Skip to content

win32_open_items.md

Latest commit

 

History

History
560 lines (482 loc) · 40.8 KB

win32_open_items.md

File metadata and controls

560 lines (482 loc) · 40.8 KB
Raw

Win32 / multiplatform — open items & temporary workarounds

Status: living tracker
Created: 2026-07-17
Rule: Every temporary workaround that future work must remove belongs here as OPEN.
When the proper fix lands, mark the item CLOSED, move it into §Closed (sorted), and keep the full history.
Do not list permanent non-goals as OPEN workarounds—list them under §Non-goals.

Filename

Name Use
win32_open_items.md (this file) Preferred: leftovers and temporary workarounds
win32_leftovers.md OK alias if you prefer
win32_bug_tracker.md Prefer for defect IDs only; this file is broader

Related docs

Doc Role
win64_art_port.md Product phases / feasibility
win32_filesystem.md Option H path model
win32_tls_jit_entrypoints.md TLS / managed ABI / quick / JIT design (draft)
win32_libcore_os_natives.md Os/Linux natives: Implemented / Needed / ENOSYS
tools/verify/win64_phase*/RESULT.md Gate evidence

How to maintain

Add when you land a temporary path (compile stub, force-interpreter, fake DLL name, wine-only gate, etc.):

### W-XXX — short title
- **State:** OPEN
- **Kind:** workaround | leftover | debt | host-gap
- **Area:** art | libcore-stub | icu | packaging | linux-host | docs | …
- **Symptom / why:** …
- **Current behavior:** …
- **Proper fix:** …
- **Code anchors:** `path:line` or symbol names
- **Blocked on / design doc:** …
- **Opened:** YYYY-MM-DD

Close when fixed:

- **State:** CLOSED (YYYY-MM-DD) — one-line how

Then move the full item into §Closed (keep history; sort by ID prefix then number). Do not leave CLOSED bodies under Temporary workarounds / leftovers / host / design.

IDs: W- workaround, L- leftover/product gap, H- host/validation gap, D- docs/process. Numbers are stable; do not reuse.

Snapshot (2026-07-17)

Bucket Summary
Phases 0–3 Gate-complete (P3 G12 real Win10 + wine)
Phase 4 Wine complete; host re-run still recommended
PE libcore/ICU/openjdk Product-default real PE (icu/javacore/openjdk); NIO.2 non-goal; NetProbe OK
Quick/JIT/TLS Partial: rSELF=r15; nterp N-1 + PE 0x328; ClassLoader empty-path fixed (defer nterp until finished_starting); float methods excluded; residual empty Hello stdout under nterp; switch/-Xint green
Linux multiplatform Native dalvikvm -showversion OK; imageless Hello e2e not re-gated here

Temporary workarounds (must be removed later)

W-001 — Force interpreter invoke (quick entrypoints effectively disabled)

  • State: CLOSED (product default uses quick invoke)
  • Kind: workaround (removed as product default)
  • Area: art / invoke
  • Symptom / why: Win64 used to force interpreter invoke until quick path was smoke-validated.
  • Current behavior: On _WIN32, invokable non-proxy methods use art_quick_invoke_* (MS entry → SysV body, rSELF=r15) by default, matching Linux. Opt-out with ART_WIN64_QUICK_INVOKE=0 forces EnterInterpreterFromInvoke. Debugger/-Xint still force interpreter via normal ART paths.
  • Proper fix: Done for product default. Residual: broader host re-run; optional delete of the env force path later.
  • Code anchors: vendor/art/runtime/art_method.cc; quick_entrypoints_x86_64.S Win prologues; win32_tls_jit_entrypoints.md §12b / §17.8
  • Blocked on: n/a (default ON as of 2026-07-19)
  • Opened: 2026-07-16 (Phase 2)
  • Updated: 2026-07-19 — product default ON (Linux-like); opt-out ART_WIN64_QUICK_INVOKE=0

W-002 — No managed GS / Thread base on Windows (InitCpu no-op for GS)

  • State: OPEN (partial — rSELF path is product default; residual JNI attach / non-invoke entries)
  • Kind: workaround / design debt
  • Area: art / TLS
  • Symptom / why: Linux x86_64 uses ARCH_SET_GS so quick/nterp use %gs:OFFSET. Windows GS is TEB.
  • Current behavior: InitCpu does not touch GS (correct). Asm uses THREAD_* macros: r15 base on _WIN32, GS on Linux. rSELF published in art_quick_invoke_* (default ON). Nterp N-1 (rREFS=rbp) product default ON (§17.8).
  • Proper fix: Keep rSELF=r15; audit remaining managed entries (JNI return, attach, trampolines) publish rSELF; then close when full matrix is green without GS.
  • Code anchors: thread_x86_64.cc; asm_support_x86_64.S THREAD_*; nterp.cc (default ON; opt-out ART_WIN64_NTERP=0); design §6 / §12b / §15 / §16 / §17 / §17.8
  • Opened: 2026-07-16
  • Updated: 2026-07-19 — product default rSELF + nterp ON; residual attach/publish audit
  • Design: win32_tls_jit_entrypoints.md §15 N-1 LOCKED, §17 register-map lock; FS-self §16 reject; §17.8 defaults ON

