Patchwork Ping: [Patch] aarch64: Thunderx specific memcpy and memmove

login
register
mail settings
Submitter Steve Ellcey
Date May 9, 2017, 9:45 p.m.
Message ID <1494366305.9224.26.camel@cavium.com>
Download mbox | patch
Permalink /patch/20361/
State New
Headers show

Comments

Steve Ellcey - May 9, 2017, 9:45 p.m.
On Tue, 2017-05-09 at 08:45 +0530, Siddhesh Poyarekar wrote:
> On Wednesday 03 May 2017 07:31 PM, Szabolcs Nagy wrote:
> > 
> > if it turns out that a single generic memcpy does not work
> > it makes more sense to me to organize the code differently:
> > if we expect the generic memcpy to diverge from the thunderx
> > one then it's better not to use the same code with ifdefs, but
> > keep them separate, so the thunderx variant can be maintained
> > independently by whoever cares about thunderx.
> If that is the case then I think Steve might be better off posting a
> patch with the thunderx implementation being independent of the stock
> aarch64 implementation while Wilco does his investigation.  That way
> we
> don't scramble for a patch late in the 2.26 cycle - there's about a
> month and a half left.
> 
> Siddhesh

That sounds reasonable to me.  Here is a patch that contains a separate
memcpy_thunderx implementation.  I still have some (minor) changes to
the generic memcpy.S file.  One change is to use macros for the
function names so that the generic multiarch memcpy can include the
standard non-multiarch version.  The other is to change a couple of
internal labels to external labels.  This change isn't absolutely
necessary but it is helpful in the thunderx memcpy where the branches
are slightly different and I would like to keep the thunderx memcpy and
the generic memcpy as similar as possible so that when a change happens
in one or the other it is easy to compare the two versions.  I don't
believe using different label types affects the generated code at all
and personally, I find named labels easier to read than the internal
numbered labels.  Being able to compare the two memcpy's is also why I
kept the THUNDERX ifdef in memcpy_thunderx.S even though it is always
defined there, so that the intended differences are explicit when
comparing the two versions of memcpy.

Tested on the top-of-tree sources with no regressions.

Steve Ellcey
sellcey@cavium.com


2017-05-09  Steve Ellcey  <sellcey@caviumnetworks.com>

	* sysdeps/aarch64/memcpy.S (MEMMOVE, MEMCPY): New macros.
	(memmove): Use MEMMOVE for name.
	(memcpy): Use MEMCPY for name.  Change internal labels
	to external labels.
	* sysdeps/aarch64/multiarch/Makefile: New file.
	* sysdeps/aarch64/multiarch/ifunc-impl-list.c: Likewise.
	* sysdeps/aarch64/multiarch/init-arch.h: Likewise.
	* sysdeps/aarch64/multiarch/memcpy.c: Likewise.
	* sysdeps/aarch64/multiarch/memcpy_generic.S: Likewise.
	* sysdeps/aarch64/multiarch/memcpy_thunderx.S: Likewise.
	* sysdeps/aarch64/multiarch/memmove.c: Likewise.
Steve Ellcey - May 18, 2017, 9:48 p.m.
On Tue, 2017-05-09 at 14:45 -0700, Steve Ellcey wrote:

> 2017-05-09  Steve Ellcey  <sellcey@caviumnetworks.com>
> 
> 	* sysdeps/aarch64/memcpy.S (MEMMOVE, MEMCPY): New macros.
> 	(memmove): Use MEMMOVE for name.
> 	(memcpy): Use MEMCPY for name.  Change internal labels
> 	to external labels.
> 	* sysdeps/aarch64/multiarch/Makefile: New file.
> 	* sysdeps/aarch64/multiarch/ifunc-impl-list.c: Likewise.
> 	* sysdeps/aarch64/multiarch/init-arch.h: Likewise.
> 	* sysdeps/aarch64/multiarch/memcpy.c: Likewise.
> 	* sysdeps/aarch64/multiarch/memcpy_generic.S: Likewise.
> 	* sysdeps/aarch64/multiarch/memcpy_thunderx.S: Likewise.
> 	* sysdeps/aarch64/multiarch/memmove.c: Likewise.

Ping.

Steve Ellcey
sellcey@cavium.com
Siddhesh Poyarekar - May 19, 2017, 7:41 a.m.
On Wednesday 10 May 2017 03:15 AM, Steve Ellcey wrote:
> That sounds reasonable to me.  Here is a patch that contains a separate
> memcpy_thunderx implementation.  I still have some (minor) changes to
> the generic memcpy.S file.  One change is to use macros for the
> function names so that the generic multiarch memcpy can include the
> standard non-multiarch version.  The other is to change a couple of
> internal labels to external labels.  This change isn't absolutely
> necessary but it is helpful in the thunderx memcpy where the branches
> are slightly different and I would like to keep the thunderx memcpy and
> the generic memcpy as similar as possible so that when a change happens
> in one or the other it is easy to compare the two versions.  I don't
> believe using different label types affects the generated code at all
> and personally, I find named labels easier to read than the internal
> numbered labels.  Being able to compare the two memcpy's is also why I
> kept the THUNDERX ifdef in memcpy_thunderx.S even though it is always
> defined there, so that the intended differences are explicit when
> comparing the two versions of memcpy.
> 
> Tested on the top-of-tree sources with no regressions.

