 3538 // Check for any forward references that are left. If we find any, error 3539 // out. 3540 if (!forwardRefPlaceholders.empty()) { 3541 SmallVector<std::pair<const char *, Value>, 4> errors; 3542 // Iteration over the map isn't deterministic, so sort by source location. 3543 for (auto entry : forwardRefPlaceholders) 3544 errors.push_back({entry.second.getPointer(), entry.first}); 3545 llvm::array_pod_sort(errors.begin(), errors.end()); 3546 3547 for (auto entry : errors) { 3548 auto loc = SMLoc::getFromPointer(entry.first); 3549 emitError(loc, "use of undeclared SSA value name"); 3550 } 3551 return failure(); 3552 } 3553 3554 return success(); 3555 } 3556 3557 //===----------------------------------------------------------------------===// 3558 // SSA Value Handling 3559 //===----------------------------------------------------------------------===// 3560 3561 void OperationParser::pushSSANameScope(bool isIsolated) { 3562 blocksByName.push_back(DenseMap<StringRef, std::pair<Block *, SMLoc>>()); 3563 forwardRef.push_back(DenseMap<Block *, SMLoc>()); 3564 3565 // Push back a new name definition scope. 3566 if (isIsolated) 3567 isolatedNameScopes.push_back({}); 3568 isolatedNameScopes.back().pushSSANameScope(); 3569 } 3570 3571 ParseResult OperationParser::popSSANameScope() { 3572 auto forwardRefInCurrentScope = forwardRef.pop_back_val(); 3573 3574 // Verify that all referenced blocks were defined. 3575 if (!forwardRefInCurrentScope.empty()) { 3576 SmallVector<std::pair<const char *, Block *>, 4> errors; 3577 // Iteration over the map isn't deterministic, so sort by source location. 3578 for (auto entry : forwardRefInCurrentScope) { 3579 errors.push_back({entry.second.getPointer(), entry.first}); 3580 // Add this block to the top-level region to allow for automatic cleanup. 3581 moduleOp.getOperation()->getRegion(0).push_back(entry.first); 3582 } 3583 llvm::array_pod_sort(errors.begin(), errors.end()); 3584 3585 for (auto entry : errors) { 3586 auto loc = SMLoc::getFromPointer(entry.first); 3587 emitError(loc, "reference to an undefined block"); 3588 } 3589 return failure(); 3590 } 3591 3592 // Pop the next nested namescope. If there is only one internal namescope, 3593 // just pop the isolated scope. 3594 auto ¤tNameScope = isolatedNameScopes.back(); 3595 if (currentNameScope.definitionsPerScope.size() == 1) 3596 isolatedNameScopes.pop_back(); 3597 else 3598 currentNameScope.popSSANameScope(); 3599 3600 blocksByName.pop_back(); 3601 return success(); 3602 } 3603 3604 /// Register a definition of a value with the symbol table. 3605 ParseResult OperationParser::addDefinition(SSAUseInfo useInfo, Value value) { 3606 auto &entries = getSSAValueEntry(useInfo.name); 3607 3608 // Make sure there is a slot for this value. 3609 if (entries.size() <= useInfo.number) 3610 entries.resize(useInfo.number + 1); 3611 3612 // If we already have an entry for this, check to see if it was a definition 3613 // or a forward reference. 3614 if (auto existing = entries[useInfo.number].first) { 3615 if (!isForwardRefPlaceholder(existing)) { 3616 return emitError(useInfo.loc) 3617 .append("redefinition of SSA value '", useInfo.name, "'") 3618 .attachNote(getEncodedSourceLocation(entries[useInfo.number].second)) 3619 .append("previously defined here"); 3620 } 3621 3622 // If it was a forward reference, update everything that used it to use 3623 // the actual definition instead, delete the forward ref, and remove it 3624 // from our set of forward references we track. 3625 existing.replaceAllUsesWith(value); 3626 existing.getDefiningOp()->destroy(); 3627 forwardRefPlaceholders.erase(existing); 3628 } 3629 3630 /// Record this definition for the current scope. 3631 entries[useInfo.number] = {value, useInfo.loc}; 3632 recordDefinition(useInfo.name); 3633 return success(); 3634 } 3635 3636 /// Parse a (possibly empty) list of SSA operands. 3637 /// 3638 /// ssa-use-list ::= ssa-use (`,` ssa-use)* 3639 /// ssa-use-list-opt ::= ssa-use-list? 3640 /// 3641 ParseResult 3642 OperationParser::parseOptionalSSAUseList(SmallVectorImpl<SSAUseInfo> &results) { 3643 if (getToken().isNot(Token::percent_identifier)) 3644 return success(); 3645 return parseCommaSeparatedList([&]() -> ParseResult { 3646 SSAUseInfo result; 3647 if (parseSSAUse(result)) 3648 return failure(); 3649 results.push_back(result); 3650 return success(); 3651 }); 3652 } 3653 3654 /// Parse a SSA operand for an operation. 3655 /// 3656 /// ssa-use ::= ssa-id 3657 /// 3658 ParseResult OperationParser::parseSSAUse(SSAUseInfo &result) { 3659 result.name = getTokenSpelling(); 3660 result.number = 0; 3661 result.loc = getToken().getLoc(); 3662 if (parseToken(Token::percent_identifier, "expected SSA operand")) 3663 return failure(); 3664 3665 // If we have an attribute ID, it is a result number. 3666 if (getToken().is(Token::hash_identifier)) { 3667 if (auto value = getToken().getHashIdentifierNumber()) 3668 result.number = value.getValue(); 3669 else 3670 return emitError("invalid SSA value result number"); 3671 consumeToken(Token::hash_identifier); 3672 } 3673 3674 return success(); 3675 } 3676 3677 /// Given an unbound reference to an SSA value and its type, return the value 3678 /// it specifies. This returns null on failure. 3679 Value OperationParser::resolveSSAUse(SSAUseInfo useInfo, Type type) { 3680 auto &entries = getSSAValueEntry(useInfo.name); 3681 3682 // If we have already seen a value of this name, return it. 3683 if (useInfo.number < entries.size() && entries[useInfo.number].first) { 3684 auto result = entries[useInfo.number].first; 3685 // Check that the type matches the other uses. 3686 if (result.getType() == type) 3687 return result; 3688 3689 emitError(useInfo.loc, "use of value '") 3690 .append(useInfo.name, 3691 "' expects different type than prior uses: ", type, " vs ", 3692 result.getType()) 3693 .attachNote(getEncodedSourceLocation(entries[useInfo.number].second)) 3694 .append("prior use here"); 3695 return nullptr; 3696 } 3697 3698 // Make sure we have enough slots for this. 3699 if (entries.size() <= useInfo.number) 3700 entries.resize(useInfo.number + 1); 3701 3702 // If the value has already been defined and this is an overly large result 3703 // number, diagnose that. 3704 if (entries[0].first && !isForwardRefPlaceholder(entries[0].first)) 3705 return (emitError(useInfo.loc, "reference to invalid result number"), 3706 nullptr); 3707 3708 // Otherwise, this is a forward reference. Create a placeholder and remember 3709 // that we did so. 3710 auto result = createForwardRefPlaceholder(useInfo.loc, type); 3711 entries[useInfo.number].first = result; 3712 entries[useInfo.number].second = useInfo.loc; 3713 return result; 3714 } 3715 3716 /// Parse an SSA use with an associated type. 3717 /// 3718 /// ssa-use-and-type ::= ssa-use `:` type 3719 ParseResult OperationParser::parseSSADefOrUseAndType( 3720 const std::function<ParseResult(SSAUseInfo, Type)> &action) { 3721 SSAUseInfo useInfo; 3722 if (parseSSAUse(useInfo) || 3723 parseToken(Token::colon, "expected ':' and type for SSA operand")) 3724 return failure(); 3725 3726 auto type = parseType(); 3727 if (!type) 3728 return failure(); 3729 3730 return action(useInfo, type); 3731 } 3732 3733 /// Parse a (possibly empty) list of SSA operands, followed by a colon, then 3734 /// followed by a type list. 3735 /// 3736 /// ssa-use-and-type-list 3737 /// ::= ssa-use-list ':' type-list-no-parens 3738 /// 3739 ParseResult OperationParser::parseOptionalSSAUseAndTypeList( 3740 SmallVectorImpl<Value> &results) { 3741 SmallVector<SSAUseInfo, 4> valueIDs; 3742 if (parseOptionalSSAUseList(valueIDs)) 3743 return failure(); 3744 3745 // If there were no operands, then there is no colon or type lists. 