Skip to content

Silinx 15.27.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 08:45
· 455 commits to main since this release

What's new in 15.27.0

The breadboard opens onto KiCad: KiCad's 74xx and 4xxx and SparkFun's parts found by part number, drawn as they look,
run in Try it by their VHDL or Verilog models, and a circuit exported to KiCad. A project is worked on together: a
Collaboration tab, a branch per person, Merge Contribution… and live sharing that follows Try it. Every drawing zooms and
pans as the Schematic editor does, and the Lattice and Gowin boards simulate after synthesis and Place & Route.

In short:

  • Board Emulator: a push button clicked just before a power cycle is let go by it, no longer seen by the design
    after the reset.
  • KiCad parts in a circuit: a KiCad catalogue in the Parts panel (KiCad's 74xx and 4xxx, SparkFun's breadboard
    parts, Arduino boards, LEDs, switches and push buttons), Browse KiCad Libraries… and Import KiCad Part…, each part
    drawn as it looks and run in Try it by its VHDL or Verilog model; Export to KiCad.
  • More ICs: 30 more gate ICs, three-state buffers, the bus transceivers 74242, 74243 and 74245 on three-state
    buses, and 23 flip-flops, latches and registers (the three-state 74173, 74373, 74374, 74573 and 74574 among them),
    each written once in VHDL.
  • Working together: a Collaboration tab, a branch per person and Merge Branch…, Merge Contribution… (also as MCP
    tools), text files merged; live sharing follows Try it, each participant keeping their own wires' style; GitHub Checks
    with a fine-grained token, and signing out of GitHub asks first.
  • Drawings: every drawing's view gestures are the Schematic editor's (wheel and pinch zoom about the pointer, a
    trackpad's scroll pans, Safari's pinch on the Mac); side panels resizable; the circuit editor's Properties tidied.
  • Boards and devices: the iCE40 and ECP5 designs simulate, after synthesis and after Place & Route too; the Tang
    Nano 9K's DDR, serialisers, LVDS and DSPs in the open toolchain; the Spartan-6's Map by nextpnr by default; Place &
    Route of a very large design with Optimization Speed in half the time.
  • ISE: the breadboard and blinky examples build with no warnings again; Export to Xilinx ISE of a new project, of an
    instance named with a reserved word and of a VHDL top with integer ports; a circuit's unwired chip outputs and tied
    presets build with no warnings.
  • Projects: a Silinx zip imported back as it was exported; the Design Summary no longer jumps in front; VHDL
    simulation of pull-ups and pull-downs.

In detail:

  • WebKit's failed load retried as CI's log says it (#4464): its address came as http: /127.0.0.1:…, which
    the retry of #4428 did not take; both forms are taken now.
  • CI's UI tests in four parts per interface (#4463): a part of three reached the step's 40 minutes (ci.yml,
    test/ci-workflow.test.js, test/test-shard.test.js).
  • The UI tests wait 30 s for an imported project (#4462): every wait for a project opened by Import (21 of them, the
    pattern of #4438) had the 10 s default, too short on a slow CI runner.
  • The short-click test scrolls BTN0 into view (#4461): on CI's runner the board reached below the window and the
    wait for BTN0 under its centre timed out; the button is scrolled into view first, and a timeout says what is there.
  • The Merge Contribution UI tests wait 30 s for the imported copy (#4438): importing the contributed copy took more
    than the 10 s default wait on a slow CI runner.
  • A power cycle lets a held push button go at once (#4437): a button clicked while the compiled engine's run was in
    flight and let go waited for that run before its release was counted, and a power cycle in between kept it held for
    the first run after the reset, long enough on a slow page for the design's debouncer to take the old click
    (core/emulate-press.js: the board's time gone back releases it before anything else).
  • The Board Emulator's short-click test presses the button it reads (#4437): the #3109 test sets the clock rate by
    its own selector, waits until the run at that rate is in and the button is under its centre before the click, reads
    the shown emulator's button, checks that the press landed and says the board's state if the button is up too soon;
    in CI's ISE interface it read the button up after 0 or 18 cycles, which no local run reproduced.
  • A UI test run once more when WebKit failed to load from the local server (#4428, #4439): a test whose only page errors
    are WebKit's refused fetch ("due to access control checks") and failed module import from the test's own server runs
    once more in a fresh page, as one whose page crashed (webkitLoadFailure in test/ui/harness.js); a second time
    fails it, and any other error fails it at once.
  • micro80 routed by router 1 on the Basys 3 runs on Linux only (#4429): test/nextpnr-xc7-cyinit.test.js joins the
    toolchain's heaviest tests (HEAVY in scripts/test-shard.mjs), which CI runs on Linux: on a slow Windows runner its
    micro80 test reached the 15 minutes a test may take.
  • The ECP5 and Gowin projects' UI tests expect their Emulate Post-… rows (#4426): since #2299 an ECP5 project has
    Emulate Post-Synthesis Model and Emulate Post-Place & Route Model, a Gowin one Emulate Post-Synthesis Model (its
    routed design not modelled yet); the tests check those rows and the absence of the others.
  • The Circuit editor's UI tests list #4397's new ICs (#4425): the library's groups now hold 50 74LS gates, 17 4000
    gates and 38 sequential ICs (the flip-flops, latches and registers 7473 to 74574, 4013, 4027), which the phase 2 and
    MSI tests expect; the phase 2 test watches the 20 Hz clock tick for up to 3 seconds
    rather than 12 looks (a slow page missed its 1).
  • The tests that measure blinky follow its new debounce (#4424): blinky's two 2-bit debounce counters (#4377) in place
    of 22-bit ones make it 40 flip-flops and 22 carry cells smaller on the LCMXO2-1200HC (test/machxo2-measure.test.js:
    114 LUTs, 92 flip-flops, 42 carry cells), and moved speed_ctrl.vhd's state_register process to line 101
    (test/twr-paths.test.js).
  • The private-data history check: 3fa4fcb's message allowed (#4423): the commit that allowed the dummy address of
    test/github-checks.test.js (#4364) quotes it in its own message, which the history check reads too; the address is
    allowed in that message as well.
  • VHDL to Verilog: the lint's silinx: ignore comments kept (#4414): the translator carried no comment over, so a
    VHDL source's lint suppressions (-- silinx: ignore latch, -- silinx: ignore data-as-clock) were lost and its
    Verilog warned where it did not. A suppression on a process's or an if / elsif branch's line, or alone on the line
    above it, now ends that statement's Verilog line (if (en) // silinx: ignore latch,
    always @(posedge clk) begin // silinx: ignore data-as-clock). The translation's lint is then as clean as its
    source's. #4397's ICs written once in VHDL no longer need their Verilog patched.
  • Three-state registers and latches (#4397, group C): core/circuit-model-ics.js adds the 74173 (a 4-bit D
    register, its M and N and its G̅1 and G̅2 enables, its clear active high), the 74373 and 74573 (8-bit latches) and the
    74374 and 74574 (8-bit registers), written once in VHDL as the other ICs of the group are. An IC's threeState (its
    outputs, their enables) turns its outputs off once an enable is not at its level. The core leaves them out, and the
    engine and the IC's simulation model apply it. In the digital model a disabled output is high, and the circuit's top
    takes the enables as it does the 74595's (now several enables). Its outputs are three-state for Check (#4409). The
    catalogue's KiCad logic ICs that behave go from 158 to 174.
  • VHDL to Verilog: a process with several asynchronous controls (#4412): a clocked process whose asynchronous
    controls come before its clock's edge, as a flip-flop's preset and clear (if pre_n = '0' … elsif clr_n = '0' … elsif rising_edge(clk) …), was refused by the translator. One reset was the only form it read. Such a process now
    becomes one always @(posedge clk or negedge pre_n or negedge clr_n) block with the same if / else if chain, which
    simulates as the VHDL does. #4397's ICs, written in VHDL with their Verilog translated, need it.
  • The bus transceivers (#4397, group B): core/circuit-ics.js adds the 74242 (inverting), 74243 and 74245, their
    A and B pins inputs and three-state outputs at once. Each bit is a buffer each way, enabled by GAB̅ and GBA, or by
    DIR and OE̅. These pins are inout in the ICs' HDL and read as a net's readers in the circuit's HDL. A transceiver's
    two ways never conduct at once, so their loop through its pins is marked as no loop for the lint. KiCad's
    74LS242, 74LS243, 74LS245 and 74HC245 behave as them, whatever pin type their symbols give A. The catalogue's
    KiCad logic ICs that behave go from 132 to 136.
  • Three-state outputs on one net are a bus (#4409): Check Circuit reported two 74125 outputs on one net as Short circuit. A three-state output (tri on its pin) is now counted apart. Several on a net are a bus. With a plain
    output or a supply rail on the net, Check gives a warning, a short whenever one is enabled.
  • KiCad's DIP?16* footprint filters read (#4397): a KiCad symbol with no footprint and a DIP filter written
    the way KiCad's 74xx and 4xxx write it (DIP?16*, DIP?20*, with no width) is a DIP part. It gets as many pins as
    the filter says, its rows 7.62 mm apart up to 20 pins and 15.24 mm from 24 (or as the filter says). Before, only
    DIP*W7.62mm* was read, and 143 of KiCad's logic ICs were no part. They are drawn only for now. Their behaviour
    comes with #4397's next changes.
  • The three-state buffers (#4397, group B): core/circuit-ics.js gives the gates enables. A gate drives its
    output while all its enable pins are at their active level, and its output is off (Z) once one is not (X while one is
    unknown). This works in the engine, in the ICs' HDL, and in the digital model synthesis reads: there, a buffer's
    level while enabled, else the net's other drivers' or its default, as the 74595's outputs. It adds 11 ICs with their
    datasheets' pinouts. KiCad's _Split symbols are read as their chip (74LS244_Split as the 74244). The catalogue's
    KiCad logic ICs that behave go from 105 to 132. The 74242, 74243 and 74245 transceivers, whose pins are inputs and
    outputs at once, come next.
