The word reaches the data-only mask through userDataRegisterReads (the stage
layout's user-data registers). Dropping those bits from the mask makes the
draw reuse the old stage (it writes 127, expected 191) and verify mode abort.
LLM assistance used.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A data-only hit (#494) reused a cached stage only when every user data word
of the stage was unchanged. A stage records the user words it depends on when
it is captured: the words its SRT walk reads as values or bases
(SrtUserDataUse), the words its binding layout forwards into shader data
(userDataRegisterReads), and the vertex fetch table words of its vertex stage
info. A cached stage is now a candidate when only words outside that set
changed; its captured memory is validated as before.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
MCOMMIT becomes STC plus NOP padding and RDPRU goes through a
runtime-checked stub returning the host TSC for MPERF/APERF while
other selectors yield zero, with converter and execution tests.
docs/dev/debt/relinker.md records the new lowering and its unverified
assumptions.
APS5_COARSE_TIMED_WAITS is read under _WIN32 only; on Linux the condition
variable path always ends a sleep when a completion arrives.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
PS4 titles import both and do not load without them; every call throws and
the debt files list them.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
agc_driver_resource_cache_replaced_texture clears the sampled cache
under a template that takes the fast proof and requires the next proof
to fail. agc_driver_resource_cache_lookup_changed checks that a proof
past the audit interval takes the full walk and keeps an unchanged
template, then host-imports the surface (no lookup runs) and requires
the template to stop proving current. Both run over a write-watched
surface so the template takes the fast proof.
The fast proof of a cached ShaderResources proves each sampled texture
from write stamps, keys and the pending registry. It never asks whether
the sampled lookup would still return the bound Texture, so a template
whose surface memory is unchanged keeps binding an old object:
- After the sampled cache drops or replaces the entry (#6492 marks the
texture released), Find keeps serving the template until the next
Insert, and its fast proof passes.
- When the lookup's answer changes without any lookup running (a
snapshot whose surface became host-imported is viewed through its
storage image from then on), nothing replaces the entry and the
template keeps the snapshot until it leaves the resource cache.
The fast proof now fails with 'replaced texture' when a bound texture
is released. A template whose last full walk is older than
APS5_TEMPLATE_AUDIT_MS (default 1000, 0 turns it off) takes the full
walk instead of the fast proof ('audit'), so a different lookup answer
fails the proof and the template is rebuilt.
The entry stated that image atomics on 32_SINT and 32_FLOAT act on the
raw dword with no measurement behind it. On gfx1100 the 13 integer image
atomics return and store the same raw-dword results on 32_UINT, 32_SINT
and 32_FLOAT, while 16_UINT and 8_8_8_8_UINT controls differ. gfx1100
has no float image atomics, so that part stays open.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Prevent decimal-comma LC_NUMERIC from emitting invalid JSON and silently truncating decimal-point input. Keep 17-digit output, signed underflow and overflow rejection without changing the process locale.
AI-assisted: yes (OpenAI Codex). Windows WinLibs GCC15.2 Release: original shared Json2 fails the final locale regression; fixed guest_json and guest_json_numeric_locale pass (2/2, no skips). Final conventions pass. Only relevant runtime/test targets built; full build/full suite and game/console/Linux/macOS behavior unverified. No performance or coverage increase claimed.
sceNpGetAccountCountry, sceNpGetAccountDateOfBirth, sceNpGetAccountLanguage
and sceNpGetParentalControlInfo take an online id that matches no user,
sceNpGetUserIdByOnlineId and sceNpCheckPlus report NP as signed out, and
sceNpNotifyPlusFeature is declared and throws.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Track the handles sceMoveOpen issues. A handle that was never opened,
is closed or was opened before sceMoveTerm now throws instead of
succeeding, since the console's INVALID_HANDLE result and its order
against INVALID_ARG are not known. List the no-controller path under
Silent stubs, as libSceHmd and libSceBluetoothHid do, and check the
handle values in the test.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Arbzg9c2DA4DEFKVSvRgYv
The Linux entry stub now passes each traced import through libc.prx's
TraceImport_nid_no_patch before it calls the guest entry, and stores the
returned thunk in the import's slot. The same imports are traced as on
Windows: functions bound through a PLT slot or a GOT slot, without an
addend, and not weak. TraceImport_nid_no_patch is bound eagerly through
an added GLOB_DAT relocation, so a libc.prx without it fails at load.
Each slot's page is made writable with mprotect first, so slots inside a
PT_GNU_RELRO range can be replaced.
--trace-imports now requires libc.prx among the libraries on Linux too,
and is refused with --macos.