3746 if (valueIDs.empty()) 3747 return success(); 3748 3749 SmallVector<Type, 4> types; 3750 if (parseToken(Token::colon, "expected ':' in operand list") || 3751 parseTypeListNoParens(types)) 3752 return failure(); 3753 3754 if (valueIDs.size() != types.size()) 3755 return emitError("expected ") 3756 << valueIDs.size() << " types to match operand list"; 3757 3758 results.reserve(valueIDs.size()); 3759 for (unsigned i = 0, e = valueIDs.size(); i != e; ++i) { 3760 if (auto value = resolveSSAUse(valueIDs[i], types[i])) 3761 results.push_back(value); 3762 else 3763 return failure(); 3764 } 3765 3766 return success(); 3767 } 3768 3769 /// Record that a definition was added at the current scope. 3770 void OperationParser::recordDefinition(StringRef def) { 3771 isolatedNameScopes.back().recordDefinition(def); 3772 } 3773 3774 /// Get the value entry for the given SSA name. 3775 SmallVectorImpl<std::pair<Value, SMLoc>> & 3776 OperationParser::getSSAValueEntry(StringRef name) { 3777 return isolatedNameScopes.back().values[name]; 3778 } 3779 3780 /// Create and remember a new placeholder for a forward reference. 3781 Value OperationParser::createForwardRefPlaceholder(SMLoc loc, Type type) { 3782 // Forward references are always created as operations, because we just need 3783 // something with a def/use chain. 3784 // 3785 // We create these placeholders as having an empty name, which we know 3786 // cannot be created through normal user input, allowing us to distinguish 3787 // them. 3788 auto name = OperationName("placeholder", getContext()); 3789 auto *op = Operation::create( 3790 getEncodedSourceLocation(loc), name, type, /*operands=*/{}, 3791 /*attributes=*/llvm::None, /*successors=*/{}, /*numRegions=*/0); 3792 forwardRefPlaceholders[op->getResult(0)] = loc; 3793 return op->getResult(0); 3794 } 3795 3796 //===----------------------------------------------------------------------===// 3797 // Operation Parsing 3798 //===----------------------------------------------------------------------===// 3799 3800 /// Parse an operation. 3801 /// 3802 /// operation ::= op-result-list? 3803 /// (generic-operation | custom-operation) 3804 /// trailing-location? 3805 /// generic-operation ::= string-literal `(` ssa-use-list? `)` 3806 /// successor-list? (`(` region-list `)`)? 3807 /// attribute-dict? `:` function-type 3808 /// custom-operation ::= bare-id custom-operation-format 3809 /// op-result-list ::= op-result (`,` op-result)* `=` 3810 /// op-result ::= ssa-id (`:` integer-literal) 3811 /// 3812 ParseResult OperationParser::parseOperation() { 3813 auto loc = getToken().getLoc(); 3814 SmallVector<ResultRecord, 1> resultIDs; 3815 size_t numExpectedResults = 0; 3816 if (getToken().is(Token::percent_identifier)) { 3817 // Parse the group of result ids. 3818 auto parseNextResult = [&]() -> ParseResult { 3819 // Parse the next result id. 3820 if (!getToken().is(Token::percent_identifier)) 3821 return emitError("expected valid ssa identifier"); 3822 3823 Token nameTok = getToken(); 3824 consumeToken(Token::percent_identifier); 3825 3826 // If the next token is a ':', we parse the expected result count. 3827 size_t expectedSubResults = 1; 3828 if (consumeIf(Token::colon)) { 3829 // Check that the next token is an integer. 3830 if (!getToken().is(Token::integer)) 3831 return emitError("expected integer number of results"); 3832 3833 // Check that number of results is > 0. 3834 auto val = getToken().getUInt64IntegerValue(); 3835 if (!val.hasValue() || val.getValue() < 1) 3836 return emitError("expected named operation to have atleast 1 result"); 3837 consumeToken(Token::integer); 3838 expectedSubResults = *val; 3839 } 3840 3841 resultIDs.emplace_back(nameTok.getSpelling(), expectedSubResults, 3842 nameTok.getLoc()); 3843 numExpectedResults += expectedSubResults; 3844 return success(); 3845 }; 3846 if (parseCommaSeparatedList(parseNextResult)) 3847 return failure(); 3848 3849 if (parseToken(Token::equal, "expected '=' after SSA name")) 3850 return failure(); 3851 } 3852 3853 Operation *op; 3854 if (getToken().is(Token::bare_identifier) || getToken().isKeyword()) 3855 op = parseCustomOperation(resultIDs); 3856 else if (getToken().is(Token::string)) 3857 op = parseGenericOperation(); 3858 else 3859 return emitError("expected operation name in quotes"); 3860 3861 // If parsing of the basic operation failed, then this whole thing fails. 3862 if (!op) 3863 return failure(); 3864 3865 // If the operation had a name, register it. 3866 if (!resultIDs.empty()) { 3867 if (op->getNumResults() == 0) 3868 return emitError(loc, "cannot name an operation with no results"); 3869 if (numExpectedResults != op->getNumResults()) 3870 return emitError(loc, "operation defines ") 3871 << op->getNumResults() << " results but was provided " 3872 << numExpectedResults << " to bind"; 3873 3874 // Add definitions for each of the result groups. 3875 unsigned opResI = 0; 3876 for (ResultRecord &resIt : resultIDs) { 3877 for (unsigned subRes : llvm::seq<unsigned>(0, std::get<1>(resIt))) { 3878 if (addDefinition({std::get<0>(resIt), subRes, std::get<2>(resIt)}, 3879 op->getResult(opResI++))) 3880 return failure(); 3881 } 3882 } 3883 } 3884 3885 return success(); 3886 } 3887 3888 /// Parse a single operation successor. 3889 /// 3890 /// successor ::= block-id 3891 /// 3892 ParseResult OperationParser::parseSuccessor(Block *&dest) { 3893 // Verify branch is identifier and get the matching block. 3894 if (!getToken().is(Token::caret_identifier)) 3895 return emitError("expected block name"); 3896 dest = getBlockNamed(getTokenSpelling(), getToken().getLoc()); 3897 consumeToken(); 3898 return success(); 3899 } 3900 3901 /// Parse a comma-separated list of operation successors in brackets. 3902 /// 3903 /// successor-list ::= `[` successor (`,` successor )* `]` 3904 /// 3905 ParseResult 3906 OperationParser::parseSuccessors(SmallVectorImpl<Block *> &destinations) { 3907 if (parseToken(Token::l_square, "expected '['")) 3908 return failure(); 3909 3910 auto parseElt = [this, &destinations] { 3911 Block *dest; 3912 ParseResult res = parseSuccessor(dest); 3913 destinations.push_back(dest); 3914 return res; 3915 }; 3916 return parseCommaSeparatedListUntil(Token::r_square, parseElt, 3917 /*allowEmptyList=*/false); 3918 } 3919 3920 namespace { 3921 // RAII-style guard for cleaning up the regions in the operation state before 3922 // deleting them. Within the parser, regions may get deleted if parsing failed, 3923 // and other errors may be present, in particular undominated uses. This makes 3924 // sure such uses are deleted. 3925 struct CleanupOpStateRegions { 3926 ~CleanupOpStateRegions() { 3927 SmallVector<Region *, 4> regionsToClean; 3928 regionsToClean.reserve(state.regions.size()); 3929 for (auto ®ion : state.regions) 3930 if (region) 3931 for (auto &block : *region) 3932 block.dropAllDefinedValueUses(); 3933 } 3934 OperationState &state; 3935 }; 3936 } // namespace 3937 3938 Operation *OperationParser::parseGenericOperation() { 3939 // Get location information for the operation. 