diff mbox

aarch64: thunderx2 memmove performance improvements

Message ID 20190415145311.GA14156@bell-sw.com
State Superseded
Headers

Commit Message

Anton Youdkevitch April 15, 2019, 2:53 p.m. UTC 
  Wilco,

Thanks a lot for your comments and suggestions.
I attached the upated patch.

On 4/12/2019 21:03, Wilco Dijkstra wrote:
> Hi Anton,
>
> This looks like a good cleanup! A few comments and suggestions for
> improvements:
>
> 0. The diff is quite large due to tab/space changes. Would it be possible to split
> this into a separate patch?
Sure, I will then send the tab/space cleanup patch later.

> 1. There are a few cases where ldp or stp could be used, but isn't, eg:
>
> +	str	B_q, [dst], #16
> +	ldp	H_q, I_q, [src], #32
> +	str	C_q, [dst], #16
>
> Why not do stp B_q, C_q, [dst], 32?
This is the remains of trying to manually schedule the instructions.
Fixed.

> 2. There are a lot of writeback instructions used in cases where this isn't
> strictly required. Have you noticed this actually improves performance? Even
> if required for the main loop, it is best to reduce them where possible:
But there are no writebacks in the main loop - they are only in the loop's
tails, aren't they?

> +L(loop128_exit0):
> +	ldp	F_q, G_q, [srcend, -64]
> +	ldp	H_q, I_q, [srcend, -32]
> +	stp	B_q, C_q, [dst], #32
> +	stp	D_q, E_q, [dst], #32
> +	stp	F_q, G_q, [dstend, -64]
> +	stp	H_q, I_q, [dstend, -32]
> +	ret
>
>   L(loop128_exit1):
> +	ldp	B_q, C_q, [srcend, -64]
> +	ldp	D_q, E_q, [srcend, -32]
> +	stp	F_q, G_q, [dst], #32
> +	stp	H_q, I_q, [dst], #32
> +	stp	B_q, C_q, [dstend, -64]
> +	stp	D_q, E_q, [dstend, -32]
> +	ret
>
> Here dst is not used but incremented twice.
Right, good catch, thanks!
Done.

> 3. Missed optimization:
>
> +L(dst_unaligned_tail):
> +	ldp	C_q, D_q, [srcend, -64]
> +	ldp	E_q, F_q, [srcend, -32]
> +	stp	A_q, B_q, [dst], #32
> +	stp	H_q, I_q, [dst], #32
> +	add	dst, dst, tmp1
> +	str	G_q, [dst, -16]
> +	stp	C_q, D_q, [dstend, -64]
> +	stp	E_q, F_q, [dstend, -32]
>   	ret
>
> Surely this could use str	G_q, [dst, tmp1] if we change the writeback on the stp?
Done.

> 4. Unrolling can be more efficient:
>
>   L(loop128):
> +	ldp	F_q, G_q, [src], #32
> +	ldp	H_q, I_q, [src], #32
> +	stp	B_q, C_q, [dst], #32
> +	stp	D_q, E_q, [dst], #32
> +	subs	count, count, 64
> +	b.lt	L(loop128_exit1)
> +	ldp	B_q, C_q, [src], #32
> +	ldp	D_q, E_q, [src], #32
> +	stp	F_q, G_q, [dst], #32
> +	stp	H_q, I_q, [dst], #32
> +	subs	count, count, 64
> +	b.ge	L(loop128)
> +L(loop128_exit0):
> The idea of unrolling 2x is to only have a single loop branch. Using a single branch
> makes it easier to remove all the writebacks too - you only need 2 rather than 8!
The idea was to minimize the length of the loop tail. For branchless 128
bytes per iteration loop the branchless tail needs to read 128 bytes.
For the branched loop as the one above the tail processes only 64
bytes. And I don't see how I can avoid writebacks of up to 127 bytes
in the branchless tails for your version.