  • ICs written once in VHDL (#4397, group C): core/circuit-model-ics.js adds 18 flip-flops, latches and registers.
    Each has its datasheet's pinout and a core, a VHDL model of the chip written by small generators (a J-K or D
    flip-flop, a D latch, an S̅-R̅ latch). Try it runs the core live (coreModel in core/circuit-live-model.js), with
    the chip's power and the reading of its inputs. The circuit's HDL (core/circuit-hdl-core.js) instantiates the core
    in its simulation model and in its digital one, and its Verilog is Silinx's translation of the VHDL, never written by
    hand. The translator drops the lint's silinx: ignore comments (#4414), so the generator adds them back to the
    Verilog. The catalogue's KiCad logic ICs that behave go from 136 to 158.
  • KiCad's other gates as library ICs (#4397, group A): core/circuit-ics.js adds 20 gates of the 74 series (7407,
    7409, 7412, 7413, 7415, 7421, 7422, 7426, 7428, 7433, 7437, 7438, 7440, 74132, 74133, 74136, 74386, and the AND-OR-INVERTs
    7451, 7454 as the 74LS54 and 7455) and 10 of the
    4000 series (4002, 4012, 4023, 4025, 4049, 4050, 4072, 4073, 4075, 4077). Each has its datasheet's pinout, and the
    gates now have the buffer, XNOR and AND-OR-INVERT functions and a DIP of 16 pins. They run in Try it, Check, the HDL and synthesis,
    and the MCP's circuit_parts and edit_circuit have them, as the other gates do. KiCad's open collector outputs that
    its symbols call outputs (its 74LS22's) fit an open collector IC. The catalogue's KiCad parts that behave as ICs go
    from 73 to 105 (KiCad's standard TTL 7454 differs from the 74LS54 and stays drawn only). Show inside names the pins of the 4023, 4025, 4073 and 4075 (one of their gates has an input past
    its output, and its lines would cross) and of the AND-OR-INVERTs.
  • The KiCad catalogue's Arduino boards (#4394): scripts/kicad-catalogue.mjs adds the boards of the KiCad PCM's
    KiCad Library for Arduino Modules v4.2.0 (CC-BY-SA-4.0) whose pins fit the 0.1 inch grid: 32 sockets, but no
    shield. They are drawn only, because Silinx does not simulate a microcontroller. Like KiCad's own footprints, none
    of their drawings is kept, only their pads' places. A footprint whose pads are named by its pins (A0, 5V, D13) and
    that has no pad 1 takes its first named through-hole pad as its origin (originPad in core/kicad-mod.js). It used
    to be "no pad 1 through the board". The same rule adds SparkFun's triple LED.
  • KiCad's own basic parts in the catalogue (#4385): scripts/kicad-catalogue.mjs takes them from KiCad 9.0.0's
    Device, Transistor_BJT, Regulator_Linear, Switch and Display_Character libraries, each with the through-hole
    footprint Silinx picks for it (a 2.54 mm TO-92, not the 1.27 mm one). The import writes it as the symbol's
    Silinx.Footprint. Their previews are a box round their pads: KiCad's footprints are CC-BY-SA-4.0, so none of their
    drawings is kept. A pad within 0.1 mm of the 0.1 inch grid is on it (it was 0.01 mm), so KiCad's capacitors (2.50 and
    5.00 mm) fit. That also adds SparkFun's 5 A relay and its 12 mm push button, which behaves as the library's button.
  • A circuit IC's case on a concatenation, qualified by a subtype of its width (#4380): the ICs' VHDL (the digital
    architecture and the simulation's model) wrote case std_logic_vector'(s1 & s0) is, whose subtype is not locally
    static, as VHDL wants of a case's expression (XST's HDLCompiler:803, on the breadboard example in the ISE nightly);
    now case slv2'(s1 & s0) is, its subtype slv2 is std_logic_vector(1 downto 0) declared in the architecture. The
    breadboard example's HDL regenerated.
  • blinky's debounce in ticks, not clock cycles (#4377): speed_ctrl's counters cnt_f and cnt_s counted
    CLK_HZ / 50 cycles (20 ms), and no fixed width fitted every board's clock: 21 bits never reached it above 104.8 MHz
    (#3922), 22 left bit 21 constant at 100 MHz (XST's WARNING:Xst:1293 - FF/Latch <u_speed/cnt_s_21> has a constant value of 0 on the Nexys3 and the Atlys). They now count the prescaler's ticks (80 a second) while the button
    differs from its debounced value, under a tick decision box as the divider's (its VHDL translates to Verilog):
    DEBOUNCE = 3 ticks, 2-bit counters, every bit used on every board. A button is
    taken once held 3 ticks, 25 to 37.5 ms (it was 20 ms); a press held 2 ticks is not. A unit test checks the counters'
    width against DEBOUNCE - 1 at every board's clock, and simulates the chart's VHDL and Verilog. The chart's register
    widths are integers; a width given by a generic's expression is #4378.
  • Spartan-6: a design with ports without a LOC implemented in seconds again (#4374): since nextpnr's packer became
    the Spartan-6's Map (#3140) the pads of the ports with no LOC were placed by nextpnr's simulated annealing, 30–50 s a
    seed for 62 of them (on Windows test/s6-pad-paths.test.js timed out); they take Silinx's pads again, as with Silinx's
    Map: a 64-bit shift register on the Nexys3 in 5 s instead of 105 s.
  • VHDL simulation: a bit-string literal's weak values (#4370): a bit-string literal with L, H or W (b"LHW1",
    x"L1") keeps them as a string literal does since #3926 (before: 01X1, 00000001); the #396 test expects the
    weak values to_string shows since #3926 (LH0XWX1X, before 010XXX1X), which failed the nightly run.
  • The Board Emulator's short-click test on a slow page (#4369): a button clicked is held down until the design
    saw 100 cycles of it (#3109), and a slow page's run in flight, up to a quarter of a second's cycles (#3727), may take
    the board past them before the test looks: the button then rightly up. The test reads the button with the board's
    time (never up before the 100 cycles), and a press longer than a real one is let go at once, its LED toggled; the
    nightly's failure in the ISE interface was this, not blinky's debouncers (#3922), whose design it does not use.
  • A closed editor's side panels let go of (#4365, #4366): each side panel's handle (web/js/side-resize.js, #4316)
    listened to the narrow-screen media query with a listener of its own, never removed, which kept every closed
    editor and viewer alive (the implemented design, the ASM chart, the Board Emulator: 34 to 119 listeners and up to
    1000 nodes a time); one listener now serves every panel, each held weakly. The handle is placed in the frame after
    the panel is resized, not in the ResizeObserver's callback, where WebKit reported "ResizeObserver loop completed
    with undelivered notifications" as a page error.
  • The private-data check of main's history (#4364): the dummy address in a made-up rate limit's message of
    test/github-checks.test.js, in one commit's version of it only (gone since), is on the allow-list with its reason
    (scripts/check-private-data.mjs). Test: private-data (that file only; another address in it still found).
  • The Circuit editor's UI tests (#4367, #4368): circuit-look's ⋮ menu ends with the KiCad group (#4282's exports)
    and circuit's Parts panel with the KiCad catalogue (#4305), as intended; their expected lists say so, in English and
    in Portuguese.
  • xc7-timing-vivado on Windows (#4371): the placement's Tcl test reads tclsh's output and files with LF, as Tcl
    writes CRLF on Windows; its standard output is made CRLF on every system, so the case is tested everywhere, and
    placementCounts on a CRLF report is tested too.
  • A constraints file not in the zip is missing, but the device's default (#4356): #4288's fix stopped counting
    silinx.json's constraints as missing at all; the import (the server's importProjectEntries, File ▸ Import Project…
    and the MCP's import_project; the browser edition's projectFromZip) now leaves out only the device's default
    path, constraints/top<ext> (.ucf, .pcf, .lpf, .cst, .xdc), which a new project names before its file is
    written. Tests: project-zip, project-store and mcp-transfer (a file of its own and the default, on a Spartan-3E and
    an iCE40), server (its test of a constraints file the zip does not contain, failing since #4288), the UI test
    projects (the warning, in both interfaces).
  • KiCad parts that behave as the library's LED, slide switch and push button (#4349): core/circuit-kicad-parts.js
    knows them by Silinx.Model (led, spdt, button) or by what their symbols say. It matches their pins by name
    (an LED's A and K, whatever their numbers), by number (a button's) or by place (a slide switch's middle pin, its
    common). Each pin keeps its footprint's place and the library part's number, so the simulation, Check and the HDL
    treat them as the library's. An LED's colour is its own. Their drawing shows the LED lit, the switch's lever and the
    button's cap. Part info says what they behave as. Import KiCad Part… and import_part offer the three models. The
    catalogue's SparkFun LEDs and slide switches are listed as the LED and the slide switch of the library.
  • The KiCad catalogue's previews drawn as the parts (#4346): scripts/kicad-catalogue.mjs keeps each SparkFun
    footprint's drawing (as #4315 draws it, in grid steps, rounded, about 42 kB) and its pads' places in
    core/kicad-catalogue-data.js (CATALOGUE_LOOKS). catalogueDrawing (core/kicad-catalogue.js) makes a part of it
    for the preview, a DIP's for a part drawn only on a DIP, and drawnPartSvg (web/js/circuit-draw.js) draws it. The
    drawings are SparkFun's (CC-BY-4.0): THIRD-PARTY-NOTICES.md has their entry and licence.
