[codethinki]

partial_load loses an element when the remaining bytes are at the top of a __memcpy_chunks branch: the second copy is only _Chunk bytes wide, so the bytes in between are never copied.

• 1-byte elements, 4 remaining: bytes 0, 1 and 3 are copied
• 2-byte elements, 4 remaining: bytes 0-3 and 6-7 are copied

std::array<signed char, 4> mem {1, 2, 3, 4};
std::size_t n = make_value_unknown(4uz);
partial_load<vec<signed char, 16>>(mem.begin(), n); // [1, 2, 0, 4, 0, ...]

With vec<char, 64> this hits every tail of 4, 20, 36 and 52 elements. The same code ships with libstdc++ in GCC 16.2.

Every branch now copies the same width from the front and the back, which covers (k, 2k] bytes for any chunk size.

tests/loads.cpp: partial_load_every_size loads every n from 0 to V::size with an unknown n (a known n takes the single memcpy path). It fails without the fix and passes with it for signed char, short, int, long long, float and double at widths 4, 8, 16, 20, 32, 36 and 64 on x86-64 and x86-64-v3 with GCC 16.2.1.

[codethinki]

Did not write the fixes, claude opus 5.5 did. However I code reviewed and approve as far as I can understand it.

If you have any questions or comments I'll reply / fix it as soon as I can.