Comments

Wilco Dijkstra April 18, 2019, 3:22 p.m. UTC | #1 
Hi Anton,

Thanks, that's much easier to follow. A few more comments:

@@ -103,188 +101,14 @@ ENTRY_ALIGN (MEMMOVE, 6)
 	DELOUSE (1)
 	DELOUSE (2)
 
+	add	srcend, src, count
+	cmp	count, 16
+	b.ls	L(memcopy16)
 	sub	tmp1, dstin, src
 	cmp	count, 96
 	ccmp	tmp1, count, 2, hi
 	b.lo	L(move_long)

That's 7 instructions, so the memcpy label ends up at an odd alignment.

+	str	B_q, [dst, 16]
 	stp	C_q, D_q, [dst, 32]
-	str     A_q, [dst, 64]
+	str	G_q, [dst, 64]
+	str	A_q, [dstin]
+	str	E_q, [dstend, -16]

You can use STP rather than STR for B and G by changing the STP of C/D.

+	/* 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:
+	ldr	G_q, [src, 48]
+	str	B_q, [dstend, -16]
+	ldr	B_q, [src, 32]
+	str	A_q, [dstend, -32]
+	ldr	A_q, [src, 16]
+	str	D_q, [dstend, -48]
+	ldr	D_q, [src]
+	str	C_q, [dstend, -64]
+	str	G_q, [dstin, 48]
+	str	B_q, [dstin, 32]
+	str	A_q, [dstin, 16]
+	str	D_q, [dstin]
+3:	ret

3 LDP+ 3 STP opportunities here.

>> The idea of unrolling 2x is to only have a single loop branch. Using a single branch
>> makes it easier to remove all the writebacks too - you only need 2 rather than 8!
> The idea was to minimize the length of the loop tail. For branchless 128
> bytes per iteration loop the branchless tail needs to read 128 bytes.
> For the branched loop as the one above the tail processes only 64
> bytes. And I don't see how I can avoid writebacks of up to 127 bytes
> in the branchless tails for your version.

Well you can check whether you need to process more than 64 bytes using separate
code before the loop or in the tail. Jumping into the 2nd half of the loop may be possible
if you want to save code (but then again this memcpy is already quite large...).

Wilco
Anton Youdkevitch April 30, 2019, 12:27 p.m. UTC | #2 
Wilco,

Thanks a lot for you comments and suggestions.

On 18.04.2019 18:22, Wilco Dijkstra wrote:
> Hi Anton,
> 
> Thanks, that's much easier to follow. A few more comments:
> 
> @@ -103,188 +101,14 @@ ENTRY_ALIGN (MEMMOVE, 6)
>   	DELOUSE (1)
>   	DELOUSE (2)
>   
> +	add	srcend, src, count
> +	cmp	count, 16
> +	b.ls	L(memcopy16)
>   	sub	tmp1, dstin, src
>   	cmp	count, 96
>   	ccmp	tmp1, count, 2, hi
>   	b.lo	L(move_long)
> 
> That's 7 instructions, so the memcpy label ends up at an odd alignment.
I added .p2align before memcpy entry to make it 16 byes
aligned.

> +	str	B_q, [dst, 16]
>   	stp	C_q, D_q, [dst, 32]
> -	str     A_q, [dst, 64]
> +	str	G_q, [dst, 64]
> +	str	A_q, [dstin]
> +	str	E_q, [dstend, -16]
> 
> You can use STP rather than STR for B and G by changing the STP of C/D.
Done.

> +	/* 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:
> +	ldr	G_q, [src, 48]
> +	str	B_q, [dstend, -16]
> +	ldr	B_q, [src, 32]
> +	str	A_q, [dstend, -32]
> +	ldr	A_q, [src, 16]
> +	str	D_q, [dstend, -48]
> +	ldr	D_q, [src]
> +	str	C_q, [dstend, -64]
> +	str	G_q, [dstin, 48]
> +	str	B_q, [dstin, 32]
> +	str	A_q, [dstin, 16]
> +	str	D_q, [dstin]
> +3:	ret
> 
> 3 LDP+ 3 STP opportunities here.
Done.

