Es40 port from gdwnldsksc (using claude) - #129
Merged
RaymiiOrg merged 41 commits intoJul 11, 2026
Merged
Conversation
Stage 1 of the ES40-Emu revival port: bump CMAKE_CXX_STANDARD from 11 to 17. Required by the MAME-derived VGA code and the asmjit-based JIT that later stages bring in. Verified: clean Release build, SRM boots to P00>>> (test/rom/test.sh). Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 2: support files the later S3Trio64/VGA/JIT ports depend on: mame_shims.hpp (bitmap/rectangle/device stubs), emu/emu.hpp, address_map.hpp, coretmpl.hpp, attotime.hpp, xtal.hpp, logmacro.hpp, diag_rpcc.hpp (RPCC wall-clock backdoor) and the i2c_spd memory SPD EEPROM device used by ARC memory detection. Only i2c_spd.cpp is compiled so far; nothing references the new code yet. Verified: clean build, SRM boots to P00>>>. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 3 of the ES40-Emu port: - System gains the Tsunami CCHIP MPD register (0x0C0) wired to an I2C bus with SPD EEPROMs generated from the configured memory size (ARC memory detection), plus the firmware-triggered reset request API (RequestSystemReset/IsSystemResetRequested; processing arrives with the CPU-core stage). - PCIDevice ported wholesale from ES40-Emu HEAD: diag_printf with RPCC stall exclusion (g_diag_excluded_ns), IO/ROM/MEM BAR sizing and alignment fixes, and the page-split DMA loops (superset of the existing AXPbox implementation, adds end-of-chunk validation). Verified: clean build, SRM boots to P00>>>. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…RCR reset Stage 4 of the ES40-Emu port: - Flash ported wholesale: AMD flash aliasing fixes (ignore address bit 2, out-of-bounds reads return 0xFF), HasBootFirmware() CPQ-signature detection, GetFlashBytes(), FlushIfDirty() and quiesce-flush in check_state(). This is what lets the LFU on the firmware update CD program AlphaBIOS/ARC into flash. - DPR ported wholesale (UTC-correct faketime handling, supersedes the previous implementation). - System.LoadROM() boots the SRM console from a flashed firmware image when one is present (self-decompressor run from flash contents), falling back to the existing cl67srmrom.exe / decompressed.rom path. - TIG SRCR0/1 writes (LFU exit) flush dirty flash and request a system reset; TIG mod_info now stored (new ModInfo state field). Verified: clean build, SRM boots to P00>>>, log diff clean. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ard, FDC, IDE, DMA Stage 5 of the ES40-Emu port, all converted to std::thread/std::mutex/ std::shared_mutex (no Poco-style wrappers): - AliM1543C: 8259 PIC rework (IRR/ISR/IMR model with pic_set_line), rudimentary SuperIO with PnP, TOY clock offset model with arc_year_compat (20-year ARC offset, default off) and faketime, get_interval_period_ns() for the CPU timer poll. The AXPbox-only 'timezone' option is re-implemented as an initial TOY offset seed (guest clock shows host local time by default). - Serial: disabled/raw_mode/null_attach options, telnet subnegotiation handling, baud-rate staged delivery, IIR-driven pic_set_line. AXPbox bind 'address' option, accept-guard on join and 5 ms poll re-applied. - Keyboard: Win2K/2210 keyboard+mouse fixes, kbd IRQ rework, mouse capture plumbing, input hardening (kbdLock as std::mutex). - FloppyController: near-rewrite; FDC works on AlphaBIOS/Windows NT (interrupt handling, no-drive abort, DMA transfers). - AliM1543C_ide: ES40-Emu HEAD (NT/Win2K boot blockers fixed, live alt-status reads, shared_mutex register locks, run(index) threads). The AXPbox fresh-value-read and sector-calculation fixes were already transplanted upstream. - DMA: send/recv_data with length, needed by the new FDC. - CSemaphore::set() now saturates at max instead of throwing, matching the ES40-Emu semaphore semantics the ported devices rely on. - test/rom/axp_correct.log refreshed: the new Serial emits one extra telnet negotiation triplet in the connection preamble (only change). Verified: clean build, SRM boots to P00>>>, log diff clean. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 6 of the ES40-Emu port. Wholesale port of the CPU family:
AlphaCPU, AlphaCPU_{ieeefloat,vaxfloat,vmspal}, all cpu_*.hpp dispatch
headers, datatypes and config_debug. Brings the revival's interpreter
correctness work: LDBU/LDWU zero-extension per HRM, HW_LD sign
extension, EV68 GH translation-buffer masks (fixes ARC screen
corruption and memory test), IEEE S-float rpack canonicalization,
STS/STF/FTOIS conversions, FPCR/inf/zero handling, ITB_IAP/IA/IS,
CAS-backed LL/SC model (cpu_lock/cpu_take_lock/cpu_clear_lock on
CSystem, EV6 64-byte lock ranges), data page cache, native VMS PALcode
improvements, and the wall-clock-pinned RPCC/interval-timer model
(guest clocks track real time). The old host-FP path (es40_float.hpp)
is removed; the softfloat path is always used.