W-003 — Quick entrypoint SETUP frames int3 on Windows

  • State: OPEN (partial — SETUP_SAVE_REFS_ONLY / ALL_CALLEE_SAVES un-int3'd on Win)
  • Kind: workaround (hard fail if other paths still stubbed)
  • Area: art / quick asm
  • Symptom / why: Apple still traps; Win now builds frames via shared SETUP with THREAD_STORE for top quick frame.
  • Current behavior: SETUP_SAVE_REFS_ONLY_FRAME / SETUP_SAVE_ALL_CALLEE_SAVES_FRAME no longer int3 on _WIN32. C++ helpers called as sysv_abi (ART_QUICK_ENTRYPOINT_ABI).
  • Proper fix: Smoke the invoke→interpreter bridge; close when no Win SETUP path traps and product default uses quick invoke.
  • Code anchors: asm_support_x86_64.S; ART_QUICK_ENTRYPOINT_ABI in macros + entrypoints
  • Depends on: W-001 validation
  • Opened: 2026-07-16
  • Updated: 2026-07-18 — Win SETUP enabled

W-004 — LOAD_RUNTIME_INSTANCE PE helper call (vs GOT)

  • State: OPEN (acceptable interim; still temporary vs ideal PE codegen)
  • Kind: workaround
  • Area: art / asm
  • Current behavior: Win path calls art_Runtime_instance_ptr with shadow space instead of @GOTPCREL.
  • Proper fix: Keep helper or RIP-relative import of Runtime::instance_ consistently for JIT + hand asm (document single sequence).
  • Code anchors: LOAD_RUNTIME_INSTANCE in asm_support_x86_64.S
  • Opened: 2026-07-16

W-008 — Product smoke always passes -Xint / imageless / -Xno-sig-chain

  • State: OPEN
  • Kind: workaround (policy flags)
  • Area: packaging / product CLI
  • Current behavior: Host package scripts and wine runners force interpreter + no boot image + no sigchain.
  • Proper fix: After W-001–W-003, drop forced -Xint for default product scripts (keep as opt-in debug). Imageless may remain until boot image (separate track).
  • Code anchors: tools/win64/host_package/package_win64_phase3.sh, tools/verify/win64_phase*/run_*.sh
  • Opened: 2026-07-16

W-010 — Sigchain is a Windows stub; VEH owns faults

  • State: OPEN (may become permanent architecture — reclassify if so)
  • Kind: workaround → candidate permanent design
  • Area: art / exceptions
  • Current behavior: sigchain_windows.cc stub; crash/null/SO paths use VEH + minidump in runtime_windows.cc.
  • Proper fix: Document as permanent WinNT model or deepen VEH/sigchain integration tests; remove any code that still assumes Linux sigchain interposition.
  • Code anchors: vendor/art/runtime/multiplatform/windows/sigchain_windows.cc, runtime_windows.cc
  • Opened: 2026-07-16

W-011 — Expanded InterpreterJni shorty coverage for PE (bypass missing quick/JNI stubs)

  • State: OPEN
  • Kind: workaround
  • Area: art / jni
  • Current behavior: Extra shorty paths in interpreter JNI so Phase-2/3 natives work without full quick/JNI trampoline matrix.
  • Proper fix: After real PE libcore + entrypoints, audit whether generic path suffices; trim PE-only shorty explosion if redundant.
  • Code anchors: interpreter JNI / generic paths (Phase-2 RESULT notes: FJ/IJ/VLJ/…)
  • Opened: 2026-07-16

W-012 — ResolveJniEntryPoint without art_jni_dlsym_lookup_stub (%gs)

  • State: OPEN
  • Kind: workaround
  • Area: art / jni
  • Current behavior: Windows avoids GS-based JNI lookup stub.
  • Proper fix: Fold into final managed-self / JNI trampoline design (W-002).
  • Code anchors: Phase-2 RESULT; JNI resolution paths under _WIN32
  • Opened: 2026-07-16

W-013 — dlmalloc WIN32 / low-4GB / MORECORE choices for imageless ART

  • State: OPEN (may stay as permanent Win allocator policy)
  • Kind: workaround / platform policy
  • Area: art / heap
  • Current behavior: WIN32 mmap/MORECORE path + low-4g constraints for compressed refs / imageless bring-up (see Phase-2 root-cause notes).
  • Proper fix: Re-validate against real Win10 under load; document as permanent if correct; remove any Linux-only assumptions left in comments.
  • Code anchors: dlmalloc Win path / Phase-2 docs in win64_art_port.md §9c
  • Opened: 2026-07-16

W-014 — Stack bounds via VirtualQuery + clamp (Wine-safe estimates)

  • State: OPEN
  • Kind: workaround
  • Area: art / threads
  • Current behavior: Win GetThreadStack uses VirtualQuery + clamps; ProtectStack skipped/no-op on _WIN32 in places to avoid mprotect of wrong regions.
  • Proper fix: TEB-accurate stack bounds + optional guard pages when trusted; re-enable protection carefully.
  • Code anchors: vendor/art/runtime/thread.cc Win branch; ProtectStack #ifdef _WIN32
  • Opened: 2026-07-16

W-017 — openjdk hybrid excludes NIO.2 / async / UNIXProcess; epoll via select

  • State: OPEN
  • Kind: workaround / incomplete port
  • Area: openjdk / nio
  • Current behavior: Phase B2 builds AOSP NIO channel natives with Winsock CRT-fd shims; epoll_* emulated with select; NIO.2 UnixNativeDispatcher/WatchService/async EPollPort not registered.
  • Proper fix: Keep NIO.2 non-goal; deepen channel/options matrix; optional IOCP epoll later if needed.
  • Code anchors: tools/verify/win64_libcore_icu/CMakeLists.txt (_OJ_SRCS filters); compat/src/win64_socket_posix.c
  • Opened: 2026-07-17

W-024 — Restore original @CriticalNative / @FastNative surfaces after JIT/TLS/entrypoints

