Skip to content
graphviz-pure

Renderer port (crates/graphviz-pure-render)

Copied at build time from docs/porting/render.md.

The output formats are produced by a function-by-function port of the Graphviz 16.1.0 emit pipeline, so SVG, xdot and JSON share one traversal exactly as in C. Node, label, record and HTML geometry come from `graphviz_pure_geom::node_size` / `graphviz_pure_geom::html_label` (the init half of `shapes.c`, `labels.c`, `htmltable.c`); positions come from the layout.

Module map

Section titled “Module map”
Rust C source
emit/model.rs the post-layout Agraphinfo_t/Agnodeinfo_t/Agedgeinfo_t state: GD_clust tree, ND_coord/lw/rw/ht, ND_shape_info, ED_spl, textlabel_t/textspan_t, place_root_label/place_graph_label fallbacks, agfstout order (head id, then sequence)
emit/job.rs lib/gvc/gvrender.c (gvrender_*, gvrender_resolve_color, gvrender_set_style), obj_state_t stack (push_obj_state/pop_obj_state), job geometry from gvRenderJobs (init_job_pad/margin/dpi/viewport/pagination, setup_page, emit_graph scale)
emit/emit.rs lib/common/emit.c: emit_graph, emit_page, emit_background, emit_view (outputorder), emit_clusters, emit_node, emit_edge, emit_begin_edge/emit_end_edge, emit_edge_graphics (multicolor, parallel colors, tapered, ortho rounded corners), emit_edge_label, emit_attachment, initObjMapData/getObjId/preprocessTooltip, layers (init_layering, parse_layers, parse_layerselect, selectedLayer, *_in_layer), parseSegs, findStopColor, stripedBox; emit_label from labels.c
emit/shapes.rs shapes.c gencode: poly_gencode, point_gencode, record_gencode/gen_fields, checkStyle, stylenode, penColor, findFill, Mcircle_hack, round_corners (rounded, diagonals, cylinder), wedgedEllipse
emit/corners.rs round_corners shape cases: note, tab, folder, box3d, component and all SBOL shapes
emit/arrows.rs arrows.c: arrow_flags, arrow_match_name, arrow_gen, every arrow type with its miter/penwidth corrections, Bezier
emit/html.rs htmltable.c emit half: emit_html_label, emit_html_tbl, emit_html_cell, emit_html_rules, doBorder, setFill, emit_htextspans, initAnchor/endAnchor
emit/geom.rs dotneato_closest, splitBSpline, computeoffset_p/qr, get_gradient_points (utils.c), taper (taper.c), ellipticWedge (ellipse.c)
emit/output.rs output.c: attach_attrs_and_arrows, set_record_rects, rec_attach_bb, write_plain
emit/num.rs gvprintdouble, gvprintnum, agxbuf_trim_zeros, %.Ng
emit/attrs.rs late_*, mapbool, parse_style, strtod/strtol
svg_engine.rs plugin/core/gvrender_core_svg.c (incl. gradients, layers, anchors, gv_xml_escape)
xdot_engine.rs xdot half of plugin/core/gvrender_core_dot.c (per-state buffers, penwidth/textflags state, gradients)
json.rs plugin/core/gvrender_core_json.c and the xdot parser of lib/xdot/xdot.c
agwrite.rs lib/cgraph/write.c (-Tdot, -Txdot, -Tcanon) over the cgraph dictionaries recorded by the parser (Graph::dicts: bindattrs, attrstmt/setattr, unviewsubgraphsattr, sym->print, ProtoGraph label)
lib.rs format dispatch, safe_dcl/agxset bookkeeping (sync_dicts), dot’s mark_lowclusters edge membership

graphviz-pure-layout/src/nop.rs ports neato’s Nop path (init_nop, user_pos, user_spline, nop_init_graphs, nop_init_edges, chkBB, compute_bb, neato_translate, neato_set_aspect, place_graph_label) for -n, -n1, -n2, -Knop/nop1/nop2. Edges without a usable pos go through the neatogen spline_edges_ machinery (ports, equivalence classes, straight or splines= routing, clipping).

