Entry 4

A model of the CD-ROM drive: libcd and the opening stream run natively in the oracle

2026-10-144mtoolingenginevideoaudio

Filescrates/rugrats-hle/src/cdrom.rshw.rslib.rs

The oracle's libcd hooks of 3 are gone: rugrats-hle now has a model of the drive behind 0x1f801800-0x1f801803 (cdrom.rs), and libcd runs as it does on the console. Commands take parameters from the FIFO and answer with INT3, then INT2 for the slow ones (Init, Pause, Stop, SeekL/P, GetID, MotorOn, ReadTOC), each response waiting until the previous is acknowledged. ReadN and ReadS deliver INT1 per sector at the mode's speed; the request register's bit 7 loads the data FIFO with 2048 bytes (mode bit 5 clear) or 0x924 from the header (set), drained by the CPU or DMA 3. With mode bit 6 set, real-time audio sectors that pass the filter go to the sound side and never raise INT1. The game's read_file shortcut stays as the default (--real-cd turns it off).

The opening stream now runs: 1500 frames, ReadS three times, 318 XA audio sectors to the sound side. The game sends Getstat 1, Setloc 5, MotorOn 2, Init 1, Demute 4, Setmode 6, SeekL 5, ReadS 3 in that stretch.

Two faults fixed on the way. An interrupt that arrived while the hand-written renderer at 0x80010000 was running crashed the handler on lw v0, 92(gp) with gp = 1: the asm uses gp as a scratch register. On the console the BIOS longjmps into the setjmp buffer ResetCallback gave HookEntryInt (0x800bbad0: ra, sp, fp, s0-s7, gp as libc's setjmp at 0x800b2758 stores them), so the dispatcher runs on libetc's own stack (sp 0x800bcab0) with the C gp. The harness now enters the dispatcher with those registers. And libetc writes DICR by halfword (0x1f8010f6); the DMA and interrupt registers now take narrower accesses as parts of the word.