The patch looks fine.  Please coordinate with Szabolcs and Wilco on the
way forward.  You could either commit now and let Wilco rebase on top of
your changes or wait till he is done with his analysis and then figure
out the next step.

Siddhesh

PS: You don't really need the USE_THUNDERX macros in your memcpy
anymore, but it's not a big deal if that is what you prefer to keep
track of generic memcpy changes since you own that code.
Siddhesh Poyarekar - May 25, 2017, 6:42 a.m.
On Wednesday 24 May 2017 10:34 PM, Szabolcs Nagy wrote:
> - i don't know if you plan to make more changes to the thunderx
> memcpy, if prefetching is the only change then it's likely that
> we can agree on a generic version that's good enough.  if you
> do plan to make further changes, then keep in mind that we try
> to have same/similar generic memcpy across c runtimes and if
> your change is good for generic we might not be able to use the
> code outside of glibc (so newlib, bionic, freebsd,.. memcpy
> would diverge)

Steve, if that is desirable then please consider contributing the code
to cortex-strings[1].

> - non-thunderx systems are affected: static linked code using
> memcpy will start to go through an indirection (iplt) instead
> of direct call. if there are complaints about it or other ifunc
> related issues come up, then again we will have to reconsider it.

They could use a library built with --disable-multiarch.  The only place
I can see this happening is on systems that currently need bespoke
images, e.g. raspberry pis or similar form factors.  Since they're
building custom images anyway, it shouldn't be too hard to add a glibc
built with --disable-multiarch for them.

Removing multiarch completely is not an option since there's thunderx
*and* falkor (yes yes, coming soon, I promise!) with their own routines
and perhaps more in future.

Siddhesh

[1] https://launchpad.net/cortex-strings
Steve Ellcey - May 25, 2017, 4:22 p.m.
On Wed, 2017-05-24 at 18:04 +0100, Szabolcs Nagy wrote:

> - i don't know if you plan to make more changes to the thunderx
> memcpy, if prefetching is the only change then it's likely that
> we can agree on a generic version that's good enough.  if you
> do plan to make further changes, then keep in mind that we try
> to have same/similar generic memcpy across c runtimes and if
> your change is good for generic we might not be able to use the
> code outside of glibc (so newlib, bionic, freebsd,.. memcpy
> would diverge)

I am not looking at any more changes to the thunderx memcpy but someone
here is working on a thunderx2 memcpy so I do expect a third version
of memcpy at some point.

> - non-thunderx systems are affected: static linked code using
> memcpy will start to go through an indirection (iplt) instead
> of direct call. if there are complaints about it or other ifunc
> related issues come up, then again we will have to reconsider it.
> 
> so the patch can go in with an understanding that it may go out.

I understand that we may have to look at issues like this in the
future, though building glibc without multiarch enabled should provide
an alternative to those who don't want IFUNC functionality.

Steve Ellcey
sellcey@cavium.com
pinskia@gmail.com - May 25, 2017, 4:28 p.m.
On Wed, May 24, 2017 at 11:42 PM, Siddhesh Poyarekar
<siddhesh@gotplt.org> wrote:
> On Wednesday 24 May 2017 10:34 PM, Szabolcs Nagy wrote:
>> - i don't know if you plan to make more changes to the thunderx
>> memcpy, if prefetching is the only change then it's likely that
>> we can agree on a generic version that's good enough.  if you
>> do plan to make further changes, then keep in mind that we try
>> to have same/similar generic memcpy across c runtimes and if
>> your change is good for generic we might not be able to use the
>> code outside of glibc (so newlib, bionic, freebsd,.. memcpy
>> would diverge)
>
> Steve, if that is desirable then please consider contributing the code
> to cortex-strings[1].

One memcpy does not fit all micro-arch.  Just look at x86, where they
have many different versions and even do selection based on cache size
(see the current discussion about the memcpy regression).

>
>> - non-thunderx systems are affected: static linked code using
>> memcpy will start to go through an indirection (iplt) instead
>> of direct call. if there are complaints about it or other ifunc
>> related issues come up, then again we will have to reconsider it.

Just to answer this.  This is true on x86 and PowerPC already so there
should be no difference on aarch64 than those two targets.

>
> They could use a library built with --disable-multiarch.  The only place
> I can see this happening is on systems that currently need bespoke
> images, e.g. raspberry pis or similar form factors.  Since they're
> building custom images anyway, it shouldn't be too hard to add a glibc
> built with --disable-multiarch for them.
>
> Removing multiarch completely is not an option since there's thunderx
> *and* falkor (yes yes, coming soon, I promise!) with their own routines
> and perhaps more in future.

ThunderX2T99 version should be posted by the end of next week.