  • Open Project's Stored In, the GitHub line (#4344): a project in a GitHub repository shows its storage, then
    "↔ GitHub: owner/repo": the repository it is linked to (opened from, stored in), not a second place its files are
    kept; the tip unchanged.
  • Open from GitHub Repository… first (#4343): on the Start page's git group (after the browser edition's Sign in
    with GitHub…) and in the Collaboration tab's Repository box (a project with no repository, or none open), Open
    from GitHub Repository… comes before Store Project in a GitHub Repository…, which applies to an open project only.
  • The pages' CSP lets the KiCad catalogue fetch from gitlab.com (#4335): https://gitlab.com joins connect-src
    (server/csp.js, the server's header and the static editions' <meta>), as #4312's fetch through GitLab's API
    needs. A test checks that every catalogue file's browser URL is allowed.
  • The GW1N's other primitives in gowin_pack's port (#4333): Silinx's JavaScript port of Apicula's gowin_pack writes
    them as Apicula 0.32 does, byte for byte on a test design each (the toolchain package's changelog lists them). A
    differential pair takes an IO_LOC for each of its two ports.
  • A flaky wait in the live sharing's MCP test (#4330): test/mcp-collab.test.js's #3355 test waited for the
    refused change (a NUL byte) to be undone at the MCP client's with a read that failed on any error; until the
    owner's undo reaches the client, which may come after the browser follower's, follow_project reads back the
    client's own change and says it is not shown (binary content), as the interface's follower does. That error,
    and only that one, now counts as "not yet" (any other still fails the test), and the test asserts it right after
    the write, read back in the same call.
  • The Start tab alone with no project (#4322): #4257 hid the whole tab row while no project is open; the Start
    tab now stays, alone and active, in both interfaces (the ISE style's row, the Silinx interface's activity bar with
    its Start icon, on a narrow screen too), so that the panel looks the same as with a project; Design, Files,
    Libraries, Teaching and Collaboration come back when a project opens, as before.
  • The git commands' names (#4320): Initialize Repository… (File ▸ Git, the Collaboration tab, the Start page),
    which puts the open project in a GitHub repository, is Store Project in a GitHub Repository… (Portuguese: Guardar o
    Projeto num Repositório GitHub…), and Open from Repository… is Open from GitHub Repository… (Abrir do Repositório
    GitHub…): the commands, their dialogs' titles, the hints that name them ("… puts it in one"), the MCP's git_init
    and git_clone descriptions, the README and the guides.
  • A symbol's file in a .kicad_symdir named as KiCad 10 names it (#4318): a symbol taken out of its library into a
    file of its own (the KiCad catalogue's import) gets its name escaped as KiCad's EscapeString does for file names
    (/ as {slash}, \, ", <, >, |, :, tabs and line ends too: symbolFileName in core/kicad-edit.js).
    A PCM package's .kicad_symdir files named that way are read too.
  • Every side panel's width dragged (#4316): a handle over the panel's inner edge (web/js/side-resize.js, mounted
    by mountSideFold and the schematic's mountSchPanels, so every side panel has it), dragged by the mouse, a pen or
    a finger, ← → from the keyboard when focused; between 160 and 720 px, the canvas keeping 200 px; remembered per
    panel (silinx.<editor>.<panel>.width); a double-click restores the default width. Folded, the panel keeps its
    strip and has no handle; on a narrow screen (the panel over the canvas behind ☰) the handle works too, the panel at
    most 90 % of the editor. The width is the view's only: no MCP tool.
  • A KiCad part drawn only by its footprint's drawing (#4315): core/kicad-mod.js reads a footprint's F.Fab
    shapes (F.SilkS when it has none): lines, rectangles, circles, arcs (KiCad 5's too) and polygons, filled or not.
    core/circuit-kicad.js keeps them with the part, in grid steps from pad 1, turned as its pins are, and finds an
    LED's colour. web/js/circuit-draw.js draws them: the closed shapes are the body, the lines are its marks. A part
    whose footprint is known by its name only keeps its dashed box. Its name goes over the body when it fits, else
    above it (U2 Red no longer runs over the body).
  • Open Project's Stored In (#4314): the list has a Stored In column, before Last Used: Workspace Folder (the
    server's, or the folder the browser edition opened) or Browser Storage, and GitHub: owner/repo under it for a
    project in a GitHub repository (cloned, or one Initialize Repository… gave a remote); its tip says it in words. The Top column, which said little, is gone. The
    server's list of projects (GET /api/projects) gives each one's stored and github (server/project-places.js:
    its origin's, with the system's git; its record's, with GitHub's API; none with git off), the browser edition's its
    storage and its repository; the words are core/project-place.js's. The MCP's list_projects says the same
    (stored, github, and its line's "Kept in the workspace folder…").
  • The KiCad catalogue's KiCad parts with no local server (#4312): gitlab.com's raw files send no CORS header, so
    the browser's edition fetches them through GitLab's API (browserUrl in core/kicad-catalogue.js), which serves
    the same files with one; their SHA-256 is checked as before. The local server keeps the catalogue's URLs.
  • The Parts panel's test knows the KiCad catalogue's group (#4310): test/ui/circuit-parts.test.js listed the panel's
    groups without the folded KiCad catalogue added by #4305, and so failed on main.
  • Import KiCad Part…: a PCM .zip or a .kicad_symdir folder listed once (#4307): when the file input fired twice,
    both reads listed the package's libraries, so each one appeared twice. The table is now cleared once the read is
    done, for the last pick only.
  • The KiCad catalogue (#4305, #4278): core/kicad-catalogue-data.js (written by scripts/kicad-catalogue.mjs) says
    where each of 212 parts is: its library file's URL and SHA-256 at a pinned version (KiCad 9.0.0's 74xx and 4xxx,
    SparkFun's v10.0.0), what it is in Silinx (as which IC, or drawn only) and its footprint's file. It holds none of a
    library's content. The Parts panel's KiCad catalogue group (folded, opened by a search) imports a part with a click
    (core/kicad-catalogue.js): its symbol and those it extends, each a file of its own in
    parts/<library>.kicad_symdir/, its footprint beside it, its origin written in it. Every file is checked against its
    SHA-256, and a file there with the same text is kept. Its files come through the local server (the catalogue's only,
    kept for offline use) or, in the browser's edition, from SparkFun's CORS host. Portuguese too.
  • MCP (#4305): part_libraries with catalogue (searched by query); import_part with part.
  • SILINX_KICAD_OFFLINE (#4305): the local server fetches none of KiCad's libraries and uses only what it kept.
  • The Circuit section hidden while a part or a wire is selected (#4303): the circuit editor's Properties showed
    the circuit's section under the selection's. It now shows only the selection's, and the circuit's again once
    nothing is selected (web/js/circuit-side.js).
  • Show inside: an IC's leads unfilled (#4299): the leads from a gate IC's pins to its gates' inputs and from their
    outputs to their pins (#4224) are bent polylines, and their style set no fill, so SVG's default black fill closed
    each one into a black triangle or polygon (in both themes, both symbol standards, the print and the exported image).
    Each lead is drawn with fill="none" and the editor's style of the leads (circ-gate-line) sets fill: none.
  • A Silinx property set on a KiCad symbol written on one line (#4297): its new property took the whole text before
    it on that line as its indent, writing the symbol's beginning twice, and the library could not be read again
    (a "(" not closed). It now takes the line's leading white space only.
  • Safari's pinch (#4296): macOS Safari sends a trackpad's pinch as gesture events (gesturestart, gesturechange,
    gestureend, their scale), not as a Ctrl + wheel. canvas-gestures.js's safariPinch zooms by each gesturechange's
    scale over the last one, about the pointer, the page itself not zoomed; where a browser sends a Ctrl + wheel too,
    the first of the two to come in the gesture zooms, the other is dropped; on iPadOS, whose two fingers on the screen
    send gesture events too, these are twoFingerView's (a finger down on the drawing: the gesture events zoom nothing),
    its trackpad's pinch the gesture events'. web/js/view-gestures.js wires it on every drawing.
  • Each server call its own project's KiCad parts (#4294): the project's parts were kept in one registry for the
    whole server. Two MCP calls (or HTTP requests) on two projects, run at once, could check, simulate or make the HDL of
    a circuit with the other project's parts, or none. Each call now has a store of its own (server/parts-scope.js);
    the browser, one project open at a time, keeps the module's.
  • A drawing's view gestures in a module of their own (#4291, #405): the schematic editor's wheel (zoom about the
    pointer, as smoothly as the wheel or the pinch moves; a trackpad's scroll pans) moved as it was from
    web/js/sch-mouse.js to web/js/canvas-gestures.js, for the other editors to share; no change of behaviour.
  • A drawing's view gestures, shared (#4291): web/js/view-gestures.js wires the schematic editor's wheel
    (canvas-gestures.js), its two fingers (ui.js twoFingerView), and a drag of the view by the middle button or by Space
    and the left one (in the capture phase: the editor's tools see none of it) on every editor and viewer of a drawing:
    the schematic (its own Space and middle button kept), the circuit, the ASM chart (its own Space kept), the state
    diagram, the Petri net, the symbol editor (the drawing scrolled in its frame, centred by its margins so that all of
    it can be reached zoomed in; one finger on its background drags it), the RTL schematic viewer and the FPGA view (two
    fingers new to both), and the ink pad; each keeps its zoom's limits. The waveform viewer keeps its own (a time axis:
    the wheel scrolls its rows).
  • The constraints path is not a missing file (#4288): silinx.json's constraints is where the project's
    constraints file is or will be (a new project's constraints/top.ucf, not written yet), not a file the project lists:
    the import of a Silinx zip (the server's importProjectEntries, File ▸ Import Project… and the MCP's
    import_project; the browser edition's projectFromZip) no longer counts it as missing when the zip has none, as the
    .xise leaves it out (#4276); the files silinx.json lists are still checked. Tests: project-zip and mcp-transfer (a new
    project's Silinx zip exported and imported back, nothing missing, no warnings), project-store and server-api, the UI
    tests projects and standalone (the server and the browser editions, in both interfaces).
