[PATCH 0/5] Added optimized memcpy/memmove/memset for A64FX

[naohirot@fujitsu.com]

Hi Wilco-san,

Let me focus on the macro " shortcut_for_small_size" for small/medium, less than
512 byte in this mail. 

> From: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
> > Yes, I implemented for the case of 1 byte to 512 byte [9][10].
> > SVE code seems faster than ASIMD in small/medium range too [11][12][13].
> 
> That adds quite a lot of code and uses a slow linear chain of comparisons. A small
> loop like used in the memset should work fine to handle copies smaller than
> 256 or 512 bytes (you can handle the zero bytes case for free in this code rather
> than special casing it).
> 

I compared performance of the size less than 512 byte for the following five
implementation cases.

CASE 1: liner chain
As mentioned in the reply [0] I removed BTI_J [1], but the macro " shortcut_for_small_size"
stays linear chain [2]
A64FX performance is 4-14 Gbps [3].
The other arch implementations call BTI_J, so performance is degraded.
.
[0] https://sourceware.org/pipermail/libc-alpha/2021-April/125079.html
[1] https://github.com/NaohiroTamura/glibc/commit/7d7217b518e59c78582ac4e89cae725cf620877e
[2] https://github.com/NaohiroTamura/glibc/blob/7d7217b518e59c78582ac4e89cae725cf620877e/sysdeps/aarch64/multiarch/memcpy_a64fx.S#L176-L267
[3] https://drive.google.com/file/d/16qo7N05W526H9j7_9qjm-_Q7gZmOXwpY/view

CASE 2: whilelt loop such as memset
I tested "whilelt loop" implementation instead of the macro " shortcut_for_small_size".
And after having tested, I commented out "whilelt loop" implementation [4]
Comparing with the CASE 1, A64FX performance degraded from 4-14 Gbps to 3-10 Gbps [5]. 
Please notice that "whilelt loop" implementation cannot be used for memmove,
because it doesn't work for backward copy.
On the other hand, the macro " shortcut_for_small_size" works for backward copy, because
it loads up to all 512 byte of data into z0 to z7 SVE registers at once, and then store all data.

[4] https://github.com/NaohiroTamura/glibc/commit/77d1da301f8161c74875b0314cae34be8cb33477#diff-03552f8369653866548b20e7867272a645fa2129c700b78fdfafe5a0ff6a259eR308-R318
[5] https://drive.google.com/file/d/1xdw7mr0c90VupVkQwelFafQHNkXslCwv/view

CASE 3: binary tree chain
I updated the macro " shortcut_for_small_size" to use binary tree chain [6][7].
Comparing with the CASE 1, the size less than 96 byte degraded from 4.0-6.0 Gbps
to 2.5-5.0 Gbps, but the size 512 byte improved from 14.0 Gbps to 17.5 Gbps.

[6] https://github.com/NaohiroTamura/glibc/commit/5c17af8c57561ede5ed2c2af96c9efde4092f02f
[7] https://github.com/NaohiroTamura/glibc/blob/5c17af8c57561ede5ed2c2af96c9efde4092f02f/sysdeps/aarch64/multiarch/memcpy_a64fx.S#L177-L204
[8] https://drive.google.com/file/d/13w8yKdeLpVbp-uJmCttKBKtScya1tXqP/view

CASE 4: binary tree chain except up to 64 byte
I handled up to 64 byte so as to return quickly [9].
Comparing with the CASE 3, the size less than 64 byte improved from 2.5 Gbps to
4.0 Gbps, but the size 512 byte degraded from 17.5 Gbps to 16.5 Gbps [10].

[9] https://github.com/NaohiroTamura/glibc/commit/77d1da301f8161c74875b0314cae34be8cb33477#diff-03552f8369653866548b20e7867272a645fa2129c700b78fdfafe5a0ff6a259eR177-R184
[10] https://drive.google.com/file/d/1lFsjns9g_7fySAsvx_RVS9o6HSrk6ir9/view

CASE 5: binary tree chain except up to 128 byte
I handled up to 128 byte so as to return quickly [11].
Comparing with the CASE 4, the size less than 128 byte improved from 4.0-6.0 Gbps
to 4.0-7.0 Gbps, but the size 512 byte degraded from 16.5 Gbps to 16.0 Gbps [12].

[11] https://github.com/NaohiroTamura/glibc/commit/fefc59f01ecfd6a207fe261de5ab133f4409d687#diff-03552f8369653866548b20e7867272a645fa2129c700b78fdfafe5a0ff6a259eR184-R195
[12] https://drive.google.com/file/d/1HS277_qQUuEeZqLUo0H2XRlFhOhIdI_o/view

In conclusion, I'd like to adopt the CASE 5 implementation, considering the
performance balance between the small size (less than 128 byte) and medium size
(close to 512 byte).

Thanks.
Naohiro