>
> Siddhesh
>
> [1] https://launchpad.net/cortex-strings
Ramana Radhakrishnan - May 25, 2017, 4:43 p.m.
On Thu, May 25, 2017 at 5:28 PM, Andrew Pinski <pinskia@gmail.com> wrote:
> On Wed, May 24, 2017 at 11:42 PM, Siddhesh Poyarekar
> <siddhesh@gotplt.org> wrote:
>> On Wednesday 24 May 2017 10:34 PM, Szabolcs Nagy wrote:
>>> - i don't know if you plan to make more changes to the thunderx
>>> memcpy, if prefetching is the only change then it's likely that
>>> we can agree on a generic version that's good enough.  if you
>>> do plan to make further changes, then keep in mind that we try
>>> to have same/similar generic memcpy across c runtimes and if
>>> your change is good for generic we might not be able to use the
>>> code outside of glibc (so newlib, bionic, freebsd,.. memcpy
>>> would diverge)
>>
>> Steve, if that is desirable then please consider contributing the code
>> to cortex-strings[1].
>
> One memcpy does not fit all micro-arch.  Just look at x86, where they
> have many different versions and even do selection based on cache size
> (see the current discussion about the memcpy regression).

No it doesn't but that's not an excuse for putting in multiple copies without
the engineering to see if generic memcpy's are improved. Otherwise it's
just bloatware.


regards
Ramana
Wilco Dijkstra - May 25, 2017, 5:49 p.m.
Andrew Pinski <pinskia@gmail.com> wrote:
> 
> One memcpy does not fit all micro-arch.  Just look at x86, where they
> have many different versions and even do selection based on cache size
> (see the current discussion about the memcpy regression).

Given the number of micro architectures already existing, it would be a really
bad situation to end up with one memcpy per micro architecture...

Micro architectures will tend to converge rather than diverge as performance
level increases. So I believe it's generally best to use the same instructions for
memcpy as for compiled code as that is what CPUs will actually encounter
and optimize for. For the rare, very large copies we could do something different
if it helps (eg. prefetch, non-temporals, SIMD registers etc).

> >> - non-thunderx systems are affected: static linked code using
> >> memcpy will start to go through an indirection (iplt) instead
> >> of direct call. if there are complaints about it or other ifunc
> >> related issues come up, then again we will have to reconsider it.
>
> Just to answer this.  This is true on x86 and PowerPC already so there
> should be no difference on aarch64 than those two targets.

An ifunc has a measurable overhead unfortunately, and that would no longer
be trivially avoidable via static linking. Most calls to memcpy tend to be very
small copies. Maybe we should investigate statically linking the small copy part
of memcpy with say -O3?

Cheers,
Wilco
Siddhesh Poyarekar - May 25, 2017, 7:26 p.m.
On Thursday 25 May 2017 11:19 PM, Wilco Dijkstra wrote:
> Given the number of micro architectures already existing, it would be a really
> bad situation to end up with one memcpy per micro architecture...

It's not just per micro-architecture...

> Micro architectures will tend to converge rather than diverge as performance
> level increases. So I believe it's generally best to use the same instructions for
> memcpy as for compiled code as that is what CPUs will actually encounter
> and optimize for. For the rare, very large copies we could do something different
> if it helps (eg. prefetch, non-temporals, SIMD registers etc).

... because as you say, micro-architectures may well converge over time
to some extent, but you will still end up having multiple memcpy
implementation taking advantage of different features in aarch64
architecture over time.  For example, SVE routines vs non-SVE routines.
You'll need both and looking at how x86 has evolved, there will be much
more to come.

> An ifunc has a measurable overhead unfortunately, and that would no longer
> be trivially avoidable via static linking. Most calls to memcpy tend to be very
> small copies. Maybe we should investigate statically linking the small copy part
> of memcpy with say -O3?

Sure, that might be something to look at as a data point, but again
getting rid of multiarch is not the option for desktop/server
implementations, especially if micro-architecture specific routines give
measurable gains over generic implementations in the general case, i.e.
dynamically linked programs that need to run out of the box and
optimally on multiple types of hardware.  Static binaries unfortunately
become the edge case here.

Siddhesh
Ramana Radhakrishnan - May 25, 2017, 9:04 p.m.
On Thu, May 25, 2017 at 8:26 PM, Siddhesh Poyarekar <siddhesh@gotplt.org> wrote:
> On Thursday 25 May 2017 11:19 PM, Wilco Dijkstra wrote:
>> Given the number of micro architectures already existing, it would be a really
>> bad situation to end up with one memcpy per micro architecture...
>
> It's not just per micro-architecture...
>
>> Micro architectures will tend to converge rather than diverge as performance
>> level increases. So I believe it's generally best to use the same instructions for
>> memcpy as for compiled code as that is what CPUs will actually encounter
>> and optimize for. For the rare, very large copies we could do something different
>> if it helps (eg. prefetch, non-temporals, SIMD registers etc).

>
> ... because as you say, micro-architectures may well converge over time
> to some extent, but you will still end up having multiple memcpy
> implementation taking advantage of different features in aarch64
> architecture over time.  For example, SVE routines vs non-SVE routines.
> You'll need both and looking at how x86 has evolved, there will be much
> more to come.

SVE in the ARM world is architectural and not micro-architectural in the context
of this discussion :) .

