Halve the grown bound (not below 1024 entries) once a window of the
bound's uses times a shrink window was served by at most a quarter of
it, and reset the bound on Clear(). The shrink window starts at 8 and
doubles (up to 1024) each time growth undoes the last shrink, so a set
of keys that comes back over a longer span stops the bound from
oscillating. The entries over the bound leave at the next insert, with
the other evictions.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01XBXM4pgkMHYEtqN1EQni2P
Main's #7269 keeps the synchronous settle's buffer, memory and fence
per device in ImportSettle, so clearing one device's imports releases
only that device's objects. This branch's settle batches move there
too: each device's ImportSettle holds its batch target, memory and
batches with their fences, so switching devices no longer drops a
device's batch handles, and clearing a device waits for and destroys
only its own batches. The per-import settle state is renamed
pendingSettles to stay clear of main's per-device map, and a submitted
settle records the device whose batch it is in.
USAGE.md describes the refused posted unlocks and their log line, and
says that the icon queries and a damaged or unwritable unlock or
statistics file stop the game with an error. The debt entries of
libSceNpTrophy2 and libSceNpUniversalDataSystem no longer say that any
posted id is stored. Both USAGE.md and the libSceNpTrophy2 entry say
that without trophy data the platinum is not unlocked when the other
trophies are.
With the title's trophy data loaded, an _UnlockTrophy posted through UDS
no longer records an id the data does not have, the platinum, or a
progressive trophy. Each refusal logs the trophy and the reason, returns
false and calls no listener.
The statistics path and the platinum award now use an internal record
that keeps part 1's behaviour, so a progressive trophy still unlocks
through its statistic and the platinum with the last trophy that names
it. Without trophy data every id of 0 or more is recorded as before.
A title whose uds file defines a stat with an aggregation other than
latest gets one log line naming the stat. Posting an event for it
through the exports stores no stat and unlocks no trophy.
Events other than the trophy events go through the title's
sce_sys/uds/udsNN.ucp: stats_definition.json gives the statistics that
keep their latest uint32 or uint64 value, and stats_extraction.json the
event property ($.name, dotted for nested objects) that sets each one.
Statistics are kept per user and service label in
_trophy/<user id>_<service label>.stats.txt next to the unlocks.
After each event that sets a statistic, every trophy's unlockCondition
(ge, gt, le or lt against its targetValue) is checked on the values that
event set and the stored values of the other statistics, so a later
event setting the same statistic cannot hide what this one reached.
Reached trophies unlock through the same path as _UnlockTrophy.
Progressive trophies report their statistic as progress, and
APS5_TRACE_UDS prints each statistic an event sets. A missing udsNN.ucp
is ignored and a malformed one is logged once.
Events and property objects keep their typed properties instead of
dropping them. PostEvent records an _UnlockTrophy event with an int32
_trophy_id for the user and service label of its context, logs why a
trophy event is ignored, and prints every posted event with
APS5_TRACE_UDS.
Unlocks are kept in _trophy/<user id>_<service label>.txt in the title's
run directory, one "<trophy id> <RTC tick>" line per trophy, and each new
unlock logs "[trophy] <title id> unlocked trophy <id> for user <id>".
GetTrophyInfo, GetTrophyInfoArray, GetGroupInfo, GetGroupInfoArray and
GetGameInfo report the recorded unlocks, their times, the unlocked
counts by grade and the progress percentage. An unlock also unlocks the
platinum the trophy names once every trophy naming that platinum is
unlocked. Unlock listeners let a later notification subscribe. Removing a
listener waits for its running callback, and a listener removed during a
dispatch is not called by it.
The title's sce_sys/trophy2/trophyNN.ucp is read on first use with a
bounds-checked .ucp reader and a small JSON reader. tropconf.json gives
each trophy's grade, group, hidden and reward flags, platinum and
unlock condition, and tropmeta_<tag>.json its texts in the console
language, then en-US, then the file's default language. Contexts
remember their user and service label, so the queries find the file.