Fidelity

Section titled “Fidelity”

The render corpus isolates the renderers: C lays every tests/wild, tests/fixtures and tests/render graph out (dot -Tdot), then both C and the port render that file in no-layout mode (neato -n2). See scripts/gen-render-golden.sh and gv-diff render (crates/graphviz-pure-diff/src/render.rs, ratchet tests/golden/render/baseline.json, test crates/graphviz-pure-diff/tests/render_fidelity.rs). The SVG generator comment is the only line ignored. tests/render/*.gv are 21 feature graphs (all arrow types, styles, shapes, gradients, HTML, records, URLs, labels, rotation, size/ratio, layers, output order).

Byte-identical files, original 508-file corpus (before = legacy renderers with -n2 support, after = this port; plain formats exclude 4 nondeterministic C references):

format before after
svg 36 / 508 507 / 508
json 0 / 472 472 / 472
json0 0 / 508 507 / 508
xdot 2 / 508 507 / 508
dot 207 / 508 507 / 508
plain 276 / 504 501 / 504
plain-ext 260 / 504 501 / 504

All 21 feature graphs are identical in every format. -Tcanon (agwrite on the parsed input) matches C on 493 of 508 inputs (454 before).

Known differences

Section titled “Known differences”
  • -Tplain of labels changed by htmlEntityUTF8/latin1ToUTF8 that need no quoting: canon in write_plain (lib/common/output.c) interns the text, agstrcanon returns it unquoted and agstrfree frees it before it is printed, so C prints whatever the freed memory holds (gv16_graphs_Latin1, b60; the charset_* Latin-1 cases, gallery__index_7, Symbol, b34, b56 change from run to run and are skipped, UNSTABLE in gv-diff). The port prints the converted label; the garbage is not emulated.
  • HTML cell STYLE="dashed" borders: C’s xdot_style walks past the end of a one-element style list and prints adjacent string literals (undefined behaviour); not reproduced (tests/fuzz-open/034-html-cell-dashed-style-overread).
  • Images (image=, shapefile=, <IMG>) are never loaded; C draws them only when the file exists.
  • The SVG generator comment names graphviz-pure.
  • Gradient and HTML anchor ids are numbered per render_dot_all call, like C’s process-wide statics for a single invocation.

Charset

Section titled “Charset”

C keeps the input’s bytes and converts only label text. graphviz_pure_core::charset gives the port the same model with Strings: bytes that are not part of valid UTF-8 become the code points U+10FF00 + byte (decode), and encode turns them back into bytes. The CLI decodes its input and encodes its output, and gv-diff reads inputs and C references the same way, so attribute values, names, -Tdot/-Tcanon/-Txdot strings, SVG titles and ids, tooltips and URLs carry C’s bytes (SVG always declares encoding="UTF-8"; gv_xml_escape leaves bytes >= 0x80 alone, and so does the port).

  • make_label, HTML fallbacks and tooltips (preprocessTooltip) convert with latin1ToUTF8 (charset=latin1) or htmlEntityUTF8 (else) on the encoded bytes, once, where C does. htmlEntityUTF8 reproduces C’s recovery for invalid UTF-8 (the offending byte is taken as Latin-1, possibly leaving invalid UTF-8, kept as raw bytes) and its once-per-graph warning.
  • Text width is measured per encoded byte, with textspan_lut.c’s once-per-process “no value for width” warnings.
  • HTML labels: expat gets the graph’s encoding. Latin-1 bytes are decoded to UTF-8, invalid UTF-8 is a fatal error, and BIG-5 is an unknown encoding, so every Big5 HTML label fails. A failed HTML label is plain text with its unconverted source afterwards (lp->html cleared): -Tplain quotes it.
  • charset=big5: labels convert like UTF-8; make_simple_label keeps the byte after a 0xA1..=0xFE byte with it (so a \ there is not an escape). Unknown charsets warn Unsupported charset and act as UTF-8.
  • -Txdot string lengths count bytes; -Tjson runs latin1ToUTF8 over every string of a Latin-1 graph (so label text, already UTF-8, is converted again, as in C) and escapes only ", \, /, \b, \f, \n, \r, \t; other control characters and bytes are written as is.
  • The charset of the graph being processed is kept per thread (charset::begin_layout/begin_render, called by layout_graph_with_options and render_graph_named); NodeSizeEnv::new and -n2 labels use it.
  • Not reproduced: HTML lexer error messages (Error: not well-formed ..., unknown encoding); input that really contains U+10FF00..=U+10FFFF is written out as raw bytes.

Debugging aid

Section titled “Debugging aid”

scripts/c-probe.c prints C’s exact (%.17g) node coordinates, sizes and shape vertices after -n2, for chasing last-bit floating-point differences.