  • State: OPEN
  • Kind: workaround / debt (must revert multipath demotions)
  • Area: art / libcore / JNI ABI
  • Symptom / why: Official AOSP libcore marks many natives @CriticalNative or @FastNative (Math/StrictMath were @FastNative → @CriticalNative in AOSP; see libcore d021f1d8475c). Win64 multipath cannot yet honor those ABIs because:
    1. W-001 forces interpreter invoke (no quick/managed entrypoints / TLS).
    2. Interpreter JNI historically lacked full CriticalNative shorty coverage (partially papered by W-019 for Math DD/DDD/…).
    3. Product demotions for bring-up: Math.ceil / Math.floor are pure Java (ART-WinNT comments in Math.java; natives unregistered in Math.c) even though AOSP exposes them as @CriticalNative natives.
    4. Win Math.c uses a PE-specific registration table (gMethodsWin CriticalNative-shaped pointers) vs Linux FAST_NATIVE_METHOD macro table — temporary dual path until PE trampolines match AOSP.
  • Current behavior:
    • Annotations remain on most Math/StrictMath/etc. methods, but ceil/floor are non-native pure Java.
    • Win64 relies on interpreter CriticalNative shorty dispatch + correct CriticalNative binding (no JNIEnv*) rather than real quick CriticalNative stubs.
    • FastNative methods stay Runnable on the interpreter bridge (W-019 supporting fix) instead of true FastNative entrypoints.
  • Proper fix (when JIT / TLS / quick entrypoints land — W-001–W-003 + win32_tls_jit_entrypoints.md):
    1. Restore every multipath demotion of methods that were originally @CriticalNative / @FastNative in AOSP (starting with Math.ceil / Math.floor → native + @CriticalNative again).
    2. Re-register matching natives in Math.c (and any other demoted tables) on both PE and ELF with AOSP-correct CriticalNative/FastNative ABIs.
    3. Prefer real ART CriticalNative/FastNative trampolines over InterpreterJni shorty expansion; then trim PE-only shorty explosion (W-011) and dual gMethodsWin paths where redundant.
    4. Audit libcore for other ART-WinNT pure-Java / ABI demotions of Critical/Fast natives (not just Math) and revert once entrypoints are product-default.
  • Exit criteria:
    • No pure-Java stand-ins for AOSP @CriticalNative/@FastNative Math (ceil/floor native again).
    • Wine + Linux smokes for Math/HashMap/conscrypt without relying on pure-Java ceil or incomplete CriticalNative shorty lists.
    • Document that W-019 temporary ABI/shorty fixes are superseded (may remain as defensive fallback until deleted).
  • Code anchors:
    • vendor/libcore/ojluni/src/main/java/java/lang/Math.java (ART-WinNT pure-Java ceil/floor)
    • vendor/libcore/ojluni/src/main/native/Math.c (gMethodsWin / no ceil|floor register; Linux FAST_NATIVE_METHOD)
    • vendor/art/runtime/interpreter/interpreter.cc (InterpreterJniGeneric CriticalNative shorties)
    • vendor/art/runtime/art_method.cc (Win32 force interpreter invoke)
    • AOSP history: d021f1d8475c FastNative→CriticalNative Math; multipath f16cd44db5fe pure-Java ceil/floor; b9265e7b5da6 CriticalNative register fix; art 7ea144b073 / 4c17423714 interpreter Critical/FastNative bridge
  • Blocked on: W-001–W-003 implementation (JIT/TLS/quick entrypoints)
  • Related: W-019 (CLOSED temporary Math ABI fix), W-011 (InterpreterJni shorty expansion)
  • Opened: 2026-07-17

Product leftovers (not single-line workarounds)

No open product leftovers. Closed L- items live under §Closed.

Host / validation gaps

H-001 — Phase 4 re-run on real Windows host

  • State: OPEN
  • Kind: host-gap
  • Gap: Wine Phase 4 PASS (incl. multiplatform rebuild 2026-07-17). Real Win10 Phase-4 subset (gcstress, threadheavy, handleleak, crash native/abort) not re-proven with multiplatform PE.
  • Exit criteria: Host logs under tools/verify/win64_phase4/evidence/host/ (or successor) OVERALL PASS.
  • Opened: 2026-07-16

H-002 — Phase 3 G12 with multiplatform-built PE (not only pre-migration tree)

  • State: OPEN
  • Kind: host-gap
  • Gap: Authoritative G12 used earlier host package; multiplatform in-tree PE rebuild should re-package and smoke on Win10 when convenient.
  • Opened: 2026-07-17

H-003 — Wine is not product acceptance

  • State: OPEN (policy reminder, not a code fix)
  • Kind: host-gap / process
  • Note: Keep wine as agent01 oracle; product claims need real Windows for VEH/TEB/network.
  • Opened: 2026-07-16

Non-goals (do not track as OPEN workarounds)

Item Decision
Windows NIO.2 (sun.nio.fs) Non-goal for now (win32_filesystem.md)
WSL2 / Wine as product runtime Rejected
Win32 x86 product SKU Out of scope (x64 first)
Full Android framework / zygote / binder Out of scope
In-process dual JIT ISA (x64+Arm64EC) Rejected in TLS/JIT draft

If product reopens a non-goal, add an L- item and link the decision.