The difference in the ARM world compared to the x86 world is the number
of micro-architectures that target the same architectural baseline.
Pushing in a memcpy
for every single micro-architecture out there will make the library a
maintenance
nightmare !

And we also need to see some numbers which compare the relative performance
of the routines being put in compared to the generic memcpy otherwise things
will not improve.  Atleast something like this routine is X % better than the
generic memcpy.


regards
Ramana
Florian Weimer - May 25, 2017, 9:12 p.m.
On Thu, May 25, 2017 at 11:04 PM, Ramana Radhakrishnan
<ramana.gcc@googlemail.com> wrote:
> The difference in the ARM world compared to the x86 world is the number
> of micro-architectures that target the same architectural baseline.
> Pushing in a memcpy
> for every single micro-architecture out there will make the library a
> maintenance nightmare !

In this case, you should reconsider putting the string functions into
the vDSO.  This will push the implementation to the kernel, but it
does more than just shifting the work: the kernel has more direct
means to provide hardware capabilities, and it also can use
just-in-time code generation (which we want to avoid in glibc).

Thanks,
Florian
Siddhesh Poyarekar - May 26, 2017, 5:34 a.m.
On Friday 26 May 2017 02:34 AM, Ramana Radhakrishnan wrote:
> SVE in the ARM world is architectural and not micro-architectural in the context
> of this discussion :) .

Yes, I did not think of SVE as a micro-architecture detail.  I used SVE
as an example to show that multiarch =/=> micro-architectures.

> The difference in the ARM world compared to the x86 world is the number
> of micro-architectures that target the same architectural baseline.
> Pushing in a memcpy
> for every single micro-architecture out there will make the library a
> maintenance
> nightmare !

Nor am I arguing that micro-architectures ==> multiarch.  My comments
have only been pointing out that the IFUNC cost for multiarch is here to
stay and there will likely never be consensus to drop it given the
innovations happening in the ARM server space.

> And we also need to see some numbers which compare the relative performance
> of the routines being put in compared to the generic memcpy otherwise things
> will not improve.  Atleast something like this routine is X % better than the
> generic memcpy.

Yes, and I understand Steve had pointed out the benefits of his
implementation in the original post.  If it turns out that the
implementation is optimal for the general case, by all means merge it
with the generic one but that's not a reason to drop multiarch.

Siddhesh
pinskia@gmail.com - May 26, 2017, 5:38 a.m.
On Thu, May 25, 2017 at 10:34 PM, Siddhesh Poyarekar
<siddhesh@gotplt.org> wrote:
> On Friday 26 May 2017 02:34 AM, Ramana Radhakrishnan wrote:
>> SVE in the ARM world is architectural and not micro-architectural in the context
>> of this discussion :) .
>
> Yes, I did not think of SVE as a micro-architecture detail.  I used SVE
> as an example to show that multiarch =/=> micro-architectures.
>
>> The difference in the ARM world compared to the x86 world is the number
>> of micro-architectures that target the same architectural baseline.
>> Pushing in a memcpy
>> for every single micro-architecture out there will make the library a
>> maintenance
>> nightmare !
>
> Nor am I arguing that micro-architectures ==> multiarch.  My comments
> have only been pointing out that the IFUNC cost for multiarch is here to
> stay and there will likely never be consensus to drop it given the
> innovations happening in the ARM server space.
>
>> And we also need to see some numbers which compare the relative performance
>> of the routines being put in compared to the generic memcpy otherwise things
>> will not improve.  Atleast something like this routine is X % better than the
>> generic memcpy.
>
> Yes, and I understand Steve had pointed out the benefits of his
> implementation in the original post.  If it turns out that the
> implementation is optimal for the general case, by all means merge it
> with the generic one but that's not a reason to drop multiarch.

One more comment about static linking.  In the server world nobody
static links any more.  It is not something people do at all.  It is
hard to do in a reasonable fashion any more due to how things like DNS
lookup in glibc.

Thanks,
Andrew

>
> Siddhesh
Siddhesh Poyarekar - May 26, 2017, 5:42 a.m.
On Friday 26 May 2017 02:42 AM, Florian Weimer wrote:
> In this case, you should reconsider putting the string functions into
> the vDSO.  This will push the implementation to the kernel, but it
> does more than just shifting the work: the kernel has more direct
> means to provide hardware capabilities, and it also can use
> just-in-time code generation (which we want to avoid in glibc).

You will still need an indirection to access the VDSO function, so it
does not solve the problem Wilco was referring to.  Micro-architecture
explosion is currently just a theory (which both Wilco and I think is
invalid since we will eventually come up with a small enough set of
functions that cater to a variety of micro-architectures), so this would
be worth worrying about only if we find that we're exceeding the number
of IFUNC implementations in x86.

