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