md5: add assembly implementation for loongarch64

This change can improve md5 performance by using a hand-optimized
assembly implementation of the inner loop of md5 calculation.
This

md5: add assembly implementation for loongarch64

This change can improve md5 performance by using a hand-optimized
assembly implementation of the inner loop of md5 calculation.
This implementation refered to md5-x86_64.pl and made more effort
to reorder instructions for separating data dependencies as much
as possible.

Test with:
$ openssl speed md5

3A5000
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes
md5 45061.04k 130440.75k 291105.28k 421101.23k 484639.27k 488320.43k
md5-modified 47179.95k 139015.57k 308836.69k 445963.26k 512540.67k 518215.00k
+5% +7% +6% +6% +6% +6%

3A6000
type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes
md5 60070.06k 161822.76k 325817.60k 438017.02k 486864.21k 492243.31k
md5-modified 62827.74k 170294.04k 343795.03k 463324.50k 515831.13k 520060.93k
+5% +5% +6% +6% +6% +6%

Signed-off-by: Min Zhou <zhoumin@loongson.cn>
Co-authored-by: Xi Ruoyao <xry111@xry111.site>

Reviewed-by: Shane Lontis <shane.lontis@oracle.com>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/21704)

show more ...