Closed

Summary (details below; do not delete history):

  • W-005 — Combined PE JNI stub DLL aliased as libjavacore/libopenjdk/libicu_jni (2026-07-17) — product packaging uses stage_native_modules.sh (real PE only); libcombined is legacy non-product
  • W-006 — Minimal NativeConverter / ICU version shims (not full ICU4C) (2026-07-17) — product uses real icu_jni NativeConverter + icuuc/icui18n + icudt; native_converter.c obsolete and removed from libcombined; charset stub no longer product path
  • W-007 — Classic sockets / poll via Winsock select (not full Os/NIO) (2026-07-17) — permanent WinNT design: classic Os sockets use Winsock + select()-based poll/timeouts (not CRT-fd WSAPoll)
  • W-009 — Phase-1 grade compat POSIX/pthread stubs (2026-07-17) — hot paths hardened; remaining ENOSYS is intentional Linux-only surface
  • W-015 — openjdkjvm memory exports minimal PE surface (2026-07-17) — product ships comprehensive standalone libopenjdkjvm.dll
  • W-016 — ICU needs external ICU_DATA / icudt72l.dat for wine smoke (2026-07-17) — product always stages run/icu/icudt72l.dat via tools/win64/stage_run_assets.sh (same class as boot.jar); libicu_jni defaults ICU_DATA to run/icu when unset
  • W-018 — NetProbe StructLinger NPE (getsockopt SO_LINGER incomplete in javacore Win bridge) (2026-07-17) — implemented getsockoptLinger/setsockoptLinger in win_net_natives; NetProbe wine PASS
  • W-019 — Math @CriticalNative / FastNative double ABI on Win64 (2026-07-17) — Math.ceil/floor/sqrt + HashSet wine PASS after interpreter CriticalNative DD/DDD
  • W-020 — FileChannelImpl.map0 pointer truncation on Win64 (LLP64) (2026-07-17) — ptr_to_jlong(mapAddress) instead of (jlong)(unsigned long)
  • W-021 — Default KeyStore type Android-compatible (AndroidCAStore) (2026-07-17)
  • W-022 — Product default CA bundle (AndroidCAStore cacerts) (2026-07-17)
  • W-023 — OkHttp Http(s)Handler on bootclasspath + ASCII IDN/Normalizer multipath (2026-07-17)
  • L-001 — Real PE libcore / openjdk / ICU module build (2026-07-17)
  • L-002 — boringssl / conscrypt / SSL PE (2026-07-17) — product TLS stack green under wine (providers + SSLContext.init + HTTPS GET)
  • L-003 — Process/exec, rich locale, zip edge, UDP/IPv6 matrix (2026-07-17)
  • L-004 — Shrink or replace multi-name DLL staging (2026-07-17) — product ships one PE soname each: libicu_jni/libjavacore/libopenjdk/libopenjdkjvm/libcrypto/libssl/libjavacrypto (+ icuuc/icui18n); short-name twins removed from packaging
  • L-005 — Linux multiplatform imageless Hello / boot.jar CI gate (2026-07-17)
  • L-006 — phase1.cmake / generated Win graph pure-vendor consistency (2026-07-17)
  • D-001 — Shared boot.jar via runtime OS selection (2026-07-17)

W-005 — Combined PE JNI stub DLL aliased as libjavacore/libopenjdk/libicu_jni

  • State: CLOSED (2026-07-17) — product packaging uses stage_native_modules.sh (real PE only); libcombined is legacy non-product
  • Kind: workaround
  • Area: libcore-stub / packaging
  • Symptom / why: Full ojluni + ICU4C PE ports not built; ART InitNativeMethods still dlopens those sonames.
  • Current behavior: tools/win64/jni_stubs/libcombined.dll copied to six names (libjavacore.dll, libopenjdk.dll, libicu_jni.dll, and short names). ~160 Java_* exports, hand-written (~2.3k LOC).
  • Proper fix: Real PE modules (or fewer real DLLs) from Soong/bp2cmake Win64 graph: javacore, openjdk, icu_jni + icuuc/i18n, etc.; stop multi-name aliasing of one stub.
  • Code anchors: tools/win64/jni_stubs/build_combined.sh, tools/win64/host_package/package_win64_phase3.sh, stage scripts in phase2 RESULT
  • Opened: 2026-07-16 (Phase 2; expanded Phase 3)

W-006 — Minimal NativeConverter / ICU version shims (not full ICU4C)

  • State: CLOSED (2026-07-17) — product uses real icu_jni NativeConverter + icuuc/icui18n + icudt; native_converter.c obsolete and removed from libcombined; charset stub no longer product path
  • Kind: workaround
  • Area: icu
  • Current behavior: Phase-3 package historically used native_converter.c stubs. Phase A progress: real PE icuuc.dll / icui18n.dll / icu_jni.dll now build from AOSP sources (tools/verify/win64_libcore_icu/) and can replace stub libicu_jni in build/win64_phase1. libjavacore/libopenjdk still combined stubs (may still register overlapping charset helpers until removed).
  • Proper fix: Default package/install to real ICU PE only; remove charset exports from libcombined; verify full data (ICU_DATA / icudt) vs stubdata; complete L-001 for javacore/openjdk.
  • Code anchors: tools/verify/win64_libcore_icu/, tools/win64/jni_stubs/native_converter.c
  • Opened: 2026-07-16
  • Progress: 2026-07-17 — real ICU PE + CoreProbe wine OK with hybrid package

W-007 — Classic sockets / poll via Winsock select (not full Os/NIO)

  • State: CLOSED (2026-07-17) — permanent WinNT design: classic Os sockets use Winsock + select()-based poll/timeouts (not CRT-fd WSAPoll)
  • Kind: workaround → permanent platform design
  • Area: libcore-stub / net
  • Symptom / why: Full AOSP libcore.io.Linux PE not used on Win64; real Win10 rejected CRT _open_osfhandle + WSAPoll (WSAEINVAL on accept poll).
  • Fix / design:
    • Product libjavacore Win bridge (win_net_natives.c) implements classic socket surface with select() for poll, SO_TIMEOUT waits, and connect write-readiness.
    • NIO epoll path similarly select-emulated in compat/src/win64_socket_posix.c (bounded FD_SETSIZE).
    • 2026-07-17: registered bind/connect SocketAddress overloads for InetSocketAddress (AF_UNIX still out of product scope).
  • Evidence:
    • Host G12 (2026-07-16): net/dns/goldenapp PASS after select poll fix (tools/verify/win64_phase3/evidence/host/ANALYSIS_20260716T205926.md).
    • Wine (2026-07-17): NetProbe, DnsProbe, UdpProbe, AsyncCloseProbe, GoldenApp, SocketAddressProbe PASS.
  • Non-goals residual: AF_UNIX SocketAddress; full AOSP libcore_io_Linux.cpp (L-001 closed with Win bridge map); NIO.2.
  • Code anchors: tools/win64/jni_stubs/win_net_natives.c, register_libcore_io_Linux_win.cpp, compat/src/win64_socket_posix.c
  • Opened: 2026-07-16
  • Closed: 2026-07-17

W-009 — Phase-1 grade compat POSIX/pthread stubs

  • State: CLOSED (2026-07-17) — hot paths hardened; remaining ENOSYS is intentional Linux-only surface
  • Kind: workaround → platform compat layer (ongoing shrink is maintenance, not open product gap)
  • Area: compat
  • Fix / evidence:
    • pthread_rwlock_* now real SRWLOCK shared/exclusive (was CRITICAL_SECTION exclusive-only) — ART Mutex/ReaderWriterMutex ABI rebuilt into product art.dll.
    • uname uses RtlGetVersion + computer name; clock_gettime(CLOCK_MONOTONIC) via QPC; pthread_setname_np/getname_np via SetThreadDescription when available.
    • Socket-aware poll/epoll already select-based (W-007); mmap/mprotect/pthread mutex/cond already real Win32.
    • Wine: dalvikvm -showversion, CoreProbe, NetProbe, GoldenApp PASS after ART rebuild.
  • Residual (not OPEN product work): fork/ptrace/sendfile/tgkill etc. remain ENOSYS; further shrink only when a product path needs them.
  • Code anchors: compat/src/win64_posix_stubs.c, compat/include/pthread.h
  • Opened: 2026-07-16 (Phase 0/1)
  • Closed: 2026-07-17

W-015 — openjdkjvm memory exports minimal PE surface

  • State: CLOSED (2026-07-17) — product ships comprehensive standalone libopenjdkjvm.dll
  • Kind: workaround
  • Area: art / openjdkjvm
  • Fix / evidence:
    • Product PE from tools/verify/win64_libcore_icu/openjdkjvm_memory_standalone.c: memory/GC + file I/O + sockets + raw monitors + time (JVM_* set used by hybrid openjdk).
    • Added JVM_ActiveProcessorCount.
    • ART-tree openjdkjvm_memory_windows.cc remains minimal ART-heap-only helper for experimental ART-linked builds; not the product soname path.
    • Wine CoreProbe/GoldenApp/NetProbe with staged libopenjdkjvm PASS.
  • Code anchors: tools/verify/win64_libcore_icu/openjdkjvm_memory_standalone.c; stage via stage_native_modules.sh
  • Opened: 2026-07-16
  • Closed: 2026-07-17

W-016 — ICU needs external ICU_DATA / icudt72l.dat for wine smoke

  • State: CLOSED (2026-07-17) — product always stages run/icu/icudt72l.dat via tools/win64/stage_run_assets.sh (same class as boot.jar); libicu_jni defaults ICU_DATA to run/icu when unset
  • Kind: workaround
  • Area: icu / packaging
  • Symptom / why: Linked stubdata alone yields u_init U_FILE_ACCESS_ERROR under wine; full data file works.
  • Current behavior: Stage run/icu/icudt72l.dat and set ICU_DATA=run/icu (or absolute path). Register.cpp also calls udata_setCommonData(&U_ICUDATA_ENTRY_POINT) on Win.
  • Proper fix: Package full ICU data by default in host package scripts; verify embedded data path or always set ICU_DATA in runners.
  • Code anchors: vendor/icu/android_icu4j/libcore_bridge/src/native/Register.cpp; build/win64_phase1/run/icu/
  • Opened: 2026-07-17
  • Progress: 2026-07-17 — package_win64_phase3.sh fails if icudt72l.dat missing; phase3/4 runners and install_into_phase1 default/export ICU_DATA=run/icu

W-018 — NetProbe StructLinger NPE (getsockopt SO_LINGER incomplete in javacore Win bridge)

  • State: CLOSED (2026-07-17) — implemented getsockoptLinger/setsockoptLinger in win_net_natives; NetProbe wine PASS
  • Kind: leftover / bug
  • Area: libcore-stub / net
  • Symptom / why: NetProbe fails: StructLinger.isOn() on null from linger get.
  • Proper fix: Implement linger get/set in win_net_natives / Linux Os bridge returning real StructLinger.
  • Code anchors: tools/win64/jni_stubs/win_net_natives.c; NetProbe client path
  • Opened: 2026-07-17

