port-from-es40

---

name: port-from-es40 description: Port upstream ES40-Emu commits into axpbox. Use whenever the user asks to bring in, sync, or port changes from the es40 repo (github ES40-Emu/es40, local clone to /tmp/es40). Covers delta computation, wholesale-replace vs manual patching, the axpbox invariants (std:: threading, .hpp, clang-format, rebrand), and the verification gates.

Porting upstream ES40-Emu changes into axpbox

axpbox (this repo) tracks the upstream ES40-Emu emulator. Upstream lives in a sibling repo https://github.com/ES40-Emu/es40 (git clone that locally to /tmp/es40/). Ported ranges so far: everything up to upstream 7d94c9e (2026-07-05), then 7d94c9e..7140555 = v0.75.1 (2026-07-09). When asked to port "new upstream changes", the base of the new range is the tip of the last ported range — check this file's history section (bottom) and git -C /tmp/es40 log --oneline to find it.

NEVER git apply / git cherry-pick upstream patches. They never apply: axpbox is clang-formatted, renames headers .h → .hpp, rewrote threading to std::, merged the two upstream binaries into one, and moved files. Follow the recipe below instead.

Step 0 — survey the range

E=/tmp/es40           # upstream clone
BASE=<last-ported-upstream-commit>     # e.g. 7140555
git -C $E fetch --all                  # only if the user says they updated it
git -C $E log --oneline $BASE..HEAD    # list commits to port
git -C $E diff --stat $BASE..HEAD      # list files touched

Group the commits into themes (JIT, timing, disk, net, config, GUI, build, docs). Plan roughly one axpbox commit per theme. Every commit must build in all lanes and pass the SRM test.

Files to ALWAYS skip (axpbox does not have them): src/Visual Studio/*, configure.ac, config_vms.h, config_win32.h, .github/ISSUE_TEMPLATE/*. Upstream CMakeLists.txt changes almost never apply — axpbox has its own; read the upstream hunk and decide if the idea (a new source file, a new dependency) is needed. axpbox versioning is its own (project(AXPBox VERSION x.y.z) in CMakeLists.txt) — NEVER adopt upstream's version number.

Step 1 — file mapping (upstream path → axpbox path)

• src/X.cpp → src/X.cpp (same name).
• src/X.h → src/X.hpp (ALL headers).
• src/es40.cfg → repo-root es40.cfg (plus per-test copies in test/rom/, test/vms/, test/nt/, test/arc/).
• src/es40-cfg.cpp → src/es40-cfg.cpp, but its main() is named main_cfg() (single axpbox binary; src/Main.cpp dispatches axpbox run → main_sim, axpbox configure → main_cfg).
• New upstream headers: create as .hpp, rename the include guard (__X_H__ → __X_HPP__), give it the axpbox license header (see Step 4), and add nothing to CMake — file(GLOB ...) picks up sources automatically (re-run the cmake configure step after adding a new .cpp).

Step 2 — choose a strategy per file

Compute two diffs for each file F in the range:

S=<scratch dir>
# (a) what upstream changed:
git -C $E diff $BASE..HEAD -- src/F        # use: git --no-pager diff --no-ext-diff
# (b) what axpbox changed relative to the last ported upstream state:
git -C $E show $BASE:src/F > $S/F.base
diff -u $S/F.base <(git show <preformat>:src/F_axpbox_path)

For (b) you need the axpbox snapshot from BEFORE the repo-wide clang-format, or the diff drowns in formatting noise. For the original port that snapshot is 3e35a80~1. If the file was clang-formatted since, fall back to diff -w -B against the current file and read carefully. ALSO check for axpbox commits made after that snapshot: git log --oneline <snapshot>..HEAD -- src/F — their changes are part of the axpbox delta too.

Pick the strategy:

• Manual edit (M) — when (a) is small (< ~100 changed lines). Apply upstream's hunks by hand with the Edit tool onto the current axpbox file, translating style (see Step 3).
• Wholesale replace (W) — when (a) is large (hundreds+ of lines, many commits of churn) AND (b) is small/enumerable. Recipe:
  1. git -C $E show HEAD:src/F > src/F (into axpbox).
  2. sed-rename every #include "X.h" to "X.hpp". Verify none remain: grep -n '\.h"' src/F | grep -v '\.hpp"'.
  3. Re-apply the axpbox delta. Fastest reliable way: save diff (b) to a file and run patch --no-backup-if-mismatch -F3 src/F < delta.diff, then open every .rej and hand-apply what failed. Read the result around every merge point.
  4. Deliberately DROP delta hunks that upstream has now superseded (e.g. a config option upstream removed) — list these in the commit message.
  5. clang-format-14 -i --style=file src/F (clang-format is installed as clang-format-14, no bare clang-format).