3940 auto srcLocation = getEncodedSourceLocation(getToken().getLoc()); 3941 3942 auto name = getToken().getStringValue(); 3943 if (name.empty()) 3944 return (emitError("empty operation name is invalid"), nullptr); 3945 if (name.find('\0') != StringRef::npos) 3946 return (emitError("null character not allowed in operation name"), nullptr); 3947 3948 consumeToken(Token::string); 3949 3950 OperationState result(srcLocation, name); 3951 3952 // Parse the operand list. 3953 SmallVector<SSAUseInfo, 8> operandInfos; 3954 if (parseToken(Token::l_paren, "expected '(' to start operand list") || 3955 parseOptionalSSAUseList(operandInfos) || 3956 parseToken(Token::r_paren, "expected ')' to end operand list")) { 3957 return nullptr; 3958 } 3959 3960 // Parse the successor list. 3961 if (getToken().is(Token::l_square)) { 3962 // Check if the operation is a known terminator. 3963 const AbstractOperation *abstractOp = result.name.getAbstractOperation(); 3964 if (abstractOp && !abstractOp->hasProperty(OperationProperty::Terminator)) 3965 return emitError("successors in non-terminator"), nullptr; 3966 3967 SmallVector<Block *, 2> successors; 3968 if (parseSuccessors(successors)) 3969 return nullptr; 3970 result.addSuccessors(successors); 3971 } 3972 3973 // Parse the region list. 3974 CleanupOpStateRegions guard{result}; 3975 if (consumeIf(Token::l_paren)) { 3976 do { 3977 // Create temporary regions with the top level region as parent. 3978 result.regions.emplace_back(new Region(moduleOp)); 3979 if (parseRegion(*result.regions.back(), /*entryArguments=*/{})) 3980 return nullptr; 3981 } while (consumeIf(Token::comma)); 3982 if (parseToken(Token::r_paren, "expected ')' to end region list")) 3983 return nullptr; 3984 } 3985 3986 if (getToken().is(Token::l_brace)) { 3987 if (parseAttributeDict(result.attributes)) 3988 return nullptr; 3989 } 3990 3991 if (parseToken(Token::colon, "expected ':' followed by operation type")) 3992 return nullptr; 3993 3994 auto typeLoc = getToken().getLoc(); 3995 auto type = parseType(); 3996 if (!type) 3997 return nullptr; 3998 auto fnType = type.dyn_cast<FunctionType>(); 3999 if (!fnType) 4000 return (emitError(typeLoc, "expected function type"), nullptr); 4001 4002 result.addTypes(fnType.getResults()); 4003 4004 // Check that we have the right number of types for the operands. 4005 auto operandTypes = fnType.getInputs(); 4006 if (operandTypes.size() != operandInfos.size()) { 4007 auto plural = "s"[operandInfos.size() == 1]; 4008 return (emitError(typeLoc, "expected ") 4009 << operandInfos.size() << " operand type" << plural 4010 << " but had " << operandTypes.size(), 4011 nullptr); 4012 } 4013 4014 // Resolve all of the operands. 4015 for (unsigned i = 0, e = operandInfos.size(); i != e; ++i) { 4016 result.operands.push_back(resolveSSAUse(operandInfos[i], operandTypes[i])); 4017 if (!result.operands.back()) 4018 return nullptr; 4019 } 4020 4021 // Parse a location if one is present. 4022 if (parseOptionalTrailingLocation(result.location)) 4023 return nullptr; 4024 4025 return opBuilder.createOperation(result); 4026 } 4027 4028 Operation *OperationParser::parseGenericOperation(Block *insertBlock, 4029 Block::iterator insertPt) { 4030 OpBuilder::InsertionGuard restoreInsertionPoint(opBuilder); 4031 opBuilder.setInsertionPoint(insertBlock, insertPt); 4032 return parseGenericOperation(); 4033 } 4034 4035 namespace { 4036 class CustomOpAsmParser : public OpAsmParser { 4037 public: 4038 CustomOpAsmParser(SMLoc nameLoc, 4039 ArrayRef<OperationParser::ResultRecord> resultIDs, 4040 const AbstractOperation *opDefinition, 4041 OperationParser &parser) 4042 : nameLoc(nameLoc), resultIDs(resultIDs), opDefinition(opDefinition), 4043 parser(parser) {} 4044 4045 /// Parse an instance of the operation described by 'opDefinition' into the 4046 /// provided operation state. 4047 ParseResult parseOperation(OperationState &opState) { 4048 if (opDefinition->parseAssembly(*this, opState)) 4049 return failure(); 4050 return success(); 4051 } 4052 4053 Operation *parseGenericOperation(Block *insertBlock, 4054 Block::iterator insertPt) final { 4055 return parser.parseGenericOperation(insertBlock, insertPt); 4056 } 4057 4058 //===--------------------------------------------------------------------===// 4059 // Utilities 4060 //===--------------------------------------------------------------------===// 4061 4062 /// Return if any errors were emitted during parsing. 4063 bool didEmitError() const { return emittedError; } 4064 4065 /// Emit a diagnostic at the specified location and return failure. 4066 InFlightDiagnostic emitError(llvm::SMLoc loc, const Twine &message) override { 4067 emittedError = true; 4068 return parser.emitError(loc, "custom op '" + opDefinition->name + "' " + 4069 message); 4070 } 4071 4072 llvm::SMLoc getCurrentLocation() override { 4073 return parser.getToken().getLoc(); 4074 } 4075 4076 Builder &getBuilder() const override { return parser.builder; } 4077 4078 /// Return the name of the specified result in the specified syntax, as well 4079 /// as the subelement in the name. For example, in this operation: 4080 /// 4081 /// %x, %y:2, %z = foo.op 4082 /// 4083 /// getResultName(0) == {"x", 0 } 4084 /// getResultName(1) == {"y", 0 } 4085 /// getResultName(2) == {"y", 1 } 4086 /// getResultName(3) == {"z", 0 } 4087 std::pair<StringRef, unsigned> 4088 getResultName(unsigned resultNo) const override { 4089 // Scan for the resultID that contains this result number. 4090 for (unsigned nameID = 0, e = resultIDs.size(); nameID != e; ++nameID) { 4091 const auto &entry = resultIDs[nameID]; 4092 if (resultNo < std::get<1>(entry)) { 4093 // Don't pass on the leading %. 4094 StringRef name = std::get<0>(entry).drop_front(); 4095 return {name, resultNo}; 4096 } 4097 resultNo -= std::get<1>(entry); 4098 } 4099 4100 // Invalid result number. 4101 return {"", ~0U}; 4102 } 4103 4104 /// Return the number of declared SSA results. This returns 4 for the foo.op 4105 /// example in the comment for getResultName. 4106 size_t getNumResults() const override { 4107 size_t count = 0; 4108 for (auto &entry : resultIDs) 4109 count += std::get<1>(entry); 4110 return count; 4111 } 4112 4113 llvm::SMLoc getNameLoc() const override { return nameLoc; } 4114 4115 //===--------------------------------------------------------------------===// 4116 // Token Parsing 4117 //===--------------------------------------------------------------------===// 4118 4119 /// Parse a `->` token. 4120 ParseResult parseArrow() override { 4121 return parser.parseToken(Token::arrow, "expected '->'"); 4122 } 4123 4124 /// Parses a `->` if present. 4125 ParseResult parseOptionalArrow() override { 4126 return success(parser.consumeIf(Token::arrow)); 4127 } 4128 4129 /// Parse a `:` token. 4130 ParseResult parseColon() override { 4131 return parser.parseToken(Token::colon, "expected ':'"); 4132 } 4133 4134 /// Parse a `:` token if present. 4135 ParseResult parseOptionalColon() override { 4136 return success(parser.consumeIf(Token::colon)); 4137 } 4138 4139 /// Parse a `,` token. 4140 ParseResult parseComma() override { 4141 return parser.parseToken(Token::comma, "expected ','"); 4142 } 4143 4144 /// Parse a `,` token if present. 4145 ParseResult parseOptionalComma() override { 4146 return success(parser.consumeIf(Token::comma)); 4147 } 4148 4149 /// Parses a `...` if present. 4150 ParseResult parseOptionalEllipsis() override { 4151 return success(parser.consumeIf(Token::ellipsis)); 4152 } 4153 4154 /// Parse a `=` token. 