W-019 — Math @CriticalNative / FastNative double ABI on Win64

  • State: CLOSED (2026-07-17) — Math.ceil/floor/sqrt + HashSet wine PASS after interpreter CriticalNative DD/DDD
  • See also: W-024 — full restore of AOSP Critical/FastNative after JIT/TLS/entrypoints (ceil/floor pure-Java demotion still OPEN debt)
  • Kind: workaround / runtime ABI
  • Area: libcore Math / ART interpreter JNI (Win64 -Xint)
  • Root cause: Official AOSP CriticalNative is fine on Linux quick/generic-JNI. Win64 multipath forces ArtMethod::Invoke through the interpreter; InterpreterJniGeneric only handled CriticalNative shorties II/I/Z/ZI. Math.ceil is shorty DD ((D)D), so dispatch fell through and crashed. Secondary: registering Math_*_jni(JNIEnv*,jclass,jdouble) under CriticalNative is the wrong ABI.
  • Fix (landed source): interpreter CriticalNative DD/DDD/FF/J; Math.c gMethodsWin → Math_ceil(jdouble) etc.; posix stubs localtime_r/mingw_gettimeofday for art rebuild.
  • Exit criteria: MathProbe + SslProviderProbe wine PASS with rebuilt art.dll.
  • Code anchors: vendor/art/runtime/interpreter/interpreter.cc, vendor/libcore/ojluni/src/main/native/Math.c, compat/src/win64_posix_stubs.c
  • Opened: 2026-07-17
  • Progress: 2026-07-17 — root cause + source fix; full art PE rebuild running

W-020 — FileChannelImpl.map0 pointer truncation on Win64 (LLP64)

  • State: CLOSED (2026-07-17) — ptr_to_jlong(mapAddress) instead of (jlong)(unsigned long)
  • Kind: bug / ABI
  • Area: openjdk NIO / boot classpath ZIP mmap
  • Root cause: AOSP FileChannelImpl_map0 returned (jlong)(unsigned long)mapAddress. On Win64 LLP64 unsigned long is 32-bit, so mapped addresses like 0x6ffff… were truncated (high bits zeroed). Memory.peekByteArray then crashed in CRT (fault_addr=0xff0e0eec pattern) while VMClassLoader clinit mapped boot.jar for ClassPathURLStreamHandler.
  • Symptom chain: Security.getProviders → provider class load → BootClassLoader.loadClass → findLoadedClass path / resource handlers → ZIP mmap via NIO → AV. Earlier W-019-style AV signature was coincidental.
  • Also fixed nearby (supporting): Win64 -Xint keeps natives on interpreter bridge (ShouldStayInSwitchInterpreter + ArtInterpreterToInterpreterBridge → InterpreterJni); FastNative stays Runnable; specialized static LLL shorty.
  • Exit criteria: SecStep17 BootClassLoader.loadClass + SecStep3 Security.getProviders wine PASS.
  • Code anchors: vendor/libcore/ojluni/src/main/native/FileChannelImpl.c; vendor/art/runtime/interpreter/interpreter.cc; interpreter_common.cc
  • Opened: 2026-07-17
  • Closed: 2026-07-17

W-021 — Default KeyStore type Android-compatible (AndroidCAStore)

  • State: CLOSED (2026-07-17)
  • Kind: config / compatibility
  • Area: JCA / conscrypt SSL defaults
  • Root cause: Win64 multipath deferred BouncyCastle, so keystore.type=BKS could not resolve. Security.initializeStatic() also omitted keystore.type, so KeyStore.getDefaultType() fell back to desktop jks, which is not registered. KeyManagerFactory.init(null,null) → KeyStore.getInstance("jks") failed and SSLContext.init aborted.
  • Fix: default keystore.type=AndroidCAStore (HarmonyJSSE/TrustedCertificateKeyStoreSpi, empty-loadable); restore loading security.properties on Windows after W-020; mirror in build_conscrypt_win64.sh and boot.jar resource.
  • Exit criteria: KeyStoreProbe + SslProviderProbe sslcontext.init=ok under wine.
  • Opened/Closed: 2026-07-17

W-022 — Product default CA bundle (AndroidCAStore cacerts)

  • State: CLOSED (2026-07-17)
  • Kind: packaging / product asset
  • Area: TLS trust / AndroidCAStore
  • Root cause: Android TrustedCertificateStore reads $ANDROID_ROOT/etc/security/cacerts/<subject_hash_old>.N. Product previously shipped empty dirs, so SSLContext.init worked but trust set was empty.
  • Fix: generate Mozilla/system PEM bundle into OpenSSL hash_old layout (tools/win64/generate_cacerts.sh), hermetic assets under tools/win64/assets/cacerts, stage via stage_run_assets.sh as required asset (with boot.jar / icudt72l.dat). LocaleData hard-coded fallback so OpenSSLX509Certificate date parsing works without full ICU4J resource bundles in boot.jar.
  • Exit criteria: TrustStoreProbe AndroidCAStore.size>=50 and acceptedIssuers>=50 under wine with ANDROID_ROOT=run.
  • Opened/Closed: 2026-07-17

W-023 — OkHttp Http(s)Handler on bootclasspath + ASCII IDN/Normalizer multipath

  • State: CLOSED (2026-07-17)
  • Kind: packaging / compatibility
  • Area: java.net URL / HTTPS
  • Root cause: Android resolves http/https via com.android.okhttp.HttpHandler/HttpsHandler, not packaged in multipath boot.jar. After packaging, pure-ASCII OkHttp/TLS paths still required ICU4J StringPrep/Normalizer tables not present in boot.jar.
  • Fix: tools/bootjar/build_okhttp_win64.sh merges repackaged OkHttp+okio into boot; IDN.toASCII and java.text.Normalizer short-circuit pure-ASCII; product ICU data preferred over stub in libicu_jni Register.cpp; cacerts already staged.
  • Exit criteria: HttpsProbe handler resolution + https://example.com/ status 200 under wine.
  • Opened/Closed: 2026-07-17

