Skip to content

Merge release/0.6 into develop - #818

Merged
peiikai merged 25 commits into
developfrom
chore/merge-release-0.6-into-develop
Nov 16, 2026
Merged

peiikai merged 25 commits into
developfrom
chore/merge-release-0.6-into-develop

Conversation

@peiikai

@peiikai peiikai commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Brings the v0.6.1 fixes on release/0.6 into develop:

One conflict, in runtime/ui/update_tags.h: develop had generalized the rolling-build lookup (newest_rolling for nightly and experimental), while release/0.6 only reworded the comment. Kept develop's code and took the comment change. tools/test_update_tags.cpp passes. develop's own files (version, nightly.yml) are unchanged.

🤖 Generated with Claude Code

peiikai and others added 25 commits November 13, 2026 23:10
…e it got (#779)

On Windows on ARM (x64 emulation, Windows 11 26H1) the game code generator still
stops with 0xC0000409 although it runs with Control Flow Guard off. Both
generators reach unicorn's first jump table emulation within seconds; the sdk
one just has no log line before it, so its log stays empty.

unicorn leaves its generated code with VCRUNTIME140's longjmp from a setjmp
that saved no frame. That longjmp fast-fails for it when CFG is on (#337),
and when a shadow stack is present it always unwinds with RtlUnwindEx, which
can't walk the generated code. So the generator is now also started with
user-mode shadow stacks off (falling back to CFG off alone where Windows
refuses that policy).

gen.py now logs a first line with the CPU it runs on (emulated or not) and
whether CFG and shadow stacks are on in its process, and a line before its
first unicorn run, so a failure's last log line shows how far it got.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The test build with shadow stacks off still stopped at the first jump table
emulation on Windows on ARM. VCRUNTIME140's longjmp fast-fails on a jmp_buf
with frame 0 (what unicorn's setjmp wrapper saves) whenever its CFG check
pointer isn't its own no-op. Under x64 emulation on ARM, Windows sets that
pointer for x64 code even with CFG off (it catches calls into ARM64EC code),
so turning CFG off for the process can't help there.

tools/win_longjmp.py points unicorn.dll's longjmp import at a plain register
restore (what VCRUNTIME140's longjmp does after that check with no frame and
no shadow stack). The runtime keeps starting the generator with shadow stacks
off, which this needs.

Also fixes the generator's mitigation line (ctypes passed the process
pseudo-handle as a C int), and adds whether VCRUNTIME140's CFG check pointer
was set by Windows. CI runs unicorn with both longjmps on Windows x64 and
under emulation on the ARM runner (tools/test_win_longjmp.py).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eader (#779)

On the Windows on ARM runner the x64 Python reports platform.machine() ARM64,
so the longjmp replacement was skipped there; it would have been for the
reporter's lasgen too. The module's PE machine decides now.

The runner doesn't reproduce #779 (VCRUNTIME140's longjmp works there and its
CFG check pointer is its own), and the generator's process reads shadow
stacks as on under emulation. So the replacement covers both ways
VCRUNTIME140's longjmp fails: unicorn's setjmp saves the shadow stack pointer
in the jmp_buf's frame field, and longjmp pops the shadow stack back to it
(incsspq) before the plain register restore. Without a shadow stack both are
plain register save / restore.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the constant buffer words promoted to push constants in both stages, the
fragment shader's start after the vertex shader's (a block at a nonzero
offset). On an Adreno X2-90 with Qualcomm's Windows driver most of the picture
was black. LARECOMP_VK_PUSH_VS=0 (fragment words only, from offset 0) drew
right, so it is now the default on that driver (by driver ID), with a log line
and a row in docs/vulkan-gpu-defaults.md.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
)

With fragment words alone (the last commit's default) the world drew but the
file select's text didn't on the reporter's Adreno X2-90. With
LARECOMP_VK_PUSH_CB=0 everything drew. c1 stays inlined, as in those runs.
LARECOMP_VK_PUSH_VS stays as a switch for narrowing it down.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rator)

Game files copied from a Mac to an SD card carry a hidden ._<name> beside every
file. The Android import copied them all (4,011 on the RP5), and the room
shuffle read ._Lv07EagleTower.bntx as a dungeon's textures: Chaos seeds failed
with "'File' object has no attribute 'textures'".

- Android import: ._* and .DS_Store aren't copied or counted.
- Seed generator: one os.listdir filter in las_headless.py drops them for every
  listing (tools/rando and LAS-Randomizer), whatever copied them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Video tab's Resolution row takes its size from the renderer's scale,
which is only worked out when a game frame is drawn. The start screen
draws none, so the scale stayed at 1x and every choice showed 1920x1080.
On the start screen the size is now worked out from the settings and the
window (output_scale_estimate, the same math output_scale uses). In game
nothing changes. Shared code: Metal and Vulkan, desktop and touch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
MSYS2's pkgconf 3.0.11 installs pkg-config and the target-prefixed names as
links without .exe; only pkgconf.exe keeps the extension. Bash still runs
them, but Python's shutil.which on Windows only matches PATHEXT names, so
the delete guard and data move tests built without SDL3 flags and failed to
link on windows-arm64. Fall back to pkgconf.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Test builds can now name windows-arm64 (-f platforms=windows-arm64,linux-aarch64). Windows on ARM is built only
then and never ships on a release; it lives on the nightly page. experimental.yml is gone: its test-arm path failed
to publish, and its rolling build would overwrite the Experimental page's notes that point old builds to nightly.
Docs and comments that named it now point at nightly.yml / release.yml.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Windows: start the generator with shadow stacks off too, and log where it got (#779)
Skip a Mac's hidden ._ files in game files (Android import, seed generator)
CI: find SDL3 through pkgconf in the delete guard and data move tests
Start screen: show the real size for each Resolution choice
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ARM test builds through release.yml; remove experimental.yml
update_tags.h: kept develop's newest_rolling (nightly + experimental); only the
comment now says experimental.yml is retired, as release/0.6 does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Vulkan: no constant-buffer push constants on Qualcomm's own driver (#779)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@peiikai
peiikai merged commit 3583ff3 into develop Nov 16, 2026
8 checks passed
@peiikai
peiikai deleted the chore/merge-release-0.6-into-develop branch November 16, 2026 01:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant