Skip to content

Silinx 15.25.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 11:22
· 793 commits to main since this release

What's new in 15.25.0

The 7-series and three more families come into the Silinx toolchain: the Basys 3 and Nexys A7 (Artix-7), the
Lattice ECP5 (ULX3S) and the Gowin GW1N (Tang Nano 9K), from HDL to the board, in the browser too. Projects are shared
live and graded on GitHub, and the drawing editors get a movable Try it, the Symbol Editor and arrayed instances.

In short:

  • The 7-series in the Silinx toolchain (experimental): the Basys 3 and the Nexys A7 implemented with no Vivado,
    in the application and in the browser, every example to its bitstream; the logic analyzer on them and on the
    Spartan-6.
  • More families: the Lattice ECP5 (ULX3S) and the Gowin GW1NR-9C (Tang Nano 9K), synthesized, placed and routed,
    their bitstreams written and their boards programmed from Silinx; the Board Emulator draws them, and the Basys 3 and
    the Nexys A7.
  • Sharing: one Share Project… for a copy or live (followers and editors, live text editing, reconnection on its
    own), short links that can expire, a QR code; Git and GitHub in every edition (history, pull and push, a drawing's
    changes and conflicts object by object, GitHub Checks in the console).
  • Teaching: a Teaching tab, guided labs written in Silinx, exercises packaged for students and graded
    automatically (GitHub Classroom annotations).
  • Drawing editors: the Try it panel resizable and docked below or on the left, the drawing fitted to its view
    whenever the layout changes; a state's and a transition's outputs set in a table; schematic junction dots
    highlighted and movable; arrayed instances and the Symbol Editor; the Design pane in natural order (x_2 before x_10).
  • Verification and timing: Verify Assertions, Equivalence and Property; SystemVerilog concurrent assertions
    simulated; the critical path shown in the drawings and the reports.
  • Silinx HLS: AXI4 bursts, pipelined loops over several cycles, a sub-module per call; long statement chains
    accepted; the corpus measured against Vitis HLS 2025.2 (68 of 73 designs match).
  • Boards: the Numato Mimas V2 and Elbert V2's displays on the right pins, their flash programmed; the examples keep
    time on every board (their clock as a generic).
  • Robustness and security: deeply nested or very long designs handled without stack overflows; paths, links,
    shared projects and pulled repositories checked; messages in the console, not in toasts.

The full list follows.

In detail:

  • The project's description told by its 64-bit file id (#3878): isProjectFileAt (server/projects.js) compares
    the files' device and file id as bigint: Windows's file ids are past a Number's precision, and two files of a
    folder could round to the same one (the release CI's Windows run, test/drawing-nesting.test.js #3559). Test:
    test/project-store.test.js.
  • CI: WebKit's jobs on Ubuntu 24.04 (#3854, #3855): browsers and webkit-standalone run on ubuntu-24.04:
    ubuntu-latest became Ubuntu 26.04 (its image of 2026-11-11), whose libicu 78 Playwright's WebKit cannot load
    (libicudata.so.74), so WebKit stopped at its start. Test: test/ci-workflow.test.js.
  • The nightly CI of 14 November, fixed (#3856, #3857, #3858, #3860, #3861, #3862, #3863): the symbol editor's own
    clone replaced by core/util.js's (#1406's one copy); the Spartan-6's Yosys script's BSCAN clock-driver step (#3766)
    in test/synth-open.test.js; blinky's step assignment at its new line (the examples' generics) in
    test/twr-paths.test.js; on Windows, the temporary folder's absolute path refused as a drive letter in
    test/server-paths.test.js; the VHDL LRM's section number in a commit message allowed by the private-data scan.
  • The Elbert V2's digit enables right to left (#3847). Numato's diagram wires Enable[0] (P120) to the leftmost
    digit and Enable[2] (P124) to the rightmost. The table had an in Numato's order, so an example's an<0> (the
    ones) lit the leftmost digit. an is now P124, P121, P120, its an<0> the rightmost digit as on the other boards.
    Tests: the examples' digit pins and ISE's build warnings on the Elbert V2, and ISE's build of the clock for it.
  • The student's copy and the Classroom template from the interface and the MCP (#3841, part of #3191):
    • core/grade-package.js:
      • starterProject, the starter project from the statement: the top module with the statement's ports and an
        empty body, on the teacher's board and device;
      • templateFiles, the template's files, made in the page or on the server.
    • server/classroom.js makeTemplate writes the same files. silinx classroom template makes the starter project
      from the statement when none is given (--language vhdl|verilog).
    • In the interface, the Exercise document's ⋮ menu has Export Student Copy and Classroom Template…
      (web/js/exercise-package.js; Portuguese labels).
    • In the MCP server, grade template writes the template to a new folder of the exchange folder.
    • Tests: grade-package (the starter compiled with the statement's ports and graded, in VHDL and Verilog; the files
      the same as makeTemplate's, nothing hidden in them), classroom (the command with no starter), mcp-grade, UI
      exercise (both interfaces, Portuguese); docs/CLASSROOM.md and docs/GRADING.md.
  • The Classroom template's text in core/ (#3841, part of #3191): the workflow's and the README's text
    (workflowYaml, readmeText, CHECKOUT, SETUP_NODE, WORKFLOW) moved as it was from server/classroom.js to
    core/classroom-template.js, so the browser editions can make the template too. server/classroom.js exports the
    same names. Tests: classroom-template (the old module's names, the workflow, the README), classroom.
  • Shared projects, phase 4a: a follower's and an editor's connection made again (#3835, part of #3180):
    • When: the connection is lost once the project has arrived, and nothing ended it on purpose: the owner did not
      stop sharing, the follower did not leave, no limit was passed, and the room did not refuse it (close codes 1003,
      1008, 1009). A new try follows after 1, 2, 4, 8 and then every 15 s, for 5 minutes at most, and at once when the
      browser is back online (core/collab-client.js followRoom's reconnect, RECONNECT). A room its owner stopped
      sharing meanwhile ends the session, with the room's reason (NOT_SHARED, the room's own words, core/collab-room.js).
    • The same participant: the page keeps the same document and client id. The owner sends the whole document
      again, as at any join. Who is in comes back on both sides: the others' states are forgotten at the loss, clocks
      included, and this participant's own is sent with a newer clock, so neither side ignores it.
    • An editor's changes: its own updates are kept. When the owner's whole document arrives, the editor sends,
      signed, what of them the owner's document lacks (its updates past the owner's clock of this client; its deletions
      with them). This also covers changes sent into a connection that was already lost before its close was noticed.
      Nothing else of the document is sent, so the owner's edit checks (#3527) refuse nothing.
    • A connection that never opens: it now fails when its socket closes before it opens, or after 10 s, instead of
      waiting forever.
    • Security: the keys and the signatures are unchanged. A rejoin is a join, checked as one by the room. A client id
      another participant holds is still refused, so the old connection must be gone first: the relay's pings end it. An
      editor's resent changes are signed with its link's key and checked as any edit.
    • The page (web/js/collab-follow.js): "Reconnecting to …" in the toolbar while lost (Portuguese: "A
      religar a …"), with no one else shown. The console says "WARNING: The connection to 's project … was
      lost: reconnecting …" and "Reconnected to 's project …". A room the owner's page hosts (WebRTC) is not
      reconnected yet: that is phase 4b.
    • The MCP (follow_project): its follow reconnects the same way. Meanwhile its result says so (connected: false), and a write is merged then and sent once the connection is back.
    • Tests: test/collab-reconnect.test.js (the rejoin, an editor's changes made offline and lost on the way, a
      deletion, a follower sending nothing, the ends that are not made again, given up, close() while lost),
      mcp-collab (#3835: a follow cut and back, a write meanwhile on the disk, no warning), UI collab-reconnect (an
      editor's page cut and back in both interfaces; a follower's page in Portuguese).
  • Reference from the Project and Expected from the Reference (#3831, part of #3191): core/grade-author.js.
    • projectReference: the project's design as the exercise's reference. A file inside the exercise's folder is listed
      by its path, so it follows the design as the teacher edits it. A drawing's HDL, or a file outside the folder, is
      written in.
    • expectedFromReference: a vectors test's rows filled with the reference's outputs. The inputs, the comments and
      the header stay as written. Each output keeps its column's style (decimal by default, binary, 0x…), and an X or
      Z of the reference is -.
    • Both are in the Exercise document's ⋮ menu, with Portuguese labels, and in the MCP server's grade tool
      (reference, expected). The MCP runs the reference in a grading process (server/grade.js expectedOf).
    • Tests: grade-author (the counter's rows against the hand-written ones, then graded; the styles; an X; the
      problems; a drawing's reference), mcp-grade, UI exercise (both interfaces, Portuguese); docs/GRADING.md.
  • Rename of a module and its symbol file (#3829, core/symbol-layout.js renameSymbolFile): the interface's Rename…
    (web/js/rename.js) and the MCP server's rename_module (server/rename-module.js) write the module's new name in its
    symbol file and move a file named after the module with it, beside its source. It was left as it was: the module had
    no symbol any more (its instances marked out of date) and the old file was the symbol of no module.
  • Try it on an arrayed instance (#3827, core/schlive.js): the instance's pins (the Try it panel's list, a pin's
    value) read from its copies reg_3 … reg_0: a pin sliced over the copies, or left open, their values side by side
    (the high end's copy first, as on its bus), a shared one its value. It was found by its name, which no copy has: a
    pin left open had no value.
  • MCP (#3827): simulate_drawing simulates a schematic (.sch.json) as its Try it does, with the project's modules:
    per cycle its input markers set and its clock inputs cycled, then its outputs and every module instance's pins (an
    arrayed one's, its copies' values side by side) (server/mcp-simulate-schematic.js).
  • The hierarchy's names in natural order (#3819). The Design pane's roots (the Implementation view's top first, the
    Simulation view's), Simulate's choice of the test bench, Set Top Module…, the design link's module list, Verify
    Equivalence's reference, a schematic's module palette and its instance's Module list sorted the names as text
    (x_1, x_10, x_2): they now compare the numbers in a name by their value (x_1, x_2, x_10), the rest case-insensitively
    as before, through one comparator (core/hier-order.js). A module's instances, listed in their source's order, are
    sorted by their names the same way (then by their modules'). The MCP's project_info gives the roots of both views in
    the same order (hierarchy).
  • The RTL and Technology Schematic's translation test: a junction dot an object (#3818): since #3806 a junction dot is
    an object of its own in the read-only schematic views too (clicked, it is selected); test/ui/open-toolchain.test.js's
    #130 test lists it among the objects it clicks (both views), and its texts are translated.
  • Boolean Algebra editor: the Inputs and Outputs text boxes wider (#3815): 280 px rather than 140 px, shrinking and
    then wrapping on a narrow toolbar. test/ui/truthtable.test.js.
  • State diagram editor: a state's and a transition's outputs as a table of the machine's outputs (#3814). The
    Outputs text box ("z=1, y=2'b10") of a state's and of a transition's Properties is replaced by a table: one row per
    output of the machine, its name and width (read-only) and its value (editable). An empty value leaves the output unset:
    its default in a state, the state's value on a transition (shown greyed in the field). A value is checked as it is
    typed (a constant that fits in the output's width): a value refused is marked, and when the field is left the console
    says why and the value in force comes back. An output renamed in the machine's Properties keeps its values in the
    states and transitions, and one removed takes its values with it. Files are saved as before (an unset output left
    out), so the generated VHDL and Verilog are the same for the same values (core/fsm-outputs.js). The MCP server's new
    set_fsm_outputs sets a state's or a transition's outputs with the same checks.
  • Schematic editor: a selected wire's junction dots highlighted with it, and a junction dot movable (#3806). The
    dots on the selected wires are drawn in the selection's colour. A click selects a dot. Dragging it, or the arrow keys,
    moves it: each wire ending there has its end segment moved across, as a dragged segment is, and stretched along. A jog
    is added instead where something else sits on that segment (another wire's end, a net name). A pin or an I/O marker
    under the dot stays where it is, joined to the dot by a new wire. The connections stay the same: a move that would
    short two named nets is refused, and one that changes them is reported. Each move is one Undo step
    (core/schjunction.js, web/js/sch-junction.js). The MCP server's new move_junction does the same move.
  • The drawings zoomed to Full View when the layout changes (#3804). web/js/layout-fit.js (#1203) now also serves
    the state diagram, ASM chart, Petri net and symbol editors and the waveform, each with its own Zoom to Full View;
    the Split Editor's split and unsplit (web/js/split-view.js) signal it as the Tasks bar's toggles do. A drawing hidden
    at the toggle is fitted once shown, if its size is not the one it had when last shown; the user's own resizes (the
    window, a divider) fit nothing.
  • The drawing editors' bottom panel docked on the left (#3802): web/js/dock-panel.js docks the Try it panel
    (and the FSM's Tables, the Petri net's Analysis) below the drawing or on its left, between the Design pane and the
    canvas, its right edge the one that resizes it (← → on it); a place kept on the right by #3797 reads as the left.
    Whenever the panel moves (its title bar dragged, its ⇆ button, or a narrow screen putting it below), the drawing is
    zoomed to fit the area it is left with, as Zoom to Full View, once the layout has the new size; a plain resize keeps
    each editor's behaviour. Tips in Portuguese; unit and UI tests (both interfaces) updated.
  • The ASM chart's Try it without its fold button (#3800): the ▾ / ▴ button that left only the panel's title bar is
    removed, with its kept setting (silinx.asm.sim.folded) and its Portuguese; the panel is resized by its edge and
    moved instead (#3797), as the state diagram's and the Petri net's.
  • The drawing editors' Try it green (#3798): the Try it panel's title bar in the schematic's Try it green, its
    title as the schematic's badge (bold green, a lit dot), in the light and dark themes; the FSM's Tables and the Petri
    net's Analysis, the same panel, stay grey.
  • The drawing editors' bottom panel resized and moved (#3797): one module for the three editors
    (web/js/dock-panel.js): its inner edge dragged (a finger or a pen too; ↑ ↓ / ← → on the focused edge) sizes it
    between 15% and 85% of the drawing area, the drawing kept in view (the FSM fitted again); its title bar dragged, the
    place it will go shown, docks it below the drawing or on its right, as its ⇆ button does. Docked rather than
    floating, so that it never covers the drawing being tried. Its place and size are kept per editor kind
    (silinx.<kind>.trypanel), when it is closed and opened again and after a reload; the FSM's Tables and the Petri
    net's Analysis, the same panel, follow them. On a narrow screen it is always below, still resizable. Its tips in
    Portuguese.
  • The schematic's Try it named as its button (#3795): the live simulation's badge, its Properties panel's title
    and its hint say "Try it" (Portuguese: Experimentar), as the button that starts it and the ASM, FSM and Petri net
    editors' panels, not "Trying it out".
  • The FSM editor locked while Try it is open (#3793): a state is not moved, an arrow not bent, no transition drawn,
    no state added and nothing edited in place; Delete, Cut, Paste, Undo / Redo and the toolbar's State, Delete, Arrange,
    Encoding and Coding Style are disabled; a state or a transition is still selected, its properties shown with their
    fields disabled and their buttons hidden, as in a read-only diagram. The view is still panned and zoomed, and the
    simulation stepped. A text being typed in place is applied when Try it opens; closing Try it unlocks the diagram.
  • The Symbol Editor's status line (#3792) has a short how-to at its left, as the Schematic editor's tool tip there
    (none when read only); in Portuguese.
  • Update Symbol in a module instance's right-click menu (#3791), after Edit Symbol: what its Properties' button
    does (web/js/sch-modsym.js updateSymbol, shared by both: one Undo step, the same toast); disabled, its tip "The symbol
    is up to date", when it is.
  • Edit Symbol in a module instance's Properties (#3791), beside Open module source: its module's symbol in the
    Symbol Editor; hidden in a read-only schematic.
  • The Symbol Editor's drag (#3789): no text of the drawing selected; while a pin is dragged, its handle under the
    pointer and a dashed bar across the side's leads where it will be dropped. Reset to Default moved from the ⋮ menu
    (now gone, it held nothing else) to the toolbar, beside ◀ Earlier and Later ▶.
  • The schematic of a table: complemented inputs as the gates' inverted inputs (#3786). In every form but the NAND
    and NOR ones (which stay NAND or NOR only), a complemented input of an AND, OR, NAND or NOR gate is one of its
    inverted inputs, the library's ANDnBk, ORnBk, NANDnBk and NORnBk as in ISE's (core/logic-sch.js); INVs are left for
    the inputs of an XOR / XNOR (the XORnBk symbols have no ISE counterpart) and for an output that is a lone complement
    (one INV, no longer an INV and a BUF). Tests: every form of many functions, one inverted input per complement under
    an AND / OR, the INVs only where no gate fits, the HDL compiled, elaborated and simulated to the table in VHDL and
    Verilog with the warnings asserted, the ISE export with library symbols only (test/logic-sch-bubbles.test.js).
  • The NAND and NOR forms: a one-input NAND / NOR written with its inputs tied (#3785). formatExpr writes it
    /(a.a) ((a·a)', /(a + a)) as the HDL already did (a nand a) and as the schematic draws its inverter (a NAND2
    with its inputs tied); the forms' tooltips say so. In the schematic of a table whose outputs are in several forms, the
    complemented inputs of an output in the NAND (NOR) form are NAND (NOR) gates of its own, not the rails' INVs
    (core/logic-sch.js). Tests: every function of up to 3 inputs and random ones of 4 to 6, each notation parsed back
    with NAND (NOR) gates only and equal to the table; the schematics NAND / NOR only, their HDL compiled, elaborated and
    simulated to the table in VHDL and Verilog, the warnings asserted (test/logic-gate-forms.test.js).
  • A design and simulation project, in the core and the MCP server (#3780, part 1): a project's kind
    (core/project-kind.js): "kind": "design" in silinx.json is a project of drawings and simulation only, with no
    board, no constraints and no HDL shown (the core keeps making and simulating the drawings' HDL). The MCP server's
    new_project makes one (design_only), project_info shows its drawings and no HDL, and the flow's and the HDL's
    tools refuse it with the messages the interface will show (server/mcp-design-only.js); set_design_properties
    makes it an FPGA project with a board (or a device), and the reverse with design_only when every source is a
    drawing. The local server's and the browser edition's implementation refuse it too; its tool descriptor is
    DESIGN_ONLY (Check Design; the Design and Simulate tasks).
  • A design and simulation project in the interface (#3780, part 2): New Project's Project Kind (its Project
    Settings page then the Top-Level Source Type alone, a drawing); Design Properties' board list with "Design and
    simulation only (no FPGA)" first, both ways (the reverse refused in the console while the project has HDL or C
    sources of its own). In such a project the Design pane has no [VHDL] / [Verilog] badges and its row is "Design",
    the Files tabs list no HDL, an HDL file opened is its drawing, an HDL file added or imported is refused with the MCP
    server's message, New Source offers the drawings and the labs with no language, the Test Bench Wizard writes no