L-001 — Real PE libcore / openjdk / ICU module build

  • State: CLOSED (2026-07-17)
  • Kind: leftover
  • Area: build / libcore / icu
  • Gap: Win64 product still on libcombined / incomplete hybrid PE product PE from AOSP + multipath hybrids; no libcombined aliasing.
  • Exit criteria: PE DLLs built from AOSP sources without libcombined aliasing; GoldenApp + charset/locale smoke still pass. Met.
  • Fix / evidence:
    • Product stages only real PE via tools/win64/stage_native_modules.sh (rejects libcombined): libicu_jni, libjavacore, libopenjdk, libopenjdkjvm, icuuc, icui18n (+ optional crypto under L-002).
    • Hybrid libjavacore includes AOSP Register surface + Memory, NetworkUtilities, NativeBN (libcrypto), ExpatParser (static vendor/external/expat), AsynchronousCloseMonitor, OsConstantsHolder (multipath), Win Os bridge (win_fs/win_net/register map).
    • Hybrid libopenjdk ships AOSP NIO/zip/fdlibm surface + win_close NET_* AsyncClose wrappers (NIO.2 non-goal).
    • Wine gates (2026-07-17): GoldenApp (golden.ok/net.ok/done), CoreProbe (charset=true), LocaleProbe, plus L-001 probes Bn/Xml/AsyncClose/OsConstants/Dns/Net/Io.
  • Opened: 2026-07-17
  • Closed: 2026-07-17
  • Progress / residual (not exit blockers):
    • Full AOSP libcore_io_Linux.cpp remains excluded by design for Win64; product Os surface is the Win bridge map (win32_libcore_os_natives.md: needed=0, 82 implemented, 44 ENOSYS).
    • cbigint unused in graph; Linux-only android_system_OsConstantsHolder.cpp replaced by multipath Win TU.
    • Crypto/TLS productization tracked under L-002; NIO.2 non-goal.
    • Details: tools/verify/win64_libcore_icu/RESULT.md

L-002 — boringssl / conscrypt / SSL PE

  • State: CLOSED (2026-07-17) — product TLS stack green under wine (providers + SSLContext.init + HTTPS GET)
  • Kind: leftover
  • Area: crypto
  • Gap: Win64 TLS/crypto PE incomplete product PE + boot packaging complete for HTTPS smoke.
  • Exit criteria: HTTPS/crypto golden or explicit non-goal. Met (wine HttpsProbe status 200 + SslProviderProbe).
  • Fix / evidence:
    • PE: libcrypto / libssl / libjavacrypto from hybrid CMake; staged single-soname product names.
    • Boot: tools/bootjar/build_conscrypt_win64.sh + build_okhttp_win64.sh → OpenSSLProvider/JSSE + OkHttp handlers + security.properties (AndroidCAStore).
    • Trust: product run/etc/security/cacerts (121 roots) via stage_run_assets.sh.
    • Wine (2026-07-17 reverify after ART/compat rebuild):
      • SslProviderProbe.done=ok (AndroidOpenSSL digests/AES-GCM/SSLContext.init)
      • HttpsProbe.done=ok (https://example.com/ status 200; handlers Http/HttpsURLConnectionImpl)
  • Residual (non-exit / optional): boringssl win-x86_64 ASM acceleration; BouncyCastle/BKS; full ICU4J IDNA tables for non-ASCII hosts; broader HTTPS golden matrix on real Win10.
  • Code anchors: tools/verify/win64_libcore_icu/CMakeLists.txt; tools/bootjar/build_conscrypt_win64.sh; tools/bootjar/build_okhttp_win64.sh; tools/win64/stage_run_assets.sh
  • Opened: 2026-07-17
  • Closed: 2026-07-17

L-003 — Process/exec, rich locale, zip edge, UDP/IPv6 matrix

  • State: CLOSED (2026-07-17)
  • Kind: leftover
  • Area: libcore / openjdk hybrid
  • Gap: Phase-3 product matrix for process/exec, locale (without full ICU4J bundles), zip edges, UDP IPv4, dual-stack IPv6 Os.socket bind.
  • Fix:
    • win_process_natives.c CreateProcess UNIXProcess + openjdk OnLoad register
    • InterpreterJni 12-slot path for multi-arg natives (forkAndExec, sendtoBytes)
    • UDP recvfrom InetSocketAddress holder fill; multicast GroupReq/IpMreqn
    • ZipFile CEN: Windows heap-read + DirectByteBuffer mirror (mmap CEN invalid under wine)
    • LocaleProbe uses Calendar/String case without ICU DecimalFormatSymbols bundles
    • Ipv6Probe: Os.socket AF_INET6 bind on raw :: (avoid reverse-DNS hang)
    • Gate: tools/verify/win64_phase3/run_l003_wine.sh — OVERALL PASS
  • Exit criteria: Process/UDP/locale/zip/IPv6 gates documented + wine green met.
  • Non-goals / host residual: TCP IPv4-mapped dual-stack under wine; full ICU Collator resources; zip STORED empty-dir edges beyond DEFLATED multi-entry.
  • Code anchors: win_process_natives.c, win_net_natives.c, ZipFile.java (Win CEN), FileInputStream.c available0, interpreter.cc 12-slot, probes under tools/verify/win64_phase3/src/
  • Opened: 2026-07-17
  • Closed: 2026-07-17

L-004 — Shrink or replace multi-name DLL staging

  • State: CLOSED (2026-07-17) — product ships one PE soname each: libicu_jni/libjavacore/libopenjdk/libopenjdkjvm/libcrypto/libssl/libjavacrypto (+ icuuc/icui18n); short-name twins removed from packaging
  • Kind: leftover / packaging debt
  • Depends on: L-001, W-005
  • Fix: CMake OUTPUT_NAME for hybrid targets; stage_native_modules.sh stages only product names and deletes short twins; install rejects short-name reappearance
  • Opened: 2026-07-17