Siddhesh

Patch

diff --git a/sysdeps/aarch64/memcpy.S b/sysdeps/aarch64/memcpy.S
index 29af8b1..88a3b90 100644
--- a/sysdeps/aarch64/memcpy.S
+++ b/sysdeps/aarch64/memcpy.S
@@ -59,7 +59,14 @@ 
    Overlapping large forward memmoves use a loop that copies backwards.
 */
 
-ENTRY_ALIGN (memmove, 6)
+#ifndef MEMMOVE
+# define MEMMOVE memmove
+#endif
+#ifndef MEMCPY
+# define MEMCPY memcpy
+#endif
+
+ENTRY_ALIGN (MEMMOVE, 6)
 
 	DELOUSE (0)
 	DELOUSE (1)
@@ -71,9 +78,9 @@  ENTRY_ALIGN (memmove, 6)
 	b.lo	L(move_long)
 
 	/* Common case falls through into memcpy.  */
-END (memmove)
-libc_hidden_builtin_def (memmove)
-ENTRY (memcpy)
+END (MEMMOVE)
+libc_hidden_builtin_def (MEMMOVE)
+ENTRY (MEMCPY)
 
 	DELOUSE (0)
 	DELOUSE (1)
@@ -169,8 +176,8 @@  L(copy_long):
 	ldp	C_l, C_h, [src, 48]
 	ldp	D_l, D_h, [src, 64]!
 	subs	count, count, 128 + 16	/* Test and readjust count.  */
-	b.ls	2f
-1:
+	b.ls	L(last64)
+L(loop64):
 	stp	A_l, A_h, [dst, 16]
 	ldp	A_l, A_h, [src, 16]
 	stp	B_l, B_h, [dst, 32]
@@ -180,12 +187,12 @@  L(copy_long):
 	stp	D_l, D_h, [dst, 64]!
 	ldp	D_l, D_h, [src, 64]!
 	subs	count, count, 64
-	b.hi	1b
+	b.hi	L(loop64)
 
 	/* Write the last full set of 64 bytes.  The remainder is at most 64
 	   bytes, so it is safe to always copy 64 bytes from the end even if
 	   there is just 1 byte left.  */
-2:
+L(last64):
 	ldp	E_l, E_h, [srcend, -64]
 	stp	A_l, A_h, [dst, 16]
 	ldp	A_l, A_h, [srcend, -48]
@@ -256,5 +263,5 @@  L(move_long):
 	stp	C_l, C_h, [dstin]
 3:	ret
 
