RISC-V memcpy routine uses misaligned loads and stores

I tried this code:

// Using the riscv64gc-unknown-none-elf target.
// Should be reproduced by:
#[inline(never)]
fn do_copy(len: usize) {
  // len is 16.
  core::ptr::copy_nonoverlapping(0x237c0 as *const u8, 0x43adc as *mut u8, len);
}

I expected to see this happen: Byte copy should succeed without missing bytes. A byte copy loop using write_volatile works fine.

Instead, this happened:

Debugging output over UART (with annotations)

stream read 16 [87, 79, 79, 70, 66, 79, 79, 84, 16, 0, 0, 0, 3, 0, 0, 0] (.data)
ptrs: 0x237c0 (.data) 0x43adc (stack, 4-byte aligned)
head (stack): [87, 79, 79, 70, 0, 0, 0, 0, 16, 0, 0, 0, 3, 0, 0, 0]

Meta

rustc --version --verbose:

rustc 1.91.0-nightly (2e2642e64 2025-08-16)
binary: rustc
commit-hash: 2e2642e641a941f0a1400c7827fd89aa86fef8f4
commit-date: 2025-08-16
host: x86_64-unknown-linux-gnu
release: 1.91.0-nightly
LLVM version: 21.1.0

Start by reproducing the provided copy_nonoverlapping example for riscv64gc-unknown-none-elf and inspect the generated RISC-V behavior for the misaligned destination. Compare it with the working write_volatile byte loop; done means the 16-byte copy preserves every byte at the shown addresses.