4155 ParseResult parseEqual() override { 4156 return parser.parseToken(Token::equal, "expected '='"); 4157 } 4158 4159 /// Parse a '<' token. 4160 ParseResult parseLess() override { 4161 return parser.parseToken(Token::less, "expected '<'"); 4162 } 4163 4164 /// Parse a '>' token. 4165 ParseResult parseGreater() override { 4166 return parser.parseToken(Token::greater, "expected '>'"); 4167 } 4168 4169 /// Parse a `(` token. 4170 ParseResult parseLParen() override { 4171 return parser.parseToken(Token::l_paren, "expected '('"); 4172 } 4173 4174 /// Parses a '(' if present. 4175 ParseResult parseOptionalLParen() override { 4176 return success(parser.consumeIf(Token::l_paren)); 4177 } 4178 4179 /// Parse a `)` token. 4180 ParseResult parseRParen() override { 4181 return parser.parseToken(Token::r_paren, "expected ')'"); 4182 } 4183 4184 /// Parses a ')' if present. 4185 ParseResult parseOptionalRParen() override { 4186 return success(parser.consumeIf(Token::r_paren)); 4187 } 4188 4189 /// Parse a `[` token. 4190 ParseResult parseLSquare() override { 4191 return parser.parseToken(Token::l_square, "expected '['"); 4192 } 4193 4194 /// Parses a '[' if present. 4195 ParseResult parseOptionalLSquare() override { 4196 return success(parser.consumeIf(Token::l_square)); 4197 } 4198 4199 /// Parse a `]` token. 4200 ParseResult parseRSquare() override { 4201 return parser.parseToken(Token::r_square, "expected ']'"); 4202 } 4203 4204 /// Parses a ']' if present. 4205 ParseResult parseOptionalRSquare() override { 4206 return success(parser.consumeIf(Token::r_square)); 4207 } 4208 4209 //===--------------------------------------------------------------------===// 4210 // Attribute Parsing 4211 //===--------------------------------------------------------------------===// 4212 4213 /// Parse an arbitrary attribute of a given type and return it in result. This 4214 /// also adds the attribute to the specified attribute list with the specified 4215 /// name. 4216 ParseResult parseAttribute(Attribute &result, Type type, StringRef attrName, 4217 SmallVectorImpl<NamedAttribute> &attrs) override { 4218 result = parser.parseAttribute(type); 4219 if (!result) 4220 return failure(); 4221 4222 attrs.push_back(parser.builder.getNamedAttr(attrName, result)); 4223 return success(); 4224 } 4225 4226 /// Parse a named dictionary into 'result' if it is present. 4227 ParseResult 4228 parseOptionalAttrDict(SmallVectorImpl<NamedAttribute> &result) override { 4229 if (parser.getToken().isNot(Token::l_brace)) 4230 return success(); 4231 return parser.parseAttributeDict(result); 4232 } 4233 4234 /// Parse a named dictionary into 'result' if the `attributes` keyword is 4235 /// present. 4236 ParseResult parseOptionalAttrDictWithKeyword( 4237 SmallVectorImpl<NamedAttribute> &result) override { 4238 if (failed(parseOptionalKeyword("attributes"))) 4239 return success(); 4240 return parser.parseAttributeDict(result); 4241 } 4242 4243 /// Parse an affine map instance into 'map'. 4244 ParseResult parseAffineMap(AffineMap &map) override { 4245 return parser.parseAffineMapReference(map); 4246 } 4247 4248 /// Parse an integer set instance into 'set'. 4249 ParseResult printIntegerSet(IntegerSet &set) override { 4250 return parser.parseIntegerSetReference(set); 4251 } 4252 4253 //===--------------------------------------------------------------------===// 4254 // Identifier Parsing 4255 //===--------------------------------------------------------------------===// 4256 4257 /// Returns if the current token corresponds to a keyword. 4258 bool isCurrentTokenAKeyword() const { 4259 return parser.getToken().is(Token::bare_identifier) || 4260 parser.getToken().isKeyword(); 4261 } 4262 4263 /// Parse the given keyword if present. 4264 ParseResult parseOptionalKeyword(StringRef keyword) override { 4265 // Check that the current token has the same spelling. 4266 if (!isCurrentTokenAKeyword() || parser.getTokenSpelling() != keyword) 4267 return failure(); 4268 parser.consumeToken(); 4269 return success(); 4270 } 4271 4272 /// Parse a keyword, if present, into 'keyword'. 4273 ParseResult parseOptionalKeyword(StringRef *keyword) override { 4274 // Check that the current token is a keyword. 4275 if (!isCurrentTokenAKeyword()) 4276 return failure(); 4277 4278 *keyword = parser.getTokenSpelling(); 4279 parser.consumeToken(); 4280 return success(); 4281 } 4282 4283 /// Parse an optional @-identifier and store it (without the '@' symbol) in a 4284 /// string attribute named 'attrName'. 4285 ParseResult 4286 parseOptionalSymbolName(StringAttr &result, StringRef attrName, 4287 SmallVectorImpl<NamedAttribute> &attrs) override { 4288 Token atToken = parser.getToken(); 4289 if (atToken.isNot(Token::at_identifier)) 4290 return failure(); 4291 4292 result = getBuilder().getStringAttr(extractSymbolReference(atToken)); 4293 attrs.push_back(getBuilder().getNamedAttr(attrName, result)); 4294 parser.consumeToken(); 4295 return success(); 4296 } 4297 4298 //===--------------------------------------------------------------------===// 4299 // Operand Parsing 4300 //===--------------------------------------------------------------------===// 4301 4302 /// Parse a single operand. 4303 ParseResult parseOperand(OperandType &result) override { 4304 OperationParser::SSAUseInfo useInfo; 4305 if (parser.parseSSAUse(useInfo)) 4306 return failure(); 4307 4308 result = {useInfo.loc, useInfo.name, useInfo.number}; 4309 return success(); 4310 } 4311 4312 /// Parse a single operand if present. 4313 OptionalParseResult parseOptionalOperand(OperandType &result) override { 4314 if (parser.getToken().is(Token::percent_identifier)) 4315 return parseOperand(result); 4316 return llvm::None; 4317 } 4318 4319 /// Parse zero or more SSA comma-separated operand references with a specified 4320 /// surrounding delimiter, and an optional required operand count. 4321 ParseResult parseOperandList(SmallVectorImpl<OperandType> &result, 4322 int requiredOperandCount = -1, 4323 Delimiter delimiter = Delimiter::None) override { 4324 return parseOperandOrRegionArgList(result, /*isOperandList=*/true, 4325 requiredOperandCount, delimiter); 4326 } 4327 4328 /// Parse zero or more SSA comma-separated operand or region arguments with 4329 /// optional surrounding delimiter and required operand count. 4330 ParseResult 4331 parseOperandOrRegionArgList(SmallVectorImpl<OperandType> &result, 4332 bool isOperandList, int requiredOperandCount = -1, 4333 Delimiter delimiter = Delimiter::None) { 4334 auto startLoc = parser.getToken().getLoc(); 4335 4336 // Handle delimiters. 4337 switch (delimiter) { 4338 case Delimiter::None: 4339 // Don't check for the absence of a delimiter if the number of operands 4340 // is unknown (and hence the operand list could be empty). 4341 if (requiredOperandCount == -1) 4342 break; 4343 // Token already matches an identifier and so can't be a delimiter. 4344 if (parser.getToken().is(Token::percent_identifier)) 4345 break; 4346 // Test against known delimiters. 4347 if (parser.getToken().is(Token::l_paren) || 4348 parser.getToken().is(Token::l_square)) 4349 return emitError(startLoc, "unexpected delimiter"); 4350 return emitError(startLoc, "invalid operand"); 4351 case Delimiter::OptionalParen: 4352 if (parser.getToken().isNot(Token::l_paren)) 4353 return success(); 4354 LLVM_FALLTHROUGH; 4355 case Delimiter::Paren: 4356 if (parser.parseToken(Token::l_paren, "expected '(' in operand list")) 4357 return failure(); 4358 break; 4359 case Delimiter::OptionalSquare: 4360 if (parser.getToken().isNot(Token::l_square)) 4361 return success(); 4362 LLVM_FALLTHROUGH; 4363 case Delimiter::Square: 4364 if (parser.parseToken(Token::l_square, "expected '[' in operand list")) 4365 return failure(); 4366 break; 4367 } 4368 4369 // Check for zero operands. 