-END (memcpy)
-libc_hidden_builtin_def (memcpy)
+END (MEMCPY)
+libc_hidden_builtin_def (MEMCPY)
diff --git a/sysdeps/aarch64/multiarch/Makefile b/sysdeps/aarch64/multiarch/Makefile
index e69de29..78d52c7 100644
--- a/sysdeps/aarch64/multiarch/Makefile
+++ b/sysdeps/aarch64/multiarch/Makefile
@@ -0,0 +1,3 @@ 
+ifeq ($(subdir),string)
+sysdep_routines += memcpy_generic memcpy_thunderx
+endif
diff --git a/sysdeps/aarch64/multiarch/ifunc-impl-list.c b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
index e69de29..c4f23df 100644
--- a/sysdeps/aarch64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
@@ -0,0 +1,51 @@ 
+/* Enumerate available IFUNC implementations of a function.  AARCH64 version.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <assert.h>
+#include <string.h>
+#include <wchar.h>
+#include <ldsodefs.h>
+#include <ifunc-impl-list.h>
+#include <init-arch.h>
+#include <stdio.h>
+
+/* Maximum number of IFUNC implementations.  */
+#define MAX_IFUNC	2
+
+size_t
+__libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
+			size_t max)
+{
+  assert (max >= MAX_IFUNC);
+
+  size_t i = 0;
+
+  INIT_ARCH ();
+
+  /* Support sysdeps/aarch64/multiarch/memcpy.c and memmove.c.  */
+  IFUNC_IMPL (i, name, memcpy,
+	      IFUNC_IMPL_ADD (array, i, memcpy, IS_THUNDERX (midr),
+			      __memcpy_thunderx)
+	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_generic))
+  IFUNC_IMPL (i, name, memmove,
+	      IFUNC_IMPL_ADD (array, i, memmove, IS_THUNDERX (midr),
+			      __memmove_thunderx)
+	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic))
+
+  return i;
+}
diff --git a/sysdeps/aarch64/multiarch/init-arch.h b/sysdeps/aarch64/multiarch/init-arch.h
index e69de29..3af442c 100644
--- a/sysdeps/aarch64/multiarch/init-arch.h
+++ b/sysdeps/aarch64/multiarch/init-arch.h
@@ -0,0 +1,23 @@ 
+/* Define INIT_ARCH so that midr is initialized before use by IFUNCs.
+   This file is part of the GNU C Library.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <ldsodefs.h>
+
+#define INIT_ARCH()				\
+  uint64_t __attribute__((unused)) midr =	\
+    GLRO(dl_aarch64_cpu_features).midr_el1;
diff --git a/sysdeps/aarch64/multiarch/memcpy.c b/sysdeps/aarch64/multiarch/memcpy.c
index e69de29..9f73efb 100644
--- a/sysdeps/aarch64/multiarch/memcpy.c
+++ b/sysdeps/aarch64/multiarch/memcpy.c
@@ -0,0 +1,39 @@ 
+/* Multiple versions of memcpy. AARCH64 version.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Define multiple versions only for the definition in libc.  */
+
+#if IS_IN (libc)
+/* Redefine memcpy so that the compiler won't complain about the type
+   mismatch with the IFUNC selector in strong_alias, below.  */
+# undef memcpy
+# define memcpy __redirect_memcpy
+# include <string.h>
+# include <init-arch.h>
+
+extern __typeof (__redirect_memcpy) __libc_memcpy;
+
+extern __typeof (__redirect_memcpy) __memcpy_generic attribute_hidden;
+extern __typeof (__redirect_memcpy) __memcpy_thunderx attribute_hidden;
+
+libc_ifunc (__libc_memcpy,
+            IS_THUNDERX (midr) ? __memcpy_thunderx : __memcpy_generic);
+
+# undef memcpy
+strong_alias (__libc_memcpy, memcpy);
+#endif
diff --git a/sysdeps/aarch64/multiarch/memcpy_generic.S b/sysdeps/aarch64/multiarch/memcpy_generic.S
index e69de29..041a779 100644
--- a/sysdeps/aarch64/multiarch/memcpy_generic.S
+++ b/sysdeps/aarch64/multiarch/memcpy_generic.S
@@ -0,0 +1,42 @@ 
+/* A Generic Optimized memcpy implementation for AARCH64.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* The actual memcpy and memmove code is in ../memcpy.S.  If we are
+   building libc this file defines __memcpy_generic and __memmove_generic.
+   Otherwise the include of ../memcpy.S will define the normal __memcpy
+   and__memmove entry points.  */
+
+#include <sysdep.h>
+
+#if IS_IN (libc)
+
+# define MEMCPY __memcpy_generic
+# define MEMMOVE __memmove_generic
+
+/* Do not hide the generic versions of memcpy and memmove, we use them
+   internally.  */
+# undef libc_hidden_builtin_def
+# define libc_hidden_builtin_def(name)
+
+/* It doesn't make sense to send libc-internal memcpy calls through a PLT. */
+	.globl __GI_memcpy; __GI_memcpy = __memcpy_generic
+	.globl __GI_memmove; __GI_memmove = __memmove_generic
+
+#endif
+
+#include "../memcpy.S"
diff --git a/sysdeps/aarch64/multiarch/memcpy_thunderx.S b/sysdeps/aarch64/multiarch/memcpy_thunderx.S
index e69de29..5ac9e34 100644
--- a/sysdeps/aarch64/multiarch/memcpy_thunderx.S
+++ b/sysdeps/aarch64/multiarch/memcpy_thunderx.S
