Follows mlockall (#4386) for XPSemu (PPSA97358, QEMU os-posix.c), which
imports munlockall (NpLBpgVV7PU) from libkernel.prx. Predicted stop:
FAIL: unresolved ELF import NpLBpgVV7PU from libkernel.prx
The NID is computed with tools/nid_names.py and matches aerolib.
Follows FreeBSD 11.4 sys/vm/vm_mmap.c sys_munlockall
(8c639945a44588f6cfdddfb057bb5aad88735da0): checks PRIV_VM_MUNLOCK,
clears wire state and returns 0. To match, munlockall unlocks ranges
locked by mlock (through VirtualUnlock on Windows, tolerating
ERROR_NOT_LOCKED, and munlockall on Linux).
In rect-list tessellation evaluation, gl_TessCoord coordinates (x, y) generated by hardware tessellators are floating-point values in [0.0, 1.0]. Direct integer truncation via OpConvertFToS causes coordinates that evaluate slightly below 1.0 (e.g. 0.999999f) due to hardware fixed-point representation to truncate to 0 instead of 1, collapsing rectangle vertices along the diagonal.
Adding +0.5f rounding prior to OpConvertFToS ensures correct mapping to {0, 1} quad vertex indices.
Implement __modti3, the signed 128-bit integer modulo runtime helper complementing __divti3, __udivti3 and __umodti3. Add unit test coverage in guest_crt_tests and document zero-divisor exception behavior in technical debt.
A tessellation draw whose domain, partitioning or output topology the
reference path does not support was refused with only VGT_SHADER_STAGES_EN,
so a log did not show which configuration was refused. The refusal now
prints VGT_TF_PARAM with TYPE, PARTITIONING and TOPOLOGY, and
VGT_LS_HS_CONFIG with NUM_PATCHES, HS_NUM_INPUT_CP and HS_NUM_OUTPUT_CP.
The fields come before the explanation, so the reason that the
APS5_CAPTURE_TRACE log cuts at 256 characters still holds all of them.
Field and enumerator names are those of Mesa's
src/amd/registers/gfx103.json (enums VGT_TESS_TYPE, VGT_TESS_PARTITION and
VGT_TESS_TOPOLOGY). A value with no name there is printed as a number only.
The accepted configurations are unchanged.
The new UnsupportedTessellationFieldTests drives DecodeState into the
refusal with power-of-two triangle and quad patches, an isoline domain,
line and point output and unnamed TYPE, PARTITIONING and TOPOLOGY values,
and checks the decoded fields of both registers, including the order of
the input and output control-point counts.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
macOS has no /proc/self/maps or MAP_FIXED_NOREPLACE, so the arena never came up and every guest allocation threw. Enumerate host regions with mach_vm_region and reserve free chunks with a fixed mach_vm_map, which fails instead of replacing. This skips dyld, the Rosetta runtime, and ranges the kernel never lets a process map (0x10_0000_0000-0x70_0000_0000). The POSIX commit/protection paths now build on macOS unchanged. Adds standalone guest_arena tests (x86_64 under Rosetta and arm64).
A dump is fully installed, so scePlayGoGetSupportedOptionalChunk and
scePlayGoGetOptionalChunk report every optional chunk (language pack) like
scePlayGoGetLanguageMask reports every language. They returned an empty
bitmask, so PPSA01748 hid every text language its playgo-scenario.json
lists as a chunk language, English included.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
guest_video_out_cursor calls each of the four cursor functions with a handle,
a cursor index, an image and a position, and requires the "not implemented"
error. Without the exports it does not link.
sceVideoOutCursorEnable, sceVideoOutCursorDisable,
sceVideoOutCursorSetImageAddress and sceVideoOutCursorSetPosition are imported
by Prospero Win's (PPSA99995) Vulkan module, so on Windows, where imports are
bound at load, the title cannot load without them. They report that they are
not implemented when called.