  • The Design Summary drawn with the reports as they are when it is opened (#4287): its refreshes are numbered and
    only the latest draws, an older one answered after it (older reports) dropped; Design Summary (openSummary)
    resolves once the summary is drawn, not before its reports are read. The UI test gowin-device (#2303) read the table
    drawn at the project's opening, its Status column empty; test/ui/summary.test.js holds the reports' requests and
    answers them out of order.
  • KiCad's libraries fetched by the local server (#4285, #4278): the PCM's official index and its library packages,
    and KiCad 9.0.0's 74xx and 4xxx symbol libraries (core/kicad-pcm.js, server/kicad-libraries.js). They come from
    an allowed list of hosts only, over HTTPS, with each redirect checked and the size limited. A package is read only
    once its SHA-256 is the index's (KiCad's own libraries: one pinned). Only a zip's symbols/ and footprints/ are
    inflated, with no path out of it. They are kept in the configuration folder for offline use, and the index is fetched
    again after a day.
  • Browse KiCad Libraries… (#4285): a dialog of the Parts panel, on the local server only. It searches the libraries
    and shows each one's version, licence and size. A library picked shows what its symbol libraries hold; those ticked
    are imported with the footprints their symbols name and their source, author and licence. Replacing files asks first.
  • Import KiCad Part… takes a PCM package's .zip and a .kicad_symdir folder (#4285).
  • MCP (#4285): part_libraries (the libraries, a search, what one holds); import_part with library, a .zip or
    a .kicad_symdir folder, and libraries to choose.
  • A 3-state output fits a library IC's output (#4285): the 74595's in KiCad's libraries (tri_state) now behave as
    Silinx's 74595. A footprint named PCM_<library>:<name>, as the PCM's installed libraries are, is found in
    parts/<library>.pretty/.
  • A KiCad part's module run live (#4284, #4278): core/circuit-live-model.js gives each such part a copy of its
    module of its own in Silinx's HDL simulator. The circuit's simulation (core/circuit-sim.js) forces the part's
    input levels on the module's inputs, runs it a step, and drives its outputs' levels after the part's delay (10 ns),
    its registers keeping their state. The live run and the generated HDL's simulation agree vector by vector (an
    inverter and a toggle flip-flop, VHDL and Verilog); a flip costs about what a 7404's does.
  • A module that cannot run live (#4284): its elaboration's errors, or a $finish or a failure as it starts, are said
    by Check Circuit and simulate_drawing ("U1 (TEST-INV): its module test_inv cannot run in Try it: its simulation
    stopped (finish)"), and its pins drive nothing. The warning "Try it does not run it" of #4283 is gone.
  • Where: the editor's Try it and Check (the project's compiled HDL), the MCP's check_drawing, edit_circuit and
    simulate_drawing (core/drawing-sync.js projectHdl).
  • A KiCad part modelled by a module of the project (#4283, #4278): Silinx.HDL on a symbol names a VHDL entity or
    a Verilog module of the project. Its ports are the part's pins (by their names, or as Silinx.Ports maps them; the
    supply pins are no ports). The circuit's generated HDL instantiates it in its simulation's model and in its digital
    one (core/circuit-kicad-hdl.js, core/circuit-hdl.js, core/circuit-digital.js), and the module is not copied
    into the circuit's file. Check Circuit, given the project's modules (the editor's, check_drawing's, edit_circuit's),
    refuses a module missing, a port without a pin, a pin without a port, a direction or a width that differ. It warns
    "U1 (TEST-INV) is simulated by its module test_inv in the HDL only: Try it does not run it".
  • In the editor (#4283): such a part has a whole outline, its tip and Part info name its HDL module, and Import
    KiCad Part… has an HDL module column. import_part's symbols.<name>.hdl and ports write the properties. Portuguese
    too.
  • The library's ICs as KiCad symbols (#4282, #4278): core/kicad-export.js writes each of the 48 ICs as a KiCad
    symbol at KiCad 9's file version. A gate IC has a unit per gate and a unit of its supply; an MSI or sequential IC has
    one unit. The pins sit on the 2.54 mm grid, an active low one named ~{CLR}, with a DIP footprint and its behaviour
    in Silinx.Model and Silinx.Family. Silinx reads every one of them back as that IC, by its properties and by its
    part number alone.
  • A project's KiCad parts as a PCM package (#4282): a .zip with the PCM schema's required metadata, the project's
    parts/ files as they are and, ticked, the library's ICs.
  • The editor's ⋮ menu (#4282): Export ICs as KiCad Symbols… (all of them, or this circuit's) and Export Parts as
    KiCad Package…, downloaded, their lines in the console. Portuguese too.
  • MCP (#4282): export_part, to the exchange folder.
  • The circuit editor's KiCad parts (#4281, #4278): a circuit opened reads its project's KiCad parts. The Parts panel
    lists them last (KiCad parts (project)), by part number and description. The canvas draws a part that behaves as an IC
    as that IC under its own number, and a part drawn only with a dashed body. The Part info adds the KiCad symbol and file,
    the part number, the datasheet (a link), the footprint, the source, author and licence.
  • Import KiCad Part… (#4281): the Parts panel's dialog picks a .kicad_sym or a .kicad_mod, lists a library's
    symbols with what each one is, and lets each one's Model, Family and Footprint be set, with the source, author and
    licence given once. It imports the file as the MCP's import_part does (core/kicad-import.js, the same checks and
    messages), each symbol said in the console. Every label is in Portuguese too.
  • A project's KiCad parts (#4280, #4278): the .kicad_sym libraries, .kicad_symdir folders and .kicad_mod
    footprints of a project's parts/ folder are its circuits' parts, named parts/<library>.kicad_sym#<symbol>.
    Nothing converted is stored: they are read again whenever a circuit is (core/circuit-kicad.js). A symbol behaves
    as an IC of the library when its part number is one or its property Silinx.Model names one, and its pins are that
    IC's, pin by pin. It is then simulated, checked and made into HDL as that IC, under its own name; Silinx.Family
    makes its inputs another family's. Any other symbol is drawn only: its pins on its footprint's places (a .kicad_mod,
    a DIP's or a pin header's name), not simulated, with Check Circuit's warning "U2 (TEST-SENSOR) is drawn only: it is
    not simulated and its pins drive nothing". A symbol with no through-hole footprint is no part, its reason said.
  • What KiCad has no field for, in the symbols' properties (#4280): Silinx.Model, Silinx.Family,
    Silinx.Footprint, Silinx.Source, Silinx.Author and Silinx.Licence are written in the library's text
    (core/kicad-edit.js), the rest of the file as it was, so the file stays valid in KiCad.
  • MCP (#4280): import_part copies a library or a footprint of the exchange folder into parts/, writes the
    properties given and says each symbol (as which IC, drawn only, or no part and why). circuit_parts with a project
    lists its KiCad parts and the symbols that are no part. edit_circuit, check_drawing, simulate_drawing,
    related_view and the drawings' HDL files take them.
  • KiCad's libraries read in core, the first step to KiCad parts in the circuit editor (#4279, #4278):
    core/sexpr.js parses KiCad's S-expressions with limits on the file's size, the lists' depth and the number of
    tokens (a malformed file is an error with its line); core/kicad-sym.js reads a .kicad_sym library (KiCad 6 to
    10) or a .kicad_symdir folder into its symbols: their properties, units and De Morgan bodies, pins (number, name,
    electrical type, style) and graphics, a derived symbol (extends) resolved from its parent, in another file of the
    folder too; KiCad 5's .lib is refused with what to do. core/kicad-mod.js reads a .kicad_mod footprint's pads,
    a through-hole part's pins on the 0.1 inch grid, and a footprint's name as its package (DIP, pin header, SMD).
    Tested on hand-written fixtures; no KiCad file is shipped.
  • The .xise lists the constraints file only when it is there (#4276): exportXise takes exists(path) and leaves
    out a constraints file it says is not there (a new project's constraints/top.ucf, not written yet); with no exists,
    listed as before. Every export passes it: File ▸ Export Project (.zip)… to Xilinx ISE (the server's
    exportProjectZip, the MCP's export_project, the browser edition's), the export.xise route and the .xise sync.
    Tests: server-units (exportXise), project-zip and mcp-transfer (a new project exported and imported back, no
    warnings; its UCF written, listed and imported with it), server-api (the route and the sync), the UI tests projects
    and standalone (the server and the browser editions, in both interfaces). The same project's Silinx zip imported back
    still names its constraints file as missing: #4288.
  • The ISE export's UI test of an arrayed instance after #3870 (#4275): test/ui/sch-array-ise.test.js expects the
    instance reg(3:0) written to ISE's schematic as reg_1(3:0) (reg, a reserved word ISE's sch2hdl refuses), the
    export's exact console lines (its warning on the renamed instance among them) and reg_1(3:0) imported back, its
    console with no warnings (it said a constraints file missing until #4276); it still expected reg(3:0), failing every
    run since #3870.
  • nextpnr's error when it stops with a WebAssembly exception (#4260): nextpnr-himbaechel-xc7 stopping on an
    ERROR: line throws an exception with no message; nextpnr's runner now turns it into nextpnr's last ERROR: line,
    and a run in a worker thread of its own (the local server's best of 3 seeds) fails with it instead of resolving with
    no files. The same for every family's nextpnr. Tests: packages/silinx-fpga/test/nextpnr-io.test.js,
    test/xc7-seeds.test.js, test/nextpnr-xc7-cyinit.test.js.