    skeleton and simulates the bench it makes, Source Properties, the schematic's Properties, the state diagram's Coding
    Style and the truth table's HDL notations ask no language, and Check Syntax is Check Design; the process tree, the
    tasks (Design and Simulate), the menus and both toolbars have no Design Summary, implementation, emulation, pin
    planning or programming (web/js/design-only.js, web/js/project-kind-ui.js, the DESIGN_ONLY descriptor).
  • The logic analyzer on the 7-series (#3767, part of #3483): Implement inserts the capture core with a BSCANE2 on USER1
    (core/flow-ide.js: the 7-series among the families whose open flow places the BSCAN; the report's link named by the
    family's BSCAN). The board's clock (#3481) as XDC: its PACKAGE_PIN, IOSTANDARD and create_clock (core/ila-insert.js
    boardClockXdc), a port of the design on its pin refused. The core's samples written on the clock's falling edge on the
    7-series (core/ila.js: the same words; the block RAM's write port timed on that edge by nextpnr, no hold violation), its
    memory planned as synth_xilinx infers it (RAMB18E1s, a RAMB36E1 two, distributed RAM where Yosys's costs choose it: the
    same counts on 18 shapes); capture_settings's block RAMs by the family's profile (none counted against the device on the
    7-series). The 7-series' tree has the logic analyzer (web/js/tool-descriptors.js); reading a capture on a board says it is
    the Spartan-3E's boards only yet (core/ila-link.js captureLinkError, #3773), in the server and the browser. Tests:
    ila (the core simulated on the 7-series in VHDL and Verilog, its memory planned and synthesized), ila-insert (the wrapper
    with its BSCANE2 synthesized, no warning), ila-link, tool-descriptors, mcp-capture-xc7 (blinky's capture to a bitstream on
    the Basys 3, no warning; gates on the board's clock; the capture read on the board refused), UI capture (the dialog on the
    Basys 3, both interfaces, Portuguese).
  • The logic analyzer on the Spartan-6 (#3766, part of #3483): Implement inserts the capture core with a
    BSCAN_SPARTAN6 on USER1 (core/flow-ide.js: the Spartan-6 among the families whose open flow places the BSCAN, the
    report's link named by the family's BSCAN). The toolchain's part: Silinx's Map packs it onto a BSCAN site, on the site
    of its JTAG chain (BSCAN_X0Y<chain - 1>: on the Spartan-6 the site decides the chain, as ISE's bitgen shows), the XDL
    writes its JTAG_CHAIN; nextpnr-spartan6 0.5.1 places and routes it there (its chip database's BSCAN BELs, its packer's
    BSCAN, pmnmalo/nextpnr-spartan#43, #45). DRCK, the JTAG clock, stays on the general routing (no
    BUFG: no switch from the fabric into a BUFGMUX is in the bit database). The dialog's note names the Spartan-6. Tests:
    s6-bscan (synthesis, Map, placement, XDL on both parts), mcp-capture-s6 (blinky's capture to a bitstream on the Nexys3
    and the Atlys, with both Maps, no warning); checked against ISE (DRC, the same design's bitstream byte-identical).
  • The FSM editor's Machine section only with nothing selected (#3764): a state or a transition selected shows its
    own section in the Properties panel, no longer the Machine section under it.
  • A new transition's condition typed in the Properties panel kept (#3763): the Condition field clicked while the
    transition's text was still open on the drawing lost the focus (the text closed drew the panel anew); and every
    field of the panel kept only its first key, writing to the diagram's objects as they were when the panel was drawn,
    no longer the diagram's after a change. The fields write to the diagram as it is now; a text in place left as it was
    changes nothing, and a field being typed in keeps the focus when the panel is drawn anew.
  • A pull request runs three more of the repository's checks (#3761). test/unused-exports.test.js,
    test/url-pathname.test.js and test/web-markup-interpolations.test.js scan the whole repository, as the others of
    REPO_WIDE do: a change that broke them passed its related-tests and failed only in the nightly and the full CI
    (#3728, #3732, #3734). They run for every change now (about 1 s together).
  • An arrayed instance (#3755, core/sch-array.js): its name base(msb:lsb) (or base[msb:lsb]), accepted by Rename
    and the Properties for a module's instance; the netlist widens its pins on a bus N times as wide (or, alone on a net
    with an arrayed output, makes it one); Check Schematic says a width mismatch as before for any other width, and an error
    for an output shared by the copies. The generated VHDL and Verilog write the copies one by one, base_i, each pin its
    slice of the bus (a bus named from another index, or bits of a bus, too); compiled, elaborated and simulated in both.
  • MCP (#3755): place_instance places a module's instance in a schematic (with its symbol), arrayed with a range,
    its pins connected to nets by name; its check's errors and warnings said.
  • A module's symbol file, <module>.sym.json beside its source (#3755). It holds the pins' order and sides, the
    spacing, the width and height and the name's place (core/symbol-layout.js); the project's modules read it (the
    interface's and the server's), its pins kept in step with the module's ports (a new port added on its default side, one
    gone dropped). A module without one is drawn as before.
  • The Symbol Editor (#3755), a document of its own (web/js/symbol-editor.js): the symbol drawn as the schematic draws
    it, a pin selected and dragged to its new place, ◀ Earlier / Later ▶, the Properties (the pin's Side and Position, the
    symbol's spacing, width, height and name), Reset to Default in its ⋮ menu; the main toolbar's Undo and Redo; its
    Properties panel foldable and behind ☰ on a narrow window; its status line its pins and size; its messages in the
    console; in Portuguese.
  • An instance keeps the layout it was placed with (#3755), as it keeps its ports (#1183): a changed symbol marks it
    out of date (Check Schematic's warning, its Properties' note) until Update Symbol, the wires following their pins. Its
    port map is written in the module's order, whatever the drawing's, so a symbol redrawn never changes the HDL.
  • MCP (#3755): read_symbol and edit_symbol (a module's symbol, as the Symbol Editor reads and edits it) and
    update_symbols (a schematic's Update Symbol).
  • The browser's flow worker: a family's own tools chosen by comparing its name (#3753). The worker looked up the
    family's tools (the iCE40's, the ECP5's, the Gowin's) in a Map by the name the page sent (CodeQL's
    js/unvalidated-dynamic-method-call, still after #3662); it now compares the name with those three, and any other
    family is the same error naming it.
  • The 7-series' Vivado timing script: --filter takes examples' names (#3752). scripts/xc7-timing-vivado.mjs built a
    regular expression from its --filter argument (CodeQL's js/regex-injection). It now takes names separated by | or
    ,, in parentheses or not, with ^ and $ as before (^(blinky|pong)$ matches the same examples); anything else
    is an error naming the filter.
  • The Cmod S6 drawn in the Board Emulator (part of #3745): web/js/emulator-cmod-s6.js, registered in
    BOARD_LAYOUTS, after the photo of its reference manual: 2.6" x 0.7", the DIP pins along both long edges (1..24 and
    25..48), the micro USB on the left end, the FPGA in the middle, LD3..LD0 in a column, BTN0 and BTN1, the header J4
    on the right end; the board has no power switch, the emulator's is by the USB, its LED the POWER LED; no maker's
    logo; I/O Pin Planning's Board view draws it too. The UI tests check the board's size and parts (on a design of its
    own: 4 LEDs and 2 buttons), where they are, that each control is what sits at its centre, in both interfaces and in
    Portuguese; test/web-emulator.test.js asks a layout for at least three kinds of parts, or all the board has.
  • The Elbert V2 drawn in the Board Emulator (Fixes #3745): web/js/emulator-elbert-v2.js, registered in
    BOARD_LAYOUTS, after Numato's connection diagram and the board's photo: the green PCB, the DC barrel jack, the
    power select jumper, the mini USB, P8 and the JTAG header (P3) down the left edge, the three LM317s, the flash, the
    PIC with its ICSP and FWUP, the 12 MHz crystal, the FPGA, the VGA, the stereo jack and the micro SD along the top,
    SW1..SW6 top right over the upright DIP switch (DPSwitch0 at the top, ON to the left), the 3 digits bottom left, each
    where its enable's pin is (P120 the left one, S3; P124 the right one, S1), D8..D1 left to right, the headers P4, P2,
    P6, P1 along the bottom edge and P5 under the display; the board has no power switch, the emulator's is by the USB;
    no maker's logo. The UI tests check the board's size, its parts and where they are, the silkscreen, each control at
    its centre, in both interfaces and in Portuguese.
  • The iCEBreaker + PmodVGA drawn in the Board Emulator (part of #3745): registered in BOARD_LAYOUTS with the
    iCEBreaker's layout (web/js/emulator-icebreaker.js), which draws the PmodVGA over PMOD1A and PMOD1B when the
    board has the VGA's pins: its two headers, its VGA connector on its far edge, the colours' and syncs' pins. The UI
    tests check the board's size and parts, the module over the board, that each control is what sits at its centre, in
    both interfaces.
  • The Mimas V2 drawn in the Board Emulator (part of #3745): web/js/emulator-mimas-v2.js, registered in
    BOARD_LAYOUTS, after Numato's connection diagram and photo: the black PCB, the DC jack and the mini USB on the left
    edge, the digits S1..S3 (SevenSegmentEnable[2..0]) and D1..D8 (LED0..LED7) over P6..P9, SW1..SW6 (Switch0..5) on
    the right, the DIP switch's positions 1..8 (DPSwitch7..0) in the bottom-right corner; the board has no power switch,
    the emulator's sits by the USB; no maker's logo; I/O Pin Planning's Board view draws it too. The UI tests check the
    board's size and parts, where they are, that each control is what sits at its centre, in both interfaces and in
    Portuguese.
  • The Papilio One drawn in the Board Emulator (part of #3745): web/js/emulator-papilio-one.js, registered in
    BOARD_LAYOUTS, after the board's photo in its datasheet: the red PCB, the DC jack, the PWR LED and the mini USB
    along the top edge, the FT2232 with its crystal, RX / TX LEDs and JTAG header, the jumpers, the 32 MHz oscillator,
    the FPGA and the SPI flash, Wing2 (C) down the left edge, Wing1 (B, A) down the right; the board has no LEDs,
    switches or buttons of its own, so the wings' pins 0..15 carry the design's ports beside them; the board has no power
    switch, the emulator's is by the jack, its LED the PWR LED; no maker's logo. The UI tests check the board's size, the
    ports on the wings, the silkscreen, in both interfaces.
  • The Spartan-3 Starter Board drawn in the Board Emulator (part of #3745): web/js/emulator-s3-starter.js,
    registered in BOARD_LAYOUTS, after the component locations of its user guide (UG130, figure 1-2): the green PCB,
    A1 and A2 along the top edge, the VGA connector, the power jack and the RS-232 DB9 (J2) down the left edge, B1 and
    the PS/2 port (J3) on the right, the FPGA, the Platform Flash, the JTAG headers J7 and J5, PROG and DONE; the
    board has no power switch, the emulator's is by the power jack, its LED the POWER LED; no maker's logo; I/O Pin
    Planning's Board view draws it too. The UI tests check the board's size and parts, where they are, that each control
    is what sits at its centre, in both interfaces and in Portuguese.
  • The Spartan-3A Starter Kit drawn in the Board Emulator (part of #3745): web/js/emulator-s3a.js, registered in
    BOARD_LAYOUTS, after the photos of its user guide (UG334): the 5 V jack and the power switch in the top-left
    corner, the VGA, PS/2 (J28), RS-232 DCE (J36) and DTE (J27) connectors and the audio jack along the top edge, the
    SUSPEND switch, Ethernet (J32) and USB on the left edge, the FX2 connector (J17) and the headers J20 and J18 on the
    right edge, the differential headers J2 and J15, the LCD in the bottom-left corner, LED7..LED0 over SW3..SW0, the
    rotary knob among the four buttons in the bottom-right corner; no maker's logo; I/O Pin Planning's Board view draws
    it too. The UI tests check the board's size and parts, where they are, that each control is what sits at its centre,
    in both interfaces and in Portuguese.
  • The Tang Nano 9K drawn in the Board Emulator (part of #3745): web/js/emulator-tangnano9k.js, registered in
    BOARD_LAYOUTS for tangnano9k and tangnano9k-vga, after the board's photo and pinout, turned a quarter: the
    USB-C and S1 / S2 on the left end, LED1..LED6 by the FPGA, the 27 MHz oscillator, the flash, the LCD connectors, the
    HDMI on the right end, the headers J7 and J6 along the long edges with their pins' numbers; J7's breadboard (the
    4-digit display, or the VGA connector) and J6's (LD0..LD7, SW0..SW7, BTN0..BTN3 under their pins) wired to them; the
    board has no power switch, the emulator's is by the USB-C; no maker's logo; I/O Pin Planning's Board view draws it
    too. The UI tests check the board's size and parts, where they are, that each control is what sits at its centre, in
    both interfaces and in Portuguese.
  • The ULX3S drawn in the Board Emulator (part of #3745): web/js/emulator-ulx3s.js, registered in
    BOARD_LAYOUTS for ulx3s and ulx3s-vga, after the board's top view in its repository: the GPIO headers J1 and
    J2 down the long edges, US1, the audio jack, GPDI1 and US2 along the top, D7..D0 and F1 / F2 top left, the DIP
    switch SW1 and PWR top right, UP / LEFT / DOWN / RIGHT bottom right, the FPGA, SDRAM, LCD and JTAG headers, the SD
    slot; the breadboard left of it wired to J1 (the 4-digit display, or the VGA connector); the board has no power
    switch, the emulator's is by US2; no maker's logo; I/O Pin Planning's Board view draws it too. The UI tests check the
    board's size and parts, where they are, that each control is what sits at its centre, in both interfaces and in
    Portuguese.
  • VHDL: a default of an interface list that uses an earlier object of the same list refused (#3744). VHDL-93 (LRM
    4.3.2.1) and XST do not allow it: a generic's default that uses an earlier generic of its entity or component, a
    port's that uses an earlier port, a subprogram parameter's that uses an earlier parameter. Check Syntax reports
    "Initializing expression of generic 'tick' is not valid in this context as it contains reference to other interface
    objects (LRM 4.3.2.1): it uses generic 'clk_hz' (line 3)." at the use; a package's constant, a selected name and a
    generic in a port's default stay accepted. The Design pane's VHDL translation of a Verilog module writes such a
    parameter's default with the earlier parameter's default in its place (D : integer := 8 * 2, not W * 2).
  • Board Emulator: constraints made for another board said, their regeneration offered (#3742). Before it runs, the
    emulator checks the constraints' pins against the board with Implement's check (boardConstraints in
    web/js/implement.js, Implement's checkConstraints now calls it): a warning lists the pins that are not on the board,
    and the dialog Constraints do not match the board offers to regenerate them; declined, the emulator does not run and
    says why. A port is the serial terminal's only in the serial pin's direction (core/uart.js uartBit): an input on
    the board's TX pin (the Basys2's clock pin B8 on a Numato Mimas V2) is no longer taken as the terminal's RX, nor held
    high as an idle RX.
  • The Elbert V2's pin table as the other boards' (#3741, as the Mimas V2's, #3709). Numato's SevenSegment[0]
    is the decimal point and [7..1] are the segments a..g (its demo's "abcdefgh"). The table had them in Numato's order
    as seg, and the digit enables as seg_en, so an example's seg<0> (a) took the decimal point's pin and its an
    and dp went to header pins. The table now has seg (a..g), dp and an (the 3 digits, Enable[0..2]; an
    example's an<3> goes on a header pin), all active low (common anode, active-low enables, Numato's demo).
  • The 7-series timing measured with Vivado: the netlist named after its top (#3738): scripts/xc7-timing-vivado.mjs's
    measure.tcl read Silinx's netlist as netlist.edf, and Vivado's link_design -top found no top module in it
    (Project 1-68); it reads it as <top>.edf (blinky on the Basys 3: Vivado 198.8 MHz, nextpnr 241.6 MHz).
    test/xc7-timing-vivado.test.js.
  • The Nightly's corpus jobs run again (#3734, #3735, #3736). #3543 had stopped exporting fetchYosysTests and
    YOSYS_TEST_DIRS from scripts/fetch-yosys-tests.mjs, which test/corpus/run-corpus.mjs imports dynamically by
    the pieces of its path ("fetchYosysTests is not a function"): both exported again. test/unused-exports.test.js
    sees a module named by a bare file name (a piece of a path.join) as used, and a new check fails on a name taken
    from a dynamic import (const { a } = await import(…)) that its module does not export; web/js/wizards.js's
    three exceptions leave its list, now seen used.
  • The Nightly's VlogHammer run tells Yosys's alumacc bug from a mismatch (#3733). wideexpr_47035's
    6'sb101111 - $signed(-((3'sb111) << s3)) gives 0xfff3 in Silinx's simulation and in Yosys's front end, and
    0xffeb in Yosys's netlist: Yosys's alumacc merges the 3-bit negation into the 16-bit subtraction, its wrap lost
    (YosysHQ/yosys#3278, open). Silinx's open synthesis gives the right value (its 0 - x, #743). A module whose Yosys
    netlist differs, but agrees with synth -noalumacc, is now yosys-bug, reported, and fails no run.
  • test/proto-names.test.js runs on Windows (#3730). It took the repository's folder from its file URL's
    pathname (/D:/…, read as D:\D:\… on Windows): fileURLToPath now, as test/url-pathname.test.js asks, which
    had failed on every operating system's unit tests since #3560.
  • The markup check reads the boards drawn as the real ones (#3728). test/web-markup-interpolations.test.js
    failed on the Basys 3's and Nexys A7's drawings (#3706) and the Spartan-3 / 3A Starter's, Mimas V2's and Cmod S6's (#3745):
    their connectors' helpers (pmodSide, unit, expTop, pins, header40: numbers, and a name through text,
    which escapes it) join the Board Emulator's helpers, a shift (i >> 1) counts as a number in a coordinate, the Cmod S6's
    PIN_X as DIG_X, its dip as markup made by a template.
  • A push button clicked while the worker runs a slice of cycles (#3727). The compiled engine's worker runs a
    slice at a time, up to a quarter of a second's cycles when the page is slow: a click during one reaches the design
    when that run is over, but its release was counted from the page's older time, so it was let go as the run came
    back and the design saw no press (the UI test of #3109 timed out in CI). A press made during a run in flight is now
    counted from that run's end (core/emulate-press.js: press(key, { now, busy }), tick(now, { settled })).
  • The private-data check passes on GitHub's hosted runners' accounts (#3726). runneradmin (macOS and Windows)
    and its short 8.3 name RUNNER~1, which a Windows fix's commit message named (#3611), are no one's home folder,
    as runner (Linux) was not already: CI's scan of the history failed on them.
  • Share Project's button after Close Project (#3719). The main toolbar reads New Project, Open Project, Close
    Project, Share Project, Print, in both interfaces.
  • Share Project: a button of the main toolbar (#3716). Its share icon (three joined nodes), its tip Share Project
    (Partilhar Projeto), right before Close Project, enabled with a project open, in both interfaces: Share Project…'s
    window (#3698). It leaves the Start page's Project Commands, whose git group (after a line, only where there is
    one) is now Sign in with GitHub… (the browser edition's, when not signed in), Initialize Repository…, Open from
    Repository….
  • The PROM bitstream as a .bin on a board programmed by its maker's flash tool (#3711). These boards' USB reaches
    only their SPI flash, through Numato's tool. The FPGA then boots from the flash, which needs StartUpClk Cclk:
    the project's JtagClk bitstream never ends its start-up there. On such a board (programmer.vendorTool on its USB, an
    SPI flash: core/vendor-flash.js), the option file of the PROM bitstream asks bitgen for a binary
    (-g Binary:yes), and the build log ends with Programming file: …: the tool, the .bin and its command. The
    Configure Target Device panel's note starts with the same line, and the MCP server's run_process bitstream
    returns it in its log.
  • The examples' clock as a generic, their counters sized by it (#3710). Each clocked example's top has a CLK_HZ
    generic (50 MHz, the Basys2's), and its timing generics come from it: in Verilog as their defaults
    (TICK = CLK_HZ, STABLE = CLK_HZ / 50…); in VHDL, whose generics' defaults cannot use another generic (VHDL-93,
    as XST reads it), with a default of 0, "from CLK_HZ", taken in an architecture constant (TICK_N := or_else(TICK, CLK_HZ)). The values at 50 MHz are the ones before, and the test benches still set them. The
    Verilog counters (the clock's tick, pong's step, the debouncers', the displays') are sized by a constant clog2
    function. vga's vga_sync makes its 25 MHz pixel enable from a DIV generic. blinky's ASM chart has 21-bit
    debounce counters (2 000 000 cycles at 100 MHz).
  • The inputs' level as a generic (#3710). BTN_LOW and SW_LOW (false: the Basys2's) invert the buttons and
    switches where a board's read 0 when pressed or on (the Mimas V2's, the Elbert V2's, now marked active low).
  • A project for another board gets the board's values (#3710). New Project from an example for another board
    (the server, the browser edition and the MCP server's new_project_from_example), Design Properties' board change
    and set_design_properties write the board's clock and levels as the defaults of the top's generics
    (core/board-generics.js). The console says so, and Undo gives the file back.
  • The Mimas V2's pin table as the other boards' (#3709). Numato's SevenSegment[0] is the decimal point and
    [7..1] are the segments a..g. The table had them in Numato's order as seg, and the digit enables as seg_en, so
    an example's seg<0> (a) took the decimal point's pin and its an and dp had no pin. The table now has seg
    (a..g), dp and an (the 3 digits; an example's an<3> has no pin). It also records the active levels: the
    display, the push buttons and the DIP switches are active low, and the DIP switches are named sw, as the Board
    Emulator's switches.
  • A port the board has no pin for, on a free header pin (#3709). Silinx refused ISE's build of a board project
    with a port bit that has no pin. The clock example's an<3> on the Mimas V2 (3 digits) had none. The board's
    constraints now put such a bit on a free pin of the board's headers (the Mimas V2's, the Elbert V2's, the Spartan-3A
    Starter Kit's), with a comment in the UCF, and the build warns about it.
  • The simulation models of the project's tool (#3707): the Simulation view's SSim Simulator group lists the Post-… models
    the project's tool generates (simModelsOf, as Emulate on Board since #3704), and so do the Tasks interface's model
    choices, the Simulate task's bar's and SSim's toolbar's (modelChoices). The MCP server's simulate (a model) and
    run_process (a …_model process) refuse a model the flow does not make, with one message ("the iCE40 flow makes no
    post-synthesis simulation model …"), where run_process ran the synthesis alone and simulate said the model was not
    made. Tests: an iCE40, a Basys 3 and a Basys2 project's Simulation processes and Tasks model choices, in Portuguese, in
    both interfaces; the MCP's refusals, and its check against the interface's tools on every board.
  • The Basys 3 and the Nexys A7 drawn in the Board Emulator (#3706): web/js/emulator-basys3.js and
    web/js/emulator-nexys-a7.js, registered in BOARD_LAYOUTS: the PCB with LD15..LD0 over SW15..SW0 along the bottom
    edge, the cross of five buttons, the 4-digit (Basys 3) or two 4-digit displays (Nexys A7), the VGA, USB-UART, USB
    host and Pmod connectors, no maker's logo; I/O Pin Planning's Board view draws them too. The UI tests check each
    board's size and parts, where they are, that each control is what sits at its centre, in both interfaces and in
    Portuguese.
  • Emulate on the 7-series, the ECP5 and the Gowin (#3704): the Silinx toolchain's 7-series flow, AMD Vivado's, the
    ECP5's, the Gowin's and the experimental family's no longer leave out Emulate on Board and the Emulate task: the emulator
    runs the HDL, not the bitstream, on the board's resources and the constraints' pins, which it already read in every format.
    Emulate on Board lists the models the flow generates (simModelsOf): the RTL model only on the flows with no simulation
    models (the iCE40's too). Tests: every example on the Basys 3, the Nexys A7, the ULX3S and the Tang Nano 9K on both engines,
    its warnings the wizard's (the ULX3S's four switches missing for isqrt and micro80); gates' switches and stopwatch's
    buttons on the display; a Basys 3 project's Emulate on Board, toolbar step and Tasks interface's Emulate, the emulator
    reacting, in Portuguese, in both interfaces.
  • One import and one export of a project's zip, the software chosen in the window (#3703). Import / Export Silinx
    Project (.zip)… and Import / Export Xilinx ISE Project (.zip)… are Import Project (.zip)… and Export Project (.zip)…
    (Importar / Exportar Projeto (.zip)…), in the File menu and in the Start page's Project Commands. Their window
    (web/js/project-zip-dialogs.js) starts with a Software: row: Silinx (the whole project) or Xilinx ISE (its .xise and
    its sources; the import's project folder row with it), AMD Vivado and Intel Quartus disabled, their reason their tip
    (core/project-software.js). The import preselects the software of the files picked (a .xise, then a silinx.json, read
    from the zip's entry names); a Vivado or Quartus project's zip, or a zip of no project, is said under the fields. The
    choice last made is kept per viewer. The messages and documents that named the old entries name the new ones; the
    MCP tools import_project and export_project keep their names and kinds, their descriptions name the new entries.
  • The Start page's Project Commands, in both interfaces: Import / Export Silinx Project (.zip)…, Import / Export Xilinx ISE Project (.zip)…, a line, then Share Project Copy…, Open from Repository…, Initialize Repository…, Sign in with GitHub… (and the ISE style's Share Project Live…), each shown where it was before (#3699).
  • The Teaching tab keeps its Labs and Exercises box only (#3699).
  • Share Project… and Share Module… (#3698). Share Project Copy… and Share Project Live… are one action, Share
    Project… (Partilhar Projeto…), in the File menu, the Start page's sharing group (in both interfaces) and the Design
    tree's project and device menus; Share Module Copy… and Share Module Live… are Share Module… (Partilhar
    Módulo…), in a module's menu and a test bench's (web/js/share-how.js). Its window, Share Project or Share Module,
    starts with its Share: row: A copy (the link, full or short, its QR code) or Live (the name asked, then the links);
    choosing the other way swaps the window for the other's, and the way chosen is kept for the next one (the copy by
    default). Live is disabled, its reason its tip, where the project cannot be shared live: a Silinx opened from a file,
    an edition that does not share, while following another's project, a test bench. The console's messages say Share
    Project (or Share Module). The MCP tools keep their names; their titles and descriptions name the new action.
  • The Vivado timing script's unused exports (#3695): prepareDesign and nextpnrNumbers of scripts/xc7-timing-vivado.mjs are used only in their file and no longer exported; the unused-exports test passes on main again.
  • A UI test's regular expression escaped completely (#3692): test/ui/project-name.test.js escaped only - in
    the constant it builds a regular expression from (CodeQL's alert 182); it escapes every special character now,
    backslashes included. Test: ui/project-name, in both interfaces.
  • The 7-series' occupied slices: a slice counted only when it holds something (#3683). The count read nextpnr's FASM,
    where an empty half of a used CLB tile still has its clock-inversion default bit. The details and measures are in the
    toolchain's changelog.
  • The private-data check reads a .tgz with no tar inside (#3682): scripts/check-private-data.mjs scans a
    .tgz that is a plain gzip stream as a .gz's content (it was read as a tar and passed with nothing found), and
    refuses a .tar with no header at all as an archive it cannot read. A real tarball, and an empty one, are read as
    before. Test: private-data.
  • MCP server: the failed results that quote a design labelled "Untrusted content" (#3681): the messages of a
    design that does not elaborate or build quote its names and files, so these results are labelled as the other
    results holding a project's own text are (#3528), with untrusted in their structured content: check_syntax when
    it fails, simulate when the simulation does not start (behavioural or against a model of the flow), run_process
    when the process fails, verify when the check fails (its counterexample's replay prints the design's messages) or
    cannot run, and c_simulation, c_cosimulation and c_synthesis_report when they fail. Their successes are unchanged. Test: mcp-untrusted.
  • The browser's flow worker: a family's own tools looked up among its own only (#3662). The worker took the
    family's tools (the iCE40's, the ECP5's, the Gowin's) from an object by the name the page sent: a name of an object's
    own (constructor) ran another function and the flow went on without them; another name failed with a TypeError
    (CodeQL's js/unvalidated-dynamic-method-call). Now a Map, and a family with no tools of its own in the browser is an
    error naming it.
  • The local server's project paths: their containment where CodeQL sees it (#3648-#3661). CodeQL's js/path-injection
    alerts on server/projects.js and server/fs-error.js were false positives: every path is the project's folder (the
    name checked against the project name's rule) or a path that safeJoin keeps inside it. Since #3538 the name's check
    is a function of core/project-name.js, which the analysis does not follow. projectDir now also checks that the
    resolved folder is inside the workspace, and safeJoin that the resolved path starts with the project's folder, before
    server/paths.js's check through symbolic links; a file's trash folder is named after the project's folder. No
    change of behaviour; a test shows every file route refusing a path out of the project.
  • The standalone Sign in with GitHub… test no longer fails at times (#3644). It compared localStorage's keys before
    and after the sign-in, and the tasks' arrangement, kept whenever a document comes to the front, was at times written
    in between; it now leaves that key out.
  • MCP server: a schema's maximum checked as its minimum is (#3641): the argument check of #3535 did not know
    maximum, which project_link's expires_days now declares (#3629), and its test refused the schema; a value above
    it is said in the one message (expires_days: at most 3650). Test: mcp-args (the keyword, project_link's expiry),
    mcp-gist (the message).
  • The QR code of a sharing dialog's link (#3637): a small encoder of its own, core/qr.js (byte mode, the smallest version that holds the link, the highest error correction that still fits it, the mask of least penalty; no network), tested by decoding its symbols back with jsQR (a development dependency, the tests' only) at every version from 1 to 40 and with each mask. Share Project Copy… draws the link it shows (web/js/link-qr.js): the full link's; the short link's once made (Show QR Code makes its gist, which Copy Link then copies, no second one); a link over 2953 bytes, none, said once in the console. Share Project Live… and Share Module Live… draw their follow link's. Enlarged in a dialog of its own; an image of its SVG (the page's CSP as it was). The MCP's project_link and share_project take qr: the link's QR code as an SVG (qr), a full link too long for one a warning. Portuguese labels.
  • Short Links… (#3635, web/js/short-links.js): the account's Silinx gists (core/gist-link.js listGists: only those whose description and single file are Silinx's, secret), the newest first, with their dates; this browser remembers the links it made (the project, the link: kept where its sign-in is), so that those rows show the project and can be copied again, a gist made elsewhere said to be; Delete Selected (deleteGists: any other gist refused before anything is deleted) and Delete Expired, each said in the console; opened from Copy Link's Manage short links… too (Portuguese labels). The server's /api/gists (its GH_TOKEN), the browser edition's through its sign-in; the MCP's short_links and delete_short_link. README, SECURITY.md.
  • Share Module Live… (#3631). In a module's menu of the Design tree, after Share Module Copy… (Partilhar Módulo em
    Direto…), in the editions that share, disabled while following another's project: the project shared live as Share
    Project Live… shares it (the name asked first; the sharing already running otherwise), the module's source opened in
    the owner's page, which the followers follow, and its dialog's links opening on the module (…#follow=…&top=<module>,
    the editors' #edit=…&top=…; core/collab-link.js withTop). A follower who opens such a link opens the module once
    the project arrived, browsing alone on it (Follow , in the toolbar, follows the owner's editor). The MCP tool share_project
    takes top (a module of the design: its links), and follow_project says a link's module.
  • Follow brings the owner's editor back even when the owner has not moved (#3631). Following again after
    browsing alone waited for the owner's next move: the view last followed was remembered as already shown.
  • The grader's hang test after the loop budget (#3630): a submission that hangs inside one clock edge is still
    Time Limit Exceeded, now for whichever limit the machine reaches first. On a fast machine that is the loop budget
    (#3525): "loop iteration limit exceeded: more than 4194304 iterations and calls in one evaluation … at 10 ns
    (src/counter4.v line 3)", naming the process and the line. On a slower one it is the test's time limit: "the test
    took more than 3 s". test/grade-server.test.js accepts either reason for the hang and checks the budget's exactly.
    A second submission, whose every clock edge stays within the budget, keeps the time limit's reason covered.
  • A short link's expiry (#3629, core/gist-link.js): kept in the gist beside the encrypted design and authenticated with it (AES-GCM's additional data, the gist's version 2; #3223's links, version 1, open as before, with no expiry), and in the gist's description in the clear; opening an expired link is refused, saying when it expired; making a short link deletes the user's Silinx gists that have expired (every page of their gists read first; a gist not only Silinx's never touched; a failure of the cleaning never fails the new link). Copy Link's Expires: list (Portuguese labels); the MCP's project_link short takes expires_days, its answer the date and the gists deleted, and open_link refuses an expired link in the same words. README, SECURITY.md.
  • The sharing actions renamed (#3628). Copy Link to Project… → Share Project Copy… (Partilhar Cópia do Projeto…),
    Copy Link to This Module… → Share Module Copy… (Partilhar Cópia do Módulo…), Share Project… → Share Project Live…
    (Partilhar Projeto em Direto…): in the File menu, the Start page, the Teaching tab, the Design tree's and the Simulation
    hierarchy's menus, the dialogs' titles, the console's messages and the documentation. The MCP tools keep their names
    (project_link, share_project); their titles and descriptions use the new ones.
  • test/timings.json: the files measured by the Nightly of 2026-11-12 (run 43150623867) added to the release's, the
    test-shard check's 80 % of unit files measured met again; it failed on main after the night's new test files (#3621).
  • The release zip's test (#3615, #3610): the packages downloaded on their first use are the iCE40's, the ECP5's
    and the Gowin's nextpnr (#2302, #2303), none of them in the zip; the zip's budget 40 → 42 MB, for the 7-series'
    back end (nextpnr-himbaechel-xc7 and its bit database, #3146) and the HDL front end's package (#1710): 40.6 MB.
  • The other live sharing tests' servers close with a WebSocket still connected (#3614). collab.test.js's fix (its
    server's connections destroyed after its tests) is now test/server-close.js's closable, also used by collab-edit,
    collab-relay-config and mcp-collab, whose after() waited the same way for a failed test's participants (srv.close
    with closeAllConnections, which leaves the relay's WebSockets): a failure there is reported, never a hang of the
    file and of the CI's part. Tests: server-close (the old way waits with a WebSocket connected, closable's close does