> 
>>> The idea of unrolling 2x is to only have a single loop branch. Using a single branch
>>> makes it easier to remove all the writebacks too - you only need 2 rather than 8!
>> The idea was to minimize the length of the loop tail. For branchless 128
>> bytes per iteration loop the branchless tail needs to read 128 bytes.
>> For the branched loop as the one above the tail processes only 64
>> bytes. And I don't see how I can avoid writebacks of up to 127 bytes
>> in the branchless tails for your version.
> 
> Well you can check whether you need to process more than 64 bytes using separate
> code before the loop or in the tail. Jumping into the 2nd half of the loop may be possible
> if you want to save code (but then again this memcpy is already quite large...).
The check inside the loop is free as it
is done while the data are being brought from the memory.
I can move the check to prologue/epilogue and while this
can still be free the codepath will look less clear as I
need to handle 2 "asymmetric" cases in the epilogue (if
the excessive writebacks are to be avoided): one for <64
bytes and the other >=64 bytes. This also makes the tails
longer.
I can probably merge the tails making the longer case
falls through to the shorter but this makes the things even
less clear.
As there is no real performance nor clarity benefit of using
single branch loop I am inclined to leave the 128 bytes loop
as it is now. Do you think this is reasonable or I'm missing
something?
Wilco Dijkstra May 1, 2019, 2:59 p.m. UTC | #3 
Hi Anton,

> The check inside the loop is free as it
> is done while the data are being brought from the memory.

The loop is used both for small copies and for the tail of
very large copies.  The small copies might well be in the
cache, while the large copies are prefetched.

> I can move the check to prologue/epilogue and while this
> can still be free the codepath will look less clear as I
> need to handle 2 "asymmetric" cases in the epilogue (if
> the excessive writebacks are to be avoided): one for <64
> bytes and the other >=64 bytes. This also makes the tails
> longer.
> I can probably merge the tails making the longer case
> falls through to the shorter but this makes the things even
> less clear.
> As there is no real performance nor clarity benefit of using
> single branch loop I am inclined to leave the 128 bytes loop
> as it is now. Do you think this is reasonable or I'm missing
> something?

It's reasonable for now (and Szabolcs already approved your
latest version). But it is feasible to improve further given that the
memmove loop does 64 bytes per iteration, so if that is fast enough
then that may be a simpler way to handle this loop too.

Wilco
Anton Youdkevitch May 3, 2019, 8:36 a.m. UTC | #4 
Wilco,

On 5/1/2019 17:59, Wilco Dijkstra wrote:
> Hi Anton,
>
>> The check inside the loop is free as it
>> is done while the data are being brought from the memory.
>
> The loop is used both for small copies and for the tail of
> very large copies.  The small copies might well be in the
> cache, while the large copies are prefetched.
Technically you are right. I was not clear enough, perhaps.
What I meant was loads/stores have latency several times
more than branches do. So, even if the data are in the cache
the branching can be processed while loads/stores are still
in the pipeline.

>> I can move the check to prologue/epilogue and while this
>> can still be free the codepath will look less clear as I
>> need to handle 2 "asymmetric" cases in the epilogue (if
>> the excessive writebacks are to be avoided): one for <64
>> bytes and the other >=64 bytes. This also makes the tails
>> longer.
>> I can probably merge the tails making the longer case
>> falls through to the shorter but this makes the things even
>> less clear.
>> As there is no real performance nor clarity benefit of using
>> single branch loop I am inclined to leave the 128 bytes loop
>> as it is now. Do you think this is reasonable or I'm missing
>> something?
>
> It's reasonable for now (and Szabolcs already approved your
> latest version). But it is feasible to improve further given that the
> memmove loop does 64 bytes per iteration, so if that is fast enough
> then that may be a simpler way to handle this loop too.
OK, I will see what I can do.

BTW, the loop in question does 128 bytes per iteration, doesn't it?
Wilco Dijkstra May 7, 2019, 3:07 p.m. UTC | #5 
Hi Anton,

>> The loop is used both for small copies and for the tail of
>> very large copies.  The small copies might well be in the
>> cache, while the large copies are prefetched.
>
> Technically you are right. I was not clear enough, perhaps.
> What I meant was loads/stores have latency several times
> more than branches do. So, even if the data are in the cache
> the branching can be processed while loads/stores are still
> in the pipeline.