  • Share a Copy… in the Collaboration tab (#4259): web/js/collab-tab.js's Live Sharing box is named Sharing
    (Partilha: its own word, since the toolbar's presence says "Sharing", A partilhar) and has Share a Copy…, disabled
    with no project as File ▸ Share Project… is (one check, web/js/share-how.js canShare(), for the menu, the
    toolbar's button and the tab), before Share Live… or the live sharing's commands; it opens Share Project…'s window
    with A copy chosen and kept as the viewer's way (shareCopyDialog(), as Share Live…'s). The MCP's project_link already makes the copy's link. Tests: test/ui/collab-tab.test.js,
    test/ui/collab-tab-browser.test.js and test/ui/git-off.test.js (the box's name and entries, Share a Copy…
    opening the window with A copy chosen, in Portuguese, in both interfaces, no warnings).
  • No side panel tabs with no project (#4257): with no project open, the side panel's tab row is hidden in both
    interfaces (<html data-no-project>, set with the project's kind flag), and the Start page (New Project…, Open
    Project…, Open from Repository…, the recent projects) is all it shows. On a narrow screen the Silinx interface's
    activity bar goes too: the drawer opens from the header's button, the Start page from the window's edge. A project
    opened brings the tabs back, on the tab last chosen where nothing else picks one (the Silinx interface opens on its
    Design page, #1171), and closing it hides them again; in the server editions and the browser edition, for a design
    and simulation project as for an FPGA one.
  • Sign out of GitHub confirmed (#4256): web/js/git-signin.js signOut() shows a dialog first (Sign Out, and
    Cancel as the default button: Enter, Escape and Cancel keep the sign-in, its console line "Sign out of GitHub was
    cancelled: still signed in."), its text by how the user signed in: a pasted token (github.com/settings/
    personal-access-tokens linked) or the device flow (Settings ▸ Applications linked). The console's line after signing
    out says the same: forgotten in this browser, still valid on GitHub, and where to revoke it; never the token. How the
    sign-in was made is kept beside the token (web/js/git-browser.js, githubSignedInWay(): a pasted token or the
    device flow; one kept by an older Silinx: the edition's way) and forgotten with it. The local server editions have
    no Sign out (git signs in with its own credentials) and no MCP tool signs out: nothing changes there. Tests: unit
    (test/github-signin.test.js: the way kept and forgotten), UI (test/ui/git-sign-out.test.js: the dialog of each
    way, Cancel, Enter and Escape keeping the sign-in, Sign Out forgetting it, the exact console lines, from the File
    menu and the Collaboration tab, in Portuguese, in both interfaces, no warnings).
  • Test expressions escaped completely, the policy's sources compared exactly (#4247, #4248, #4249, #4250, #4251,
    #4252, #4253, #4254):
    CodeQL's alerts 188 to 192 (js/incomplete-sanitization): test/mcp-grade.test.js,
    test/mcp-new-source.test.js, test/ui/collab-relay.test.js and test/ui/new-source-paths.test.js escaped only
    . in the names they build a regular expression from, and test/symbol-standard.test.js replaced only the first
    & of an IEC qualifier; each replaces every & and escapes every special character now, backslashes included.
    Alerts 193, 194 and 195 (js/incomplete-url-substring-sanitization): test/static-editions.test.js looked the
    Content Security Policy's sources up with includes (an array's, never a substring); it compares each source
    exactly now. The tests assert the same. Tests: mcp-grade, mcp-new-source, symbol-standard, static-editions,
    ui/collab-relay, ui/new-source-paths.
  • Try it shared live (#4244): a new activity channel, try (core/collab-try.js, web/js/collab-try.js): the
    owner's page sends each open document's Try it, its editor's tryState(), at its sharing's tick when it changed; a
    follower's page puts it on its own document with setTryState(), opened and closed with the owner's. A state
    diagram's, an ASM chart's and a Petri net's Try it travels as its state (the present state, the registers, the
    marking, the inputs, the cycle, the trace); a schematic's (a Boolean algebra's, a circuit's and an RTL schematic's
    too) and a circuit's as their inputs then and the steps taken since it started, which the follower's own simulation
    replays (only what is new; a clock's cycles and its run sent as one growing step, within the relay's limits, #4173).
    What a follower takes is untrusted input: checked, capped, what is not one left out. A followed schematic can be
    tried read only. The follower's own click, key or input in a Try it is browsing alone. MCP: follow_project's
    ownerActivity.try and its lines give each document's Try it: its state, steps, inputs and outputs. Tests: unit
    (test/collab-try.test.js: the values made and taken, a schematic's and a circuit's steps replayed to the owner's
    values), follow_project's, and UI tests with the owner's and a viewer's pages for a state diagram and a schematic,
    in both interfaces, in Portuguese, no warnings.
  • The Parts panel's gate ICs in Settings ▸ Schematic Symbols' standard (#4242): their images were kept by part type
    alone, so each stayed in the standard it was first drawn in for the page's life (web/js/circuit-parts-panel.js);
    they are kept for each standard now, and the panel, drawn again on the setting's change (#4220), shows IEEE shapes or
    IEC boxes as the schematic editor's palette and Show inside do. Test: UI circuit-parts-symbols (new: a 7400 and a
    7432 in both standards, changed with the circuit open; no warnings), in both interfaces.
  • A drawing's HDL language in Source Properties only (#4240): the HDL Language choice #4063 put at the end of the
    drawings' ⋮ menus (Preferred Language, VHDL, Verilog) is gone from the schematic, state diagram, ASM chart and Petri
    net editors too (the Boolean algebra and the circuit lost it in #4159 and #4168): Source Properties choose it (the
    project's sourceLanguage; the MCP's set_source_properties, unchanged). web/js/drawing-lang-menu.js and the
    editors' menuItems that only carried it are removed; the state diagram's Try it lock keeps its own choices.
    toolbar-menu.js is no hub of scripts/related-tests.mjs any more (14 importers). Tests: UI
    drawing-source-language (was drawing-language-menu: no menu has it; a schematic, a state diagram and a circuit in
    Verilog and back by Source Properties; Portuguese; no warnings), the editors' menus' tests, in both interfaces.
  • The wires' drawing style not shared live (#4239): a schematic's and a circuit's wireStyle is a view setting
    (core/collab-model.js: view of the sch and circ kinds, viewFields, withView, withoutView), left out of
    the shared drawing's structure, so a participant's choice sends nothing; the owner's writes keep the owner's file's
    style (the page's and the MCP server's), and so do a follower's reads and Keep a Copy. A follower's or an editor's
    Wires choice (the ⋮ menu) is their view only, never an edit, a circuit's enabled read only too, and remembered for
    that shared project (web/js/collab-wire-view.js: by owner, project and drawing) and applied when the drawing opens
    again; the owner's change of style leaves theirs. MCP: follow_project write of a drawing with another wires' style
    says it was left out. Regression tests: the collab model (unit), follow_project's write, and a UI test with the
    owner and an editor (a schematic) and a follower (a circuit), both ways, in both interfaces.
  • GitHub Checks with a fine-grained personal access token (#4237): GitHub gives fine-grained tokens no Checks
    permission, so its Checks API refused them. When it refuses the token (a 403 that is not a rate limit),
    core/github-checks.js reads the commit's workflow runs and their jobs (Actions: read) and its statuses (Commit
    statuses: read) instead, in the same shape for the Collaboration tab, Git ▸ GitHub Checks and the MCP's git_checks,
    in both editions; the console says so first (the tests' verdicts, GitHub's annotations, stay unreadable), a
    permission refused as a warning, and both refused an error naming the two permissions to add. The token sign-in's
    steps list the two permissions. Regression tests against the fake GitHub (now with workflow runs, jobs, statuses and a
    fine-grained token's refusals): unit, MCP (git_checks, both back ends) and the Collaboration tab in both
    interfaces.
  • The Release waits for CI as long as it takes (#4236): the Release of v15.26.0 gave up after 150 minutes while
    its CI's second attempt (the failed jobs run again) was still running: the tags' runs go one at a time, so it had
    waited for the toolchain package's tag's run too, and the re-run passed 85 seconds later. scripts/wait-for-ci.mjs
    now has no limit of its own while a run is going (--timeout, none by default), reports a run's latest attempt
    ((attempt 2)), and gives a failed run --rerun minutes to be run again before the release fails; both Release
    workflows (the IDE's and silinx-fpga's) wait up to 340 and 345 minutes, a re-run within 30, in jobs of 360 minutes,
    GitHub's limit. Tests: wait-for-ci (the latest attempt, when to stop waiting, v15.26.0's runs replayed minute by
    minute: done at 153 minutes, where the 150 minutes' wait failed), the two Release workflows' limits.
  • A bitstream removed while the programming starts (#4235): programming the board (and verifying or programming
    the PROM) checked that the bitstream existed, then read it: a file removed in between (a rebuild, a clean) gave the
    system's raw "ENOENT: no such file or directory" instead of "bitstream not found". The read's ENOENT is now that same
    message, for a .bit, an iCE40's .bin, a Gowin's .fs, an ECP5's .bit and the PROM's image. The test made the file
    vanish at a random moment (it failed the CI of the release silinx-fpga-v0.4.3 on macOS): it now removes it exactly
    between the check and the read. Tests: programmer (program: request errors; prom: failures and request errors).
  • A follow link given while the app starts (#4234): web/js/app.js keeps the hash its start-up read: a shared
    project's link that comes after it is no longer read as a design's link (which refused it, the hash cleared), and
    web/js/link-hash.js watchLinks(target, seen) opens a link the hash gives that differs from the start-up's, once
    the links are watched (its hashchange came before the listener). The cause of test/ui/collab.test.js's "#3318 a
    follow link pasted into an open tab" timing out on a loaded runner: its follower changed the hash once the start-up
    checks were written, which may be before the app watches the links. Test: test/ui/collab.test.js "#4234" (a link
    given as the page loads, and a hash changed before the links are watched, opened then; no warnings nor errors).