    not and the client is closed); collab's and server-close's times in test/timings.json.
  • The live sharing's tests on a loaded machine (#3614): two tests of test/collab.test.js read a follower's events
    once it was closing, before its end's event came (on a loaded runner it had not), and the failure left the
    relay's connections open, so the file waited for them with no end and CI's part timed out after 90 minutes; they
    wait for the end's event, and the file's server drops every connection at its end.
  • Calls nested at most 32 levels deep (#3612): MAX_CALL_NESTING 64 → 32 (#3508's limit, not released): each
    level of a chain of calls in an if, a for and a case takes about 10 KB of the elaboration's stack, and 64 of them
    overflowed Node's default stack on CI's macOS runner (Node 22) in every run; 32 bind and simulate there in every run,
    still with 800 KB.
  • Git, a project's repository on Windows (#3611, #3613, #3616): the project's folder, the repository's top and
    the workspace compared by the system's resolution of their paths (fs.realpathSync.native, as git reports them),
    which expands a short (8.3) name such as RUNNER~1; fs.realpathSync kept it, so no project was in a repository
    and GIT_CEILING_DIRECTORIES named the workspace's folder by a form git did not match. A test repository named
    nul, a reserved name on Windows, renamed.
  • The MCP server's paths with '/' on Windows too (#3611, #3613): a result's or an error's path after
    <workspace> or <exchange folder> is written with / on every system (<exchange folder>/traffic.zip), as a
    tool's arguments take it; on Windows it had the system's \.
  • Tests that run on Windows too (#3611, #3613, #3616): a module imported by its file URL, never its path
    (config-isolation, silinx-hdl-package); the classroom template's files by their / paths; no folder named GIT~1
    (on Windows .git's own short name exists: hidden-short-names); Python 3's path as Python names it (ila-board); the
    history scan's file named with " and a backslash only where a name can have them (private-data); the child's
    account's home in USERPROFILE too, which Windows's os.homedir reads (config-isolation); a filter's path with /
    in git's configuration, where a backslash is an escape (git); #3412's expected paths with / after the exchange
    folder's name, as the results now give them (mcp-exchange-paths).
  • A written or renamed file's name checked as it was given (#3611): the import gate's check of a file the
    interface or the MCP writes (PUT /file, write_file) and of a rename's new name (rename_file, Rename) reads
    the name before the system joins it: on Windows \ is a folder's separator, so src/a\b.vhd became
    src/a/b.vhd, which the gate let through.
  • The design corpus's expectations recorded again (#3606). The Nightly's designs job failed on 13 designs whose
    results had changed with main: the synthesis input's line numbers in Yosys's "Replacing memory" warnings (neorv32,
    verilog-axis-fifo, secworks-aes, itc99-b15, itc99-b17), and the sizes and periods that sliceEstimate's change (#3389)
    moved: itc99-b11 81 slices from 89 but 36 flip-flops from 31, itc99-b15 1336 slices from 1233 with its 20 ns now
    met (#3697), epfl-bar, epfl-router, itc99-b07, b08, iwls05-sasc and usb-phy smaller. Recorded with run.mjs --update
    on those designs (the 'local' times left as they were), the README's table from them.
  • A file's path: one rule in the browser and on the server (#3598). Read, write, rename and delete (the interface's
    and the MCP's file tools) take a path by filePath (core/import-gate.js) in both editions: src/../x.v is refused
    (the server read it as x.v) and src\y.v too (the browser's store took it for src/y.v), with the same message,
    'src\y.v': invalid path: backslash in the name; ./src//a.v is still src/a.v.
  • Silinx HLS: the dataflow graph's passes follow a chain with a stack of their own (#3595): the nodes a module
    needs (Graph.needed), the sums' balancing, the wide multiplications' splitting and a pipeline's stages' values
    walk the graph with an explicit stack (postorder, bottomUp in core/c-graph.js) instead of one call per node, in
    the same order: a chain of thousands of updates of one variable is made and co-simulated in VHDL and Verilog; the
    HLS corpus's HDL unchanged.
  • The toasts that were messages moved to the console (#3591): logWarning and logInfo beside logError in
    web/js/console.js (each translated, its tab or the console's page in front); the busy refusals of iMPACT,
    the open families' programming, Implement Design and the processes, "The design is being implemented", a failed programming's
    "Failed — see console" (its ERROR line instead), Rename Symbol's refusals, the stimuli's save error, Verify's "no
    other module to be the reference", the top's choice with no module, a disabled process's reason, Stop with nothing
    running, Find in Project with no project, Select the top module first, Paste without the clipboard, and the pin
    planner's Auto-assign that assigns nothing (a WARNING); a menu action's error and the editor's "No quick fix
    here." through ui.js's message, which the app sends to the console (setMessageSink). test/web-toasts.test.js
    checks that no toast of kind error or warning is left; the i18n scan reads the console's messages too; UI tests in
    test/ui/console-messages.test.js.
  • Sequential ifs read back as an ASM chart in linear time (#3582): reading HDL as a chart (core/asm-trees.js) copied
    the statements after an if into both of its branches. They are now one subtree shared by the branches (the first
    branch is built before them, so the errors come in the statements' order), and every walk of a state's tree (merging
    the processes, the next-state resolution, the defaults, the renaming) takes a shared subtree once and keeps what it
    does not change. A state's tree is drawn as before up to 1024 boxes written out (the generator's limit, #3554); past
    it, a box reached again is one box reached from both branches. The layout's order of the states (core/asm-layout.js)
    visits each box of a block once. A state with more than 256 decisions on a path (if and case statements one after
    the other or nested) is refused with "more than 256 decisions on a path of a state … too many for a chart" instead
    of overflowing the stack. 30 sequential ifs read in about 20 ms, 200 in 60 ms. The 268 distinct conversions from HDL
    and the 258 layouts the ASM, FSM, diagram-corpus, C-to-ASM, labs and MCP tests make are byte-identical to main's,
    and 3000 random charts with boxes reached again and loops are laid out the same. Tests:
    test/asm-from-hdl-seq-ifs.test.js (chains of 16, 30 and 200 decisions generated by Silinx and 30 and 200 written by
    hand, in VHDL and Verilog, read in a bounded time and simulated cycle by cycle against their HDL, the chart and its
    HDL in both languages; a chain of 8 read as before; the limit's message; the MCP server's convert_module).
  • The interface-parity test: Show VHDL / Verilog Beside no longer listed as the ISE style's own; both interfaces reach
    them now; the test failed on main (#3579).
  • The UI test of a project hosted by the owner's page, on a Mac behind a VPN (#3567): its two pages, with no STUN
    server, offered each other only an mDNS name (.local, which this Mac's Chrome did not resolve) or, without it, only
    the default route's address (a VPN's tunnel, which never reaches itself), and never connected: the test timed out
    waiting for the hosted project. The UI tests' Chrome takes webrtc: true (setupUi), as if the pages had the
    camera's permission, so their candidates are every interface's address, and they connect on any machine.
  • Names of Object.prototype's are a design's own names (#3560): the Verilog front end's and Silinx HLS's tables
    of built-ins (math.h's functions, the bit casts, the headers and their macros, the pragmas' values, the stdint
    types and sizes, the stream types and manipulators, hls::fft's and hls::FIR's constants and fields) have no
    prototype: wire toString; is a declaration, and an undeclared toString(a) is the C check's error "the function
    'toString' is not declared", where cToHdl and the C simulation failed. A test checks that every table of the front
    ends read with a computed key has no prototype, but those keyed by Silinx's own words.
  • A drawing's deeply nested expression is the box's error, not a stack overflow (#3559): an ASM chart's, a state
    diagram's or a Petri net's condition or action (core/asm-expr.js) and a truth table's expression (core/logic.js
    parseExpr) nested a thousand levels deep threw RangeError: Maximum call stack size exceeded in the parser, the
    checks and the generators. They now nest at most 128 levels (parentheses, unary operators, a chain's operators: half
    the HDL front end's MAX_NESTING, #3395, as the HDL generated from them nests up to twice as deep and still compiles);
    a deeper one is one error at its box or cell, "expression nested more than 128 levels deep (parentheses,
    operators)", whatever its depth. Their tokenizers read the text in linear time (100,000 characters took minutes).
    Convert to Schematic of assign y = a ^ b ^ … with 10,000 terms overflowed the stack too: a chain is flattened
    without recursing (core/schdoc-gates.js), and the stubs of a schematic laid out without ELK are checked against the
    wires and pins near them only (core/schdoc-place.js; a chain of 2000 inverters took 19 s, now 0.6 s). The
    expressions of the repository's tests and examples parse the same, and the examples' and test designs' 68
    conversions to a schematic are the same, with and without ELK. Tests: test/drawing-nesting.test.js (the limit's
    charts, state diagrams and truth tables checked, their HDL compiled and simulated as the plain ones in VHDL and in
    Verilog; one level more and 100,000 the exact errors; the MCP server's check_drawing and convert_module).
  • The limits of a C source (#3558, core/c-limits.js):
    • an array of at most 2^22 elements, all its dimensions (checkC; a test bench's array of a size known when it runs
      and malloc's in the C simulation): 100 million elements were accepted, the C simulation making an object per
      element (minutes, gigabytes); 2^22 is twice the bits of the largest device's block RAMs and 140 times the HLS
      corpus' largest array;
    • an array in the hardware (a memory of the module, core/c-mem.js; the values of a combinational module's array,
      core/c-dataflow.js) of at most 2^23 bits: its contents are written in the HDL word by word, an array of values
      is a value per element (a million ints took 2 GB, four million ran out of memory);
    • the parser's nesting (core/c-parse.js): parentheses, a call's arguments, indexes, ?: and assignments at most 256
      levels deep, as the HDL's, and statements (blocks, if and else if, loops, switch) too; about 600 parentheses
      overflowed the parser; an else if chain past the limit is skipped with its if, one error;
    • an expression at most 1024 operators deep from its statement (a chain of a left-associative operator nests in its
      left operand: x + x + … of 10 000 terms overflowed the checks, the simulation and the synthesis): the deeper
      part replaced by 0 with one error per expression, the tree walked with a stack of its own;
    • a macro's expansion at most 65,536 tokens per macro the line names, macros nested at most 256 deep
      (core/c-lex.js): a macro doubling 16 times threw a RangeError (out.push(...) of its tokens) after a time
      doubling per level; its tokens are now pushed one by one;
    • what no limit catches (functions calling each other at the far end of long expressions, a type nested 20 000
      deep): the stack's overflow itself, one error of checkC, simulateC and cToHdl instead of a RangeError.
    • Statements nested in blocks took 2^depth in the synthesis (the floating-point units' scan of ALLOCATION pragmas,
      core/c-float-units.js, scanned each block again from every block around it): once each.
      The HLS corpus and the repository's C designs make the same HDL. Tests: test/c-limits.test.js (each limit at its
      value and one past it, the exact errors of the check, the C simulation and the synthesis in VHDL and Verilog, the
      ones accepted co-simulated in both), test/mcp-hls.test.js (c_simulation, c_synthesis_report, c_cosimulation and
      c_generated on three sources past the limits: their one error).
  • The .sch / .sym, .wcfg and PNML readers in linear time (#3557): regular expressions that went back over the
    text, quadratic in a file the import gate accepts, are scans that read each character a bounded number of times,
    with the same results: an ISE schematic's or symbol's start-tag attributes on a long name with no '='
    (core/isesch.js, the scan in core/xml-attrs.js; 80 kB took 12.6 s), a VERSION 5 symbol's SYMPIN lines over many
    blank lines (100,000: 29 s), a wave configuration's comments, instructions and CDATA left unclosed (core/wcfg.js,
    one scanner for the whole file; 160 kB took 4 s), and a PNML file's <net>, <place>, <transition> and <arc>
    with no end (core/petri.js fromPnml, the searches in core/pnml-scan.js; 200 kB took 5 s). A character reference
    past Unicode (&#x110000; on) threw RangeError: Invalid code point: in a schematic or symbol it reads as U+FFFD
    (that reader takes any text), in a wave configuration it is the file's error ("malformed XML: invalid character
    reference '…'"). The repository's schematics read the same; random documents read the same as with the
    expressions. Tests: test/xml-readers-linear.test.js (each case at n and 2n in a child process with a time limit:
    about twice the time, the same result; the errors exact; the scans against the expressions they replace).
  • Verilog macros' expansion in linear time (#3556): the lexer (packages/silinx-hdl/src/verilog/lexer.js) replaced
    each use of a macro by rebuilding the whole text, a cost of the file's length per use. It reads the text through a
    window: a use makes its body the window and sets the text after it aside, brought back as the reading needs it; a
    name joined to many bodies (f4_`N) is read on rather than read again. Its limits are the depth of nested expansions
    (1000) and the text the macros add (8 MB, over a
    file and its includes), each one error that stops the expansion; it counted uses (100,000). Every Verilog file of
    the repository lexes to the same tokens, positions and expanded text. Tests: test/verilog-macro-expansion.test.js
    (100,000 and 200,000 uses, in a child process with a time limit, no warnings, about twice the time; a name joined to
    100,000 times, a body of 200,000 uses, an argument of 1 MB; a macro used in its own body, 1000 and 1001 nested
    expansions, a text doubling at each level: the errors exact).
  • A limit on an object's size (#3555): only an array type's number of elements was limited (4M). The
    elaboration now limits an object's bits to MAX_BITS (224: a vector's width, an array's elements times their
    width, a record's fields, a string's characters at 8 bits each, the result of an expression such as a
    concatenation or a replication) and an array's elements, counting those of every level of an array of arrays, to
    MAX_ELEMENTS (2
    22, the old limit of one array). Both are checked before any value is made, and each gives one
    error of the file at the declaration or the expression, for Check Syntax, the simulation and the synthesis input.
    The repository's largest object is 40 960 bits (a memory of 4096 elements of 10 bits) and the supported devices'
    largest block RAM 4.9 Mbits. A vector of the limit's width costs 2 MB, and an array of 4M elements about 2 s and
    600 MB. A vector aggregate is built as runs of equal bits, not one entry per bit; a replication is built by
    doubling and a long bit literal is read at once (each was quadratic in its width). An aggregate's choices outside
    the array are no longer walked, and a slice of more than 4M elements (a constant one, or one of run-time bounds
    in the simulation) is an error. The repository's HDL files and the design corpus elaborate, check, simulate and
    make their synthesis input as before. Tests: test/wide-objects.test.js (VHDL and Verilog: at the limits, no
    warnings, simulated, the synthesis input; one past them and the issue's cases, the exact errors, at once; the new
    literals and replications against the old bit-by-bit forms); test/grade-server.test.js's Memory Limit Exceeded
    design, within the new limits.
  • ASM charts: a state's decisions walked once each (#3554). The validation and the generators walked a block's
    decisions as a tree: where both branches of a decision reached the same next box, that box was walked once per path,
    so a chain of n decisions took 2^n steps (16: 0.9 s, 30: out of memory, in the validation). The walk now shares a
    subtree reached again (core/asm.js sharedWalk; a subtree with a loop, an error, is walked again on each path for
    20000 boxes, then shared too), keeps its own stack, and every walk of the trees (treeParts, treeNodes, the
    counts) takes each subtree once; the join points (where a decision's branches meet again) are the immediate
    post-dominators found by intersection, not from sets of post-dominators. A chart is generated as before while a
    state's tree written out has 1024 boxes at most (the repository's charts, byte-identical); past it, its decisions'
    branches are written up to where they meet again and the boxes after once, as the every-cycle blocks already were, a
    state's logic 200000 statements at most ("State 'A': its logic takes more than 200000 statements, too many to
    generate"). The header lists a block's paths up to 256 of them, and counts them past it; extractTransitions,
    extractAlwaysBlocks and the conversion to a state diagram refuse more than 4096 paths, saying how many ("the state
    A has 1073741824 paths through its decisions, more than 4096 transitions for a state diagram"), and the conversion no
    longer takes a chain of more than 200 boxes for a loop. A chain of 1000 decisions: validated and generated in about
    0.3 s.
  • Constraints: a net named __proto__, constructor or toString (#3553). Its constraints are its own, in every
    format: a constraints file with such a name no longer gives its pins to the designs implemented after it, nor loses
    them (the toolchain's changelog has the details). The Constraints Wizard's attributes (applyAttributes, the MCP's
    constraints_wizard) and its pins take such a name the same way: attributes for it with no pin are refused as for any
