Port request: Trails in the Sky 1st Chapter #877

Activity

1. ricardosousa191 commented on Nov 2, 2026

   

   ricardosousa191

   on Nov 2, 2026

   Last edited by ricardosousa191

   Author

   More actions

   Update: with two local changes the game now builds, boots and plays its cutscenes on the Mac. Inside the house the furniture (table, chairs) is missing. Details below, in case they help.

   Setup: MacBook Air M3, macOS 27.2, Xcode 27.2, Metal 32023.922, main cd859c3 plus my local changes, shader chain b7cd656abd7c. Full reports attached (bug_report.py zips, no game files).

   1. Rewrite of main (patch_nso.py)
      The assertion "retry block overlaps a window 2cdeb4" comes from a compare-and-swap window on x18:

   The window at 0x2cdeb4 (to 0x2cdec8) is a byte compare-and-swap on x18 (ldaxrb ... stlxrb). Its first cbz jumps over an inline failure path (clrex; b out) onto the stlxrb, and the second cbz at 0x2cdec8 exits on success.

   The cbz at 2cdeb8 is inside the window and jumps to its own stlxrb, but the check for branches into the middle of a window treats it as coming from outside. site_window also asserts on a plain b inside a window. My local change:

   # check loop, after `wseg = next(...)`
   if x.address in {y.address for y in wseg}: continue
   
   # site_window, before `elif m == 'adrp'`
   elif m == 'b':
       tgt = x.operands[-1].imm
       if lo <= tgt < hi: newt = copy0 + (tgt - lo)
       else: ex_targets.append(tgt); newt = exits + 16 * (len(ex_targets) - 1)
       pc = copy0 + 4 * k; off = (newt - pc) >> 2
       w = 0x14000000 | (off & 0x3FFFFFF)
       dd = dis(w, pc); assert dd.operands[-1].imm == newt and dd.mnemonic == 'b'
   

   With it the rewrite of main, subsdk0 and sdk passes. I do not know if this is the right way to handle it.

   2. Address regions
      The heap check and the profile guess give an iPhone-sized block (mem_reserve 0x140000000, mem_alias 0x98000000, 2.38 GB). The address region check reports a red line: the game reaches alias +12.27 GB (an 11.27 GB region at alias +1.00 GB, returning to main+0x56d37c). With that profile the app crashed at launch:

   CRASH sig=11 addr=0x1 pc=sdk+0x15576c lr=sdk+0x60301c, bt main+0x56e1e8 main+0x56d5dc ... (right after the game's VirtualMemoryAllocator Initialize)

   I set mem_alias 0x1000000000 and mem_reserve 0x10A8000000 (the layout the region check itself ran with, Mac only) and the crash is gone.

   3. Shaders and app
      6,601 of 6,602 shaders translated. The failed one is an ArgumentOutOfRangeException with "Invalid instruction at 0x..." lines in dmsh_build.log. The pipeline archive and the Mac app (14.6 GB) were built.

   4. What happens in the game
      It boots, plays the cutscenes and gets to the inside of the house, where the table and chairs are not drawn. No crash. In the app log:

   • 3,797 times: [RED] failed to create compute pipeline: Threadgroup memory size (45056) exceeds the maximum threadgroup memory allowed (32768). It is always the same size, 44 KB against the 32 KB limit of this Mac. nvn_mtl.m returns from the dispatch when the pipeline fails and tries again on the next dispatch, so that compute pass never runs.
   • The shader is 43c2b838b11bc19c (local size 1024x1x1, shared memory 45056 bytes, 44 bytes per thread). It is the only compute shader in shader_facts.tsv above Metal's 32 KB threadgroup limit. A group of 1024 threads fits Metal, the 44 KB of shared memory does not.
   • I do not know if it is the reason for the missing furniture. Since it is one shader, it might be easy to test by making it fit in 32 KB or by skipping that dispatch and seeing if the furniture appears.
   • 2 shaders not in the shader map: f9def5d93748a39f and a370ad7eceb0897a, and once "the dispatched program has no compute shader".
   • 2 unimplemented NVN functions, once each: nvnCommandBufferSetZCullZF32CompressionEnable and nvnTextureBuilderGetDepthStencilMode.