The trophy, group and game queries report these names, grades, groups
and counts, list every trophy and group, and give progressive trophies
their target. Unlocks are not recorded yet, so every trophy reports as
locked and nothing is written to the host. A missing file is ignored
and a malformed one is logged once; either way the queries keep their
placeholders. A JSON document that nests deeper than 64 levels or holds
more than 1048576 values counts as malformed, which keeps the parsed
values to about 140 MiB.
Main keeps an import settle's objects per device in ImportSettle, next
to where this branch adds ImportRefusal. Keep both. Main's fallback for
a refused host-write alias returns after this branch's failure records.
The unused-slot case had no colors and one blend, which the pipeline
refuses before it validates shaders. A target at slot 1 with slot 0
unused is the gap a draw can have.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The fragment output check demanded a float4 (or uint4 for integer and
packed targets) at every location below the draw's MRT count, at least
one. A location with no color target, a draw without color targets or a
gap between written slots, has no attachment there, and Vulkan discards
the output whatever its type. Crimson Desert's first draw binds no color
target and exports a uint4 at location 0, and stopped there.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A progress value is {int32 type, 4 reserved bytes, uint64 value}, as
KytyPS5 lays it out, which makes the trophy details 1312 bytes and the
trophy data 32 bytes. Size checks for every query structure keep the
layout fixed, and the technical debt names its single public source.
SystemServiceConsoleLanguage in SystemService.hpp reads ANYPS5_LANGUAGE
with the same default, range and error as sceSystemServiceParamGetInt,
which now calls it. Other libraries can pick the console language
without linking libSceSystemService.
#6243 now expects a per-lane table image store to be supported, since
main's #6402 runs image stores through image tables. This branch
carries #6243's test, so it takes the same expectation.
On Linux, APS5_TRACE_WAITS and APS5_TRACE_USLEEP used an image base and a thread
id of 0, so "exe+0x..." printed absolute addresses and every site reported
"tid 0". The call site is now resolved with dladdr1(RTLD_DL_LINKMAP) and printed
relative to its module's load address ("exe+..." for the executable, the module
file name otherwise), matching guest offsets in a disassembly; the thread id is
gettid(). Descriptions are cached per call site, so polling titles do not pay a
dladdr1 per wait. Windows output is unchanged.
The tests build their working directories with tempfile.TemporaryDirectory()
and std::filesystem::temp_directory_path(), so every executable they generate
lands in the user's system temp directory. On Windows, Microsoft Defender
quarantines those generated binaries as Trojan:Win32/Wacatac.H!ml, and the
nine tests that run one fail with WinError 5 (access denied) on a clean
checkout: autorun_exit_status, optional_plt, empty_tls, tls_function_coverage,
executable_segments, windows_icon, guest_needed_modules,
guest_module_directories and windows_import_modules.
Point TMPDIR, TEMP and TMP at a build-local directory for every test
registered here, so generated executables stay inside the build tree. Test
artifacts are then removed with the build directory, and a contributor who
needs an antivirus exclusion can scope it to the build tree instead of the
system temp directory.
conventions_function_bodies_shard_4 timed out at 30 s on Windows in five runs on 17 Oct (main pushes 39076267931 and 39061030084, dispatches 39074123651 and 39044676931, relinker 39072918668). Shard 4 was the slowest shard in every Windows log (10-16 s on a quiet runner, 5-11 s for the others).
One case dominated it: test_comment_in_a_shader_is_reported built six fixture git repositories, 36 of the suite's 390 processes. Process start-up is what costs on Windows. #7314 doubled the shard count, but modulo sharding keeps that case whole, so its shard (now shard 10) still cost 0.63 s of the suite's 3.0 s locally against 0.76 s before.
The case now writes all six shaders into one fixture and runs the checker once (6 processes), still asserting each path. The slowest of the 12 shards drops to 0.39 s locally (43 processes), with no timeout raised.
Main's #6402 stores to image tables: a table with a written entry takes
the fresh path, as it does on main, and so does one without a sampled
table binding, whose entries the fresh path rejects. The shared place
step uses the binding as a pointer. Main's #6614 frees descriptor sets
from shared pools one by one, while this branch keeps only dedicated
pools, so both stay. #7101's bulk key read counts its unmapped-page
skips as unmapped, and its key words sit next to this branch's write
stamp and compared runs.
Dean