4370 if (parser.getToken().is(Token::percent_identifier)) { 4371 do { 4372 OperandType operandOrArg; 4373 if (isOperandList ? parseOperand(operandOrArg) 4374 : parseRegionArgument(operandOrArg)) 4375 return failure(); 4376 result.push_back(operandOrArg); 4377 } while (parser.consumeIf(Token::comma)); 4378 } 4379 4380 // Handle delimiters. If we reach here, the optional delimiters were 4381 // present, so we need to parse their closing one. 4382 switch (delimiter) { 4383 case Delimiter::None: 4384 break; 4385 case Delimiter::OptionalParen: 4386 case Delimiter::Paren: 4387 if (parser.parseToken(Token::r_paren, "expected ')' in operand list")) 4388 return failure(); 4389 break; 4390 case Delimiter::OptionalSquare: 4391 case Delimiter::Square: 4392 if (parser.parseToken(Token::r_square, "expected ']' in operand list")) 4393 return failure(); 4394 break; 4395 } 4396 4397 if (requiredOperandCount != -1 && 4398 result.size() != static_cast<size_t>(requiredOperandCount)) 4399 return emitError(startLoc, "expected ") 4400 << requiredOperandCount << " operands"; 4401 return success(); 4402 } 4403 4404 /// Parse zero or more trailing SSA comma-separated trailing operand 4405 /// references with a specified surrounding delimiter, and an optional 4406 /// required operand count. A leading comma is expected before the operands. 4407 ParseResult parseTrailingOperandList(SmallVectorImpl<OperandType> &result, 4408 int requiredOperandCount, 4409 Delimiter delimiter) override { 4410 if (parser.getToken().is(Token::comma)) { 4411 parseComma(); 4412 return parseOperandList(result, requiredOperandCount, delimiter); 4413 } 4414 if (requiredOperandCount != -1) 4415 return emitError(parser.getToken().getLoc(), "expected ") 4416 << requiredOperandCount << " operands"; 4417 return success(); 4418 } 4419 4420 /// Resolve an operand to an SSA value, emitting an error on failure. 4421 ParseResult resolveOperand(const OperandType &operand, Type type, 4422 SmallVectorImpl<Value> &result) override { 4423 OperationParser::SSAUseInfo operandInfo = {operand.name, operand.number, 4424 operand.location}; 4425 if (auto value = parser.resolveSSAUse(operandInfo, type)) { 4426 result.push_back(value); 4427 return success(); 4428 } 4429 return failure(); 4430 } 4431 4432 /// Parse an AffineMap of SSA ids. 4433 ParseResult parseAffineMapOfSSAIds(SmallVectorImpl<OperandType> &operands, 4434 Attribute &mapAttr, StringRef attrName, 4435 SmallVectorImpl<NamedAttribute> &attrs, 4436 Delimiter delimiter) override { 4437 SmallVector<OperandType, 2> dimOperands; 4438 SmallVector<OperandType, 1> symOperands; 4439 4440 auto parseElement = [&](bool isSymbol) -> ParseResult { 4441 OperandType operand; 4442 if (parseOperand(operand)) 4443 return failure(); 4444 if (isSymbol) 4445 symOperands.push_back(operand); 4446 else 4447 dimOperands.push_back(operand); 4448 return success(); 4449 }; 4450 4451 AffineMap map; 4452 if (parser.parseAffineMapOfSSAIds(map, parseElement, delimiter)) 4453 return failure(); 4454 // Add AffineMap attribute. 4455 if (map) { 4456 mapAttr = AffineMapAttr::get(map); 4457 attrs.push_back(parser.builder.getNamedAttr(attrName, mapAttr)); 4458 } 4459 4460 // Add dim operands before symbol operands in 'operands'. 4461 operands.assign(dimOperands.begin(), dimOperands.end()); 4462 operands.append(symOperands.begin(), symOperands.end()); 4463 return success(); 4464 } 4465 4466 //===--------------------------------------------------------------------===// 4467 // Region Parsing 4468 //===--------------------------------------------------------------------===// 4469 4470 /// Parse a region that takes `arguments` of `argTypes` types. This 4471 /// effectively defines the SSA values of `arguments` and assigns their type. 4472 ParseResult parseRegion(Region ®ion, ArrayRef<OperandType> arguments, 4473 ArrayRef<Type> argTypes, 4474 bool enableNameShadowing) override { 4475 assert(arguments.size() == argTypes.size() && 4476 "mismatching number of arguments and types"); 4477 4478 SmallVector<std::pair<OperationParser::SSAUseInfo, Type>, 2> 4479 regionArguments; 4480 for (auto pair : llvm::zip(arguments, argTypes)) { 4481 const OperandType &operand = std::get<0>(pair); 4482 Type type = std::get<1>(pair); 4483 OperationParser::SSAUseInfo operandInfo = {operand.name, operand.number, 4484 operand.location}; 4485 regionArguments.emplace_back(operandInfo, type); 4486 } 4487 4488 // Try to parse the region. 4489 assert((!enableNameShadowing || 4490 opDefinition->hasProperty(OperationProperty::IsolatedFromAbove)) && 4491 "name shadowing is only allowed on isolated regions"); 4492 if (parser.parseRegion(region, regionArguments, enableNameShadowing)) 4493 return failure(); 4494 return success(); 4495 } 4496 4497 /// Parses a region if present. 4498 ParseResult parseOptionalRegion(Region ®ion, 4499 ArrayRef<OperandType> arguments, 4500 ArrayRef<Type> argTypes, 4501 bool enableNameShadowing) override { 4502 if (parser.getToken().isNot(Token::l_brace)) 4503 return success(); 4504 return parseRegion(region, arguments, argTypes, enableNameShadowing); 4505 } 4506 4507 /// Parse a region argument. The type of the argument will be resolved later 4508 /// by a call to `parseRegion`. 4509 ParseResult parseRegionArgument(OperandType &argument) override { 4510 return parseOperand(argument); 4511 } 4512 4513 /// Parse a region argument if present. 4514 ParseResult parseOptionalRegionArgument(OperandType &argument) override { 4515 if (parser.getToken().isNot(Token::percent_identifier)) 4516 return success(); 4517 return parseRegionArgument(argument); 4518 } 4519 4520 ParseResult 4521 parseRegionArgumentList(SmallVectorImpl<OperandType> &result, 4522 int requiredOperandCount = -1, 4523 Delimiter delimiter = Delimiter::None) override { 4524 return parseOperandOrRegionArgList(result, /*isOperandList=*/false, 4525 requiredOperandCount, delimiter); 4526 } 4527 4528 //===--------------------------------------------------------------------===// 4529 // Successor Parsing 4530 //===--------------------------------------------------------------------===// 4531 4532 /// Parse a single operation successor. 4533 ParseResult parseSuccessor(Block *&dest) override { 4534 return parser.parseSuccessor(dest); 4535 } 4536 4537 /// Parse an optional operation successor and its operand list. 4538 OptionalParseResult parseOptionalSuccessor(Block *&dest) override { 4539 if (parser.getToken().isNot(Token::caret_identifier)) 4540 return llvm::None; 4541 return parseSuccessor(dest); 4542 } 4543 4544 /// Parse a single operation successor and its operand list. 4545 ParseResult 4546 parseSuccessorAndUseList(Block *&dest, 4547 SmallVectorImpl<Value> &operands) override { 4548 if (parseSuccessor(dest)) 4549 return failure(); 4550 4551 // Handle optional arguments. 4552 if (succeeded(parseOptionalLParen()) && 4553 (parser.parseOptionalSSAUseAndTypeList(operands) || parseRParen())) { 4554 return failure(); 4555 } 4556 return success(); 4557 } 4558 4559 //===--------------------------------------------------------------------===// 4560 // Type Parsing 4561 //===--------------------------------------------------------------------===// 4562 4563 /// Parse a type. 4564 ParseResult parseType(Type &result) override { 4565 return failure(!