A bundled module's import keeps its name only when the executable
exports it as a visible global or weak symbol, as the dynamic loader
can bind it. Test local, internal and hidden executable symbols of the
same name next to global, weak and protected ones.
Evaluate the existing accuracy correction only inside its established input range. Keep the native log2 result and all correction thresholds and arithmetic unchanged, and cover inputs adjacent to both range boundaries in the existing GPU test.
The empty-indirect skip and the last-drawable search filtered records by
firstVertexOrIndex < indexCount, which is 0 when the queue never set an
index limit, so a second submission of the same draw was skipped as empty.
A queue that has received neither INDEX_BUFFER_SIZE nor
INDEX_ATTRIBUTES_INDIRECT since its last reset has no index limit.
Such a draw takes the CPU record path and reads each record's index
range as written, refusing a range that overflows or is not readable.
An explicit INDEX_BUFFER_SIZE of 0 still throws.
The Windows and macOS patchers reject an entry point outside executable
PT_LOAD memory and a relocation offset or DT_PLTGOT outside PT_LOAD memory
through WindowsLoadImage::GetRva. The Linux patcher copied these values
unchecked and reported success; with a bad entry or relocation offset the
output crashed with SIGSEGV at load.
LinuxElfPatcher::Patch now requires e_entry in an executable load and every
non-NONE RELA and JMPREL r_offset and DT_PLTGOT to fit inside one PT_LOAD's
MemorySize, in the fixed-base and --lazy-binding paths. A relocation needs
the bytes ld.so writes: 4 for PC32 and 32, the output symbol's st_size (at
least 1) for COPY, and 8 otherwise. DT_PLTGOT needs 8 bytes with eager
binding and 24 with --lazy-binding, where ld.so writes GOT[1] and GOT[2].
Two test fixtures that set e_entry to a file offset outside their only
PT_LOAD now use the matching virtual address.
The PT_DYNAMIC loop used the first header and ignored any later one, so
the Linux target converted an input with two PT_DYNAMIC headers. A
second PT_DYNAMIC now fails with "Multiple PT_DYNAMIC segments" at its
file offset on every target.
Relink() rejected only ET_REL, ET_NONE and ET_CORE, so the Linux target
converted any other e_type, such as 5, 0xfe00 or 0xffff. It now rejects
every type other than ET_EXEC, ET_DYN, 0xfe10 and 0xfe18 with
"Unsupported ELF object type 0x..", after the three named rejections.
drand48 and erand48 returned the whole 48-bit state over 2^48, as FreeBSD
does. The libc.prx bundled with eleven titles returns its upper 32 bits
over 2^32, so the low 16 bits never reach the result; both functions now
do the same. The generator and the other rand48 functions are unchanged.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
v_rcp_f32 and v_rcp_iflag_f32 lower to the driver's 1/x, whose low bits
differ from the console's on NVIDIA and Intel. With APS5_EXACT_RCP set,
normal inputs with normal results take the RDNA reciprocal datapath
instead: a cubic over 32 segments of the mantissa in 32-bit integer
arithmetic, the linear term rounded to nearest even, and a 2,761-entry
table of one-ULP corrections, looked up only when the dropped bits fall
next to the rounding cut (6.25 % of inputs). This reproduces all 2^23
mantissas of an RDNA2 (gfx1031) sweep. Other inputs keep 1/x.
It is off by default: the tables add about 13k SPIR-V words to each
shader that uses it.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_017EjjFU1FjMR11oWDYs8UMo
Brings in main with #9204. The measured CP932 and CP51932 encoding rules (JIS form aliases, U+FEFF, the C1 range and U+000E) apply to those profiles only; the strict JIS X 0208 profiles keep #9204's tables.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Keep #9204's strict JIS X 0208 profiles with their SS2 and SS3 choices inside the new helpers: decodeEucJp sends the measured CP51932 profile to decodeEucJpScalar and the strict profiles to #9204's rules, and only the X0208Ss2Ss3 profile encodes JIS X 0212. The EUC-KR, UHC and Big5 UTF-8 forms decode with decodeScalarUtf8.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Zenodeon