D890 - A REWIND released resumes only a command stream held at it

Status: decided Date: 2026-10-18

orbistoun carries a submission out from what it holds when submitted, following each INDIRECT_BUFFER it reaches. A REWIND not yet released ends the stream where it stands, and the address of that REWIND is kept as held. When the title releases a REWIND (sceAgcRewindPatchSetRewindState(rewind, 1)), the stream held at it goes on: what follows it is carried out, up to and with the jump that leaves its buffer, in order behind what came before. A REWIND released that no stream is held at only has its valid bit set; the stream that reaches it later passes it.

Why: PPSA04263 feeds the GPU as a ring. Each buffer ends in a jump to a 14-dword trampoline holding a REWIND and a jump it fills in once its next buffer is written, and then it releases the REWIND. The console's command processor runs behind the title: it holds at a REWIND it has reached, and passes one already released. 7cd7387c carried a release's continuation out whenever the title released it, and the title often releases a trampoline's REWIND before it submits the buffer that leads there - so the continuation ran ahead of its predecessor, and the submit then followed the filled-in jump again. Two boot runs in three stalled at the second flip (0568dfe6 unwired it). Only a stream that reached the REWIND can resume from it.

Rejected:

• Passing over every jump in a submitted buffer. Deterministic, but the buffers the jumps lead to never run: the title's end-of-pipe label stopped short of the prologue's fence.
• A command processor thread that stalls at a REWIND and re-reads guest memory. The same order, but preparation would have to move off the guest's thread, which costs more than it buys (an earlier measurement moved it and lost a fifth of the frame rate).