On Fri, Jul 2, 2021 at 4:37 PM Ben Dooks [off-list ref] wrote:

On 02/07/2021 13:31, Matteo Croce wrote:

quoted

From: Matteo Croce <redacted>

Rewrite the generic memcpy() to copy a word at time, without generating
unaligned accesses.

The procedure is made of three steps:
First copy data one byte at time until the destination buffer is aligned
to a long boundary.
Then copy the data one long at time shifting the current and the next long
to compose a long at every cycle.
Finally, copy the remainder one byte at time.

This is the improvement on RISC-V:

original aligned:      75 Mb/s
original unaligned:    75 Mb/s
new aligned:          114 Mb/s
new unaligned:                107 Mb/s

and this the binary size increase according to bloat-o-meter:

Function     old     new   delta
memcpy        36     324    +288


Signed-off-by: Matteo Croce <redacted>
---
  lib/string.c | 80 ++++++++++++++++++++++++++++++++++++++++++++++++++--
  1 file changed, 77 insertions(+), 3 deletions(-)

Doesn't arch/riscv/lib/memcpy.S also exist for an architecture
optimised version? I would have thought the lib/string.c version
was not being used?

Yes, but this series started as C replacement for the assembly one,
which generates unaligned accesses.
Unfortunately the existing RISC-V processors can't handle unaligned
accesses, so they are emulated with a terrible slowdown.
Then, since there wasn't any riscv specific code, it was proposed as
generic code:

Discussion: https://lore.kernel.org/linux-riscv/20210617152754.17960-1-mcroce@linux.microsoft.com/ (local)

-- 
per aspera ad upstream