On the AMD Windows driver (Radeon RDNA2 iGPU, 32.0.21030.2001)
UnpackHalf2x16 of a byte-selected half converts the unmasked source
register: an SDWA BYTE_0 source of 0xfff0fff0 reads as f16 0xfff0
instead of 0x00f0, so agc_driver_sdwa_float_selectors and
agc_driver_sdwa_ldexp_f16 fail there. EmitF16BitsToF32 now decodes
sign, exponent and mantissa with integer operations, scaling
subnormals by 2^-24, which is exact for every half and leaves the
driver nothing to fold into the conversion.
BnMAMrsfVWo is _ZSt19_Throw_bad_weak_ptrv. Replace the unknown stub
with a throw of a guest std::bad_weak_ptr, and add its type info,
vtable, destructors and what().
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Treat equal BASE_ARRAY and DEPTH as a six-face cube. Measured with image_load, image_sample and image_get_resinfo on RX 6600 gfx1032 using base slices 0, 5 and 6. Partial cube ranges still throw because filtered sampling differs from load bounds.
Graphics tests and GPU snapshot/storage sampling tests pass for all six faces at each tested base.
Assisted-by: OpenAI Codex
FreeBSD's setvbuf takes a zero size with _IOFBF or _IOLBF as "pick the
buffer size". The UCRT rejects sizes below 2 for those modes through
its invalid parameter handler, so the call failed on Windows and a test
process was terminated. A zero size now drops the given buffer and
asks the host for one of BUFSIZ bytes.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Gathering every iovec into one host buffer threw for a total above
INT_MAX, which a console readv/writev accepts, and copied the whole
transfer twice. Each iovec is now read or written in place, in chunks
of at most INT_MAX bytes. A readv on a pipe returns after one read so
it does not block once the available bytes are consumed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01UTjpbMX7BUFMZZaxJE98zV
sceKernelReadv, sceKernelWritev, sceKernelPreadv and sceKernelPwritev threw on Windows. They now gather or scatter through one host buffer and issue a single read or write, with the same argument checks and error codes as the Linux path. Positioned transfers restore the file offset afterwards, as pread and pwrite do on POSIX, and report ESPIPE for handles that are not disk files. guest_file_vector now runs on Windows as well.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01UTjpbMX7BUFMZZaxJE98zV
COPY_DATA rejected source select 9 (GPU clock count) and any cache
policy bits. PPSA16064 (Minecraft Dungeons II, Unreal Engine)
calibrates its GPU timer at start-up with a submission of five
COPY_DATA packets, control 0x06016209: GPU clock source, memory
destination, 64-bit count, source and destination cache policy 3. The
submission was rejected and the title aborted.
The clock source now writes the same counter as a RELEASE_MEM
timestamp (the host steady clock in 10 ns ticks), read when the packet
runs, through the direct execute path and through ResolveStore for
stores recorded on the GPU. A 32-bit count writes its low half. The
source and destination cache policy fields are accepted and ignored,
as for WRITE_DATA and DMA_DATA. The reference clock (source select 10)
and the other unimplemented sources still throw.
agc_driver_pm4 runs the title's packet twice and checks a nonzero,
non-decreasing counter and an 8-byte resolved store, and that source
select 10 is still rejected.
MinGW's thread_local goes through emutls and winpthreads' pthread_getspecific, which holds a per-thread spinlock. A raise that redirected a thread holding it ran the handler on that same thread, and a handler reading any thread_local, scePthreadSelf for one, spun on the lock forever. A raise that finds the target inside libwinpthread-1.dll now resumes it and suspends it again, as one inside NtContinue does.
guest_raise_exception's new phase raises at a thread that calls scePthreadSelf in a loop and checks that the handler sees its own thread. It hung in every run without the retry. TechnicalDebt records that a handler can still reenter other host code the thread was interrupted in.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The guest's wchar_t is 16 bits, but wcslen, wcscmp, wcsncmp, wcscpy, wcsncpy, wcschr, wcsrchr, wcsstr, wcspbrk, wcsspn and the wmem functions forwarded to the host's wchar_t functions, which use 32-bit units on Linux. Implement them on char16_t. The wide conversions that still use the host's wchar_t are listed in TechnicalDebt.
