Release 0.4.0

This release improves compliance with the YAML test suite (thanks @ingydotnet and @perlpunk for extensive and helpful cooperation), and adds node location tracking using the parser.

Breaking changes

As part of the new feature to track source locations, opportunity was taken to address a number of pre-existing API issues. These changes consisted of:

• Deprecate c4::yml::parse() and c4::yml::Parser::parse() overloads; all these functions will be removed in short order. Until removal, any call from client code will trigger a compiler warning.
• Add parse() alternatives, either parse_in_place() or parse_in_arena():
  • parse_in_place() receives only substr buffers, ie mutable YAML source buffers. Trying to pass a csubstr buffer to parse_in_place() will cause a compile error:

    substr readwrite = /*...*/;
    Tree tree = parse_in_place(readwrite); // OK
    
    csubstr readonly = /*...*/;
    Tree tree = parse_in_place(readonly); // compile error

  • parse_in_arena() receives only csubstr buffers, ie immutable YAML source buffers. Prior to parsing, the buffer is copied to the tree's arena, then the copy is parsed in place. Because parse_in_arena() is meant for immutable buffers, overloads receiving a substr YAML buffer are now declared but marked deprecated, and intentionally left undefined, such that calling parse_in_arena() with a substr will cause a linker error as well as a compiler warning.

    substr readwrite = /*...*/;
    Tree tree = parse_in_arena(readwrite); // compile warning+linker error

    This is to prevent an accidental extra copy of the mutable source buffer to the tree's arena: substr is implicitly convertible to csubstr. If you really intend to parse an originally mutable buffer in the tree's arena, convert it first explicitly to immutable by assigning the substr to a csubstr prior to calling parse_in_arena():

    substr readwrite = /*...*/;
    csubstr readonly = readwrite; // ok
    Tree tree = parse_in_arena(readonly); // ok

    This problem does not occur with parse_in_place() because csubstr is not implicitly convertible to substr.
• In the python API, ryml.parse() was removed and not just deprecated; the parse_in_arena() and parse_in_place() now replace this.
• Callbacks: changed behavior in Parser and Tree:
  • When a tree is copy-constructed or move-constructed to another, the receiving tree will start with the callbacks of the original.
  • When a tree is copy-assigned or move-assigned to another, the receiving tree will now change its callbacks to the original.
  • When a parser creates a new tree, the tree will now use a copy of the parser's callbacks object.
  • When an existing tree is given directly to the parser, both the tree and the parser now retain their own callback objects; any allocation or error during parsing will go through the respective callback object.

New features

• Add tracking of source code locations. This is useful for reporting semantic errors after the parsing phase (ie where the YAML is syntatically valid and parsing is successful, but the tree contents are semantically invalid). The locations can be obtained lazily from the parser when the first location is queried:

  // To obtain locations, use of the parser is needed:
  ryml::Parser parser;
  ryml::Tree tree = parser.parse_in_arena("source.yml", R"({
  aa: contents,
  foo: [one, [two, three]]
  })");
  // After parsing, on the first call to obtain a location,
  // the parser will cache a lookup structure to accelerate
  // tracking the location of a node, with complexity
  // O(numchars(srcbuffer)). Then it will do the lookup, with
  // complexity O(log(numlines(srcbuffer))).
  ryml::Location loc = parser.location(tree.rootref());
  assert(parser.location_contents(loc).begins_with("{"));
  // note the location members are zero-based:
  assert(loc.offset == 0u);
  assert(loc.line == 0u);
  assert(loc.col == 0u);
  // On the next call to location(), the accelerator is reused
  // and only the lookup is done.
  loc = parser.location(tree["aa"]);
  assert(parser.location_contents(loc).begins_with("aa"));
  assert(loc.offset == 2u);
  assert(loc.line == 1u);
  assert(loc.col == 0u);
  // KEYSEQ in flow style: points at the key
  loc = parser.location(tree["foo"]);
  assert(parser.location_contents(loc).begins_with("foo"));
  assert(loc.offset == 16u);
  assert(loc.line == 2u);
  assert(loc.col == 0u);
  loc = parser.location(tree["foo"][0]);
  assert(parser.location_contents(loc).begins_with("one"));
  assert(loc.line == 2u);
  assert(loc.col == 6u);
  // SEQ in flow style: location points at the opening '[' (there's no key)
  loc = parser.location(tree["foo"][1]);
  assert(parser.location_contents(loc).begins_with("["));
  assert(loc.line == 2u);
  assert(loc.col == 11u);
  loc = parser.location(tree["foo"][1][0]);
  assert(parser.location_contents(loc).begins_with("two"));
  assert(loc.line == 2u);
  assert(loc.col == 12u);
  loc = parser.location(tree["foo"][1][1]);
  assert(parser.location_contents(loc).begins_with("three"));
  assert(loc.line == 2u);
  assert(loc.col == 17u);
  // NOTE: reusing the parser with a new YAML source buffer
  // will invalidate the accelerator.

