While hard to measure, reducing the number of possibly/likely
mis-predicted branches can generally be expected to be slightly
better.

Other than apparent at the first glance, this also doesn't grow
the function size (the alignment gap to the next function just
gets smaller).
Signed-off-by: NJan Beulich <jbeulich@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/4F218584020000780006F422@nat28.tlf.novell.comSigned-off-by: NIngo Molnar <mingo@elte.hu>

While currently there doesn't appear to be any reachable in-tree
case where such large memory blocks may be passed to memcpy(),
we already had hit the problem in our Xen kernels. Just like
done recently for mmeset(), rather than working around it,
prevent others from falling into the same trap by fixing this
long standing limitation.
Signed-off-by: NJan Beulich <jbeulich@suse.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: http://lkml.kernel.org/r/4F21846F020000780006F3FA@nat28.tlf.novell.comSigned-off-by: NIngo Molnar <mingo@elte.hu>

Support memcpy() with enhanced rep movsb. On processors supporting enhanced
rep movsb, the alternative memcpy() function using enhanced rep movsb overrides the original function and the fast string
function.
Signed-off-by: NFenghua Yu <fenghua.yu@intel.com>
Link: http://lkml.kernel.org/r/1305671358-14478-8-git-send-email-fenghua.yu@intel.comSigned-off-by: NH. Peter Anvin <hpa@linux.intel.com>

Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Link: http://lkml.kernel.org/r/201105011409.21629.bvanassche@acm.orgSigned-off-by: NIngo Molnar <mingo@elte.hu>

All read operations after allocation stage can run speculatively,
all write operation will run in program order, and if addresses are
different read may run before older write operation, otherwise wait
until write commit. However CPU don't check each address bit,
so read could fail to recognize different address even they
are in different page.For example if rsi is 0xf004, rdi is 0xe008,
in following operation there will generate big performance latency.
1. movq (%rsi),	%rax
2. movq %rax,	(%rdi)
3. movq 8(%rsi), %rax
4. movq %rax,	8(%rdi)

If %rsi and rdi were in really the same meory page, there are TRUE
read-after-write dependence because instruction 2 write 0x008 and
instruction 3 read 0x00c, the two address are overlap partially.
Actually there are in different page and no any issues,
but without checking each address bit CPU could think they are
in the same page, and instruction 3 have to wait for instruction 2
to write data into cache from write buffer, then load data from cache,
the cost time read spent is equal to mfence instruction. We may avoid it by
tuning operation sequence as follow.

1. movq 8(%rsi), %rax
2. movq %rax,	8(%rdi)
3. movq (%rsi),	%rax
4. movq %rax,	(%rdi)

Instruction 3 read 0x004, instruction 2 write address 0x010, no any
dependence.  At last on Core2 we gain 1.83x speedup compared with
original instruction sequence.  In this patch we first handle small
size(less 20bytes), then jump to different copy mode. Based on our
micro-benchmark small bytes from 1 to 127 bytes, we got up to 2X
improvement, and up to 1.5X improvement for 1024 bytes on Corei7.  (We
use our micro-benchmark, and will do further test according to your
requirment)
Signed-off-by: NMa Ling <ling.ma@intel.com>
LKML-Reference: <1277753065-18610-1-git-send-email-ling.ma@intel.com>
Signed-off-by: NH. Peter Anvin <hpa@zytor.com>

We already have cpufeature indicies above 255, so use a 16-bit number
for the alternatives index.  This consumes a padding field and so
doesn't add any size, but it means that abusing the padding field to
create assembly errors on overflow no longer works.  We can retain the
test simply by redirecting it to the .discard section, however.

[ v3: updated to include open-coded locations ]
Signed-off-by: NH. Peter Anvin <hpa@linux.intel.com>
LKML-Reference: <tip-f88731e3068f9d1392ba71cc9f50f035d26a0d4f@git.kernel.org>
Signed-off-by: NH. Peter Anvin <hpa@zytor.com>

In order to avoid unnecessary chains of branches, rather than
implementing memcpy()/memset()'s access to their alternative
implementations via a jump, patch the (larger) original function
directly.

The memcpy() part of this is slightly subtle: while alternative
instruction patching does itself use memcpy(), with the
replacement block being less than 64-bytes in size the main loop
of the original function doesn't get used for copying memcpy_c()
over memcpy(), and hence we can safely write over its beginning.

Also note that the CFI annotations are fine for both variants of
each of the functions.
Signed-off-by: NJan Beulich <jbeulich@novell.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
LKML-Reference: <4B2BB8D30200007800026AF2@vpn.id2.novell.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: cleanup

Make this file more readable by bringing it more in line
with the usual kernel style.
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: micro-optimization

This should slightly improve its performance.
Signed-off-by: NJan Beulich <jbeulich@novell.com>
LKML-Reference: <49B8F641.76E4.0078.0@novell.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

apply_alternatives uses memcpy() to apply alternatives.  Which has the
unfortunate effect that while applying memcpy alternative to memcpy
itself it tries to overwrite itself with nops - which causes #UD fault
as it overwrites half of an instruction in copy loop, and from this
point on only possible outcome is triplefault and reboot.

So let's overwrite only first two instructions of memcpy - as long as
the main memcpy loop is not in first two bytes it will work fine.
Signed-off-by: NPetr Vandrovec <petr@vandrovec.name>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

kbuild explicitly includes this at build time.
Signed-off-by: NDave Jones <davej@redhat.com>

Add unwind annotations to arch/x86_64/lib/*.S, and also use the macros
provided by linux/linkage.h where-ever possible.

Some of the alternative instructions handling needed to be adjusted so
that the replacement code would also have valid unwind information.
Signed-off-by: NJan Beulich <jbeulich@novell.com>
Signed-off-by: NAndi Kleen <ak@suse.de>

They cause quite bad performance regressions on Netburst
This is temporary until we can get new optimized functions
for these CPUs.

This undoes changes that were done in 2.6.15 and in 2.6.16-rc1,
essentially bringing the code back to 2.6.14 level. Only change
is I renamed the X86_FEATURE_K8_C flag to X86_FEATURE_REP_GOOD
and fixed the check for the flag and also fixed some comments.
Signed-off-by: NAndi Kleen <ak@suse.de>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

B stepping were the first shipping Opterons. memcpy/memset/copy_page/
clear_page had special optimized version for them. These are really
old and in the minority now and the difference to the generic versions
(using rep microcode) is not that big anyways. So just remove them.

TODO: figure out optimized versions for Intel Netburst based EM64T
Signed-off-by: NAndi Kleen <ak@suse.de>
Signed-off-by: NLinus Torvalds <torvalds@osdl.org>

Initial git repository build. I'm not bothering with the full history,
even though we have it. We can create a separate "historical" git
archive of that later if we want to, and in the meantime it's about
3.2GB when imported into git - space that would just make the early
git days unnecessarily complicated, when we don't have a lot of good
infrastructure for it.

Let it rip!