   The two shaders not in the shader map and "the dispatched program has no compute shader" look like the symptoms listed in docs/agents/symptoms.md for shaders the split did not find (at least one of them seems to be a compute shader), so they may be as likely a cause of the missing furniture as the 44 KB one. I have not dumped them yet.

   Happy to test patches or collect more logs.

   nxnative-report-01002C9022770000-2026-11-01.zip
2. github-actions

   added

   new gameAdds a game profile that main does not have yetAdds a game profile that main does not have yet

   game: 01002C9022770000A game profileA game profile

   on Nov 4, 2026
3. seedds commented on Nov 6, 2026

   

   seedds

   on Nov 6, 2026

   Owner

   More actions

   e9d71e2 gives the Mac a Switch-sized 64 GB alias region by default, for every game, so the launch crash after the 11.27 GB region at alias +1.00 GB (sig=11 addr=0x1 at sdk+0x15576c) should no longer need DM_ALIAS_SIZE. I don't have this game, so I haven't run it.

   If you retest on e9d71e2 or later without that switch, please say whether it gets past launch. If it crashes in a way that looks like a cut-off pointer, check the log for a [mem] line saying the heap and stacks reach past 2^39. This issue stays open for the rest of the port, including the missing furniture.

   👍React with 👍1Reacted by Ricardo Sousa
4. ricardosousa191 commented on Nov 6, 2026

   

   ricardosousa191

   on Nov 6, 2026

   Author

   More actions

   @seedds Thanks, i will give it a test!
5. ricardosousa191 commented on Nov 6, 2026

   

   ricardosousa191

   on Nov 6, 2026

   Last edited by ricardosousa191

   Author

   More actions

   @seedds Thanks. Quick clarification on the rewrite step: on main b4318c7 (which includes e9d71e2) the main rewrite still stops before building the app, so I can't reach the launch yet:

   AssertionError: retry block overlaps a window 2cdeb4 (dm/tools/patch_nso.py:1311)

   That is the compare-and-swap window from my first report (cbz at 0x2cdeb8 jumps to its own stlxrb, and the check treats it as a branch from outside). e9d71e2 doesn't touch patch_nso.py. As far as I can tell from the diff, PR #1002 already has the fix (it skips a branch that lands inside its own window, and site_window / site_retry_window handle a plain b), so it looks like main just needs that part of #1002.

   I've applied the same two changes locally on top of main and the game launched without crashing on the startup. I did a testing on #1002 PR too and the game worked really well with all the furniture and details working.
6. seedds

   added a commit that references this issue on Nov 7, 2026

   Branches within an exclusive window stay in its copy, and its uncondi…

   ...

   10fd970
7. seedds commented on Nov 7, 2026

   

   seedds

   on Nov 7, 2026

   Owner

   More actions

   The rewrite step should no longer stop here. 10fd970 (from PR #1171) fixes the retry block overlaps a window assertion. On update 1.0.6 (v327680) here, main failed the same way (at 2d7c64) on the old main and now rewrites. Your 1.0.7 (v393216) wasn't checked, but it is the same shape.

   Could you pull main and run the NSP script again on base + 1.0.7? If it packs, the next question is whether the game boots and plays; a log from the first start would help either way. Leaving this open until then.
8. ricardosousa191 commented on Nov 7, 2026

   

   ricardosousa191

   on Nov 7, 2026

   Last edited by ricardosousa191

   Author

   More actions

   @seedds Pulled main (4861fb0, which includes 10fd970) and ran the NSP script on base + 1.0.7: the rewrite of main now passes, and the app packs and boots. Report attached (bug_report.py zip; the current log is main at 4861fb0, chain 3ee493b2c3a4, and its .prev log is from the PR #1002 build, 65c7213, chain a4511bb2bb79, for comparison).