    other net, and change nothing every object inherits.
  • The project bundle removed (#3548): File ▸ Download Project Bundle… and Open Project Bundle… of the browser
    edition removed, with the API behind them (api-local.js exportBundle / importBundle) and their labels; the texts that
    pointed to the bundle (the console's notice of the browser edition, a full browser storage's error, the way to the
    full application, the README, Getting Started / Começar) point to File ▸ Export Silinx Project (.zip)…. In the browser
    edition, Import Silinx Project (.zip)… also takes an old .silinx.json bundle (project-store.js bundleEntries: its
    silinx.json and its text files as a zip's entries), through the import gate (#2830), with the zip import's naming
    and warnings; its dialog says so. The snapshot that brings back a replaced project when its import fails is the
    Silinx zip in the browser edition too.
  • MCP parity (#3546): find_in_project (Find in Project: the lines of the text sources and the texts of the
    drawings, an HDL file generated from its drawing found but not replaced; its Replace All, written through the import
    gate), import_source (Import Source / Add Copy of Source / Import Test Bench: files of the exchange folder copied to
    src/, sim/ or constraints/ with their association, through the import gate, a drawing named with a reserved word
    refused, a file the project has replaced only with replace) and check_syntax's lint (Edit ▸ Design Checks
    (Lint Warnings)). Their rules are the interface's: Find in Project's files and results (core/find.js findPaths,
    findInProject) and the copy of a source (core/import-source.js, out of web/js/wizards.js addSourceDialog).
  • The identifier rules and the XML escapes, one copy of each (#3545): core/ident.js (a name: a letter, then
    letters, digits and _; an identifier that may begin with _; a new file's name) and core/escape.js (the XML escapes
    of the files Silinx writes, their three sets kept as each writer had them) replace the copies of 15 modules, with
    no change of behaviour (test/ident-escape.test.js compares each helper with the copies it replaced, and fails on a
    new copy). The copies in the ASM, FSM and truth-table parsers, Silinx HLS's pragmas and three MCP tools are left for
    a follow-up.
  • A pull request's tests of a module no test imports (#3544): scripts/related-tests.mjs follows a changed module
    that no unit test (nor a helper of the tests) imports up through its importers, at most four levels
    (UNTESTED_DEPTH), to the modules the tests import, and runs their tests: a change of core/translate-to-verilog.js
    runs translate.test.js and the deep-* tests (21 files with the repository's checks, before 13), core/cpp-tb-parse.js
    the C++ tests (22, before 13), core/c-dataflow.js the HLS's (107, before 12), packages/silinx-hdl/src/vhdl/purity.js
    the elaboration's (20, before 13); a module a test imports, through core/'s old path too, keeps its own tests (core/
    schdoc.js 86, core/compile.js 296, as before). The web/ pages are not followed (their UI tests are by name).
    test/related-tests.test.js: the climb, its bound, a helper, the repository's modules of the audit.
  • Exports nothing imports (#3543): the seven exports used nowhere removed (CAPTURE_FAMILIES, LAB_DIR, compareDocs,
    translationOf, the toolchain's longFuses and powerSources, and sharing of the page's live sharing state), and
    165 names used only in their own module no longer exported, with no change of behaviour.
    test/unused-exports.test.js fails on an export nothing else uses, with its list of reviewed exceptions (the
    packages' public API is used by definition; other sessions' files at the time are listed for a follow-up).
  • The code-size test measures every function (#3542): test/code-size.test.js found a function only by a
    declaration at the start of a line (function f(…) {), so an arrow function, a function expression, a class's or an
    object's method, getter or setter, or a nested function over 300 lines passed. Its measure (test/code-size-lib.mjs)
    now reads the source's tokens (comments, strings, template literals and regular expressions skipped), pairs its
    brackets and measures every function, each nested one on its own too, from its first line to its last; it no longer
    takes a function in a template literal's generated code for one of the module's, nor a one-line function as running
    to the next closing brace. No function of the repository is over the limit with it (the largest have 300 lines), so
    no entry joined the list of known ones. Tests: test/code-size.test.js (every form of function measured, a nested one
    and a one-line arrow function, strings, templates, regular expressions, comments and statements not taken for
    functions, a 301-line arrow function, function expression and method caught).
  • The layers' directions (#3541): the New Source templates (templates.js, c-templates.js, out of web/js/) and
    the device database, the boards, the .xise reader and writer and a drawing's HDL sync (out of server/) moved to
    core/ as they were, their old modules re-exporting them; server/mcp-new-source.js and the standalone page's
    web/js/api-local.js and git-browser.js import them from core/. test/layering.test.js checks that server/
    imports nothing of web/, web/ nothing of server/, core/ neither, and the packages nothing of the IDE.
  • Errors shown as toasts (#3540): "Cannot open …" (web/js/drawings.js, drawings-tt.js, app.js; and the ASM chart's
    read of Convert to State Diagram) and "Cannot parse UCF" (pinplanner.js, now naming the constraints file) go through
    the new logError of web/js/console.js: ERROR: cannot open <file>: <reason>, the Errors tab in front. Verify
    Equivalence's, Verify Assertions' and Verify Property's dialogs (verify.js, verify-props.js) show their fields' errors
    on a line of the dialog (fieldError), no longer as toasts of the kind 'err', which had no style.
    test/web-toasts.test.js: every toast's kind has a style, and no toast says a file cannot be opened or parsed nor a
    field's error; test/ui/open-errors.test.js and test/ui/verify-props.test.js in both interfaces. The toasts that still
    carry a message: #3591.
  • The texts test/web-i18n.test.js did not scan (#3539): its label scan also reads the texts of toasts, dialogs'
    titles and messages, plain elements (div, p, span, option… but those marked data-no-i18n), textContent and the
    sentence a dialog's check returns, so that a new one with no translation fails; the 31 it found untranslated get
    their Portuguese in web/js/i18n-messages.js. The console's "No messages." (a console page, which the DOM pass
    leaves alone) is translated where it is written. A UI test (test/ui/i18n-messages.test.js) in both languages.
  • A project's name: one rule, one message, one name from a file's (#3538). The wizard checked a module's rule
    (lab-1 refused, a name of 70 letters let through to the server's "invalid project name"), compared the existing
    projects in the exact case, and the two imports made a name from the file's in two ways, one of them a name the
    server refused. core/project-name.js now holds the rule and its message ("invalid project name '…': a letter, then up
    to 63 letters, digits, - and _"), projectNameFrom and sameProject, used by the wizard, the import dialogs, the
    browser's stores, the local server, the design links and the MCP server (new_project, open_link). Creating a project
    whose name differs from an existing one only in case is refused on every system ("project 'a1' already exists"), as a
    case-insensitive file system makes it the same folder. Portuguese for the message and for Replace Project.
  • The server's file system errors by the project's path (#3537): writing a file onto a folder or through a file,
    or a rename through a file, answers 400 with the project's path ("'src/a.v' is a file, not a folder", "'src' is a
    folder, not a file") instead of a 500 with Node's message and the server's absolute path; a text that is no string is
    a 400 ("text: the whole text of the file", as the MCP's write_file); deleting a path where nothing is answers 404
    ("no file or folder 'x' in the project 'p'", as the MCP's delete_file) instead of ok; any other error of the system
    with no status is "internal server error", as the final handler says. One helper (server/fs-error.js) for the HTTP
    API and the MCP server. Tests: fs-error, server-api, mcp-sources.
  • Project paths spelled as the project spells them (#3536). On a case-insensitive file system (macOS, Windows)
    SRC/extra.vhd went into the project's src/ folder while the project's description recorded SRC/extra.vhd: the
    exported zip then held src/extra.vhd beside a silinx.json naming SRC/extra.vhd, missing on a case-sensitive
    system. Writing, reading, renaming and deleting a file now take each folder (and the file) of its path with the
    spelling the project already has (core/fold-name.js spelledAs), on the local server, the MCP's tools and the browser's
    stores alike; a rename to a name the folder has in another case is refused ("already exists"), but a rename of a file's
    or folder's own case, which the browser's workspace folder now makes through a temporary name.
  • MCP server: every tool call's arguments checked against the tool's schema (#3535): besides the missing and the
    unknown arguments, each value's type, enum, minimum, an array's items and size and an object's properties are checked
    before the tool runs (server/mcp-args.js), one message for every tool naming the argument and what it must be
    (language: one of vhdl, verilog, ports[0].msb: an integer, all the wrong ones together). A wrong value no longer
    reaches a tool: new_source wrote the Verilog template into a .vhd file for a language neither, set_design_properties
    wrote any synthesis tool into silinx.json, a number for a name gave JavaScript's own error. simulate's forces
    declare what they took already (a time or a value as a number). new_project with a variant and no board says so as
    set_design_properties does, an unknown board pointing at list_devices. Test: mcp-args (each kind of wrong value,
    nothing written; every tool's schema of the kinds checked).
  • Git's first commands where a project starts (#3534):
    • The Start page's Project Commands have Initialize Repository… after Open from Repository… (disabled with no
      project open, as the exports, and gone once the open project is a repository) and, in the browser edition, Sign in
      with GitHub… while not signed in; in both interfaces. The page is drawn again when the project becomes a repository
      and after a sign-in.
    • The right-click menu of the project's row and of the device's row has the same two, where they apply, after
      Share Project… and before the Git submenu, which keeps them (and Sign out of GitHub).
    • Each runs the File menu's command (gitStartItems, web/js/git.js): the same dialog, checks and messages.
    • Tests: UI git (the Start page and both rows, run from each, gone once a repository, Portuguese), git-browser (the
      sign-in on the Start page and both rows, a cancel, gone once signed in, back once signed out, Portuguese), projects
      and collab (the Start page's commands).
  • The Teaching tab (#3533): a page of the side panel after Libraries (web/js/teaching-page.js), in the Silinx
    and the ISE style interfaces: Labs and Exercises (Guided Labs…, Exercise…) and Links and Sharing (Copy Link to
    Project…, Open from Repository… where there is git, Share Project… in the editions that share), the same dialogs as
    the Project and File menus, which keep their entries; a project's commands are disabled with none. The Start
    page's Teaching box (#3333) is gone. Portuguese labels; a UI test in both interfaces, on a narrow screen too.
  • A schematic whose symbol's parameters are not what they should be opens (#3532): a module symbol's module that is
    not text (a number, a list, an object) is read as its text or as no module, where normalizeDoc threw a
    TypeError; so are the other parameters read as text, a module's generics (a value that is not text dropped), the
    parameters of a symbol that are not an object, a wire's points that are not points and the HDL kept from a module's
    source (its language, texts and generics). A module named as a name of Object.prototype (toString,
    constructor) is no longer found where the project has none.
  • A shared project's link given to a running Silinx is read before anything is saved (#3531): a #follow= or
    #edit= hash is parsed first (core/collab-link.js parseFollowLink): a broken one is the console's error
    ("Follow a shared project: the link to follow is broken …"), the hash cleared, nothing saved and the page not loaded
    again; in the page that shares a project, the console's warning, nothing saved; a good one asks, when a document has
    changes not saved yet, "This link follows a project someone shares. Your open documents have changes not saved yet:
    save them, then follow it?" (Save and Follow, Cancel), then saves them before the page loads again. While a project
    is followed, another link is read first too: a broken one is said and saves nothing.
  • The single file's GitHub sign-in kept in its tab only (#3530): at a file: address (Silinx.html opened from
    the disk, whose storage every local page shares) Sign in with GitHub… keeps the token in the tab's sessionStorage
    only, Remember me unticked and disabled with its reason as its tip, the dialog's note saying so; a token left in
    localStorage is not read; the site's editions as before. SECURITY.md says it. Tests: the sign-in's two ways at
    file:; the single file's dialog, its sign-in writing nothing to localStorage, in Portuguese too.
  • Live sharing: a follower's join sends the whole project to that follower only (#3529): on a join the owner
    sent the whole document and who is in to the room, and the relay passed it to every follower. Whoever had a follower's
    link could join and leave over and over to multiply the owner's upload and every follower's download. The owner now
    addresses those frames to the follower that joined (KIND.to, protocol 4: the frame marked with its client id), and
    the room (core/collab-room.js, the server's relay and a page's hosted room alike) gives the sealed frame to that
    follower only. It sends nothing to a follower that has left. Tests: collab-room (the frame to one follower, its mark
    taken off; to one gone, to nobody), collab (a follower already in gets no snapshot while five others join and leave,
    each of them with the project; an owner's change still reaches everyone).
  • MCP server: more results labelled "Untrusted content" (#3528): as read_file's are (#3363), these results are
    now labelled after their text, with untrusted in their structured content: open_link's result for a design link
    (the design's names, its warnings), project_info (its silinx.json's fields and files' names), lab on a
    project's own lab (its titles and texts, a lab file's problems), simulate when the design printed lines (report,
    $display, assertions), share_project's status with warnings and no followers, and every tool's error, which may
    name a project's or a shared project's files. An argument the schema refuses is not labelled. open_link now takes
    a design's own board, top and simTop only as names, as it does with the link's board= and top=, and refuses
    the link before anything is made. Test: mcp-untrusted (each tool's label, or none for Silinx's own text).
  • A shared project's limits hold for its whole document (#3527): a follower (the page, the MCP's follow_project)
    counts the document's bytes as the updates arrive (its encoded size, values that are no text and updates waiting for
    others included) against the 64 MB of FOLLOW_LIMITS, and leaves a room whose owner sends a part no shared project
    has (a root other than meta, project, files, drawings), not applying it, with the exact error. At the
    owner's (the page, the MCP's share_project), an editor's edit that changes another part than the files and the
    drawings, builds on changes the owner does not have, or would take the document past the limit is refused before
    it is applied, said as a warning (the share's status for the MCP). What an update changes is read before it is
    applied (core/collab-roots.js). SECURITY.md says it. Tests: collab-roots (the roots against applying the update,
    the messages, the meter), collab (a room whose owner sends another root or values that are no text: left, said),
    collab-edit (an editor's edits refused at the owner's, applied nowhere), mcp-collab (follow_project,
    share_project's status).
  • Guided labs, the simulate check (#3526, core/labs-check.js): a value's at_ns is within the limit of time_ns
    (1 ms; the lab's problem otherwise, "values": "at_ns" from 0 to 1000000 ns), and the check's simulation stops after
    2,000,000 time steps or 60 seconds, as the MCP's simulate does (in parts, the page and the server answering in
    between), the check then failing with "the simulation of did not reach N ns in 2000000 time steps (it stopped
    at M ns)" (or "in 60 s"). Before, the time came from the largest at_ns with no limit, and the simulation ran with no
    bound: the MCP's lab check never answered.
  • One budget of iterations per evaluation (#3525): the interpreter stopped a while or a VHDL loop at 10 million
    iterations, each loop on its own count (about half a minute; nested loops multiplied it). A for over a range and a
    repeat had no bound at all, and a constant function runs while the design is elaborated, before any limit of the
    simulation. The interpreter now counts every loop iteration (for, while, VHDL loop, repeat, a forever that does not
    wait) and every call of a function, procedure or task in one budget. There is one budget per constant evaluated and
    one per run of a process until it waits; nested loops and recursion share it. Past MAX_ITERATIONS (2**22, about
    1.5 s; the repository's largest evaluation is 1536) there is one error at the loop, in its file (a package's
    file), reported once: at elaboration by the declaration or the statement, never retried as a constant that is not
    one; in the simulation as the process's error. The message keeps its old start, which the grader reads as Time
    Limit Exceeded. The clock is checked every 65 536 iterations: the MCP's simulate (runBounded) gives the
    simulator its deadline, so timeout_s stops a run that never waits. The repository's HDL files and the design corpus