  • UI tests: a project's folder given again after a try that timed out (#4233): the tests that import a
    project from its folder give it through setFolder (test/ui/zip-helpers.js): in Firefox and WebKit in tries
    of 15 s, the input emptied after one that timed out (three tries), since WebKit's folder listing now and then
    never sends the input's event that Playwright waits for (the browser edition's Silinx project imported from its
    folder, flaky in WebKit).
  • The Collaboration tab (#4215): a page of the side panel after Teaching (web/js/collab-tab.js), in the Silinx
    and the ISE style interfaces: Repository (the GitHub repository, linked, the branch, the commits to push or to pull,
    Commit…, Push to GitHub (not in the browser edition, whose commit is pushed), Pull from GitHub, History…), Changes
    (each changed, new, deleted or conflicting file with the Design pane's mark; a click opens Compare) and GitHub (the
    sign-in's state with Sign in with GitHub… / Sign out of GitHub in the browser edition, GitHub Checks and its last
    result: passed, failed, running or none, the details in the console) and Live Sharing (#3181: the toolbar presence's
    label, who is in with their role, and its menu's commands, presence() of web/js/collab-presence.js; not shared:
    Share Live…, Share Project's window with Live chosen, disabled where it cannot be, its reason its tip). The same items
    and functions as the Git submenu's and the presence menu's, which stay, as do the Start page's and the File menu's. A
    followed project shows the read-only notice and its Live Sharing box; with git turned off (SILINX_GIT=off) the tab
    has its Live Sharing box only, and no tab where nothing is shared live either. The ISE style's side panel is 330 px
    wide by default (300 before) and its tabs more compact, so the six show their names; in a narrower panel the long
    names are ellipsized. Portuguese labels; UI tests in both interfaces and both editions (a follower's tab too), on a
    narrow screen.
  • The git status's remote (#4215): a project's status (the server's git and GitHub's API) gives its GitHub
    repository (remote, owner/repository, null with none), and the MCP's git_status says it ("(GitHub: owner/repo)"
    or "(no GitHub repository)").
  • Pull and Merge Contribution… merge text files by their lines (#4214, phase 3 of #4204): the server's git
    (server/git-merge.js: the conflicting files' three versions from git's index, merged by core/text-merge.js, never
    git's markers) and GitHub's REST API (core/git-rest.js, core/git-rest-merge.js) give a pull's text conflicts per
    hunk (the error's texts, resolvable), and Pull opens the merge view, then pulls again with the choices
    (resolve: { path: { h1: 'mine' | 'theirs' | 'both' | { text } } }, checked by checkChoices); a file removed on
    one side and changed on the other still pulls nothing. Merge Contribution… (core/project-merge.js) merges a text
    file both changed by its lines (merged), its conflicts per hunk (Merge Lines… opens the merge view) or the whole
    file; with no ancestor each hunk that differs is a conflict. A renamed file (on GitHub or here, in the contribution or
    the project) is found by its lines and its other side's changes merged into it (renamed). The MCP's git_pull and
    merge_contribution take a resolution per hunk, and their results and merge_preview name each hunk with its line
    and both sides' lines.
  • The merge view of text files (#4214): web/js/merge-view.js, a document of its own for text files changed on
    both sides: each file's hunks side by side (the common ancestor, mine, theirs, a conflict marked), the result below
    them, and for each conflict Mine, Theirs, Both (mine's lines, then theirs') or its lines edited by hand; its toolbar
    holds the file, Previous / Next Conflict, the caller's action and Cancel, its status line the counts, its messages
    go to the console (an action with conflicts left is refused and the first one brought into view). It writes
    nothing: openMergeView({ files, labels, action }) gives its caller the choices per hunk (or null), which the
    caller's merge applies, so that no conflict marker ever reaches a project's files. In Portuguese too.
  • A text file's three-way merge by lines (#4214, phase 3 of #4204): core/text-merge.js, a diff3 on
    core/text-diff.js's diff, the same in the server and the browser: a hunk changed on one side only taken from
    that side, one changed alike on both taken once, hunks that overlap or touch a conflict (resolved by Mine, Theirs,
    Both or a text of the user's; never conflict markers), with no ancestor every difference a conflict; the lines
    compared without their CR, the line end and the final newline merged as the file's properties; a renamed file
    found by its lines (git's 50 % score). Checked against git merge-file and on every HDL file of the repository's
    designs merged with itself.
  • Git's branches (#4213, phase 4 of #4204): New Branch… (made from the last commit, its name checked as git checks
    it, the project on it, its changes kept), Switch Branch… (a branch only GitHub has made here, following it; the
    changes committed first, the other projects' of the workspace's repository too; what it brings passes the import gate) and Merge Branch… (GitHub's version of the branch
    when it is newer; the merge of a pull, server/git.js mergeCommits, its conflicts resolved in the merge view of
    #4214 and the drawings' dialog, then a merge commit), in the server editions through the system's git
    (server/git-branch.js) and in the browser edition through GitHub's REST API (core/git-rest-branch.js: the
    branches listed, a branch made on GitHub, the two branches' merge base from GitHub's compare, a merge commit with both
    parents). The MCP server's git_branch, git_switch and git_merge do the same, through the same functions and
    messages.
  • Two live-collaboration tests that timed out now and then (#4210): test/ui/collab-sim.test.js "#4002" compared
    the viewer's simulation with the owner's as the owner's page showed it at one moment, after a run or a restart:
    the owner's Objects values are drawn at its next frame, so on a loaded runner that snapshot could be one it never
    settles on; both are now read again at each look. test/mcp-collab.test.js "#3355" wrote the MCP client's second
    change once a page saw the owner undo the refused one, not once the MCP client itself had: written over its copy
    with the NUL byte's line still in it, it merged as by Bot… (no -- ), never the text the test waited for on the
    disk; it now waits for the MCP client's copy too.
  • Project ▸ Merge Contribution… and Undo Merge (#4206): the window lists the files that differ, what the merge does
    with each, and each conflict's Mine / Theirs (a drawing's per object, silinx.json's per field). Merge is refused
    until every conflict is chosen. A file's Compare closes the window and opens the comparison, and the console's line
    opens the window again with the choices kept. Every message goes to the console: the warnings, the merge's line with
    Undo Merge's button, and the project's untrusted line with Trust Project's. In the server editions the merge runs
    through server/merge-routes.js; in the browser edition, through its back end (web/js/merge-local.js, the backup
    in the browser's storage). In Portuguese, Juntar Contribuição… and Desfazer Junção.
  • The merge of a contribution into a project (#4206, phase 2 of #4204): a contribution (a Silinx project's zip or
    folder, through the import gate) is compared with the project file by file through their common ancestor (the one
    the copy carries, else the project's snapshot of its origin, #4205), or two-way with none, said. A file one side
    changed, added or removed is taken (or kept). A drawing both changed is merged by its objects (#3197, #4202), and
    its HDL is made again. A text file both changed is a choice of Mine or Theirs, and so is each of a drawing's objects
    both changed. silinx.json is merged by its fields (the files' list by path), never taking the contribution's id,
    origin or trust.
  • Undo Merge and trust (#4206): a merge that takes any of the contribution's files makes the project untrusted
    (madeBy: "merge"). Undo Merge puts the project back as it was, unless it changed since (core/project-merge.js,
    core/merge-apply.js, server/merge-projects.js, the MCP's server/mcp-merge.js).
  • A copy's ancestor recorded (#4205, phase 1 of #4204): Export Project (.zip) as Silinx, the MCP's
    export_project and Share Project's links give a copy. The project gets an id (once), and the copy's
    silinx.json names its origin (origin: { id, rev, at }: the project's id, the revision of its files, the date).
    The zip carries the ancestor (silinx.base.json), which the project keeps too: a hidden .silinx/snapshots/ in the
    server editions, the browser's storage in the browser edition, the newest few. An import gives the copy an id of
    its own and keeps its origin and its ancestor apart (checked, never a project file), and a copy of a copy keeps
    its first origin. The interface's description never changes the id or the origin. This is what Merge
    Contribution… (#4206) will merge with (core/project-rev.js, server/project-snapshots.js,
    web/js/snapshots-local.js).
  • The Design Summary opened late stays behind (#4021): Implement brought forward opens its Design Summary once
    summary.js is loaded, which may take a while; by then the viewer may have opened another document (it stays in
    front, the summary opened behind it) or gone to another task (the summary is not opened: Implement opens it when it
    comes forward again). Two UI tests hold summary.js back (CDP's Fetch) to show it; the UI test "#1409 Tasks (dark)"
    timed out on CI this way, its simulation hidden behind the summary.
  • XST's warnings on a circuit's chips left out (#4020): a circuit on the FPGA instances each MSI or sequential chip
    with its unread outputs open, and XST warns on them (Xst:753) and on the flip-flops behind them (Xst:2677). The ISE
    run (and the MCP's run_process) leaves those two messages out when they are on a circuit's chips, found by their
    names in the circuit's generated HDL (core/circuit-xst.js); any other instance's still warn. The console says how
    many were left out (INFO:Silinx - … warning(s) of XST left out), and the synthesis report's count of warnings is
    without them.
