WinDriver/CLAUDE.md

Win31drv Project Memory

Build Environment

• DJGPP cross-compiler: ~/djgpp/djgpp/bin/i586-pc-msdosdjgpp-gcc (GCC 12.2.0)
• DJGPP binutils need libfl.so.2: stored in tools/lib/ (Makefiles set LD_LIBRARY_PATH)
• CWSDPMI zip stored in tools/cwsdpmi.zip (extracted to bin/ during build)
• DOSBox-X: flatpak run com.dosbox_x.DOSBox-X (installed as user flatpak)
• CWSDPMI.EXE in bin/ directory for DPMI support under DOSBox-X
• Config: dosbox-x.conf with S3 Trio64 machine type, 64MB RAM

Project Structure

windriver/
├── Makefile          # Top-level: builds demo, calls win31drv/Makefile
├── demo.c            # Demo program
├── dosbox-x.conf     # DOSBox-X config (S3 SVGA)
├── obj/              # Demo object files
├── bin/              # Executables + CWSDPMI.EXE
└── win31drv/         # Library
    ├── Makefile      # Builds libwindrv.a
    ├── obj/          # Library objects
    ├── neload.c/h    # NE format loader
    ├── neformat.h    # NE structures
    ├── thunk.c/h     # 32→16 bit thunking
    ├── windrv.c/h    # Main API
    ├── winstub.c/h   # Windows API stubs
    ├── winddi.h      # DDI structures
    └── wintypes.h    # Win16 types

DJGPP Portability Notes

• uint32_t is unsigned long (not unsigned int) in DJGPP — use PRIu32/PRIX32 from <inttypes.h>
• Always include <stdarg.h> explicitly for va_list/va_start/va_end
• Headers must be self-contained (include their own dependencies)

Thunking Architecture Notes

• SS == DS == DGROUP: Win3.x drivers assume SS == DS == DGROUP. VBESVGA's BBLT.ASM does PrestoChangeoSelector(SS, WorkSelector) to create a code alias for compiled blit code. thunkCall16 uses dgroupSel as SS (SP=0xFFF0) when available. Without this, the code alias has the wrong base and the CPU executes garbage.
• Register corruption with -O2 inlining: When demo.c's demoDrawing is inlined into main, DJGPP GCC 12.2.0 mishandles callee-saved registers across thunk calls in long functions. Fix: __attribute__((noinline)) on demoDrawing. Symptom: handle pointer corrupted to a ColorInfo return value (e.g. 0xFF0001F6) between Demo 2 and Demo 3.

DOSBox-X Driver Notes

• waitForEngine(): GP_STAT port 0x9AE8 bit 9 polling — S3 only (gIsS3 guard)
• S3 detection: Probe CR30 chip ID register. S3 chips: 0x81-0xE1. ET4000: 0x00. Only apply S3-specific setup (cursor disable, dispYOffset, setDisplayStart) when isS3=true AND driver is not VGA-class (1bpp/4planes).
• Pattern scratch artifact: S3 driver writes 8x8 dithered brush pattern to VRAM at fixed position (~(144,1)-(151,8)) during accelerated pattern fills. Fixed by shifting CRTC display start down 10 scanlines (dispYOffset) so the scratch area is off-screen.
• -fno-gcse required for windrv.c: With -O2 GCSE, stack layout causes issues during 16-bit driver calls. Only windrv.c needs this. See WINDRV_CFLAGS in win31drv/Makefile.
• Output DDI (polylines/rectangles) requires a physical pen from RealizeObject, not a raw LogPen16T. The pen must be in DGROUP (same as brush, drawMode, PDEVICE).
• wdrvUnloadDriver does NOT auto-call Disable — caller must handle text mode restore
• sleep() hangs under DOSBox-X because BIOS timer ticks don't advance without I/O
• Debug output: -d flag enables verbose logging in neload, winstub, thunk, and windrv
• Known issue: mode mismatch HW=800x600 vs GDIINFO=640x480

DGROUP Stack Management

• VGA.DRV ships with DGROUP[0x0A]=0xFFFF (stack bottom = top of segment → no stack). Its BitBlt prolog calls a stack check function at 0x18CA that compares available stack against [SS:0x0A]. With 0xFFFF, ALL functions fail immediately (return 0).
• Fix: patch [0x0A] to objBase after extending DGROUP. Only done when original = 0xFFFF.
• S3TRIO.DRV and VBESVGA.DRV have [0x0A]=0x0000 — no patching needed.
• Do NOT unconditionally overwrite DGROUP offsets 0x00-0x0F — VBESVGA.DRV stores driver-specific data there (0x030A at offset 0, 0x01 at offset 4).

BitBlt Source Device Rules

• For pattern-only ROPs (PATCOPY=0xF0, BLACKNESS=0x00, WHITENESS=0xFF, etc.), lpSrcDev must be NULL (0:0) per DDI spec. VGA.DRV rejects non-NULL source for pattern-only ROPs. S3TRIO.DRV tolerates it but correct behavior is NULL.
• Source dependency check: ropNeedsSrc = (((rop8 >> 2) ^ rop8) & 0x33) != 0

INT 10h ES Translation

• Different INT 10h function families use different ES:offset registers: VBE 4Fxx → ES:DI, AH=10h (palette) → ES:DX, AH=11h (fonts) → ES:BP, AH=1Bh → ES:DI
• Only specific AL subfunctions use ES as a buffer pointer; most don't
• Copy sizes must be exact (17 bytes for palette, CX*3 for DAC blocks, etc.)
• Copy direction matters: "Set" = copy-in only, "Read/Get" = copy-out only

WINFLAGS Handling

• WF_80x87 NOT used: We don't save/restore FPU state across thunk boundaries
• VGA-class drivers need WF_STANDARD: VGA.DRV's physical_enable hangs in Enhanced mode (polls VDD that doesn't exist). Auto-detected after Enable(style=1) returns 1bpp/4planes GDIINFO → repatch __WINFLAGS in all segments (0x0025→0x0015).
• SVGA drivers (S3TRIO, VBESVGA) use WF_ENHANCED normally

ET4000 Driver Notes

• ET4000.DRV from Win 3.x distribution is SZDD-compressed; decompress with msexpand (rename to .DR_ first, output is .DR without the V — rename to .DRV)
• DOSBox-X machine type: svga_et4000 for ET4000 hardware emulation
• ET4000 is 640x480 8bpp, software-rendered (no accelerator engine in DOSBox-X)
• CR30=0x00 on ET4000 → isS3=false → no S3 engine wait, no display start shift

Current Demo Status

• S3TRIO.DRV, VBESVGA.DRV, VGA.DRV, ET4000.DRV all work: Load → Enable → Draw → Disable → Unload
• Demo 1: Fill rectangles (BitBlt) — works
• Demo 2: Pixel patterns (Pixel) — works
• Demo 3: Lines/starburst (Output/Polyline) — works
• Demo 4: Screen-to-screen blit (BitBlt SRCCOPY) — works
• VGA.DRV: 640x480 4-plane 16-color mode; limited color palette but functional
• ET4000.DRV: 640x480 8bpp on svga_et4000; software-only, no hw acceleration
• Drivers stored in drivers/ directory, copied to bin/ during build