Converted to AXPbox conventions: std::thread run loop with
myThreadDead, runtime skip_memtest_hack option (replaces the
compile-time SKIP_SRM_MEMTEST default), icache config default remains
off, checked fread in RestoreState.
System side: ProcessPendingReset/ResetChipsetState (firmware-triggered
reset), drir_lock serialization for interrupt delivery, clear_ipi.
Two fixes for bugs inherited from ES40-Emu HEAD, found while debugging
SRM boot on the icache-off configuration:
- execute() must advance the PC through next_pc(): the icache-disabled
fetch path depends on next_pc() keeping pc_phys/rem_ins_in_page in
sync. Upstream's bare pc += 4 leaves pc_phys frozen, so fetches
re-read one stale physical word and SRM runs away into pattern-filled
memory during the console banner (upstream defaults icache on and
never hits this).
- vmspal_int_initiate_exception() now tracks nesting depth; virt2phys
diverts to the real PALcode vectors when the native VMS fast path is
already delivering an exception. Asynchronous (wall-clock) interrupts
could otherwise land before the guest interrupt context exists and
recurse the host stack to a segfault.
test/rom/test.sh now ignores the 'CPU n speed is' line: SRM measures
speed from the wall-clock RPCC, so the value is host-dependent.
Verified: clean build; SRM boots to P00>>> (9/10 consecutive runs
clean, one inconclusive early cut; no crashes).
Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 7 of the ES40-Emu port. Adds src/jit/jitengine.{hpp,cpp}: a
block-compiling JIT over asmjit with direct-mapped block cache, block
chaining with lazy flush revalidation, inline data-page-cache loads,
SSE floating point, and interpreter bail-out for uncompilable or
faulting operations. Runtime field offsets into CAlphaCPU are computed
at init, so no fixed layout is assumed. JIT_VERIFY (differential
check against the interpreter) and JIT_STATS remain available through
config_debug.hpp.
Build: -DES40_DISABLE_ASMJIT=OFF enables the JIT (off by default).
asmjit is a plain pinned clone in third_party/asmjit (gitignored);
CMake errors with clone instructions when it is missing.
One fix over upstream: break_seq_icache() now also clears
rem_ins_in_page. Compiled blocks write state.pc natively, so the
icache-disabled fetch cursor (pc_phys/rem_ins_in_page) is stale when
interpretation resumes; upstream's Linux JIT loops on 'halt code = 5'
with icache=false and hangs at the SRM idle loop with icache=true.
With this fix the JIT lane boots SRM to P00>>> in ~18s under both
icache settings (verified against upstream es40 as a control, which
does neither on this host).
Verified: default lane (JIT off) builds and passes test/rom/test.sh;
JIT lane builds and boots SRM to P00>>> with icache on and off;
JIT_VERIFY run: 429M compiled-block executions, 0 mismatches.
Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 8 of the ES40-Emu port: - Sym53C810 wholesale: ARC boot blockers fixed (missing register reads, STO handling, SSTAT1 real-phase update, SIDL update), NT 4.0 register readbacks, GP/Gen timer and DFE stack-drain fixes, instruction runaway guard. The residual-state short-move fix and MA raise that AXPbox carried for Tru64 were already transplanted upstream. Threading converted to std::thread. - Disk wholesale: mode_sense_page rework, do_scsi_error with info byte, media_changed reporting, CD-R/RW command handling. The AXPbox START STOP UNIT workaround is preserved: opcode 0x1b answers SCSI_OK instead of illegal-command (upstream routed it to the CD-R/RW reject group). - DiskFile: reload_file with autocreate_size support, CD diskfile registry for runtime ISO swapping (the SDL3 file dialog activates with the graphics stage), flush(). The AXPbox default-filename fallback is preserved. - DiskController/DiskDevice/SCSIBus/SCSIDevice refreshed to ES40-Emu HEAD. fopen_large added to StdAfx.hpp. Verified: clean build, SRM boots to P00>>> with clean log diff, and 'show device' lists an attached IDE ATAPI CD (OpenVMS ISO) and the floppy correctly. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 9 of the ES40-Emu port: - DEC21143 wholesale: SROM v3 with corrected ID-block CRC and no phantom MII PHY, link-up status and transmit-descriptor fixes (ARC DHCP/self-test work), RX/TX buffer-2 descriptor handling for the NT driver, RX/TX wake semaphores, MBO bit, LFU reset fix, optional packet debug. Windows dynamic wpcap loading retained under _WIN32. Threading converted to std::thread. - Ethernet base class refreshed; lockstep updated to ES40-Emu HEAD. - New devices: ES1370 sound, MPU401 MIDI, AliM1543C PMU (power management shell). - gui_win32.cpp restored from ES40-Emu (fully _WIN32-guarded; empty translation unit elsewhere). Verified: clean build, SRM test passes 3/3 with clean log diff (one earlier run showed the known rare mid-boot console restart transient, still reaching P00>>>). Authored-By: Claude Fable 5 <noreply@anthropic.com>
…64, SDL3 GUI Stage 10 of the ES40-Emu port: - VGA base class becomes the MAME-derived VGA core (register dispatch, sequencer/CRTC/GC/attribute state, pel/DAC, screen_update render path); Cirrus is excluded from the build (not ported to the new core, disabled upstream too) and its config entry reports a clear error suggesting s3. - New ibm8514a device: MAME IBM 8514/A accelerator used by the S3. - S3Trio64 wholesale: full rewrite around the MAME core. Boots the real S3 Trio64 VGA BIOS (86c764x1.bin via the s3 'rom' config key), linear framebuffer with dynamic BAR handling, 2D acceleration via the 8514/A, hardware cursor, 4MB interlaced modes, text rendering, Windows NT high/true-color background fix, pause-instead-of-destroy window handling across firmware resets. Threading converted to std::thread. - GUI layer moves to SDL3 (upstream is SDL3 too): dynamic resolution and DPI scaling, linear filtering, mouse capture (Ctrl+F10), Ctrl+Alt+End sends Ctrl+Alt+Del for NT, Ctrl+F11 opens the CD-swap file dialog (DiskFile SDL3 dialog now active), runtime scale hotkeys. gui.cpp/gui.hpp gain a base graphics_frame_update so the X11 backend keeps working. SDL 1.2 is fully removed from the build (find_package(SDL3 CONFIG), no SDL/SDL.h includes remain). Verified: clean build linking SDL3; headless SRM test passes with clean log diff; DISABLE_SDL=yes build compiles; manual boot with pci0.2=s3 + gui=sdl + the 86Box Trio64 BIOS reaches P00>>> with the SDL window on WSLg. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 11 of the ES40-Emu port: - Configurator refreshed to ES40-Emu HEAD: s3, es1370, mpu401 and ali_pmu device registration, PCI slot validation for the extra reserved slot 17 and the PMU. The cirrus entry stays registered but reports a clear configuration error pointing at s3 (device not ported to the MAME VGA core, disabled upstream too). - es40-cfg configuration wizard refreshed: VGA card question with S3 VGA BIOS path, in-memory CD option, serial disabled/none ports, display scaling questions, fixed date/time question. The AXPbox icache and skip_memtest_hack questions are preserved (upstream had commented icache out and hardcoded true). Entry point remains main_cfg for the 'axpbox configure' subcommand. - Sample es40.cfg: arc_year_compat and extended time documentation, S3 + 86c764x1.bin example replaces the cirrus example, SDL video scaling options and runtime hotkeys documented, timezone docs kept. - VERSION is now a quoted string in config.hpp (banner printf pastes it). - AlphaCPU: the first Cchip interval-timer tick now fires one second after CPU start instead of immediately. An immediate tick could land inside the SRM console's earliest initialization and make it dump registers and restart (intermittent boot-log transient seen roughly one run in five). Verified: clean build; SRM test passes 6/6 consecutive runs with clean log diff. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stage 12 of the ES40-Emu port: run clang-format (LLVM style, repo .clang-format) over all files touched by the port so the tree is uniformly formatted again; make Ethernet.hpp self-contained (include sorting exposed its reliance on the includer providing StdAfx types). README documents the C++17/SDL3 requirements, the optional asmjit JIT build, and the S3/ARC/Windows NT setup; the ARC and OpenVMS-VGA entries leave the 'not working' list. Verified after formatting: fresh configure + build of both lanes; default lane passes test/rom/test.sh 3/3 with clean diff; JIT lane boots SRM to P00>>> in 18s. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Token-stream comparison of every shared source file against ES40-Emu HEAD confirmed the port is complete (all remaining deltas are the intentional AXPbox keep-list additions) except for five small items, now applied: - Sym53C895: don't delete the shared SCSI bus in the destructor (double free; the 810 already had the fix) - TraceEngine (IDB builds): trace-map init used CPU 0's slot for every CPU, and the branch-target address was computed from the branch source - StdAfx: PRINTF_BYTE_TO_BINARY debug macros (referenced by S3 debug paths) and <cerrno> - telnet.hpp: TELOPT_BINARY fallback define for platforms without arpa/telnet.h Verified: clean build. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Captures the verification work done after the ES40-Emu port as reusable pieces: - .claude/skills/: srm-boot-test (the P00>>> regression test and its pitfalls), boot-openvms (full-guest boot from the provided 8.4 image), test-arc (AlphaBIOS from the pre-programmed flash, S3 BIOS sources, timing caveats), verify-vga-sdl (framebuffer dumps, SRM-on-VGA text check, WSLg display notes), build-lanes (default / JIT / headless builds and pre-commit gates). - test/vms/: config + telnet drivers that boot OpenVMS end-to-end (SRM boot dqa0, date prompt answer, startup markers). Verified: OpenVMS 8.4-2L1 boots from the provided image through full system startup with correct 2026 timestamps. - test/arc/: config + serial watcher for the AlphaBIOS flash boot. Verified: SRM boots from flash on the graphics console (S3 text rendering pixel-verified), nvram script loads ARC firmware; ARC waits on its first-boot blue screen for an interactive keypress. - src/gui/sdl.cpp: two env-gated debug hooks used by the skills: AXPBOX_DUMP_FB=<prefix> (periodic PPM dump of the rendered frame) and AXPBOX_AUTOKEY_ENTER=<seconds> (periodic Enter through the emulated keyboard for headless firmware prompts). Media and ROMs (6 GB disk copy, firmware, VGA BIOS, dumps) are gitignored; scripts fetch what they need. Authored-By: Claude Fable 5 <noreply@anthropic.com>
JIT (upstream 7d94c9e..7140555 series): block cache 64K->256K slots, poly-link direct chaining (2 cached successors, round-robin backpatch), register allocation with global pins (R26/R16/R27; +3 on Win64, +1 SysV), standalone emit_op shared by block and trace compilers, value caching, smaller ALU codegen, HW_LD 0xa (virtual+WrChk), JIT_VERIFY fixes, and the dormant trace tier (JIT_TRACES, off by default: measured net perf loss, kept as correctness infrastructure). Timing (upstream d12d4e8): state.cc is wall-clock in both engines (DO_RPCC unified; odd deltas capped at 1s instead of dropped - the 357 MHz boot-calibration bug), Cchip interval timer driven by CPU0 at dispatch-batch boundaries with count-preserving paced catch-up (repay lost ticks at <= 2x nominal, drop >1s backlog), 8254/PIT advanced by elapsed host time at 1.193182 MHz with analytic pit_out() for jitter-free port-61h polling, and RTC register A preset to 0x26 at reset (battery-backed semantics: SRM's boot-CPU speed calibration sees 1024Hz ticks from power-on). The first interval tick now fires immediately like upstream; the 1-second grace is dropped (the RTC boot tick and vmspal_exc_depth guard cover the early-boot transient it worked around - verified over repeated SRM boots in both lanes). Per upstream ed9bf49, the icache and mouse.enabled options are gone: icache is always on, PS/2 mouse always present/captured (superio LDN8 reports 0x01 unconditionally). Kept axpbox-specific fixes: next_pc() in execute(), rem_ins_in_page clear in break_seq_icache(), vmspal exception-nesting guard, skip_memtest_hack, fread return checks, AXPBOX_PC_SAMPLE sampler, std::thread threading. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ings Port upstream BIN/CUE support (ES40-Emu PR lenticularis39#121 plus follow-ups): a cdrom file ending in .cue is parsed as a cue sheet (single- or multi-file BIN, MODE1/2048, MODE1/2352, MODE2/2336, MODE2/2352, AUDIO), with graceful fallback to a raw image on parse failure. New DiskFileBinCue.hpp defines CueTrack/CueTrackMode; CDiskFile gains the track query API and LBA-to-file mapping; Disk READ_TOC and track reporting use real track data when a BIN/CUE image is mounted. Disk identity values now accept both spellings: serial_number/serial_num and rev_number/rev_num (upstream 11666e8) - previously the documented long forms were silently ignored. Kept axpbox-specific behavior: default image filename fallback when a disk has no file attached, and START_STOP_UNIT accepted with SCSI_OK. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…DISCONNECT fix Serialize CPU-thread CSR/rx_queue access against the NIC worker thread (upstream 4bb0161): the worker holds the register lock for its whole batch (receive_process, TX drain, interrupt summary), and ReadMem_Bar / WriteMem_Bar / SaveState / RestoreState take the same lock. Written with std::recursive_mutex + std::lock_guard rather than upstream's CMutex wrapper, preserving the recursive semantics. Raise the default rx packet queue from 100 to 1024 (upstream d8a8d4e). Sym53C810 WAIT DISCONNECT (upstream 14d613e): instead of unconditionally freeing the bus, check the current SCSI phase - free only from arbitration; if the bus is still connected, release it on the passive target's behalf via free_bus(GET_DEST()) with a diagnostic, since emulated targets never drop BSY on their own. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…attach serials Port upstream 11666e8 / 43469dc / ed9bf49 configuration hardening: - Each device class now carries an allow-list of the configuration values it actually reads; anything else in that section gets a one-line %SYS-W-UNKNOWNCFG warning at startup instead of being silently ignored. The lists are extended for axpbox-specific options: skip_memtest_hack (ev68cb), timezone (ali), rom.decompressed (tsunami), address (serial). - ';' at the start of a line now works as a line comment. - Serial ports that are not configured are synthesized as null_attach with a %SYS-W-NOSERIAL notice (the ES40 always has two UARTs). - The kept axpbox cirrus class row still resolves, and still fails with the pointer to use an s3 section instead. Configuration wizard (axpbox configure): the sym53c895 wide-SCSI option, the mouse question/mouse.enabled output, and the icache question are gone, matching the removal of those options from the emulator. Shipped configs drop their icache = false and mouse.enabled lines so they start warning-free. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Port upstream 074b5e0 / b26fc37: new sdl-section options mouse.speed (0.0 < speed <= 10.0, default 1.0, warns and falls back to 1.0 on invalid values) and mouse.invert_x / mouse.invert_y. Scaled motion is accumulated fractionally across events so sub-1.0 multipliers don't drop slow movement. The allow-list entries landed with the Configurator port. Authored-By: Claude Fable 5 <noreply@anthropic.com>