  • A flip-flop's tied controls left out of the digital model (#3975): the 7474, 7476 and 74112 have the generics
    USE_PRE1, USE_CLR1, USE_PRE2, USE_CLR2 (VHDL booleans, Verilog parameters, true by default), and their
    digital model (core/circuit-seq.js) a flip-flop for each combination (a VHDL if-generate, a Verilog generate if), with
    the preset and the clear, one of them, or neither. core/circuit-digital.js sets them on the instance: a control whose
    net is the supply's + or no one drives (a 74LS input reads it high) false, its port the constant '1'. The simulation's
    model is unchanged (the chip's). In Verilog a preset tied high used to stop the open synthesis with Yosys's "Multiple
    edge sensitive events" error (the front end's, #4332); now it is left out. docs/CIRCUITS.md follows. Tests: circuit-ff-controls (the instance's
    generics, the entity's generates, the open synthesis's exact warnings with the preset tied, the clear tied, both
    driven, the generated HDL compiled, checked and simulated against the engine, VHDL and Verilog).
  • std_logic's resolution of the weak values 'W', 'L', 'H' (#3926): a value keeps the mask of its weak bits (a
    literal, a copy, a slice); the simulator resolves the drivers per bit as IEEE 1164's resolution table (a strong
    value wins over a weak one; a W, or an L and an H, make W); 'image and to_string show L, H and W; to_x01,
    to_01, to_bit and to_bitvector make them strong, and the operators give strong values. 'U' and '-' are
    still simulated as 'X'. Tested against the standard's whole 9×9 table.
  • blinky's debouncers on a fast clock (#3922): speed_ctrl's debounce counters (cnt_f, cnt_s, in the ASM
    chart and its generated VHDL) are 22 bits wide instead of 21, so they reach DEBOUNCE - 1 (CLK_HZ / 50, 20 ms)
    on every board up to 209 MHz: on the Zybo Z7's 125 MHz (DEBOUNCE 2,500,000) the faster / slower buttons did
    nothing, and Vivado removed their logic (Synth 8-6014, Synth 8-7129), warnings now gone.
  • ISE schematic export, instances named with a reserved word (#3870): ISE's sch2hdl refuses a block named with a
    reserved word of Verilog ("is an illegal block name. It is a reserved word for verilog", in a VHDL schematic too).
    The export (core/isesch-instnames.js, called by core/isesch-export.js) writes an instance whose name, or an
    arrayed instance's base, is a word of Verilog-2005 or of VHDL, in any letter case, with _1 after it (_2 … when
    another instance has that name), as the import already names a keyword's copies, and warns
    instance 'reg(3:0)' written as 'reg_1(3:0)': 'reg' is a reserved word of Verilog, which ISE's schematic refuses as a block's name. The interface's Export Xilinx ISE Project and the MCP server's export_project share it. Tests: unit
    (test/isesch-instnames.test.js: every word of each language as a gate's name in a VHDL and a Verilog schematic,
    exported, its exact warning, read back with no warnings; another letter case; the names taken; an arrayed module's
    instance read back arrayed, its HDL compiled and elaborated with no warnings), export_project's
    (test/mcp-ise-array.test.js), and ISE's (test/ise/xise-interop.test.js, the nightly's: reg(3:0) built
    by ISE's project flow as reg_1(3:0), no warnings, and imported back).
  • The 7-series against Vivado in the weekly Quality workflow (#3834): a job of its own, before the ISE comparison's
    parts (its 22 builds about 2-3 hours; the ISE comparison takes the rest of the window and resumes), runs
    scripts/xc7-vivado-compare.mjs --guard --update on the self-hosted runner, Vivado's image there; its results
    (test/xc7-vivado-comparison.json, docs/TOOLCHAIN.md's table) are pushed to the branch quality-results with the ISE
    comparison's, each only when its comparison passed. --guard: nothing written, the job failed, when a build that
    passed in the recorded results fails now.
  • Two placements over 2000 slices at Speed (#3821): the toolchain's NEXTPNR_TRIES (silinx-fpga's changelog), so
    that the Quality comparison measures itc99-b17 at Speed within its 2-hour limit per build, which stopped it every run;
    the README and the Quality page say the counts (32 up to 100 slices, 8 up to 500, 4 up to 2000, 2 above).
  • The pins of an integer port with a range (#3820): core/auto-pins.js (the port bits with no LOC, to which the
    ISE flow gives free pins) gave every VHDL integer port 32 bits, x<0> to x<31>; XST makes an integer with a range
    only its range's bits (integer range 63 downto 0: 6; 127 downto -128: 8, signed), so ngdbuild refused the LOCs of
    the bits its netlist does not have. portWidth(t) gives XST's width (a plain integer: 32), for the pins and for a
    UCF's bus bits.
  • The Quality comparison's I/O wrapper (#3820): the corpus's wrapper (test/designs-corpus/lib.mjs wrapperPorts)
    connects such a port with those bits too: XST refused the 32-bit connection to a VHDL port of 20 or 30 bits
    (Xst:2587 … has different type in definition: ITC'99's b14, b15, b17). Both fixed, the weekly Quality run builds
    with ISE the ITC'99 designs it failed on (b04, b07, b11, b13, b14, b15, b17, b20, b21); femtorv32's instr, used before its
    declaration, stays XST's own limit, listed so on the Quality page. itc99-b15, whose wrapper XORs
    two bits fewer (1321 slices, 20.06 ns against 1336, 19.77), gets the corpus's PERIOD of 22 ns (periodNs, as #440's
    designs within a few per cent of 20 ns): its warning that 20 ns is not met would flip from run to run. Tests: test/auto-pins.test.js (the widths, the
    pins, a UCF's LOCs), test/designs-corpus.test.js (b14's wrapper: its connections, elaborated with no warning).
  • Vivado timing nextpnr's placement, the tooling (part of #3803): scripts/xc7-timing-vivado.mjs --placement places
    each example's netlist with nextpnr here (as the open flow runs it: the same arguments, now xc7ParArgs of
    open-flow-xc7.js, a BSCANE2 as the flow gives it, its design written
    with --write), and has Vivado route and time that placement: measure.tcl, after
    link_design, finds each cell's site from its tile's grid point (Vivado's GRID_POINT_X and _Y, nextpnr's tile names) and
    fixes it there (place_cell, IS_LOC_FIXED and IS_BEL_FIXED; the same constraints written as placement.xdc), then
    place_design places only what nextpnr's placement did not give (the IO buffers at their pins, a carry chain nextpnr
    moved by a MUXCY), and route_design routes it; placement.rpt counts the cells fixed and those not. The numbers go to
    each example's nextpnrPlacement in research/xc7/vivado-timing.json, Vivado's own measurement and the fit kept. It
    separates the placement's quality from nextpnr's estimate of its delays. Tests: xc7-timing-vivado (the cells nextpnr
    placed and those it made, the Tcl run by tclsh on Vivado's commands stubbed, the counts, the measurement file).
  • The 7-series: a 5LUT feeding another slice's CYINIT, its regression test (#3746): nextpnr's xilinx packer took
    a LUT pair's 5LUT whose only user was another slice's CARRY4 (its CYINIT) as feeding its own carry chain inside the
    site, so the placer put a flip-flop in the 5FF beside it, whose Q needs the same AMUX..DMUX output, and router 1
    (the browser's, when router 2 runs out of its memory) never finished micro80. The fork's fix (only the cluster's own
    CARRY4, on S or DI, counts) is in @silinx/nextpnr-himbaechel-xc7 since 0.2.0. test/nextpnr-xc7-cyinit.test.js
    holds it: a small design with that LUT pair, its flip-flop locked in that 5FF refused (it was taken, and router 1
    never finished), the design placed by nextpnr routed by router 1, and micro80 routed by router 1 on the Basys 3
    (seed 3; seed 1 routes too, in far longer).
  • nextpnr-himbaechel-spartan6 0.5.2 (#3743): the Spartan-6's nextpnr pinned at 0.5.2, whose packer has the rule of
    Silinx's Map for a carry's multiplexed outputs (#3151, pmnmalo/nextpnr-spartan#47): #3151's design, which nextpnr's
    Map (the default since #3140) stopped on, packs, places and routes. THIRD-PARTY-NOTICES follows. Tests: s6-nextpnr-pack
    (#3151's product in Verilog and VHDL by nextpnr's Map on the Nexys3, the application's flow: packed, routed on every
    seed, its three DSP48A1s, no warnings; it fails on 0.5.1).
  • The Nexys A7 (xc7a100t) in the browser edition, tested, and nextpnr's memory there measured (#3669): the browser
    edition's UI test of the 7-series (standalone-xc7) runs blinky and vga on the Nexys A7 too, to the .bit, with no
    warnings, its configuration data byte for byte the local server's flow's; the xc7a100t's bit database and chip
    database fetched once, the xc7a50t's not. nextpnr's router2 there takes 2043 MB of WebAssembly memory for blinky and
    2095 MB for vga (the same in Chrome, in WebKit and in Node: the same WebAssembly), about half of the 4 GB a page may
    have in Chrome and Safari, and a flow takes about 5 s in Chrome, 6 s in WebKit; router1, the browser's fallback when
    router2 runs out of memory, 90 MB and 164 MB (9 and 20 s). A new unit test, nextpnr-xc7-memory, guards these peaks
    (router2 within 2.5 GB, router1 within 256 MB, both routing the two examples with no warning).
  • The exports nothing used and the identifier rules' copies left by #3543 / #3545, cleaned up (#3599): the names
    of the drawings' parsers, Silinx HLS, live sharing, the MCP server and the 7-series in the browser that only their own
    module used are no longer exported, LANGS (core/asm.js) and typeName (core/c-types.js), used nowhere, are removed,
    and core/asm.js, asm-gen.js, logic.js, c-pragma.js and the MCP server's new_source, test bench and wizards take the
    identifier rules from core/ident.js; test/unused-exports.test.js's and test/ident-escape.test.js's exceptions keep
    only their reviewed ones. No change in behaviour.