(result = parser.parseType())); 4566 } 4567 4568 /// Parse an optional type. 4569 OptionalParseResult parseOptionalType(Type &result) override { 4570 return parser.parseOptionalType(result); 4571 } 4572 4573 /// Parse an arrow followed by a type list. 4574 ParseResult parseArrowTypeList(SmallVectorImpl<Type> &result) override { 4575 if (parseArrow() || parser.parseFunctionResultTypes(result)) 4576 return failure(); 4577 return success(); 4578 } 4579 4580 /// Parse an optional arrow followed by a type list. 4581 ParseResult 4582 parseOptionalArrowTypeList(SmallVectorImpl<Type> &result) override { 4583 if (!parser.consumeIf(Token::arrow)) 4584 return success(); 4585 return parser.parseFunctionResultTypes(result); 4586 } 4587 4588 /// Parse a colon followed by a type. 4589 ParseResult parseColonType(Type &result) override { 4590 return failure(parser.parseToken(Token::colon, "expected ':'") || 4591 !(result = parser.parseType())); 4592 } 4593 4594 /// Parse a colon followed by a type list, which must have at least one type. 4595 ParseResult parseColonTypeList(SmallVectorImpl<Type> &result) override { 4596 if (parser.parseToken(Token::colon, "expected ':'")) 4597 return failure(); 4598 return parser.parseTypeListNoParens(result); 4599 } 4600 4601 /// Parse an optional colon followed by a type list, which if present must 4602 /// have at least one type. 4603 ParseResult 4604 parseOptionalColonTypeList(SmallVectorImpl<Type> &result) override { 4605 if (!parser.consumeIf(Token::colon)) 4606 return success(); 4607 return parser.parseTypeListNoParens(result); 4608 } 4609 4610 /// Parse a list of assignments of the form 4611 /// (%x1 = %y1 : type1, %x2 = %y2 : type2, ...). 4612 /// The list must contain at least one entry 4613 ParseResult parseAssignmentList(SmallVectorImpl<OperandType> &lhs, 4614 SmallVectorImpl<OperandType> &rhs) override { 4615 auto parseElt = [&]() -> ParseResult { 4616 OperandType regionArg, operand; 4617 if (parseRegionArgument(regionArg) || parseEqual() || 4618 parseOperand(operand)) 4619 return failure(); 4620 lhs.push_back(regionArg); 4621 rhs.push_back(operand); 4622 return success(); 4623 }; 4624 if (parseLParen()) 4625 return failure(); 4626 return parser.parseCommaSeparatedListUntil(Token::r_paren, parseElt); 4627 } 4628 4629 private: 4630 /// The source location of the operation name. 4631 SMLoc nameLoc; 4632 4633 /// Information about the result name specifiers. 4634 ArrayRef<OperationParser::ResultRecord> resultIDs; 4635 4636 /// The abstract information of the operation. 4637 const AbstractOperation *opDefinition; 4638 4639 /// The main operation parser. 4640 OperationParser &parser; 4641 4642 /// A flag that indicates if any errors were emitted during parsing. 4643 bool emittedError = false; 4644 }; 4645 } // end anonymous namespace. 4646 4647 Operation * 4648 OperationParser::parseCustomOperation(ArrayRef<ResultRecord> resultIDs) { 4649 auto opLoc = getToken().getLoc(); 4650 auto opName = getTokenSpelling(); 4651 4652 auto *opDefinition = AbstractOperation::lookup(opName, getContext()); 4653 if (!opDefinition && !opName.contains('.')) { 4654 // If the operation name has no namespace prefix we treat it as a standard 4655 // operation and prefix it with "std". 4656 // TODO: Would it be better to just build a mapping of the registered 4657 // operations in the standard dialect? 4658 opDefinition = 4659 AbstractOperation::lookup(Twine("std." + opName).str(), getContext()); 4660 } 4661 4662 if (!opDefinition) { 4663 emitError(opLoc) << "custom op '" << opName << "' is unknown"; 4664 return nullptr; 4665 } 4666 4667 consumeToken(); 4668 4669 // If the custom op parser crashes, produce some indication to help 4670 // debugging. 4671 std::string opNameStr = opName.str(); 4672 llvm::PrettyStackTraceFormat fmt("MLIR Parser: custom op parser '%s'", 4673 opNameStr.c_str()); 4674 4675 // Get location information for the operation. 4676 auto srcLocation = getEncodedSourceLocation(opLoc); 4677 4678 // Have the op implementation take a crack and parsing this. 4679 OperationState opState(srcLocation, opDefinition->name); 4680 CleanupOpStateRegions guard{opState}; 4681 CustomOpAsmParser opAsmParser(opLoc, resultIDs, opDefinition, *this); 4682 if (opAsmParser.parseOperation(opState)) 4683 return nullptr; 4684 4685 // If it emitted an error, we failed. 4686 if (opAsmParser.didEmitError()) 4687 return nullptr; 4688 4689 // Parse a location if one is present. 4690 if (parseOptionalTrailingLocation(opState.location)) 4691 return nullptr; 4692 4693 // Otherwise, we succeeded. Use the state it parsed as our op information. 4694 return opBuilder.createOperation(opState); 4695 } 4696 4697 //===----------------------------------------------------------------------===// 4698 // Region Parsing 4699 //===----------------------------------------------------------------------===// 4700 4701 /// Region. 4702 /// 4703 /// region ::= '{' region-body 4704 /// 4705 ParseResult OperationParser::parseRegion( 4706 Region ®ion, 4707 ArrayRef<std::pair<OperationParser::SSAUseInfo, Type>> entryArguments, 4708 bool isIsolatedNameScope) { 4709 // Parse the '{'. 4710 if (parseToken(Token::l_brace, "expected '{' to begin a region")) 4711 return failure(); 4712 4713 // Check for an empty region. 4714 if (entryArguments.empty() && consumeIf(Token::r_brace)) 4715 return success(); 4716 auto currentPt = opBuilder.saveInsertionPoint(); 4717 4718 // Push a new named value scope. 4719 pushSSANameScope(isIsolatedNameScope); 4720 4721 // Parse the first block directly to allow for it to be unnamed. 4722 Block *block = new Block(); 4723 4724 // Add arguments to the entry block. 4725 if (!entryArguments.empty()) { 4726 for (auto &placeholderArgPair : entryArguments) { 4727 auto &argInfo = placeholderArgPair.first; 4728 // Ensure that the argument was not already defined. 4729 if (auto defLoc = getReferenceLoc(argInfo.name, argInfo.number)) { 4730 return emitError(argInfo.loc, "region entry argument '" + argInfo.name + 4731 "' is already in use") 4732 .attachNote(getEncodedSourceLocation(*defLoc)) 4733 << "previously referenced here"; 4734 } 4735 if (addDefinition(placeholderArgPair.first, 4736 block->addArgument(placeholderArgPair.second))) { 4737 delete block; 4738 return failure(); 4739 } 4740 } 4741 4742 // If we had named arguments, then don't allow a block name. 4743 if (getToken().is(Token::caret_identifier)) 4744 return emitError("invalid block name in region with named arguments"); 4745 } 4746 4747 if (parseBlock(block)) { 4748 delete block; 4749 return failure(); 4750 } 4751 4752 // Verify that no other arguments were parsed. 4753 if (!entryArguments.empty() && 4754 block->getNumArguments() > entryArguments.size()) { 4755 delete block; 4756 return emitError("entry block arguments were already defined"); 4757 } 4758 4759 // Parse the rest of the region. 4760 region.push_back(block); 4761 if (parseRegionBody(region)) 4762 return failure(); 4763 4764 // Pop the SSA value scope for this region. 4765 if (popSSANameScope()) 4766 return failure(); 4767 4768 // Reset the original insertion point. 4769 opBuilder.restoreInsertionPoint(currentPt); 4770 return success(); 4771 } 4772 4773 /// Region. 4774 /// 4775 /// region-body ::= block* '}' 4776 /// 4777 ParseResult OperationParser::parseRegionBody(Region ®ion) { 4778 // Parse the list of blocks. 