Step 3 — axpbox invariants (apply to every ported line)

1. No Poco-style wrappers in new code. Upstream uses CThread, CMutex, CSemaphore, SCOPED_M_LOCK, CRunnable (from src/base/). New/ported code must use std:: equivalents:
   • CThread::sleep(ms) → std::this_thread::sleep_for(std::chrono::milliseconds(ms))
   • CThread* myThread → std::unique_ptr<std::thread> myThread; started as myThread = std::make_unique<std::thread>([this]() { this->run(); }); joined as myThread->join(); myThread = nullptr;
   • thread-death flag: std::atomic_bool myThreadDead{false}; set in the run() catch block, checked in check_state().
   • CMutex + SCOPED_M_LOCK(m) → std::recursive_mutex m; + std::lock_guard<std::recursive_mutex> lock(m); (upstream CMutex IS recursive — keep recursive semantics unless you can prove no re-entry). #include <mutex> in the header.
   • Pre-existing CMutex/CSemaphore uses already in axpbox (e.g. Sym53C810, the CPU semaphore) may stay — do not churn them.
2. Preserve the axpbox-only bug fixes when replacing CPU/JIT code:
   • execute() advances the PC with next_pc(), never bare state.pc += 4 (keeps pc_phys/rem_ins_in_page in sync).
   • break_seq_icache() in AlphaCPU.hpp also clears state.rem_ins_in_page = 0;.
   • vmspal_int_initiate_exception has the vmspal_exc_depth >= 3 recursion guard (guard in AlphaCPU.cpp, ++/-- in AlphaCPU_vmspal.cpp, member in AlphaCPU.hpp).
   • skip_memtest_hack config option (AlphaCPU init + gated 0x8b000 page check in execute()), replaces upstream's #ifdef SKIP_SRM_MEMTEST.
   • fread return-value checks in RestoreState.
   • Debug env hooks (documented in README "Headless testing"): AXPBOX_PC_SAMPLE (AlphaCPU::check_state), AXPBOX_DUMP_FB, AXPBOX_AUTOKEY_ENTER, AXPBOX_KEYSCRIPT, AXPBOX_KEYPIPE, AXPBOX_AUTOMOUSE, AXPBOX_MOUSE_DEBUG (gui/sdl.cpp; the aux-cmd trace part of MOUSE_DEBUG is in Keyboard.cpp ctrl_to_mouse).
   • SDL grab handling: re-grab on FOCUS_GAINED after a compositor focus bounce (WSLg) — do not revert to plain "FOCUS_LOST => ungrab" when porting upstream sdl.cpp changes.
   • DiskFile: default-image-filename fallback (defaultFilename member) instead of upstream's FAILURE when no file is configured.
   • Disk: SCSICMD_START_STOP_UNIT case answered with do_scsi_error(SCSI_OK).
   • Configurator: the cirrus class row exists and fails with a helpful "use s3 instead" message.
   • es40-cfg: the skip_memtest_hack question.
3. Settled decisions — do not re-litigate:
   • icache is hardcoded ON (icache_enabled = true in init/reset, no config read, no configurator question) — upstream ed9bf49, user-confirmed 2026-07-09.
   • The interval timer fires immediately at thread start (next_timer_fire = start_time) — the old 1-second grace is gone; the RTC boot tick (toy_stored_data[0x0a] = 0x26) and the vmspal guard cover the early-boot transient.
   • mouse is always present/captured (no mouse.enabled).
4. Configurator allow-lists (kv_*[] arrays in src/Configurator.cpp): any NEW config key a ported device reads must be added to that device's list, and axpbox-only keys must never be lost: skip_memtest_hack (kv_ev68cb), timezone (kv_ali), rom.decompressed (kv_tsunami), address (kv_serial). Audit after porting:

   grep -rn 'myCfg->get_\(text\|num\|bool\)_value("' src/*.cpp src/gui/*.cpp \
     | grep -o 'value("[^"]*"' | sort -u
   

   then start the emulator once with every shipped config and require ZERO %SYS-W-UNKNOWNCFG (see Step 5.4).

Step 4 — rebranding rules

All HOST-side user-visible text says AXPbox, never ES40: window titles (sdl/x11/win32), the startup banner (src/banner.hpp — print_axpbox_banner, shows axpbox's own VERSION, credits authors by era: Camiel Vanderhoeven 2007-2010, Tim Stark/fsword7 2018, Tomas Glozar 2020-2023, Remy van Elst 2020-2026, gdwnldsKSC 2023-2026), configurator wizard text, serial telnet greeting, debugger greeting, stat lines.

Do NOT rename GUEST-visible or protocol identifiers: disk serial default "ES40EM00000", ramdisk model "ES40RAMDISK", the MAC seed "ES40", the es40.cfg file name, "[]ES40.CFG", hardware/SRM references ("AlphaServer ES40" is the emulated machine — keep), state file magics, include guards.

License header for files whose axpbox copy already has it, and for new files:

/* AXPbox Alpha Emulator
 * Copyright (C) 2020 Tomáš Glozar
 * Website: https://github.com/lenticularis39/axpbox
 *
 * Forked from: ES40 emulator
 * Copyright (C) 2007-2008 by the ES40 Emulator Project
 * Copyright (C) 2007 by Camiel Vanderhoeven
 * ... (GPL v2 boilerplate as in existing files)
 */

Some ported files (e.g. AlphaCPU.cpp) still carry the upstream ES40 header — match whatever the file's current axpbox copy has. Strip upstream's $Id$ / X-1.xx CVS changelog blocks from any file you wholesale-replace (keep only the \file description line).

