65816-llvm-mos/demos

llvm816 GS/OS Demo Apps

Small Apple IIgs S16 applications that build with our LLVM/clang toolchain, wrap as OMF v2.1 ExpressLoad, and launch under real GS/OS 6.0.2 in MAME.

Building / running

bash demos/build.sh helloBeep        # build the OMF
bash demos/launch.sh helloBeep       # interactive run in MAME (visible window, audio)
bash demos/test.sh helloBeep         # headless test (boots, runs, checks marker)

The launch script runs MAME with no timeout; close the MAME window when done (Esc, then Cmd-Q). The test script injects a keystroke after the demo has launched, then checks $00:0070 for the 0x99 end-of-run marker the demo writes before exit.

Demos

helloBeep.c

Three SysBeep() calls then exit. The toolbox SysBeep lives in Misc Tools; no other startup needed (Loader handles MM + TL).

helloText.c

Initialises Event Manager (so GetNextEvent can read keystrokes) and Text Tools (so WriteCString has an output device), writes two greeting lines to the text screen, waits for any keypress via GetNextEvent, beeps, exits. Uses these toolbox calls:

• MMStartUp, NewHandle (DP allocation for Event Manager)
• EMStartUp, GetNextEvent (event-driven keyboard)
• TextStartUp, SetOutputDevice, WriteCString (text I/O)
• SysBeep

helloWindow.c

Initialises the full Window Manager startup chain (Memory, QD, Event, Scheduler, Window), constructs an Apple-IIgs-Toolbox-Ref NewWindow parm block, and calls NewWindow. When NewWindow returns a real handle the demo calls SetPort, ShowWindow, MoveTo, and DrawString to put a greeting in the window, then waits for a keypress, beeps, and exits.

Three toolchain bugs needed fixing to make this work end-to-end under real GS/OS 6.0.2 (all now landed):

• omfEmit RESSPC=0 for code segments — BSS got no memory allocation; writes past the image end silently vanished. Now BSS is embedded as zeros in the LCONST data.
• Loader cRELOC at segPlacedBase=$0000 (not $1000 like our text-base) — host probes need to compute runtime addresses as link_addr - text_base + bank<<16.
• &symbol bank=0 — proper backend fix landed. LDAi16imm_bank AsmPrinter pseudo lowers to lda $BE; crt0 stores PBR to $BE
  • zero to $BF at startup, so the high half of every &symbol pointer carries the actual load bank at runtime. Toolbox pointer args now Just Work without per-wrapper PBR overrides.

orcaFrame.c

First ORCA-style desktop application. Opens a Window Manager window via startdesk() (full toolset chain), runs a TaskMaster event loop until the close box / Q key / 1000-iteration watchdog fires. Both 6.0.2 (sys602.po) and 6.0.4 (6.0.4 - System.Disk.po) launch it cleanly; fTitle works on both.

frame.c

Full port of ORCA-C's Frame.cc sample. Builds the Apple+File+Edit menu bar via the real ROM Menu Manager (NewMenu / InsertMenu / FixAppleMenu / FixMenuBar / DrawMenuBar) and renders the original "About Frame" dialog (white-filled framed rect with the 1989 Byte Works copyright text and an OK button).

minicad.c

Full port of ORCA-C's MiniCAD.cc sample. Apple+File+Edit+ Options menu bar + a windowed canvas with three seeded line-art patterns (curve-stitching, sunburst, Star of David).

reversi.c

Full Othello game ported from ORCA-C's Reversi.cc. 100-byte sentinel-bordered board, 8-direction capture detection, 1-ply AI with corner/edge weighting, QD-rendered board with black/white pieces.

qdProbe.c

Diagnostic — minimal QD/EM/WM init followed by RefreshDesktop plus ZP/SHR markers. Used to prove that WindStartUp does NOT auto-paint the desktop and that RefreshDesktop(NULL) is what actually fills SHR with the dithered desktop pattern. Run via scripts/probeQdStartup.sh.

Known limitations

• fTitle on stripped 6.0.2: orcaFrame uses fTitle and runs fine on both 6.0.2 sys602.po and 6.0.4 System.Disk.po. Earlier notes that fTitle required disk fonts were superseded — the underlying bug was a QDStartUp argument-order mistake in runtime/src/desktop.c, fixed 2026-05-16.
• Window not visually painted: WindStartUp does NOT paint the desktop on its own; RefreshDesktop(NULL) is required. Adding the call to startdesk() works for qdProbe.c but pushes the orca demos past the GS/OS Loader's silent-rejection threshold (see memory: loader-creloc-threshold).
• GS/OS Loader cRELOC threshold: anywhere from 65-90 cRELOCs the Loader silently refuses to launch ExpressLoad OMFs. The threshold is not a clean reloc-count cutoff; OMF byte layout and reloc patch offsets both matter. The ORCA ports are slim to stay under it.

What got fixed during demo authoring

Substantive toolchain bugs surfaced and fixed:

• iigsGsos.s GS/OS wrappers and genToolbox.py's 890 toolbox wrappers were pushing 4-byte Long args low-word-first. ORCA-C's PushLong macro is high-word-first. Anything passing a Long arg through the toolbox dispatcher (NewHandle, NewWindow, fopen->gsosOpen, etc.) was reading garbage parm-block pointers. Fixed in both files plus the generator; regenerated 890 wrappers.
• runtime/build.sh wasn't rebuilding iigsToolbox.s -- the .o was stale since May 4. Added the missing line.
• (short)(*(void **)dpHandle) was a double-dereference bug that read garbage at the master pointer's destination instead of the master pointer VALUE (= the DP address). Corrected to (unsigned short)(unsigned long)*(void **)dpHandle for proper block-address extraction.
• helloText used TextStartUp + WriteCString directly which crashed to the IIgs monitor; fixed by adding SetOutputDevice(1, 0) to wire stdout to the text screen.
• helloText's event loop hung forever because EMStartUp wasn't being called; fixed by adding it (with proper DP allocation).

ptr32 mode note

All address constants in the demos (text screen $00:0400, SHR RAM $E1:2000, soft switches $00:C0xx, keyboard register $00:C000) are plain C 32-bit pointers like (volatile unsigned char *)0xE12000UL. In ptr32 mode the compiler emits LONG addressing automatically; no switchToBank2-style inline asm or DBR juggling required.