  See more details in the quickstart sample. Thanks to @cschreib for submitting a working example proving how simple it could be to achieve this.
• Parser:
  • add source() and filename() to get the latest buffer and filename to be parsed
  • add callbacks() to get the parser's callbacks
• Add from_tag_long() and normalize_tag_long():

  assert(from_tag_long(TAG_MAP) == "<tag:yaml.org,2002:map>");
  assert(normalize_tag_long("!!map") == "<tag:yaml.org,2002:map>");

• Add an experimental API to resolve tags based on the tree's tag directives. This API is still imature and will likely be subject to changes, so we won't document it yet.
• Regarding emit styles (see issue #37): add an experimental API to force flow/block style on container nodes, as well as block-literal/block-folded/double-quoted/single-quoted/plain styles on scalar nodes. This API is also immature and will likely be subject to changes, so we won't document it yet. But if you are desperate for this functionality, the new facilities will let you go further. See PR#191.
• Add preliminary support for bare-metal ARM architectures, with CI tests pending implementation of QEMU action. (#193, c4core#63).
• Add preliminary support for RISC-V architectures, with CI tests pending availability of RISC-V based github actions. (c4core#69).

Fixes

• Fix edge cases of parsing of explicit keys (ie keys after ?) (PR#212):

  # all these were fixed:
  ? : # empty
  ? explicit key   # this comment was not parsed correctly
  ?    # trailing empty key was not added to the map

• Fixed parsing of tabs used as whitespace tokens after : or -. This feature is costly (see some benchmark results here) and thus it is disabled by default, and requires defining a macro or cmake option RYML_WITH_TAB_TOKENS to enable (PR#211).
• Allow tab indentation in flow seqs (PR#215) (6CA3).
• ryml now parses successfully compact JSON code {"like":"this"} without any need for preprocessing. This code was not valid YAML 1.1, but was made valid in YAML 1.2. So the preprocess_json() functions, used to insert spaces after : are no longer necessary and have been removed. If you were using these functions, remove the calls and just pass the original source directly to ryml's parser (PR#210).
• Fix handling of indentation when parsing block scalars (PR#210):

  ---
  |
  hello
  there
  ---
  |
  ciao
  qua
  ---
  - |
   hello
   there
  - |
   ciao
   qua
  ---
  foo: |
   hello
   there
  bar: |
   ciao
   qua

• Fix parsing of maps when opening a scope with whitespace before the colon (PR#210):

  foo0 : bar
  ---
  foo1 : bar  # the " :" was causing an assert
  ---
  foo2 : bar
  ---
  foo3	: bar
  ---
  foo4   	  : bar

• Ensure container keys preserve quote flags when the key is quoted (PR#210).
• Ensure scalars beginning with % are emitted with quotes ((PR#216).
• Fix #203: when parsing, do not convert null or ~ to null scalar strings. Now the scalar strings contain the verbatim contents of the original scalar; to query whether a scalar value is null, use Tree::key_is_null()/val_is_null() and NodeRef::key_is_null()/val_is_null() which return true if it is empty or any of the unquoted strings ~, null, Null, or NULL. (PR#207):
• Fix #205: fix parsing of escaped characters in double-quoted strings: "\\\"\n\r\t\<TAB>\/\<SPC>\0\b\f\a\v\e\_\N\L\P" (PR#207).
• Fix #204: add decoding of unicode codepoints \x \u \U in double-quoted scalars:

  Tree tree = parse_in_arena(R"(["\u263A \xE2\x98\xBA \u2705 \U0001D11E"])");
  assert(tree[0].val() == "☺ ☺ ✅ 𝄞");

  This is mandated by the YAML standard and was missing from ryml (PR#207).
• Fix emission of nested nodes which are sequences: when these are given as the emit root, the - from the parent node was added (PR#210):

  const ryml::Tree tree = ryml::parse_in_arena(R"(
  - - Rochefort 10
    - Busch
    - Leffe Rituel
    - - and so
      - many other
      - wonderful beers
  )");
  // before (error), YAML valid but not expected
  //assert(ryml::emitrs<std::string>(tree[0][3]) == R"(- - and so
  //  - many other
  //  - wonderful beers
  //)");
  // now: YAML valid and expected
  assert(ryml::emitrs<std::string>(tree[0][3]) == R"(- and so
  - many other
  - wonderful beers
  )");

• Fix parsing of isolated !: should be an empty val tagged with ! (UKK06-02) (PR#215).
• Fix #193: amalgamated header missing #include <stdarg.h> which prevented compilation in bare-metal arm-none-eabi (PR #195, requiring also c4core #64).
• Accept infinity,inf and nan as special float values (but not mixed case: eg InFiNiTy or Inf or NaN are not accepted) (PR #186).
• Accept special float values with upper or mixed case: .Inf, .INF, .NaN, .NAN. Previously, only low-case .inf and .nan were accepted (PR #186).
• Accept null with upper or mixed case: Null or NULL. Previously, only low-case null was accepted (PR #186).
• Fix #182: add missing export of DLL symbols, and document requirements for compiling shared library from the amalgamated header. PR #183, also PR c4core#56 and PR c4core#57.
• Fix #185: compilation failures in earlier Xcode versions (PR #187 and PR c4core#61):
  • c4/substr_fwd.hpp: (failure in Xcode 12 and earlier) forward declaration for std::allocator is inside the inline namespace __1, unlike later versions.
  • c4/error.hpp: (failure in debug mode in Xcode 11 and earlier) __clang_major__ does not mean the same as in the common clang, and as a result the warning -Wgnu-inline-cpp-without-extern does not exist there.
• Ensure error messages do not wrap around the buffer when the YAML source line is too long (PR#210).
• Ensure error is emitted on unclosed flow sequence characters eg [[[ (PR#210). Same thing for []].
• Refactor error message building and parser debug logging to use the new dump facilities in c4core (PR#212).
• Parse: fix read-after-free when duplicating a parser state node, when pushing to the stack requires a stack buffer resize (PR#210).
• Add support for legacy gcc 4.8 (PR#217).

Improvements

• Rewrite filtering of scalars to improve parsing performance (PR #188). Previously the scalar strings were filtered in place, which resulted in quadratic complexity in terms of scalar length. This did not matter for small scalars fitting the cache (which is the more frequent case), but grew in cost as the scalars grew larger. To achieve linearity, the code was changed so that the strings are now filtered to a temporary scratch space in the parser, and copied back to the output buffer after filtering, if any change occurred. The improvements were large for the folded scalars; the table below shows the benchmark results of throughput (MB/s) for several files containing large scalars of a single type:

  scalar type	before	after	improvement
  block folded	276	561	103%
  block literal	331	611	85%
  single quoted	247	267	8%
  double quoted	212	230	8%
  plain (unquoted)	173	186	8%

  The cost for small scalars is negligible, with benchmark improvement in the interval of -2% to 5%, so well within the margin of benchmark variability in a regular OS. In the future, this will be optimized again by copying each character in place, thus completely avoiding the staging arena.

• Callbacks: add operator==() and operator!=() (PR #168).

• Tree: on error or assert prefer the error callback stored into the tree's current Callbacks, rather than the global Callbacks (PR #168).

• detail::stack<>: improve behavior when assigning from objects Callbacks, test all rule-of-5 scenarios (PR #168).

• Improve formatting of error messages.

Thanks

• @ingydotnet
• @perlpunk
• @cschreib
• @fargies
• @Xeonacid
• @aviktorov
• @xTVaser