   Launch: it gets past launch with the default 64 GB alias (no DM_ALIAS_SIZE, no mem_alias edit): the log shows "alias region 64.00 GB" and there is no crash. The old Threadgroup memory (44 KB > 32 KB) red line no longer shows in the log on main, but the furniture is still missing there.

   What is still wrong on main is the whole lighting pass. Screenshots attached (Top: how the scene should look, "Exit the Sewers"; Bottom: main build on my M3). Geometry and textures are fine, but there is no dim ambient, no light pools around the torches, no fog/haze, no bloom and no shadows, and the furniture is missing indoors.

   

   The main log has these lines:

   • shader f9def5d93748a39f is not in the shader map
   • shader a370ad7eceb0897a is not in the shader map
     (prewarm: "shader keys not in the shader map (another game version) 2")
   • "the dispatched program has no compute shader (or it is not mapped)", once
   • shader 567fcd0c97a29418_b writes colour target 1, an integer format, as a float

   On the PR #1002 build the same scene renders correctly on my M3: the lighting, fog, shadows and the furniture are all there. So what is missing on main looks like the profile/shader tables and the compute and image fixes that #1002 carries (compute scratch lowering, mixed material outputs, cube image stores), not the rewriter.

   nxnative-report-01002C9022770000-2026-11-07.zip

   Update: I rebuilt on main c8bfa16 and "Mac shaders" still reports 6,601 of 6,602 (one shader fails to translate: IndexOutOfRangeException, with "Invalid instruction at 0x000008 (0x0000000000000000)", "0x000028 (0x000000003F800000)" ...; it looks like constants being read as code). The log doesn't print which key it is, so I can't tell whether it is one of the two shaders missing from the map. PR #1002's description mentions an explicit compute-code offset fix for constants after the control section, which might be related.
9. Know-It-All-Brother

   added a commit that references this issue on Nov 9, 2026

   Branches within an exclusive window stay in its copy, and its uncondi…

   ...

   bae5109
10. seedds

    added a commit that references this issue on Nov 13, 2026

    game: Trails in the Sky 2nd Chapter 1.0.4 plays on the Mac

    ...

    6cb67b8
11. seedds commented on Nov 13, 2026

    

    seedds

    on Nov 13, 2026

    Owner

    More actions

    An update related to this request: Trails in the Sky 2nd Chapter (1.0.4, the same Falcom engine) now plays on the Mac on main (6cb67b8e, game note).

    Three general changes it needed are on main too, and should reach the 1st Chapter as well:

    • Integer colour outputs chosen by each pipeline's targets (c398ff3). This fixes the same integer-target red line that PR Trails in the Sky 1st Chapter: current-main integration and GPU memory fix #1002 met on the 1st Chapter. See Integer outputs told by the pipeline's targets.
    • One-stage GLSLC shader outputs read through their section headers (0c83e68). This is needed for newer SDKs' compute shaders.
    • Two NVN functions: nvnTextureBuilderGetDepthStencilMode and nvnCommandBufferSetControlMemoryCallbackEnabled (c7edcd4).

    The 1st Chapter itself is still PR #1002, so this issue stays open.

    ❤️React with ❤️1Reacted by Ricardo Sousa
12. ricardosousa191 commented on Nov 13, 2026

    

    ricardosousa191

    on Nov 13, 2026

    Author

    More actions

    Thanks a lot, i will test it and report back about 1st chapter.
13. ricardosousa191 commented on Nov 13, 2026

    

    ricardosousa191

    on Nov 13, 2026

    Last edited by ricardosousa191

    Author

    More actions

    @seedds Tested the game for 30 minutes and still haven't found any type of glitch or texture bug. All the forniture and lighting is showing normally from what ive seen. I will continue playing to see if i encounter any glitch. Thanks!