… documentation refresh New src/banner.hpp (ported from upstream banner.h): a shared startup banner with runtime-computed padding, printed by both the run and configure subcommands. It credits the emulator's major authors by era: Camiel Vanderhoeven and the ES40 Emulator Project (2007-2010), Tim Stark / fsword7 (2018), Tomas Glozar (2020-2023), Remy van Elst (2020-2026), and Gary / gdwnldsKSC of the ES40-Emu revival (2023-2026). Shows axpbox's own version. All user-facing "ES40" text now says AXPbox: SDL/X11/Win32 window titles, the configurator banner, the serial-port telnet greeting (expected SRM test log updated to match), the interactive-debugger greeting, the memory-map header, and the MIPS statistics line. Guest-visible identity strings (ES40EM00000 disk serial, ES40RAMDISK model, the ES40-derived MAC seed) and the es40.cfg file name are intentionally unchanged. The sample es40.cfg is refreshed with the upstream 0.75.1 documentation (comment forms, unknown-value warning, mouse.speed/invert, DEC21143 queue, serial null_attach/disabled/raw_mode, disk identity values, no icache or mouse.enabled) merged with the axpbox-specific docs (rom.decompressed, skip_memtest_hack, timezone, S3 Trio64 with the 86Box BIOS). The old sample's stack of five active timezone lines (last-one-wins, leaving utc+2y in effect) is now a commented example block. README points at the sample file. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Step-by-step recipe for porting ES40-Emu commits into axpbox: range survey, file mapping (.h->.hpp, main_cfg, config locations), per-file strategy (manual vs wholesale-replace with axpbox-delta reapplication), the axpbox invariants (std:: threading, preserved bug fixes, settled icache/timer/mouse decisions, configurator allow-lists), rebranding rules (host-side AXPbox text vs guest-visible ES40 identifiers), the full verification gate sequence (build lanes, SRM diff, JIT_VERIFY, config-warning sweep, OpenVMS boot, ARC smoke), and the port history with the next range's base commit. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add debug hooks to investigate the reported Linux SDL mouse-input problem and to drive headless firmware/guest runs: - AXPBOX_MOUSE_DEBUG=1 traces every host SDL mouse-motion event (xrel/yrel/grab) and every grab transition to stderr, and logs if SDL_SetWindowRelativeMouseMode fails. This localizes a dead mouse to host-side (no motion events delivered) vs guest-side (events arrive but the guest pointer doesn't move). - AXPBOX_AUTOMOUSE=<sec> injects synthetic motion straight into the guest PS/2 path (a square pattern + periodic click), exercising KBC aux / IRQ12 / guest driver with no host input. - AXPBOX_KEYPIPE=<file> injects named keys appended to a file while the emulator runs, for interactive headless menu navigation; the AXPBOX_KEYSCRIPT key table is factored into a shared lookup and extended with the full letter/digit set and ins/home/end/f4/f5 etc. Also finish the rebrand: the GUI-install banner now says "AXPbox GUI". No change to the normal input path. Offscreen headless rendering (SDL_VIDEO_DRIVER=offscreen) with AXPBOX_DUMP_FB confirmed working. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Describe the AXPBOX_* debug environment variables (framebuffer dumps, scripted and interactive key injection, synthetic mouse motion, mouse event tracing, PC sampling), the key-name list, an example fully headless firmware run with the offscreen SDL driver, and the serial-port-waits-for-telnet startup gotcha. Authored-By: Claude Fable 5 <noreply@anthropic.com>
On WSLg (and some Wayland compositors) engaging the mouse grab bounces window focus; the FOCUS_LOST handler released the grab immediately after the user's click, so motion events kept arriving but were discarded and the guest mouse appeared dead (traced with AXPBOX_MOUSE_DEBUG: grab -> 1 followed instantly by grab -> 0, all motion at grab=0). Remember that the guest owned the mouse when focus loss forces an ungrab, and re-grab on FOCUS_GAINED. Bounded to 8 bounces per user action so a compositor that always rejects the grab cannot ping-pong; a deliberate Ctrl+F10 release or a fresh click resets the budget. Note: WSLg's Wayland backend additionally delivers no relative motion at all while relative mouse mode is active (SDL reports success, zero motion events arrive). On such hosts run the emulator under XWayland: SDL_VIDEO_DRIVER=x11 DISPLAY=:0 SDL_RENDER_DRIVER=software. Authored-By: Claude Fable 5 <noreply@anthropic.com>
