[PATCH 0/3] RISC-V: memcpy(): align dest when misaligned access is

[Kito Cheng]

LGTM, thanks for the patch :)

On Sun, Apr 27, 2025 at 7:59 PM <ma.mandourr@gmail.com> wrote:
>
> From: Mahmoud Abumandour <ma.mandourr@gmail.com>
>
> Add a code path for when source and dest are differently aligned.
>
> If misaligned access is slow or prohibited (i.e. `__riscv_misaligned_slow` or
> `__riscv_misaligned_avoid` is defined), and the alignments of the source and destination are
> different, we align the destination to do XLEN stores. This uses only one
> aligned store for every four (or eight for XLEN == 64) bytes of data.
>
> This only affects the case where source and destination are differently aligned.
> Previously, this case was handled by copying one byte at a time. The new
> approach significantly reduces the number of executed (dynamic) instructions
> (30% weighted average for `-mstrict-align`, while not affecting
> `-mno-strict-align`, which is expected).
>
> A comparison between the implementations with raw numbers is posted here:
>     https://docs.google.com/spreadsheets/d/1KDnfhYVFQYIXb23LjsgWFyWna4A_9mcRZRLm_u0zELE/edit?usp=sharing
>
> While at it, the function is generally improved to use `uintxlen_t` and
> eliminate gotos and macros.
>
> Mahmoud Abumandour (3):
>   RISC-V: memcpy() Use uintxlen_t for xlen-sized copy
>   RISC-V: memcpy() Use inline functions instead of macros and gotos
>   RISC-V: memcpy() align dest when misaligned access is prohibited
>
>  newlib/libc/machine/riscv/memcpy.c | 163 +++++++++++++++++++----------
>  1 file changed, 109 insertions(+), 54 deletions(-)
>
> --
> 2.43.0
>