Enhanced REP MOVSB feature of CPUs starting from Ivy Bridge makes
REP MOVSB the fastest way to copy memory in most of cases. However
Intel Optimization Reference Manual says: "setting the DF to force
REP MOVSB to copy bytes from high towards low addresses will expe-
rience significant performance degradation". Measurements on Intel
Cascade Lake and Alder Lake, same as on AMD Zen3 show that it can
drop throughput to as low as 2.5-3.5GB/s, comparing to ~10-30GB/s
of REP MOVSQ or hand-rolled loop, used for non-ERMS CPUs.
This patch keeps ERMS use for forward ordered memory copies, but
removes it for backward overlapped moves where it does not work.
This is just a cosmetic sync with kernel, since libc does not use
ERMS at this time.
Reviewed by: mjg
MFC after: 2 weeks
See the review for sample test results.
Reviewed by: kib (kernel part)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18401
Handling sizes of > 32 backwards will be updated later.
Reviewed by: kib (kernel part)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18387
For non-ERMS case the code used handle possible trailing bytes with
movsb first and then followed it up with movsq. This also happened
to alter how calculations were done for other cases.
Handle the tail with regular movs, just like when copying forward.
Use leaq to calculate the right offset from the get go, instead of
doing separate add and sub.
This adjusts the offset for non-rep cases so that they can be used
to handle the tail.
The routine is still a work in progress.
Sponsored by: The FreeBSD Foundation
bcopy is left alone as it is expected to be converted to a C func.
Due to header mess ALIGN_TEXT is temporarily defined explicitly in memmove.S
Reviewed by: kib
Approved by: re (gjb)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17538