AXPBOX_MOUSE_DEBUG=1 now also traces every command the guest sends to the PS/2 aux device (Keyboard.cpp ctrl_to_mouse) - shows whether a guest driver ever detects and enables the mouse (f4 = enable stream). README gains the "Mouse on WSLg / Wayland" section: the default Wayland backend on WSLg delivers no relative motion while grabbed; run via SDL_VIDEO_DRIVER=x11 DISPLAY=:0 SDL_RENDER_DRIVER=software (verified working). The verify-vga-sdl skill is updated with the corrected WSLg facts (x11 backend IS available), the offscreen-driver headless recipe, the serial-port startup gotcha, and the mouse diagnosis procedure; port-from-es40 lists the new hooks among the axpbox-only code to preserve across future ports. Known-open: AlphaBIOS moves the pointer with both axes inverted (mouse.invert_x/y compensate); Windows 2000 shows no pointer movement once booted - under investigation with the new aux trace. Authored-By: Claude Fable 5 <noreply@anthropic.com>
…E feature docs The sample es40.cfg now starts without any telnet interaction: both UARTs are null_attach (console on the VGA window), with the port/action telnet variants kept as commented alternatives and a note that a configured port waits for a client at startup. The DEC 21143 NIC section is active with adapter = "eth0" and documents the one-time Linux capture-permission grant (setcap cap_net_raw, cap_net_admin+eip) plus the Windows npcap adapter syntax; leaving adapter unset triggers an interactive chooser, so unattended runs must set it. The ES1370 sound card is enabled (NT4 guest drivers; other guests ignore it). README gains a "Serial consoles, networking, and sound" section covering these plus the mouse tuning options, window scaling hotkeys, and BIN/CUE cdrom support. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Repository guidance for Claude Code: the three build lanes and JIT debug flags, the SRM regression test and its pitfalls, pointers to the deeper verification skills, a big-picture map of the architecture (CSystem device graph, CPU interpreter + VMS PAL fast path, wall-clock timing, JIT block cache, configurator allow-lists, threading idioms), and the settled project rules (std:: over the base/ wrappers, .hpp headers, icache/mouse/timer decisions, rebrand boundaries, own versioning). Authored-By: Claude Fable 5 <noreply@anthropic.com>
gui_win32.cpp is pure GDI code and DiskFile.cpp includes commdlg.h, both of which fail to compile when NOGDI strips wingdi.h out of windows.h. Drop the unused Poco-style scoped lock macros from RWLock.hpp; AliM1543C_ide.hpp defines the std::shared_mutex versions. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adopt the CNetworkBackend abstraction (pcap + Linux TAP) from upstream while preserving the local networking features on top of it: - trace_packets tracing re-added in the DEC21143 RX/TX paths - NetworkFilter struct extends set_filter() with the full CSR6 semantics (PR/RA/PM/IF + broadcast + own-MAC fallback); the BPF builder moved from DEC21143::SetupFilter into CNetworkPcap - Windows run-time wpcap.dll loading (from System32\Npcap) moved into NetworkPcap and shared with the es40-cfg wizard, replacing the two duplicated copies; the exe no longer needs Npcap in WinPcap-compat mode - robust fgets-based adapter picker kept, now in CNetworkPcap::init - adapter descriptions kept in the configure wizard device list - new TAP config keys (type, host_ip, bridge, uplink, tap_create) added to the dec21143 config allow-list Authored-By: Claude Fable 5 <noreply@anthropic.com>