sceNpRegisterStateCallback and sceNpRegisterStateCallbackA stored nothing,
the unregister calls always succeeded and sceNpCheckCallback did nothing,
so a title that waits for its NP state callback never learned the user's
state. #1277 reports that Minecraft Preview (PPSA19639) needs the callback
(stop 2).
sceNpRegisterStateCallbackA now keeps up to 8 callbacks and returns the
slot plus 1. A null callback returns INVALID_ARGUMENT, a callback that is
already registered CALLBACK_ALREADY_REGISTERED and a ninth one
CALLBACK_MAX (0x8055001D). sceNpUnregisterStateCallbackA returns
INVALID_ARGUMENT for an id outside 1..8 and CALLBACK_NOT_REGISTERED for an
empty slot. The legacy sceNpRegisterStateCallback keeps one callback with
the same errors. The rules follow shadPS4
src/core/libraries/np/np_manager.cpp @15abf68c9b52 (RegisterStateCallbackA
and UnregisterStateCallbackAById L857-L901, the legacy pair L981-L1005,
ORBIS_NP_STATE_CALLBACK_MAX L36) and np_error.h L38.
Each registration is marked pending. sceNpCheckCallback takes the pending
registrations under the lock and calls them after releasing it, on the
calling thread, with SIGNED_OUT for the initial user 0x10000000. For the A
form this follows KytyPS5 src/libs/network.cpp @9e41d8320a27
(NpRegisterStateCallbackA L4021-L4034 marks the registration pending,
NpCheckCallback L3959-L3974 delivers it once). KytyPS5 never calls the
legacy callback (NpRegisterStateCallback and NpUnregisterStateCallback,
L4009-L4019, do nothing). The legacy callback is called as
(user, SIGNED_OUT, nullptr, userdata), as in shadPS4, which passes an NpId
only for SIGNED_IN (L823, L938-L940).
shadPS4 queues SIGNED_OUT from the user service login event instead
(DispatchPendingNpStateCallbacks L903-L962), and the same event calls the
reachability callback once with UNAVAILABLE (L922-L931, L959-L961). The
reachability callback is still not called here. These differences and the
unmeasured console behaviour are recorded in
docs/dev/debt/libSceNpManager.md.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
EVENT_WRITE (PM4 opcode 0x46) with event type SAMPLE_STREAMOUTSTATS1-3
(0x01-0x03) or SAMPLE_STREAMOUTSTATS (0x20) failed submission validation
with "EVENT_WRITE event type 1 is not implemented". Minecraft Preview
(PPSA19639, discussion #1277, stop 9) submits a 2-dword type-1 packet
while entering a world, and the GPU thread aborted intermittently.
The four values are VGT_EVENT_TYPE entries in Linux
drivers/gpu/drm/amd/include/navi10_enum.h L14155-14188 at
0fcb05b72344f40806cbe810de8872f1698bf053.
drivers/gpu/drm/amd/amdgpu/nvd.h L326-347 at the same commit documents
these events only with the sampling index (3, or
EVENT_INDEX__SAMPLE_STREAMOUTSTATS* 8-11 on gfx10), followed by
ADDRESS_LO/HI. The 2-dword index-0 packet has no destination dwords, and
what the CP does with it is neither documented nor measured. AnyPS5 does
not emulate the VGT stream-out counters, so there is nothing to sample.
Packet execution already skips EVENT_WRITE.
These event types are now accepted on the graphics queue as 2-dword
packets with event index 0. The sampling indices 3 and 8-11 throw that
stream-out statistics sampling to memory is not implemented, before the
size and reserved-bit checks. Other indices, other sizes and the compute
queue still throw.
sceAgcDcbEventWrite ignores the address for every event type except 0x38
and 0x39, so a call with one of these event types and an address now
builds a packet that validates and writes nothing. This is recorded in
docs/dev/debt/libSceAgcDriver.md with the unmeasured hardware behaviour.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The first cp1252 run timed out at 10 s on a loaded Windows runner (Build
runs 38642532107 and 38570483873); ctest still bounds the test at 120 s.
Co-authored-by: Claude <[email protected]>
Dean