L-005 — Linux multiplatform imageless Hello / boot.jar CI gate

  • State: CLOSED (2026-07-17)
  • Kind: leftover
  • Area: linux-host
  • Gap: After repo migration, host Linux verified dalvikvm -showversion only scripted gate landed.
  • Exit criteria: One scripted imageless Hello (or RESULT) on multiplatform main.
  • Fix: tools/verify/linux_hello/run_imageless_hello.sh + RESULT.md PASS on build/native/dalvikvm imageless -Xint Hello. Requires UnixFileSystem boot.jar (rejects Win64 WinNT product boot).
  • Opened: 2026-07-17
  • Closed: 2026-07-17

L-006 — phase1.cmake / generated Win graph pure-vendor consistency

  • State: CLOSED (2026-07-17)
  • Kind: leftover / build
  • Area: build
  • Gap: Residual MinDalvikVM-Archive path assumptions in product scripts pure-vendor.
  • Fix / evidence:
    • Product CMake (tools/verify/win64_phase1, win64_libcore_icu, native/, Linux verify) already resolved via ${MDVM_NATIVE_SRC_ROOT_DIR} → vendor/; phase1.cmake has no hard-coded archive absolutes.
    • tools/bootjar/build.sh no longer auto-discovers sibling MinDalvikVM-Archive(_) for ICU/annotation stubs; requires nested vendor/icu + in-tree compat/java-stubs (expanded minimal android.annotation / android.compat.annotation set).
    • MDVM_ARCHIVE remains an optional non-default escape hatch only.
    • Docs/tests scrubbed: README.md, native/{CMakeLists,generate}.sh, tools/bp2cmake CODEGEN/codegen + unit tests point at multipath vendor/.
    • Historical tools/verify/*/RESULT.md absolute archive paths left as past evidence only (not product inputs).
  • Opened: 2026-07-17
  • Closed: 2026-07-17

D-001 — Shared boot.jar via runtime OS selection

  • State: CLOSED (2026-07-17)
  • Goal (actual): one multipath boot.jar (not dual packaged jars / not “prove WinFS-on-Win and UnixFS-on-Unix” as close criteria)
  • Doc: archived/shared_bootjar_runtime_os_detection.md
  • Canonical property: dalvik.vm.multiplatform.internal.os = windows | unix
    • Long + internal intentional (not a public app API; not expected for external use)
    • Reject short dalvik.vm.mp.os (mp ambiguous)
    • Values: windows|unix (not posix, not linux) — aligns with WinNTFileSystem / UnixFileSystem
  • Injection: vendor/art/runtime/runtime.cc after PropertiesList release (PE=windows, ELF=unix if unset)
  • Detection ladder: VMRuntime.properties() → System props / os.name → default unix (VMRuntime.isWindowsOs)
  • Separators: removed from AndroidHardcodedSystemProperties; set in System.initUnchangeableSystemProperties
  • Boot: tools/bootjar/build_win64.sh stages shared jar (no WinNT-only overlay); jar embeds both FS + isWindowsOs
  • Exit criteria (met): single shared boot pipeline produces one jar used for Linux imageless Hello (L-005 PASS on shared multipath bytes)
  • Non-goals for this close: dual-host acceptance that Windows always selects WinNTFileSystem and Unix always selects UnixFileSystem under product PE/wine — those are ordinary product smoke, not D-001 scope
  • Follow-up (orthogonal): wine/host Hello on same bytes; PE art.dll inject path when PE product is rebuilt
  • Code anchors: dalvik/system/VMRuntime.java (isWindowsOs*), DefaultFileSystem.java, System.java, runtime.cc, build_win64.sh
  • Opened: 2026-07-17
  • Closed: 2026-07-17

Design notes

No open design notes. Closed D- items live under §Closed.

Suggested next closures (priority)

  1. D-001 CLOSED — single shared boot.jar (runtime OS selection); dual-host FS smoke is not the close bar.
  2. W-001–W-003 — TLS/entrypoint implementation (design draft ready); then W-024 restore Critical/FastNative surfaces.
  3. L-001 — CLOSED real PE libcore/openjdk/ICU hybrid; residual Linux TU/bridge growth optional.
  4. H-001 — host Phase-4 with multiplatform package.
  5. L-005 — Linux Hello gate CLOSED.

Maintenance checklist for future PRs

  • New #ifdef _WIN32 temporary behavior → new W- row
  • New stub JNI → update W-005 export scope or split W-
  • Gate newly green on host → close matching H-
  • Permanent design choice (e.g. VEH forever) → move from W- to documented architecture; close workaround
  • CLOSED items: move full item into §Closed (sorted by ID); keep State CLOSED history

W-025 — JIT code cache + x86_64 codegen TLS (Windows)

  • State: OPEN (J-1 green; D-1 mostly on; residual StringFactory exclude)
  • Kind: feature gap
  • Area: art / jit / compiler
  • Symptom / why: Need Linux-like JIT. Create used to soft-fail; residual NPE when both StringBuilder.toString and StringFactory.newStringFromBytes are JIT'd.
  • Current behavior: J-1 Create OK; ThreadOffsetAddr/r15; compile default ON excluding StringFactory. ART_WIN64_JIT=0 disables compile; ART_WIN64_JIT_ALLOW_STRINGFACTORY=1 re-enables residual.
  • Proper fix: Fix StringFactory/toString pair (win32_jit_memory.md §13); drop exclude.
  • Code anchors: mem_map.cc RemapAtEnd; jit.cc Win gate; assembler_x86_64.*; codegen x86_64
  • Opened: 2026-07-19
  • Updated: 2026-07-19 — default compile ON minus StringFactory; pair isolation; SysV-wrapper attempt AV (reverted)

Last snapshot: 2026-07-19 — W-001 closed (quick invoke default ON); nterp product default ON (opt-out ART_WIN64_NTERP=0); JIT default remains ART UseJitCompilation=true.