- third_party/SDL submodule pinned at release-3.4.12 (shallow); CMake falls back to a static build of it when no system SDL3 is found, so Windows needs no SDL install and the exe is self-contained - CI: check out submodules on the Linux and Windows jobs; Linux jobs drop the obsolete SDL1 packages in favor of X11/Wayland headers for the bundled SDL3 - README: Building-from-source section (Linux + Windows incl. npcap SDK and VS 2022 steps) and pcap/TAP backend documentation - banner credits: gdwnldsKSC listed by GitHub handle Authored-By: Claude Fable 5 <noreply@anthropic.com>
Guests that legitimately hit integer overflow traps (e.g. Tru64) no longer spam the console; define DEBUG_ARITH_TRAP in config_debug.hpp to get the old logging. DEBUG_LOCKS comment style aligned with upstream. Authored-By: Claude Fable 5 <noreply@anthropic.com>
EV68 HRM 4.6: an external (DMA) writer must invalidate a CPU's locked cache line. Adds a reader/writer gate (cpu_llsc_dma_gate) that makes the DMA invalidating probe + RAM write atomic against the LDx_L load/publish and STx_C test/store windows: - System: CLLSCDRAMGuard (reader) / CPCIDMAWriteGuard (writer) RAII classes, gate enter/leave, cpu_clear_external_locks() - PCIDevice::do_pci_write takes the writer guard and probes every reservation line the DMA touches - interpreter LDx_L/STx_C (cpu_defs.hpp READ_VIRT_LOCK[_F], WRITE_VIRT_COND) and vmspal HW_LD/HW_ST locked/conditional cases (cpu_pal.hpp) take the reader guard - JIT jit_read_locked/jit_stc take the reader guard Also ports the DEBUG_INSTALL_PAL_TRAP hook in GO_PAL from upstream c2bd1fb (compiles out unless DEBUG_INSTALL_VMFAULT is defined). Authored-By: Claude Fable 5 <noreply@anthropic.com>
ufdiv64 cannot represent asig*2^64 / (zsig*2^32): with zsig >= 2^31 the divisor has bit 63 set and the shifted dividend wraps, producing a garbage estimate that the correction loop walks back ~2^60 ULPs. Use SoftFloat's estimateDiv128To64 (udiv128to64) instead. Fixes SQRT results that broke network dialogs on Tru64 (upstream #136). Authored-By: Claude Fable 5 <noreply@anthropic.com>
Upstream reverted the CSR lock (originally 4bb01619 / our std::recursive_mutex myRegLock) in v0.75.4 after it caused regressions; follow suit: the NIC thread batch, ReadMem_Bar, WriteMem_Bar and Save/RestoreState run lock-free again, matching the pre-0.75.1 model. The axpbox instant-IRQ recompute (update_irq) is unaffected. Authored-By: Claude Fable 5 <noreply@anthropic.com>
Authored-By: Claude Fable 5 <noreply@anthropic.com>
The blessed log predated the AXPbox startup banner, the blank-flash auto-create (%FLS-I-CREATE), the %SYS-W-NOSERIAL warning and the new DiskFile failure text, so the comparison always failed. Also: - drop the removed 'icache' option from the test config (it now triggers %SYS-W-UNKNOWNCFG) - normalize the banner version line in test.sh (Version X.Y.Z) so version bumps don't break the comparison Authored-By: Claude Fable 5 <noreply@anthropic.com>
A config file with e.g. two serial0 sections silently created two devices; the later one wins the address range, so a stray telnet serial0 section makes startup block waiting for a connection even though an earlier serial0 says null_attach. Warn at parse time. Authored-By: Claude Fable 5 <noreply@anthropic.com>
RaymiiOrg
force-pushed
the
es40-ports-from-gdwnldsKSC
branch
from
July 11, 2026 17:05
9d09861 to
89000e3
Compare
The runner label was literally 'macos-15"' so no runner ever picked
the job up ('Requested labels: macos-15"' → waits forever). While in
there:
- macos-15 runners are arm64: drop the x86 -std=c++11/-stdlib=libc++
and /usr/local pcap paths; point PCAP at $(brew --prefix libpcap);
rename the job/artifact osx-x86 → osx/macos-arm64
- check out submodules so the bundled SDL3 builds in statically
(the old 'sdl doesnt work' comment predates SDL3)
- replace the disabled ::set-output syntax and the broken
'name=date::...' GITHUB_OUTPUT variant (which left BUILD_DATE empty
in artifact names) with the proper 'date=...' form in all jobs
Workflow validated with actionlint (no errors).
Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This is a port of most changes from https://github.com/ES40-Emu/es40 via Claude, as we have earlier discussed by email. I haven't been able to test the Windows build. The Linux build (via WSL) is able to boot Windows 2000. This should fix #95