libretroadmin 00c04c347f GIF: path 1 read accounting on the VU1 thread, and savestate paths ...

With the VU1 thread running VU1, path 1's packets are counted as unread
in gifPathReadAmount, where the space checks look, instead of the path's
own count, which nothing that checks for space reads. Path 1's buffer
could be written over packets the GS thread had yet to send, and a
realign's skipped end, with no packet data to hand over, was waited on
for good: the copy now hands such a skip to the GS on its own, and a
realign of a packet already at the start of the buffer returns, there
being nothing to move.

A savestate path whose sizes and packet do not agree -- more buffered
than the buffer holds, the tag past the data, a packet that starts
before the buffer -- fails the load and leaves the path empty, where it
realigned from outside the buffer. tests/eegifunit checks all of these,
the unending VU1 stream that deadlocked included, under a watchdog. The
stack frame check makes the sources the Makefile generates before it
compiles them, so it runs from a fresh checkout.

2026-11-09 18:51:10 +01:00

..

build.sh

GIF: path 1 read accounting on the VU1 thread, and savestate paths

2026-11-09 18:51:10 +01:00