4779 while (!consumeIf(Token::r_brace)) { 4780 Block *newBlock = nullptr; 4781 if (parseBlock(newBlock)) 4782 return failure(); 4783 region.push_back(newBlock); 4784 } 4785 return success(); 4786 } 4787 4788 //===----------------------------------------------------------------------===// 4789 // Block Parsing 4790 //===----------------------------------------------------------------------===// 4791 4792 /// Block declaration. 4793 /// 4794 /// block ::= block-label? operation* 4795 /// block-label ::= block-id block-arg-list? `:` 4796 /// block-id ::= caret-id 4797 /// block-arg-list ::= `(` ssa-id-and-type-list? `)` 4798 /// 4799 ParseResult OperationParser::parseBlock(Block *&block) { 4800 // The first block of a region may already exist, if it does the caret 4801 // identifier is optional. 4802 if (block && getToken().isNot(Token::caret_identifier)) 4803 return parseBlockBody(block); 4804 4805 SMLoc nameLoc = getToken().getLoc(); 4806 auto name = getTokenSpelling(); 4807 if (parseToken(Token::caret_identifier, "expected block name")) 4808 return failure(); 4809 4810 block = defineBlockNamed(name, nameLoc, block); 4811 4812 // Fail if the block was already defined. 4813 if (!block) 4814 return emitError(nameLoc, "redefinition of block '") << name << "'"; 4815 4816 // If an argument list is present, parse it. 4817 if (consumeIf(Token::l_paren)) { 4818 SmallVector<BlockArgument, 8> bbArgs; 4819 if (parseOptionalBlockArgList(bbArgs, block) || 4820 parseToken(Token::r_paren, "expected ')' to end argument list")) 4821 return failure(); 4822 } 4823 4824 if (parseToken(Token::colon, "expected ':' after block name")) 4825 return failure(); 4826 4827 return parseBlockBody(block); 4828 } 4829 4830 ParseResult OperationParser::parseBlockBody(Block *block) { 4831 // Set the insertion point to the end of the block to parse. 4832 opBuilder.setInsertionPointToEnd(block); 4833 4834 // Parse the list of operations that make up the body of the block. 4835 while (getToken().isNot(Token::caret_identifier, Token::r_brace)) 4836 if (parseOperation()) 4837 return failure(); 4838 4839 return success(); 4840 } 4841 4842 /// Get the block with the specified name, creating it if it doesn't already 4843 /// exist. The location specified is the point of use, which allows 4844 /// us to diagnose references to blocks that are not defined precisely. 4845 Block *OperationParser::getBlockNamed(StringRef name, SMLoc loc) { 4846 auto &blockAndLoc = getBlockInfoByName(name); 4847 if (!blockAndLoc.first) { 4848 blockAndLoc = {new Block(), loc}; 4849 insertForwardRef(blockAndLoc.first, loc); 4850 } 4851 4852 return blockAndLoc.first; 4853 } 4854 4855 /// Define the block with the specified name. Returns the Block* or nullptr in 4856 /// the case of redefinition. 4857 Block *OperationParser::defineBlockNamed(StringRef name, SMLoc loc, 4858 Block *existing) { 4859 auto &blockAndLoc = getBlockInfoByName(name); 4860 if (!blockAndLoc.first) { 4861 // If the caller provided a block, use it. Otherwise create a new one. 4862 if (!existing) 4863 existing = new Block(); 4864 blockAndLoc.first = existing; 4865 blockAndLoc.second = loc; 4866 return blockAndLoc.first; 4867 } 4868 4869 // Forward declarations are removed once defined, so if we are defining a 4870 // existing block and it is not a forward declaration, then it is a 4871 // redeclaration. 4872 if (!eraseForwardRef(blockAndLoc.first)) 4873 return nullptr; 4874 return blockAndLoc.first; 4875 } 4876 4877 /// Parse a (possibly empty) list of SSA operands with types as block arguments. 4878 /// 4879 /// ssa-id-and-type-list ::= ssa-id-and-type (`,` ssa-id-and-type)* 4880 /// 4881 ParseResult OperationParser::parseOptionalBlockArgList( 4882 SmallVectorImpl<BlockArgument> &results, Block *owner) { 4883 if (getToken().is(Token::r_brace)) 4884 return success(); 4885 4886 // If the block already has arguments, then we're handling the entry block. 4887 // Parse and register the names for the arguments, but do not add them. 4888 bool definingExistingArgs = owner->getNumArguments() != 0; 4889 unsigned nextArgument = 0; 4890 4891 return parseCommaSeparatedList([&]() -> ParseResult { 4892 return parseSSADefOrUseAndType( 4893 [&](SSAUseInfo useInfo, Type type) -> ParseResult { 4894 // If this block did not have existing arguments, define a new one. 4895 if (!definingExistingArgs) 4896 return addDefinition(useInfo, owner->addArgument(type)); 4897 4898 // Otherwise, ensure that this argument has already been created. 4899 if (nextArgument >= owner->getNumArguments()) 4900 return emitError("too many arguments specified in argument list"); 4901 4902 // Finally, make sure the existing argument has the correct type. 4903 auto arg = owner->getArgument(nextArgument++); 4904 if (arg.getType() != type) 4905 return emitError("argument and block argument type mismatch"); 4906 return addDefinition(useInfo, arg); 4907 }); 4908 }); 4909 } 4910 4911 //===----------------------------------------------------------------------===// 4912 // Top-level entity parsing. 4913 //===----------------------------------------------------------------------===// 4914 4915 namespace { 4916 /// This parser handles entities that are only valid at the top level of the 4917 /// file. 4918 class ModuleParser : public Parser { 4919 public: 4920 explicit ModuleParser(ParserState &state) : Parser(state) {} 4921 4922 ParseResult parseModule(ModuleOp module); 4923 4924 private: 4925 /// Parse an attribute alias declaration. 4926 ParseResult parseAttributeAliasDef(); 4927 4928 /// Parse an attribute alias declaration. 4929 ParseResult parseTypeAliasDef(); 4930 }; 4931 } // end anonymous namespace 4932 4933 /// Parses an attribute alias declaration. 4934 /// 4935 /// attribute-alias-def ::= '#' alias-name `=` attribute-value 4936 /// 4937 ParseResult ModuleParser::parseAttributeAliasDef() { 4938 assert(getToken().is(Token::hash_identifier)); 4939 StringRef aliasName = getTokenSpelling().drop_front(); 4940 4941 // Check for redefinitions. 4942 if (getState().symbols.attributeAliasDefinitions.count(aliasName) > 0) 4943 return emitError("redefinition of attribute alias id '" + aliasName + "'"); 4944 4945 // Make sure this isn't invading the dialect attribute namespace. 4946 if (aliasName.contains('.')) 4947 return emitError("attribute names with a '.' are reserved for " 4948 "dialect-defined names"); 4949 4950 consumeToken(Token::hash_identifier); 4951 4952 // Parse the '='. 4953 if (parseToken(Token::equal, "expected '=' in attribute alias definition")) 4954 return failure(); 4955 4956 // Parse the attribute value. 4957 Attribute attr = parseAttribute(); 4958 if (!attr) 4959 return failure(); 4960 4961 getState().symbols.attributeAliasDefinitions[aliasName] = attr; 4962 return success(); 4963 } 4964 4965 /// Parse a type alias declaration. 4966 /// 4967 /// type-alias-def ::= '!' alias-name `=` 'type' type 4968 /// 4969 ParseResult ModuleParser::parseTypeAliasDef() { 4970 assert(getToken().is(Token::exclamation_identifier)); 4971 StringRef aliasName = getTokenSpelling().drop_front(); 4972 4973 // Check for redefinitions. 4974 if (getState().symbols.typeAliasDefinitions.count(aliasName) > 0) 4975 return emitError("redefinition of type alias id '" + aliasName + "'"); 4976 4977 // Make sure this isn't invading the dialect type namespace. 4978 if (aliasName.contains('.')) 