@@ -0,0 +1,326 @@ 
+/* A Thunderx Optimized memcpy implementation for AARCH64.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* The actual code in this memcpy and memmove should be identical to the
+   generic version except for the code under '#ifdef THUNDERX'.  This is
+   to make is easier to keep this version and the generic version in sync
+   for changes that are not specific to thunderx.  */
+
+#include <sysdep.h>
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64, unaligned accesses.
+ *
+ */
+
+#define dstin	x0
+#define src	x1
+#define count	x2
+#define dst	x3
+#define srcend	x4
+#define dstend	x5
+#define A_l	x6
+#define A_lw	w6
+#define A_h	x7
+#define A_hw	w7
+#define B_l	x8
+#define B_lw	w8
+#define B_h	x9
+#define C_l	x10
+#define C_h	x11
+#define D_l	x12
+#define D_h	x13
+#define E_l	src
+#define E_h	count
+#define F_l	srcend
+#define F_h	dst
+#define G_l	count
+#define G_h	dst
+#define tmp1	x14
+
+/* Copies are split into 3 main cases: small copies of up to 16 bytes,
+   medium copies of 17..96 bytes which are fully unrolled. Large copies
+   of more than 96 bytes align the destination and use an unrolled loop
+   processing 64 bytes per iteration.
+   In order to share code with memmove, small and medium copies read all
+   data before writing, allowing any kind of overlap. So small, medium
+   and large backwards memmoves are handled by falling through into memcpy.
+   Overlapping large forward memmoves use a loop that copies backwards.
+*/
+
+#ifndef MEMMOVE
+# define MEMMOVE memmove
+#endif
+#ifndef MEMCPY
+# define MEMCPY memcpy
+#endif
+
+#if IS_IN (libc)
+
+# undef MEMCPY
+# define MEMCPY __memcpy_thunderx
+# undef MEMMOVE
+# define MEMMOVE __memmove_thunderx
+# define USE_THUNDERX
+
+ENTRY_ALIGN (MEMMOVE, 6)
+
+	DELOUSE (0)
+	DELOUSE (1)
+	DELOUSE (2)
+
+	sub	tmp1, dstin, src
+	cmp	count, 96
+	ccmp	tmp1, count, 2, hi
+	b.lo	L(move_long)
+
+	/* Common case falls through into memcpy.  */
+END (MEMMOVE)
+libc_hidden_builtin_def (MEMMOVE)
+ENTRY (MEMCPY)
+
+	DELOUSE (0)
+	DELOUSE (1)
+	DELOUSE (2)
+
+	prfm	PLDL1KEEP, [src]
+	add	srcend, src, count
+	add	dstend, dstin, count
+	cmp	count, 16
+	b.ls	L(copy16)
+	cmp	count, 96
+	b.hi	L(copy_long)
+
+	/* Medium copies: 17..96 bytes.  */
+	sub	tmp1, count, 1
+	ldp	A_l, A_h, [src]
+	tbnz	tmp1, 6, L(copy96)
+	ldp	D_l, D_h, [srcend, -16]
+	tbz	tmp1, 5, 1f
+	ldp	B_l, B_h, [src, 16]
+	ldp	C_l, C_h, [srcend, -32]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstend, -32]
+1:
+	stp	A_l, A_h, [dstin]
+	stp	D_l, D_h, [dstend, -16]
+	ret
+
+	.p2align 4
+	/* Small copies: 0..16 bytes.  */
+L(copy16):
+	cmp	count, 8
+	b.lo	1f
+	ldr	A_l, [src]
+	ldr	A_h, [srcend, -8]
+	str	A_l, [dstin]
+	str	A_h, [dstend, -8]
+	ret
+	.p2align 4
+1:
+	tbz	count, 2, 1f
+	ldr	A_lw, [src]
+	ldr	A_hw, [srcend, -4]
+	str	A_lw, [dstin]
+	str	A_hw, [dstend, -4]
+	ret
+
+	/* Copy 0..3 bytes.  Use a branchless sequence that copies the same
+	   byte 3 times if count==1, or the 2nd byte twice if count==2.  */
+1:
+	cbz	count, 2f
+	lsr	tmp1, count, 1
+	ldrb	A_lw, [src]
+	ldrb	A_hw, [srcend, -1]
+	ldrb	B_lw, [src, tmp1]
+	strb	A_lw, [dstin]
+	strb	B_lw, [dstin, tmp1]
+	strb	A_hw, [dstend, -1]
+2:	ret
+
+	.p2align 4
+	/* Copy 64..96 bytes.  Copy 64 bytes from the start and
+	   32 bytes from the end.  */
+L(copy96):
+	ldp	B_l, B_h, [src, 16]
+	ldp	C_l, C_h, [src, 32]
+	ldp	D_l, D_h, [src, 48]
+	ldp	E_l, E_h, [srcend, -32]
+	ldp	F_l, F_h, [srcend, -16]
+	stp	A_l, A_h, [dstin]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstin, 32]
+	stp	D_l, D_h, [dstin, 48]
+	stp	E_l, E_h, [dstend, -32]
+	stp	F_l, F_h, [dstend, -16]
+	ret
+
+	/* Align DST to 16 byte alignment so that we don't cross cache line
+	   boundaries on both loads and stores.  There are at least 96 bytes
+	   to copy, so copy 16 bytes unaligned and then align.  The loop
+	   copies 64 bytes per iteration and prefetches one iteration ahead.  */
+
+	.p2align 4
+L(copy_long):
+
+# ifdef USE_THUNDERX
+
+	/* On thunderx, large memcpy's are helped by software prefetching.
+	   This loop is identical to the one below it but with prefetching
+	   instructions included.  For loops that are less than 32768 bytes,
+	   the prefetching does not help and slow the code down so we only
+	   use the prefetching loop for the largest memcpys.  */
+
+	cmp	count, #32768
+	b.lo	L(copy_long_without_prefetch)
+	and	tmp1, dstin, 15
+	bic	dst, dstin, 15
+	ldp	D_l, D_h, [src]
+	sub	src, src, tmp1
+	prfm	pldl1strm, [src, 384]
+	add	count, count, tmp1	/* Count is now 16 too large.  */
+	ldp	A_l, A_h, [src, 16]
+	stp	D_l, D_h, [dstin]
+	ldp	B_l, B_h, [src, 32]
+	ldp	C_l, C_h, [src, 48]
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 128 + 16	/* Test and readjust count.  */
+
+L(prefetch_loop64):
+	tbz	src, #6, 1f
+	prfm	pldl1strm, [src, 512]
+1:
+	stp	A_l, A_h, [dst, 16]
+	ldp	A_l, A_h, [src, 16]
+	stp	B_l, B_h, [dst, 32]
+	ldp	B_l, B_h, [src, 32]
+	stp	C_l, C_h, [dst, 48]
+	ldp	C_l, C_h, [src, 48]
+	stp	D_l, D_h, [dst, 64]!
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 64
+	b.hi	L(prefetch_loop64)
+	b	L(last64)
+
+L(copy_long_without_prefetch):
+# endif
+
+	and	tmp1, dstin, 15
+	bic	dst, dstin, 15
+	ldp	D_l, D_h, [src]
+	sub	src, src, tmp1
+	add	count, count, tmp1	/* Count is now 16 too large.  */
+	ldp	A_l, A_h, [src, 16]
+	stp	D_l, D_h, [dstin]
+	ldp	B_l, B_h, [src, 32]
+	ldp	C_l, C_h, [src, 48]
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 128 + 16	/* Test and readjust count.  */
+	b.ls	L(last64)
+L(loop64):
+	stp	A_l, A_h, [dst, 16]
+	ldp	A_l, A_h, [src, 16]
+	stp	B_l, B_h, [dst, 32]
+	ldp	B_l, B_h, [src, 32]
+	stp	C_l, C_h, [dst, 48]
+	ldp	C_l, C_h, [src, 48]
+	stp	D_l, D_h, [dst, 64]!
+	ldp	D_l, D_h, [src, 64]!
+	subs	count, count, 64
+	b.hi	L(loop64)
+
+	/* Write the last full set of 64 bytes.  The remainder is at most 64
+	   bytes, so it is safe to always copy 64 bytes from the end even if
+	   there is just 1 byte left.  */
+L(last64):
+	ldp	E_l, E_h, [srcend, -64]
+	stp	A_l, A_h, [dst, 16]
+	ldp	A_l, A_h, [srcend, -48]
+	stp	B_l, B_h, [dst, 32]
+	ldp	B_l, B_h, [srcend, -32]
+	stp	C_l, C_h, [dst, 48]
+	ldp	C_l, C_h, [srcend, -16]
+	stp	D_l, D_h, [dst, 64]
+	stp	E_l, E_h, [dstend, -64]
+	stp	A_l, A_h, [dstend, -48]
+	stp	B_l, B_h, [dstend, -32]
+	stp	C_l, C_h, [dstend, -16]
+	ret
+
+	.p2align 4
+L(move_long):
+	cbz	tmp1, 3f
+
+	add	srcend, src, count
+	add	dstend, dstin, count
+
+	/* Align dstend to 16 byte alignment so that we don't cross cache line
+	   boundaries on both loads and stores.  There are at least 96 bytes
+	   to copy, so copy 16 bytes unaligned and then align.  The loop
+	   copies 64 bytes per iteration and prefetches one iteration ahead.  */
+
+	and	tmp1, dstend, 15
+	ldp	D_l, D_h, [srcend, -16]
+	sub	srcend, srcend, tmp1
+	sub	count, count, tmp1
+	ldp	A_l, A_h, [srcend, -16]
+	stp	D_l, D_h, [dstend, -16]
+	ldp	B_l, B_h, [srcend, -32]
+	ldp	C_l, C_h, [srcend, -48]
+	ldp	D_l, D_h, [srcend, -64]!
+	sub	dstend, dstend, tmp1
+	subs	count, count, 128
+	b.ls	2f
+
+	nop
+1:
+	stp	A_l, A_h, [dstend, -16]
+	ldp	A_l, A_h, [srcend, -16]
+	stp	B_l, B_h, [dstend, -32]
+	ldp	B_l, B_h, [srcend, -32]
+	stp	C_l, C_h, [dstend, -48]
+	ldp	C_l, C_h, [srcend, -48]
+	stp	D_l, D_h, [dstend, -64]!
+	ldp	D_l, D_h, [srcend, -64]!
+	subs	count, count, 64
+	b.hi	1b
+
+	/* Write the last full set of 64 bytes.  The remainder is at most 64
+	   bytes, so it is safe to always copy 64 bytes from the start even if
+	   there is just 1 byte left.  */
+2:
+	ldp	G_l, G_h, [src, 48]
+	stp	A_l, A_h, [dstend, -16]
+	ldp	A_l, A_h, [src, 32]
+	stp	B_l, B_h, [dstend, -32]
+	ldp	B_l, B_h, [src, 16]
+	stp	C_l, C_h, [dstend, -48]
+	ldp	C_l, C_h, [src]
+	stp	D_l, D_h, [dstend, -64]
+	stp	G_l, G_h, [dstin, 48]
+	stp	A_l, A_h, [dstin, 32]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstin]
+3:	ret
+
+END (MEMCPY)
+libc_hidden_builtin_def (MEMCPY)
+
+#endif
diff --git a/sysdeps/aarch64/multiarch/memmove.c b/sysdeps/aarch64/multiarch/memmove.c
index e69de29..34c6b29 100644
--- a/sysdeps/aarch64/multiarch/memmove.c
+++ b/sysdeps/aarch64/multiarch/memmove.c
@@ -0,0 +1,39 @@ 
+/* Multiple versions of memmove. AARCH64 version.
+   Copyright (C) 2017 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+/* Define multiple versions only for the definition in libc.  */
+
+#if IS_IN (libc)
+/* Redefine memmove so that the compiler won't complain about the type
+   mismatch with the IFUNC selector in strong_alias, below.  */
+# undef memmove
+# define memmove __redirect_memmove
+# include <string.h>
+# include <init-arch.h>
+
+extern __typeof (__redirect_memmove) __libc_memmove;
+
+extern __typeof (__redirect_memmove) __memmove_generic attribute_hidden;
+extern __typeof (__redirect_memmove) __memmove_thunderx attribute_hidden;
+
+libc_ifunc (__libc_memmove,
+            IS_THUNDERX (midr) ? __memmove_thunderx : __memmove_generic);
+
+# undef memmove
+strong_alias (__libc_memmove, memmove);
+#endif