Text in banner.hpp must stay ASCII (Tomas, not Tomáš) — the column padding is strlen-based and multi-byte UTF-8 breaks alignment.

Step 5 — verification gates (run in this order)

1. Build all lanes (see the build-lanes skill): build (default), build-jit (-DES40_DISABLE_ASMJIT=OFF, asmjit pinned 0bd5787b in third_party/asmjit), build-nosdl (-DDISABLE_SDL=yes). All three after every commit.
2. SRM regression (see srm-boot-test skill): cd test/rom && bash test.sh — expect literal diff clean. Run against build-jit too (copy test.sh and point it at ../../build-jit/axpbox). Pitfalls: test.sh DELETES the tracked cl67srmrom.exe, flash.rom, decompressed.rom at the end — restore with git checkout -- test/rom/ before committing. axp_correct.log contains NUL bytes (grep needs -a) and embeds the serial telnet greeting — if you change that greeting, patch the expected log binary-safely (python bytes replace, not sed).
3. JIT differential check after ANY jit/ or AlphaCPU change:

   cmake -S . -B /tmp/build-verify -DCMAKE_BUILD_TYPE=Release \
     -DES40_DISABLE_ASMJIT=OFF -DCMAKE_CXX_FLAGS="-DJIT_VERIFY"
   cmake --build /tmp/build-verify -j$(nproc)
   # run the SRM test against it; the log must show
   # "[JIT][VERIFY] ... 0 mismatches" lines and end at P00>>>
   

   Also do one throwaway compile with -DCMAKE_CXX_FLAGS="-DJIT_TRACES -DJIT_STATS" so the dormant trace tier doesn't rot.
4. Config-warning sweep: for each of test/rom, test/vms, test/nt, test/arc and a scratch dir holding the root es40.cfg:

   (cd <dir> && timeout 8 stdbuf -oL <repo>/build/axpbox run > w.log 2>&1; \
    grep -aE "SYS-W" w.log | sort -u)
   

   Expected: nothing except %SYS-W-NOSERIAL on configs that omit serial1. NOTE: output MUST go to a file (stdbuf + redirect) — piping timeout'd output loses everything to stdio buffering.
5. OpenVMS boot after CPU/disk/timing changes (see boot-openvms skill) — run on both build and build-jit lanes; expect RESULT: SUCCESS: login prompt reached and OPCOM timestamps consistent with the date the script answers.
6. ARC smoke after RTC/superio/VGA changes (see test-arc skill).
7. Configurator smoke: printf 'no\n' | build/axpbox configure shows the AXPbox banner and exits cleanly.
8. Kill strays ONLY with pkill -x axpbox (never pkill -f — it matches the invoking shell's own command line).

Step 6 — commit conventions

One thematic commit per upstream area, message = what was ported + which upstream commits + which axpbox deltas were preserved/dropped. Do not sweep in unrelated untracked files (test/nt/ may hold user disk images and logs; there may be a user stash WIP on s3-port — leave both alone). Trailer: Co-Authored-By: Claude <the current model's attribution line>.

Port history

• up to upstream 7d94c9e — 12 commits on baseline e81dce5 (2026-07-05).
• 7d94c9e..7140555 (v0.74→0.75.1) — 7 commits 29f83b4..377c362 (2026-07-09): JIT 0.75.1 + timing bundle, BIN/CUE, DEC21143 lock, Sym53C810 disconnect, configurator warnings, SDL mouse options, AXPbox rebrand + banner. Next port starts at upstream 7140555.
• 7140555..9f7554d (v0.75.1→0.75.4) — 4 commits (2026-07-11): DEBUG_ARITH_TRAP printf gating, LL/SC reservation invalidation on PCI DMA writes (cpu_llsc_dma_gate reader/writer gate + guards in interpreter/vmspal/JIT paths + DEBUG_INSTALL_PAL_TRAP hook), fsqrt64 udiv128to64 fix (Tru64), and upstream's REVERT of the DEC21143 register mutex (myRegLock removed again — follow upstream, do not re-add). Skipped: version bumps, configure.ac/m4, build.yml macOS pcap logic. Next port starts at upstream 9f7554d. NOTE: the DEC21143/NetworkBackend area diverges from ES40-Emu since the lenticularis TAP/TUN merge (NetworkPcap/NetworkTap/NetworkFilter live only in axpbox+lenticularis) — port upstream DEC21143 pcap changes into src/NetworkPcap.cpp instead.
• Pitfall: piping bash test.sh | tail can hang even after the test finishes (a lingering child keeps the pipe open) — redirect test.sh output to a file instead, then read the file.
• Post-port axpbox-only additions (2026-07-10, keep across future ports): input-injection/diagnostic hooks (AXPBOX_KEYPIPE, AXPBOX_AUTOMOUSE, AXPBOX_MOUSE_DEBUG + aux-cmd trace in Keyboard.cpp), the focus-bounce re-grab in sdl.cpp, README "Headless testing" + "Mouse on WSLg" sections.