4979 return emitError("type names with a '.' are reserved for " 4980 "dialect-defined names"); 4981 4982 consumeToken(Token::exclamation_identifier); 4983 4984 // Parse the '=' and 'type'. 4985 if (parseToken(Token::equal, "expected '=' in type alias definition") || 4986 parseToken(Token::kw_type, "expected 'type' in type alias definition")) 4987 return failure(); 4988 4989 // Parse the type. 4990 Type aliasedType = parseType(); 4991 if (!aliasedType) 4992 return failure(); 4993 4994 // Register this alias with the parser state. 4995 getState().symbols.typeAliasDefinitions.try_emplace(aliasName, aliasedType); 4996 return success(); 4997 } 4998 4999 /// This is the top-level module parser. 5000 ParseResult ModuleParser::parseModule(ModuleOp module) { 5001 OperationParser opParser(getState(), module); 5002 5003 // Module itself is a name scope. 5004 opParser.pushSSANameScope(/*isIsolated=*/true); 5005 5006 while (true) { 5007 switch (getToken().getKind()) { 5008 default: 5009 // Parse a top-level operation. 5010 if (opParser.parseOperation()) 5011 return failure(); 5012 break; 5013 5014 // If we got to the end of the file, then we're done. 5015 case Token::eof: { 5016 if (opParser.finalize()) 5017 return failure(); 5018 5019 // Handle the case where the top level module was explicitly defined. 5020 auto &bodyBlocks = module.getBodyRegion().getBlocks(); 5021 auto &operations = bodyBlocks.front().getOperations(); 5022 assert(!operations.empty() && "expected a valid module terminator"); 5023 5024 // Check that the first operation is a module, and it is the only 5025 // non-terminator operation. 5026 ModuleOp nested = dyn_cast<ModuleOp>(operations.front()); 5027 if (nested && std::next(operations.begin(), 2) == operations.end()) { 5028 // Merge the data of the nested module operation into 'module'. 5029 module.setLoc(nested.getLoc()); 5030 module.setAttrs(nested.getOperation()->getAttrList()); 5031 bodyBlocks.splice(bodyBlocks.end(), nested.getBodyRegion().getBlocks()); 5032 5033 // Erase the original module body. 5034 bodyBlocks.pop_front(); 5035 } 5036 5037 return opParser.popSSANameScope(); 5038 } 5039 5040 // If we got an error token, then the lexer already emitted an error, just 5041 // stop. Someday we could introduce error recovery if there was demand 5042 // for it. 5043 case Token::error: 5044 return failure(); 5045 5046 // Parse an attribute alias. 5047 case Token::hash_identifier: 5048 if (parseAttributeAliasDef()) 5049 return failure(); 5050 break; 5051 5052 // Parse a type alias. 5053 case Token::exclamation_identifier: 5054 if (parseTypeAliasDef()) 5055 return failure(); 5056 break; 5057 } 5058 } 5059 } 5060 5061 //===----------------------------------------------------------------------===// 5062 5063 /// This parses the file specified by the indicated SourceMgr and returns an 5064 /// MLIR module if it was valid. If not, it emits diagnostics and returns 5065 /// null. 5066 OwningModuleRef mlir::parseSourceFile(const llvm::SourceMgr &sourceMgr, 5067 MLIRContext *context) { 5068 auto sourceBuf = sourceMgr.getMemoryBuffer(sourceMgr.getMainFileID()); 5069 5070 // This is the result module we are parsing into. 5071 OwningModuleRef module(ModuleOp::create(FileLineColLoc::get( 5072 sourceBuf->getBufferIdentifier(), /*line=*/0, /*column=*/0, context))); 5073 5074 SymbolState aliasState; 5075 ParserState state(sourceMgr, context, aliasState); 5076 if (ModuleParser(state).parseModule(*module)) 5077 return nullptr; 5078 5079 // Make sure the parse module has no other structural problems detected by 5080 // the verifier. 5081 if (failed(verify(*module))) 5082 return nullptr; 5083 5084 return module; 5085 } 5086 5087 /// This parses the file specified by the indicated filename and returns an 5088 /// MLIR module if it was valid. If not, the error message is emitted through 5089 /// the error handler registered in the context, and a null pointer is returned. 5090 OwningModuleRef mlir::parseSourceFile(StringRef filename, 5091 MLIRContext *context) { 5092 llvm::SourceMgr sourceMgr; 5093 return parseSourceFile(filename, sourceMgr, context); 5094 } 5095 5096 /// This parses the file specified by the indicated filename using the provided 5097 /// SourceMgr and returns an MLIR module if it was valid. If not, the error 5098 /// message is emitted through the error handler registered in the context, and 5099 /// a null pointer is returned. 5100 OwningModuleRef mlir::parseSourceFile(StringRef filename, 5101 llvm::SourceMgr &sourceMgr, 5102 MLIRContext *context) { 5103 if (sourceMgr.getNumBuffers() != 0) { 5104 // TODO(b/136086478): Extend to support multiple buffers. 5105 emitError(mlir::UnknownLoc::get(context), 5106 "only main buffer parsed at the moment"); 5107 return nullptr; 5108 } 5109 auto file_or_err = llvm::MemoryBuffer::getFileOrSTDIN(filename); 5110 if (std::error_code error = file_or_err.getError()) { 5111 emitError(mlir::UnknownLoc::get(context), 5112 "could not open input file " + filename); 5113 return nullptr; 5114 } 5115 5116 // Load the MLIR module. 5117 sourceMgr.AddNewSourceBuffer(std::move(*file_or_err), llvm::SMLoc()); 5118 return parseSourceFile(sourceMgr, context); 5119 } 5120 5121 /// This parses the program string to a MLIR module if it was valid. If not, 5122 /// it emits diagnostics and returns null. 5123 OwningModuleRef mlir::parseSourceString(StringRef moduleStr, 5124 MLIRContext *context) { 5125 auto memBuffer = MemoryBuffer::getMemBuffer(moduleStr); 5126 if (!memBuffer) 5127 return nullptr; 5128 5129 SourceMgr sourceMgr; 5130 sourceMgr.AddNewSourceBuffer(std::move(memBuffer), SMLoc()); 5131 return parseSourceFile(sourceMgr, context); 5132 } 5133 5134 /// Parses a symbol, of type 'T', and returns it if parsing was successful. If 5135 /// parsing failed, nullptr is returned. The number of bytes read from the input 5136 /// string is returned in 'numRead'. 5137 template <typename T, typename ParserFn> 5138 static T parseSymbol(StringRef inputStr, MLIRContext *context, size_t &numRead, 5139 ParserFn &&parserFn) { 5140 SymbolState aliasState; 5141 return parseSymbol<T>( 5142 inputStr, context, aliasState, 5143 [&](Parser &parser) { 5144 SourceMgrDiagnosticHandler handler( 5145 const_cast<llvm::SourceMgr &>(parser.getSourceMgr()), 5146 parser.getContext()); 5147 return parserFn(parser); 5148 }, 5149 &numRead); 5150 } 5151 5152 Attribute mlir::parseAttribute(StringRef attrStr, MLIRContext *context) { 5153 size_t numRead = 0; 5154 return parseAttribute(attrStr, context, numRead); 5155 } 5156 Attribute mlir::parseAttribute(StringRef attrStr, Type type) { 5157 size_t numRead = 0; 5158 return parseAttribute(attrStr, type, numRead); 5159 } 5160 5161 Attribute mlir::parseAttribute(StringRef attrStr, MLIRContext *context, 5162 size_t &numRead) { 5163 return parseSymbol<Attribute>(attrStr, context, numRead, [](Parser &parser) { 5164 return parser.parseAttribute(); 5165 }); 5166 } 5167 Attribute mlir::parseAttribute(StringRef attrStr, Type type, size_t &numRead) { 5168 return parseSymbol<Attribute>( 5169 attrStr, type.getContext(), numRead, 5170 [type](Parser &parser) { return parser.parseAttribute(type); }); 5171 } 5172 5173 Type mlir::parseType(StringRef typeStr, MLIRContext *context) { 5174 size_t numRead = 0; 5175 return parseType(typeStr, context, numRead); 5176 } 5177 5178 Type mlir::parseType(StringRef typeStr, MLIRContext *context, size_t &numRead) { 5179 return parseSymbol<Type>(typeStr, context, numRead, 5180 [](Parser &parser) { return parser.parseType(); }); 5181 } 5182