NVIDIA (595.91.07) refuses VK_EXT_external_memory_host imports of
read-only or executable mappings with VK_ERROR_OUT_OF_DEVICE_MEMORY and
then fails every later vkAllocateMemory on the device. Check the range
is mapped read-write before importing on Linux, as the Windows path
already does for read-only memory, and copy it otherwise.
Refs #1022
Exclude COLOR_INVALID attachments from the effective write mask while preserving later MRT export indices and depth-only rendering. Validated with the graphics tests and Skyrim menu draws.
AI-assisted implementation and testing by OpenAI Codex.
guest_formatting called sprintf with four variadic arguments and
libc_printf with one, all in registers, and did not call sscanf, so only
snprintf had a check that fails with the old Linux register rebuild. sprintf
now formats six integers, sscanf converts eight fields and libc_printf prints
seven integers; on the old code the first gives "1 2 3 4 0 0" and the other
two crash.
Checks suggested by nauroth in #1435.
The host's EHOSTUNREACH was reported as ENETUNREACH (51); FreeBSD's
EHOSTUNREACH is 65, which libkernel's sockets already report. EADDRNOTAVAIL,
ENOBUFS, EISCONN, ENOTCONN and EACCES had no translation and became EIO;
they now map to 49, 55, 56, 57 and 13, as in libkernel/Socket/src/Udp.cpp.
guest_sce_net connects the connected client again and expects EISCONN
(0x80410138); before this change it received EIO.
A CMASK holds 4 KiB per 1024x512-pixel metablock, not one nibble per 8x8
tile of the target: 1920x1080 spans 0x6000 bytes, as titles clear it,
not 16200. The swizzled nibbles of the lower-right tiles lie past the
packed size, so the fast-clear check missed them and the expanded store
left them cleared.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A non-blocking sceNetConnect that cannot complete at once must fail with
EINPROGRESS (36, 0x80410124), as FreeBSD's connect does, and a repeated call
while it is pending with EALREADY (37). On Windows, Winsock reports the
first case as WSAEWOULDBLOCK, which was translated to EAGAIN (0x80410123); on
Linux, EINPROGRESS and EALREADY had no translation and became EIO
(0x80410105). Code that waits for the connection only when connect reports
EINPROGRESS treated both as a failed connection.
guest_sce_net connects a non-blocking socket to the loopback listener and
accepts 0 or EINPROGRESS; before this change it received EAGAIN on Windows.
sceKernelGetModuleInfoForUnwind and sceKernelGetModuleInfoFromAddr each
parsed .ehmeta and the eh_frame_hdr it names. Both now read them through
Module/EhFrame.hpp, so a bad table fails the same way in both: an
eh_frame_hdr version other than 1, an indirect or non-absptr/pcrel
encoding, a 64-bit record length or anything outside the image throws.
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01JkiyF6p7oDAaTtVe8ci5zP
tools/nid_names.py resolves 7gxh+5QubhY to sceAppContentAddcontEnqueueDownload
(compute_nid of the name matches). The stub takes that name instead of
sceAppContentUnknown00 with an APS5_EXPORT alias, so its NID is unchanged and
NotImplemented reports the real function.
Detach the device cache buffer attached by sceFontAttachDeviceCacheBuffer. The signature is taken from the IDA type database of Orbital: it takes the library, a pointer to receive the buffer and a pointer to receive its size. When no cache buffer is attached, it reports NOT_ATTACHED_CACHE_BUFFER. If the library allocated the buffer itself and no buffer pointer is passed, the memory is freed. The owned bit is cleared and device_cache_buf is reset to null.
tests/Mesh.cpp called setenv and std::aligned_alloc, which MinGW does not
provide, so agc_driver_mesh_tests did not compile on Windows. The 64 KiB
indirect record block now comes from aligned operator new, and the
environment is set with _putenv_s on Windows, as ShaderDiskCache.cpp does.
The setenv also named APS5_NO_SHADER_DISK_CACHE, which nothing reads; the
shader cache is turned off by ANYPS5_NO_SHADER_CACHE.
Suryansh Prajapati