AGC: pipelines first built after the journal reaches APS5_PIPELINE_JOURNAL_MAX_MIB are compiled again on every warm boot #8309

What happens

PipelineJournal refuses any append that would take the journal past APS5_PIPELINE_JOURNAL_MAX_MIB (default 1024 MiB), and it evicts nothing (core/libs/prx/libSceAgcDriver/Graphics/src/PipelineJournal.cpp:777 on main 3fd1cc3).

• It prints "[pipeline-journal] … reached APS5_PIPELINE_JOURNAL_MAX_MIB; later pipelines are not recorded" once. After that it only counts the refused writes.
• So the journal holds whatever a title builds first. If it fills while a level loads, the rest of that level's pipelines are never recorded, and every warm boot compiles them again at the level's entry.
• The loader reads a larger file in full: a 2 GiB journal was replayed under the default cap here. So the cap limits only what gets recorded.

Module records are almost all of the bytes. In the 2 GiB journal below:

• 924 module records take 2047.6 MiB (median 362 KiB, largest 14.7 MiB);
• the 958 pipeline records take 0.3 MiB.

#8213 notes that Ghost of Tsushima's journal is 1023 MiB, with 940 pipelines and 979 modules. So at least one more title's journal is within 1 MiB of the cap. I haven't measured that one.

Measured on PPSA11386 (007 First Light)

One run each:

1. A cold seeding run, which writes the journal.
2. A warm run from that seed.

The warm run always used the default cap.

• Boot to resume took 385 s and 384 s, including the journal replay (258 s and 263 s, 16 threads).
• In the warm run from the default-cap seed, the 280 s level entry had no shader-cache misses. The four async queues spent about 282 s each in WAIT_REG_MEM packets, while queue 0's worker built pipelines in its draw packets.

Options, none of them tested:

• a higher default cap;
• when an append would pass the cap, evicting the oldest records, or keeping the records the last replay used.

The replay keeps at most APS5_PIPELINE_JOURNAL_ENTRIES (default 1024) entries (PipelineJournal.cpp:908); the 2 GiB journal holds 958.

Storing references to the shader disk cache's entries instead of the modules wouldn't shrink this journal much. Only 4 of its 924 modules have SPIR-V identical to an artifact in the disk cache written by the same run.

Setup:

• Build: main 68323ba plus the open PRs PPSA11386 needs (fix(shader): apply cube output modifiers whenever IEEE mode is off #5799, fix(agc): a driver store over an earlier store's range takes its piece slot #6868, perf(agc-driver): join a buffer bound from inside a range across the ranges it crosses #5372, fix(agc): import committed spans of partially accessible windows #6918, fix(agc): wait for the display buffer's GPU writers before the presentation lock #7643, feat(agc): run registered merged hull programs (binary type 3) #6679), with draw profiling (APS5_PROFILE_DRAW=1) and scripted input.
• The title's two feedback-wait limits are widened in its executable as described in 007 First Light (PPSA11386): the terrain crash after Resume is a stale read of a GPU feedback buffer past the title's own wait limits #6172 (local only).
• RTX 3080, NVIDIA 615.78.08, Linux 7.2.10, GCC 16.2.1, Release.
• None of those PRs touches PipelineJournal.cpp.

AI-assisted: yes (Claude Code)