  • The 7-series' builds checked without byte-identity (#3145): npm run check:xc7 (scripts/check-xc7.mjs) builds the
    examples (or project folders) through the open flow and checks each .bit with no Vivado: its frames the ones
    Project X-Ray's fasm2frames makes of nextpnr's FASM, the round trip by bits clean, and its pins: each port's IOB at its
    XDC pin, with its direction and IO standard, and no other IOB configured. --vivado then runs
    scripts/xc7-vivado-compare.mjs, which now also reads Silinx's own .bit with Vivado's write_cfgmem (its format
    checked; a column of the comparison's table in docs/TOOLCHAIN.md). The open-flow test of every example on the Basys 3
    and the Nexys A7 checks the pins too.
  • The Spartan-6's Map: nextpnr's packer the default (part of #3140): after its parity (#3139: the same slices, the
    periods within 5 % on average, ISE's DRC clean and the .bit byte-identical to ISE's bitgen on 20 of 20 designs),
    s6MapOf gives nextpnr unless a project saved impl.s6Map silinx, and the flow packs with nextpnr unless the
    spec's s6Map is silinx (the command line and the research scripts too). Process Properties lists nextpnr first, with
    no "(on trial)" (its Portuguese entry removed), and saves the Map only when it is Silinx's; a project saved with
    nextpnr's on trial keeps it, the choice dropped on OK. The MCP tool set_process_properties's s6_map: nextpnr (the
    default) or silinx. The ISE tests check nextpnr's Map by default, Silinx's with SILINX_S6_MAP=silinx.
    docs/TOOLCHAIN.md and docs/OPEN-TOOLCHAIN.md follow. Tests: mcp-s6-map, s6-nextpnr-pack (s6MapOf), nextpnr-s6
    (Silinx's placement of the block RAMs and DSPs with s6Map silinx), examples-nextpnr-s6-pack (Silinx's Map chosen
    for the comparison), the UI tests open-toolchain (the field, in Portuguese too, Silinx's saved, the default not saved)
    and standalone (the browser build packs and places by nextpnr).
  • The MachXO2 measured (#2302): scripts/machxo2-measure.mjs builds each example on a MachXO2 (or MachXO3)
    part with the open tools, outside the IDE: the Silinx toolchain's flat Verilog (as a ULX3S project) through Yosys's
    synth_lattice -family xo2 (xo3), nextpnr-machxo2 and ecppack, YoWASP's @yowasp/nextpnr-machxo2 0.11.825 (a
    development dependency), each bitstream read back by ecpunpack the same, and prints a table per part. Every example
    builds on the LCMXO2-1200HC-4TG100C and the LCMXO3LF-6900C-5BG256C with no warning, seeds 1 to 5, the board's 25 MHz met;
    the block and distributed RAM, the global clock buffer, the oscillator (OSCH) and the PLL (EHXPLLJ) too. The clock is
    given a pin: nextpnr-machxo2 fails to route a global clock from a pin it placed itself (YosysHQ/nextpnr#1236).
    test/machxo2-measure.test.js (the script's parts; gates, blinky, OSCH and EHXPLLJ built on the 1200HC); the results in
    docs/TOOLCHAIN.md. The MachXO2 is not a family of Silinx yet.
  • What proves an iCE40 or ECP5 build (Fixes #2301): npm run check:lattice-flow (scripts/check-lattice-flow.mjs) builds
    every example on the iCEBreaker and the ULX3S as the application does, and checks each three ways: nextpnr's timing
    meets the board's clock (icetime is not in YoWASP's packages); the bitstream read back by iceunpack / ecpunpack and
    packed again by icepack / ecppack is the same bytes; and the Post-Place & Route model (#2299) runs the example's test
    bench as the project's RTL does, the same log line for line, with no warning. It prints a line per example, and the failures. docs/TOOLCHAIN.md's "How a result is checked" says
    so in place of "to be decided". Test: check-lattice-flow (blinky on both boards; a bitstream packed again otherwise and a
    model that fails its test bench, each said).
  • Simulation models of the ECP5 primitives (Part of #2299): Silinx's own functional models, from Yosys's
    techlibs/ecp5/cells_sim.v and Lattice's guides (packages/silinx-hdl/src/ecp5-prims.js, ecp5-bram.js, ecp5-dsp.js): LUT4, LUT2,
    PFUMX, L6MUX21, CCU2C, INV, VLO, VHI; TRELLIS_FF (CEMUX, CLKMUX, LSRMUX, SRMODE, REGSET, LSRMODE) and the Diamond flip-flops
    FD1P3* / FD1S3* / IFS1P3* / OFS1P3*; TRELLIS_DPR16X4, DPR16X4C, TRELLIS_RAM16X2; TRELLIS_IO, IB, OB, OBZ, BB and their pull
    variants, OBCO, ILVDS, OLVDS, IDDRX1F, ODDRX1F; DCCA, DCSC, OSCG (310 MHz / DIV), EHXPLLL (its four outputs' frequencies and
    phases as ecppll sets them, LOCK, RST), GSR, SGSR, PUR, USRMCLK; DP16KD (the widths 1 to 18 on each port, the 36-bit pseudo
    dual port, the write modes, the output register, the resets, INITVAL) and MULT18X18D (signed and unsigned, its registers).
    primitiveSources gives a design the ECP5 library when it instantiates one of them (its LUT4 and INV by their pins: Z) or
    uses library ecp5u; the netlist cells run natively (native-lattice.js); the open synthesis keeps them as the ECP5's
    cells. Tests: ecp5-prims, ecp5-sim-clocks (each primitive against values computed in the test, in Verilog and VHDL; the
    PLL against Project Trellis's ecppll), ecp5-netlist (every example's ECP5 post-synthesis netlist passes its test bench as
    its RTL does, no warning), mcp-ecp5 (check_syntax, simulate and run_process on a design with EHXPLLL and OSCG),
    hdl-boundary, silinx-hdl-package.
  • Simulation models of the iCE40 primitives (Part of #2299): Silinx's own functional models, from Yosys's
    techlibs/ice40/cells_sim.v and Lattice's guides (packages/silinx-hdl/src/ice40-prims.js, ice40-ram.js, ice40-dsp.js):
    SB_LUT4, SB_CARRY, the 20 flip-flops (SB_DFF, E, SR, R, SS, S, ESR, ER, ESS, ES and their falling-edge SB_DFFN… twins), SB_IO,
    SB_GB_IO and SB_IO_OD (PIN_TYPE's input, output, enable and DDR registers), SB_GB, SB_WARMBOOT, SB_FILTER_50NS, the LED
    drivers SB_RGBA_DRV, SB_RGB_DRV and SB_LED_DRV_CUR; the PLLs SB_PLL40_CORE, PAD, 2_PAD, 2F_CORE and 2F_PAD (the reference ×
    (DIVF + 1) / (DIVR + 1) / 2^DIVQ, PLLOUT_SELECT, BYPASS, RESETB, LOCK), SB_HFOSC (48 MHz / 2^CLKHF_DIV) and SB_LFOSC (10 kHz);
    the block RAMs SB_RAM40_4K and its NR / NW twins (the four widths of WRITE_MODE / READ_MODE, MASK, INIT_0 … F), the
    UltraPlus's SB_SPRAM256KA and SB_MAC16 (its multipliers, adders, accumulators and registers). primitiveSources gives a
    design the iCE40 library when it instantiates one of them or uses library sb_ice40_components_syn (lattice-libs.js); the
    netlist cells run natively (native-lattice.js), as UNISIM's do; the open synthesis keeps them as the iCE40's cells. Tests:
    ice40-prims, ice40-sim-clocks (each primitive against values computed in the test, in Verilog and VHDL; the PLL's frequency
    against IceStorm's icepll), ice40-netlist (every example's iCE40 post-synthesis netlist passes its test bench as its RTL
    does, no warning), mcp-ice40 (check_syntax, simulate and run_process on a design with a PLL and the oscillator),
    hdl-boundary, silinx-hdl-package.
  • The iCE40's, ECP5's and Gowin's simulation models (Part of #2299): Generate Post-Synthesis Simulation Model writes the
    family's synthesis's netlist, and Generate Post-Place & Route Simulation Model nextpnr's routed design (its --write JSON
    through Yosys's write_verilog), as build/netgen/synthesis/<top>_synthesis.v and build/netgen/par/<top>_timesim.v
    (core/flow-models-nextpnr.js, core/netlist-verilog.js): the design synthesized with the generics the test bench gives it
    (declared as the netlist's parameters), placed and routed as Place & Route does it, the board's model beside it. Each simulates
    with Silinx's models of the family's cells, the routed design's packed ones too (ICESTORM_LC, ICESTORM_RAM, TRELLIS_COMB,
    TRELLIS_RAMW, TRELLIS_FF's SD, the pins' *MUX generics), without timing; the global buffers are left out of the model and
    each pad gets a net of its own (#4360). The Simulate / Emulate Post-… rows and the model choices are shown for them (the
    Gowin's Post-Synthesis only: its routed design's ALU modes are not modelled yet); simulate and run_process accept those
    models and refuse the others with what the flow makes. Tests: ice40-routed and ecp5-routed (every example's post-place &
    route model passes its test bench, no warning), netlist-verilog, ice40-prims, ecp5-prims (the packed cells), mcp-ice40,
    mcp-ecp5, mcp-gowin (run_process and simulate of the models), mcp-simulate-models, tasks, tool-descriptors,
    test/ui/sim-models-flows.

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):

0a5e0d67ad26312e19a1ac09069a5eb3050ba84c761785e8b42fd8b13de322bf  Silinx.html
a5fe626c66b105f0b92b8d70993a35292093eb0ec45f64fb02fce36aa82d8f8f  silinx-15.27.0.zip
05ecb6d65b825a4b74d320f3bedd018c46d43748cc63d531248dff7d76e4bf47  silinx-docker-kit-15.27.0.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.