    give the same diagnostics, design checks, synthesis input and simulation as before. Tests: test/loop-budget.test.js
    (VHDL and Verilog: at the budget, one past it and to integer'high; nested loops; a package's function; a recursion
    of many calls; repeat and while; a process's run, and a budget per run; the MCP's set_top, check_syntax and
    simulate with timeout_s).
  • Git pull: a name another tracked file has but for its case refused and undone (#3524): on a case-insensitive file
    system the two names are one file, which the pull wrote over and the import gate's taking out of another spelling of
    silinx.json then deleted, so the project vanished from the workspace. Such a pull is refused on any file system, as
    a clone's gate refuses the name ("same name as … but for its case"), and undone: the branch where it was, the other
    file as it was (its changes not committed kept); a file taken out of the folder is never one another tracked path
    names. Test: git (the exact error, silinx.json and the source kept, the project opens, a later pull without the
    names).
  • Git, a repository's own configuration (#3523, server/git-repo-config.js): before any git command runs in a
    project's repository, its local configuration is read, and a setting outside the ones a clone, Silinx and the usual
    tools write (a content filter, a merge or diff driver, a signing program, a credential helper, a remote's helper, an
    included file, a rewritten address, an alias…) stops it: "the repository's own configuration sets '': Silinx
    runs no git command in it", with how to remove it. Silinx turned off a repository's hooks and its fsmonitor already;
    a folder copied into the workspace with its own .git could still name other programs, run by the status a project's
    opening asks for.
  • Git, the remote's addresses (#3523): every fetch and push URL of origin is checked to be github.com's before a
    pull and a push, not only the first of each (git pushes to every push URL).
  • Git pull, the workspace's repository (#3522): a folder with a silinx.json whose name is no valid project name is no
    project (listProjects leaves it out, as every route refused it already), and the projects a pull brings are marked
    untrusted each on its own (server/git.js broughtUntrusted): one that cannot be marked is the pull's error, naming
    it, after the others were marked. Before, such a folder ended the marking with "invalid project name", and the
    projects listed after it stayed trusted: Xilinx ISE and the board programming were not refused for them.
  • Dependabot: a week's cooldown for the GitHub Actions and the Docker image too (#3521): an action's release and
    the Vitis HLS image's base digest are proposed a week after they are published, as an npm version is (#1436);
    test/workflows.test.js checks every ecosystem's cooldown; SECURITY.md says so.
  • The private-data checks' gaps (#3520): scripts/check-private-data.mjs flags a tailnet's name (MagicDNS,
    *.ts.net), a .local / .internal name with no hyphen or digit in quotes, right after = or a colon or as a YAML
    value, and a login on a computer named with no domain (user@name) after ssh / scp / sftp / rsync / mosh, :// or as a
    YAML value; reads UTF-16 text, the files of a .zip (and the formats built on it), .tar, .tgz and .tar.gz, and
    refuses another archive, or one it cannot read, unless allowed; the history scan reads a file whose path git quotes
    under its own name (its lines were filed under the file before it); the scanner and its test are scanned too, with
    their own allow-list entries. scripts/report-failure.mjs cleans a tailnet's name out of a log line. Tests:
    private-data, report-failure.
  • A one-bit vector's pin in an XDC, PCF, LPF or CST (#3517): Implement's check of the constraints said that sw[0] of a
    std_logic_vector(0 downto 0) (the vga example's sw) was no port of the top; found running vga on the Basys 3 in the
    browser edition (#3143). The checks' move to core/constraints-check.js (#2296) takes such a port as a bus; a regression
    test keeps it so. Tests: constraints-format (XDC, PCF, LPF, CST: the one-bit vector's bit, a bit it has not, a scalar's).
  • The 7-series' timing measured with Vivado, the tooling (part of #3516): scripts/xc7-timing-vivado.mjs runs the
    examples' netlists, as Silinx's open flow makes them (Yosys's synth_xilinx, written as an EDIF), through Vivado ML
    Standard 2025.2 on the ISE runner's machine (its Docker image; --detach, --collect, one design at a time, alone
    there): link_design, place_design, route_design, then write_sdf, report_timing with each path's elements, the timing
    summary's Fmax and the routes' nodes; nextpnr's Fmax on the same netlist alongside. scripts/xc7-timing-parse.mjs
    reads the SDF, the paths and the routes; scripts/xc7-timing-fit.mjs makes the numbers of them (the cells' delays by
    nextpnr's arcs, the routes' delays by wire class) and fits the table the fork's chip database is built with (the
    slice's LUT and flip-flop stubs replaced, a delay per wire class by non-negative least squares). Only numbers are
    written (packages/silinx-fpga/research/xc7/), never Vivado's files. Tests: xc7-timing-vivado (a hand-written SDF and
    report_timing, the routes, the fit on known delays, the Tcl and the run's script).
  • The path-drawing test of the ASM chart: the comparison's path among the 10 slowest rather than the first, which follows
    the placement (nextpnr 1.5.2 puts the VHDL's counter a little ahead); its boxes the same in VHDL and Verilog (#3514).
  • A write-back after an earlier turn of a loop (#3512): the synthesis input writes x <= c ? v : x as
    if (c) x <= v (#523) only where no earlier write of the process run can have written x's bits; a loop was checked
    on its first turn's writes alone, so a second turn's x <= c ? 2 : x after the first turn's x <= 1 became
    if (c) x <= 2, and with c false the netlist kept the 1 where the design holds x's old value. A loop's body is now checked after all its turns' writes, part by part
    (packages/silinx-hdl/src/synth-verilog.js firstWritesOf): the whole signal, constant bits and slices, and the
    element at the loop's own variable (VHDL's loop; a Verilog for stepped by a constant, tested on its variable
    alone, its body not writing it), which each turn writes once (#523's x[i] <= c ? v : x[i] over i keeps its
    conditional write); any other variable index counts as the whole signal. The repository's synthesis inputs (203
    modules) are the same. Tests: test/loop-write-back-synth.test.js (the issue's case and its variants in VHDL and
    Verilog: a whole signal, a constant index, the loop's variable, another variable and an inner loop's turns, two
    signals; each synthesized with no warnings, its netlist simulated against the design).
  • A global of a sequential module in a function inlined into it (#3511): a call evaluated in place (a function with
    no loop, in a plain call, a loop, a pipelined loop or a fully unrolled one, or one with INLINE off kept inlined) ran
    the callee with its own scope only, so a global it named fell to a slot of its initial value: a read gave the reset's
    value, a write was "'g' is a constant". The module's globals (a register, the registers of a partitioned array, a
    memory) are now kept apart (globalOwn, core/c-fsmd.js) and a name the callee does not declare is looked up there
    before a constant's slot (ownGlobal, core/c-dataflow.js): the callee reads and writes the top's own storage. The HLS
    corpus and the repository's C designs make the same HDL. Tests: test/c-inline-globals.test.js (a plain call, a loop,
    a pipelined loop, a fully unrolled loop writing the global, INLINE off kept inlined, a global array of registers and
    of a memory, a callee calling another, a local of the same name: the C model, the HDL with its exact warnings, the
    co-simulation in VHDL and in Verilog).
  • Calls nested more than 32 levels deep (#3508): each call's subprogram is bound (and a constant one evaluated)
    inside its caller's binding, so a chain of functions, procedures or tasks each calling the next used the stack once
    per level and overflowed it past about 230 functions of one statement, past about 100 whose call sits in an if, a for,
    a case and parentheses (Node's default stack). The elaboration counts the subprograms being bound and refuses a call
    past MAX_CALL_NESTING (32: 32 such functions still bind with their outermost call 250 blocks deep and 30
    parentheses around each call) with one error at that call's statement, in its file (a package's), reported by the
    statement or declaration of the design that makes the outermost call, never thrown; a VHDL function bound again per
    constant argument does not bind the chain again. Recursion is not counted (a recursive function is bound once; its
    evaluation has its own limit). The repository's 256 HDL files elaborate as before. Tests: test/deep-calls.test.js
    (VHDL and Verilog: 32 chained functions computing a constant and a simulation's value, checked with no warnings,
    simulated, the synthesis input with no warnings; 33 and 20 000, the error and its place; a package's functions;
    tasks and procedures; constant arguments); test/deep-if-chain.test.js's #2577 case, the new error.
  • Nested loops in the synthesis input (#3504): the synthesis input marks the assignments that no other write to
    their signal or variable precedes (packages/silinx-hdl/src/synth-verilog.js markFirstWrites, #523: an element
    written back with its own value as a conditional write). It walked a loop's body twice (the writes of a turn before
    its next turn), each nested loop twice per walk of the enclosing one: 2^n walks of the innermost statement of n nested
    loops. The second walk marked nothing the first had not (from more writes before them, fewer assignments are first)
    and returned the same writes: the body is walked once. The repository's synthesis inputs (203 modules) are the same.
    Tests: test/deep-loops-synth.test.js (loops of one turn nested 12, 52 and 256 levels deep in VHDL and Verilog, the
    synthesis input in a child process with a time limit and no warnings; 10 nested loops and three loops of two turns
    synthesized with no warnings, their netlists simulated against the designs), test/deep-statements.test.js (VHDL's
    loops 256 levels deep make their synthesis input with no warnings, the step #3490 left out).
  • The MCP's simulate keeps only the changes it returns (#3503). Every change of every signal was recorded until the
    waveform's limit (5,000,000, about 500 MB of the heap, 2 GB of the process), though the answer holds only the last
    max_changes of the signals asked. The waves now keep those (the simulator's waveTail), the other signals' present
    value only, with every change counted (totalChanges); with no recording's limit, the values at the run's end are
    right (they stopped at the limit before). A run that writes a VCD records every change, as before.
  • If generates in the translations (#3501). The translations read an if generate's items and its else as neither
    parser writes them (items, else.items): Verilog to VHDL dropped the body with no error, and VHDL to Verilog
    refused every if generate as one "with elsif branches". Both now read the branches as the parsers write them
    (core/translate-util.js, ifGenerateBranches: then, else, an else if or elsif chain, a case generate's
    alternatives, read in a loop). To VHDL, each branch is an if generate of its own (VHDL-93: no elsif or else generate,
    as XST reads it) on its condition and the earlier ones negated, its declarations before its begin; to Verilog, an
    else if chain of named blocks, the generate … endgenerate keywords around the outermost generate only (generate
    regions do not nest). What a branch holds that has no translation is an error at its line.
  • Generate chains of thousands of branches (#3500): Verilog's parser reads an if … else if generate's branches in a
    loop (it called itself per else if), VHDL's an if … elsif generate's; the elaboration takes a chain's branch in a loop,
    and a VHDL elsif or case alternative is marked as the same construct as a Verilog else if (elseChain), so it
    elaborates in the generate's own scope (g.s, as its first alternative), no longer one scope deeper per alternative
    before it (g.g.g.s). The passes that walk a module's items (the hierarchy's tops and roots, the VHDL pure-function
    check, a schematic's source pieces, the Design pane's instances, the editor's symbols) do it in a loop
    (genChain, walkItems in packages/silinx-hdl/src/util.js), as #2577 did for the sequential chains. A chain is not
    nesting: 20 000 branches check with no warnings, simulate, make their synthesis input and translate both ways. A
    Verilog else if chain of more than 16 branches translates to VHDL in a block of constants, <label>_none_<k> (no
    branch up to k), each branch's if generate on the one before and its own condition: the earlier branches' negation
    written out in each (#3501) made the text quadratic in their number (20 000 branches: too long a string). The repository's 256 HDL
    files parse, elaborate, check and make their synthesis input as before. Tests: test/generate-chains.test.js (Verilog's
    if … else if and case generates, VHDL's if … elsif and case generates of 20 000 branches: checked with no warnings,
    simulated at their first and last branches and past them, the synthesis input with no warnings, the hierarchy;
    translated both ways, checked with no warnings and simulated as the original; 16 and 17 branches to VHDL; a VHDL
    alternative's scope), test/ui/hierarchy.test.js (the Design
    pane's instances in a chain's branches, Check Syntax, the sources in the editor; both interfaces).
  • The private-data check of the history: #3415's commit message names a made-up local address (a link refused with
    trailing dots); it is allowed in that message only, as the link tests' (#3494).
  • Statements nested more than 256 levels deep (#3490): the parsers count how deep a statement nests (Verilog: a
    statement in another, a generate construct in another; VHDL: a concurrent statement in another, a sequential one in a
    process, a subprogram or another statement), one count from the module's or architecture's outermost construct to
    its innermost statement, and refuse one past MAX_NESTING (256, as expressions, #3395) with one error at its first
    token, once for the outermost statement however deep it nests. The statement is skipped whole in a loop (to its ;
    or the end that closes it, an else of its ifs included) and the rest of the file is read. Every pass after the
    parsers recurses once per level (the elaboration, the simulator, the design checks, the synthesis input, the
    translations): 256 levels leave them a margin of 2 or more, with an expression 256 levels deep in the innermost
    statement too. A translation nests more (a branch of an if in a begin … end): designs half as deep translate. The
    repository's 256 HDL files parse as before. else if / elsif chains are read in a loop as before (#2577) and do not
    count. Tests: test/deep-statements.test.js (VHDL: if … else, loops, cases, generates, blocks; Verilog: begin … end,
    if, if … else, cases, generates; 256 levels checked with no warnings, simulated against values computed in the test,
    synthesis input with no warnings; 128 levels translated both ways and simulated; one level more and 20 000 levels,
    the error and its place, the design elaborated without the statement, simulate's error; two statements too deep;
    what the skipped statement holds).
  • Gowin: each clock at its own frequency (#3488). nextpnr-himbaechel-gowin derives no clock through an rPLL and its
    --freq applied the CST's clock to every clock, so a PLL's outputs were checked against its input's frequency and the
    oscillator's at 12 MHz. The flow now computes each rPLL output's frequency (CLKOUT, CLKOUTP, CLKOUTD, CLKOUTD3, with
    their sources and bypasses; none when the dividers are dynamic) and the oscillator's (250 MHz / FREQ_DIV) and gives them
    to nextpnr as an SDC beside the CST's clocks. The report's clock with no constraint of its own shows "no constraint"
    (and constraint: null in get_report's summary). The details are in the toolchain's changelog.
  • The FSM editor's label under the pointer wins (#3487): a later transition's arc, drawn above an earlier one's
    label, took the press with its wide invisible hit stroke, so a click selected it and a double click opened its
    condition; the pointer's handlers (web/js/fsm-pointer.js) now take the transition whose label is under the pointer
    before the arc. Test: fsm-parts (a click on t1's label and on t4's selects each; a double click on t1's label edits
    t1's condition, undone; t4's opens its own).
  • The capture on the board's clock (#3481): core/ila-insert.js takes a capture's boardClock (the board's
    oscillator, boardClockOf: its first clock resource with a period) in place of a signal of the design: the wrapper has
    a port of its own for it, silinx_board_clk, and Implement (core/flow-ide.js captureUcf) adds its pin, its
    period and, when the design's constraints give I/O standards, its own to the design's constraints; a design that has
    a port on that pin already is refused, that port being the clock to choose. The dialog (web/js/capture-ui.js)
    proposes a design's clocks (the signals its processes are clocked on) before its other signals, and the board's
    clock when it has none; the MCP server's capture_settings takes clock: "board"; the messages, the Design Summary
    and build/capture.json name it ("the board's clock mclk (B8, 20 ns)"). Tests: ila-insert (the oscillator of a
    board, the constraints, a combinational design's capture simulated on it), mcp-capture (gates implemented with its
    capture on MCLK, its port on B8, no warning; blinky's clk on B8 refused; no board), UI capture in both interfaces and
    Portuguese.
  • Share Project… in the Design tree's project and device menus (#3479):
    • The right-click menu of the project's row and of the device's row has Share Project…, after Copy Link to
      Project… (the File menu's order), in the editions that share; it opens the File menu's dialog, and is disabled
      while following another's project. A module's menu has none.
    • The Start page's Project Commands have Share Project… in the ISE style only (#3315).
    • Tests: collab (the project's and the device's menus, none in a module's, disabled while following, the Start
      page in each interface, Portuguese); projects (the Start page's commands shown); tasks (the Start page's icons shown).