A thread that lost the race to write the report passed SIGABRT on straight away, and the default action ended the process while the first thread was still writing. It now waits for the report to finish, polling with nanosleep for at most 2 seconds, before it chains. guest_call_trace_abort_race aborts from two threads at once and checks the whole report; it failed 35 times in 50 without the wait. Reported by jotitadev.
The SIGABRT handler used stdio and dladdr, so an abort raised while another thread held the stderr lock or the loader lock hung instead of ending. The report is now formatted into a stack buffer and written with write(2). Callers are named from a module list taken with dl_iterate_phdr when imports are traced, and import counts are read without a lock from a fixed record array. The abort test now holds the stderr lock in another thread, which timed out before this change.
Build the call tracer for Linux as well as Windows. The thunk area is a
memfd mapped twice, writable and executable, and threads are named by
gettid. Linux has no crash report, so the first traced import installs a
SIGABRT handler that prints the last calls and the most called imports,
then passes the signal on to the previous handler.
guest_call_trace and guest_call_trace_abort now run on Linux too.
The 73 sceNpCppWebApiUnknown<nn> stubs carried an APS5_EXPORT alias because their
names were placeholders. aerolib names all of them, and each name recomputes to
the NID a title imports, so the alias is dropped and the loader derives the NID
from the real symbol name, as the rest of the library does. Behaviour is
unchanged: every one of them throws.
The stubs are ordered by demangled name instead of by NID, so the Common template
instantiations and the V1/V5 accessors sit together, as the rest of the file
does. No symbol changes: the built .prx exports the same 359 NIDs as main.
The Windows pool closed each thread handle at creation, so Close() joined
no thread and liveWorkers was never raised before each worker lowered it.
Keep the handle and thread id, wait for the handle in Close(), and raise
liveWorkers like the POSIX path. The stack test now checks the live count
before and after CloseRegistrationPreparation().
The entry before it still says sceKernelAddReadEvent and
sceKernelAddWriteEvent take only guest sockets. The pipe entry now says it
is the exception, without editing that entry, which an open change to the
shared-memory entry next to it would conflict with.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
The registration preparation and library prebuild pools ask for 64 MiB stacks through
pthread_attr_setstacksize. With winpthreads that size reaches _beginthreadex as the initial
commit, so every pool thread commits its whole 64 MiB up front: 7 threads and 448 MiB on a
12-thread job, 9 threads and 576 MiB on the 16-thread i7-11800H, before any shader is prepared.
On Windows the pool threads are now created with STACK_SIZE_PARAM_IS_A_RESERVATION, which
keeps the 64 MiB reservation and commits pages as the stack grows. The pool code is host code
built with stack probes, so on-demand growth is safe there. Guest threads are unchanged.
agc_driver_registration_prepare_stack measures a pool thread's stack from inside a task: at
least 64 MiB reserved everywhere, and at most 8 MiB committed on Windows. On main it fails with
"registration preparation thread commits 64 MiB of its 64 MiB stack up front".
The overlap audit against origin/main and the open pull requests found 39 of
the 193 declared exports that must not be declared here:
- 14 that origin/main already exports. The nid_patcher computes the NID of a
symbol from its name, and a name ending in _nid_postfix exports the NID of
the base name, so scePadOpen, sceNetInit and the rest are already there.
- 23 that an open pull request provides: #6832, #6847, #6852, #7846, #7881,
#7889, #7890, #6977 and #7690.
- 2 whose technical-debt file an open pull request creates: #7463 and #8094.
The scope is now 30 libraries and 154 exports.
Guest SCE_BREAK assertions and debug breaks terminate in int 0x41 (0xcd 0x41), which Windows reports as a read access violation of ~0ull. Name this trap in the crash report alongside int 0x45 (abort).
A recorded comp-to-single clear stayed valid for its surface address,
size and sample count until another recognized clear replaced it, so a
surface freed and reallocated on the same memory could import the old
color for its 0x10 DCC keys. The record now also carries the image's
pipe-bank XOR, which RADV assigns per image, and is dropped when an
unmap or allocation removal covers the surface's base.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
RADV's comp-to-single shader stores user data 5 to 8 as the four dwords of
sample 0's texel: v0 from s5, v1 and v2 from s[6:7] through a 64-bit shift by
zero, and v3 from s8. RADV binds the image through a uint view of the texel's
size, up to R32G32B32A32_UINT, so a 64-bit texel takes the first two dwords
and a 128-bit texel all four. 0x10 keys of RGBA16F and RGBA32F targets now
import as a clear to the recorded color instead of being refused.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Adrià