198 more entries of "In detail" left out (a release's notes hold at most 125,000 characters): the full list is in CHANGELOG.md, https://github.com/pmnmalo/Silinx/blob/v15.25.0/CHANGELOG.md

Silinx is an ISE-style FPGA design environment that runs in the browser, made for teaching digital systems: VHDL/Verilog projects; schematics with live simulation; FSM and ASM state-machine editors, truth tables and Karnaugh maps, all kept in sync with their HDL; module and test bench wizards; behavioural and netlist simulation with waveforms; a board emulator; synthesis, place & route and bitstreams with its own toolchain (the Basys2 and Nexys2 boards) or with Xilinx ISE 14.7; and board programming, also from the browser.

Downloads

File What it is
Silinx.html Single-file edition, the same app as the site (https://pmnmalo.github.io/Silinx/). Download it and open it in a browser: design, simulation, synthesis, place & route and bitstream (the Silinx toolchain, Basys2 / Nexys2; its tools come from the site the first time). Projects are kept in the browser (File ▸ Download Project Bundle to back them up) or, in Chrome / Edge, in a folder on the disk. Programming the board's FPGA and Platform Flash over USB works from the site's address (Chrome / Edge, Chrome on Android).
silinx-<version>.zip Silinx run locally, with its dependencies: an alternative to the site (offline synthesis, projects as folders, the installed programmers, Xilinx ISE 14.7 modes). Needs Node.js 18 or newer. Unzip, then double-click Start Silinx.bat (Windows), Start Silinx.command (macOS) or start-silinx.sh (Linux) in the silinx folder; Silinx opens at http://127.0.0.1:8642. Projects are stored in ~/Silinx-projects.
silinx-docker-kit-<version>.zip Kit to build your own private Docker image of Xilinx ISE 14.7 (Windows, macOS Intel/Apple Silicon, Linux) for synthesis, place & route and bitstreams. ISE and its licence are not included and must not be redistributed: download them with your own AMD account. See ise/README.md.

The source code is attached by GitHub below (Source code zip / tar.gz).

Silinx is free software (GNU AGPL-3.0, Copyright 2026 Pedro Maló; see LICENSE, NOTICE and THIRD-PARTY-NOTICES.md). It is an independent project, not affiliated with or endorsed by AMD/Xilinx. Xilinx and ISE are trademarks of AMD.

Checking a download

The files' SHA-256 checksums (the file SHA256SUMS):

f05df828ac6f30c12786dcfe0a112ab454ff51518b1cefa4a3dd4bbc634e680b  Silinx.html
0dc191abeb6e3384ebc108a13cda4b16d8a098dd84ed0be7dcfc0762ca11f5b7  silinx-15.25.0.zip
e9a5c947bc3543bb491a8492db20415e833ff3985bed0fd70084782100827914  silinx-docker-kit-15.25.0.zip

In the folder of the downloads: sha256sum -c SHA256SUMS --ignore-missing (Linux), shasum -a 256 -c SHA256SUMS --ignore-missing (macOS), Get-FileHash <file> (Windows PowerShell, compared by eye). Each file also has a build provenance attestation (it was built by this repository's Release workflow from this tag): gh attestation verify <file> --repo pmnmalo/Silinx.