OK but if you're assuming the latency is high then what benefit
does unrolling give here? If the loop is 16-byte aligned fetch should
be optimal.

>> It's reasonable for now (and Szabolcs already approved your
>> latest version). But it is feasible to improve further given that the
>> memmove loop does 64 bytes per iteration, so if that is fast enough
>> then that may be a simpler way to handle this loop too.
> OK, I will see what I can do.
>
> BTW, the loop in question does 128 bytes per iteration, doesn't it?

The memmove loop does 64-bytes per iteration in all cases for the
overlap case:

+L(move_long):
...
+	.p2align 4
+1:
+	subs	count, count, 64
+	stp	A_q, B_q, [dstend, -32]
+	ldp	A_q, B_q, [srcend, -32]
+	stp	C_q, D_q, [dstend, -64]!
+	ldp	C_q, D_q, [srcend, -64]!
+	b.hi	1b

Wilco
diff mbox

Patch

From 1b6634e6bb89e9127e025433e9562dbc0e0c8fde Mon Sep 17 00:00:00 2001
From: Anton Youdkevitch <anton.youdkevitch@bell-sw.com>
Date: Wed, 10 Apr 2019 15:18:00 +0000
Subject: [PATCH] [AARCH64] thunderx2 memmove performance improvements

Used SIMD load/store instead of GPR for overlapping
forward move.

Reused existing memcpy implementation for small or
overlapping backward move.

Fixed the existing memcpy implementation to allow it
to deal with the overlapping case.

Simplified loop tails in the memcpy implementation -
use branchless overlapping sequence of fixed length
load/stores instead of branching depending on the
size.
---
 sysdeps/aarch64/multiarch/ifunc-impl-list.c  |   1 +
 sysdeps/aarch64/multiarch/memcpy_thunderx2.S | 372 ++++++++-------------------
 sysdeps/aarch64/multiarch/memmove.c          |   5 +-
 3 files changed, 119 insertions(+), 259 deletions(-)

diff --git a/sysdeps/aarch64/multiarch/ifunc-impl-list.c b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
index 5f43362..10ff7d4 100644
--- a/sysdeps/aarch64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/aarch64/multiarch/ifunc-impl-list.c
@@ -45,6 +45,7 @@  __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_generic))
   IFUNC_IMPL (i, name, memmove,
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx)
+	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_thunderx2)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_falkor)
 	      IFUNC_IMPL_ADD (array, i, memmove, 1, __memmove_generic))
   IFUNC_IMPL (i, name, memset,
diff --git a/sysdeps/aarch64/multiarch/memcpy_thunderx2.S b/sysdeps/aarch64/multiarch/memcpy_thunderx2.S
index 45e9a29..2f099d6 100644
--- a/sysdeps/aarch64/multiarch/memcpy_thunderx2.S
+++ b/sysdeps/aarch64/multiarch/memcpy_thunderx2.S
@@ -85,16 +85,14 @@ 
 #if IS_IN (libc)
 
 #undef MEMCPY
-#undef MEMMOVE
 #define MEMCPY __memcpy_thunderx2
+#undef MEMMOVE
 #define MEMMOVE __memmove_thunderx2
 
 
-/* Moves 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.
-   Overlapping large forward memmoves use a loop that copies backwards.
+/* Overlapping large forward memmoves use a loop that copies backwards.
+   Otherwise memcpy is used. Small moves branch to memcopy16 directly.
+   The longer memcpy cases fall through to the memcpy head.
 */
 
 ENTRY_ALIGN (MEMMOVE, 6)
@@ -103,188 +101,14 @@  ENTRY_ALIGN (MEMMOVE, 6)
 	DELOUSE (1)
 	DELOUSE (2)
 
+	add	srcend, src, count
+	cmp	count, 16
+	b.ls	L(memcopy16)
 	sub	tmp1, dstin, src
 	cmp	count, 96
 	ccmp	tmp1, count, 2, hi
 	b.lo	L(move_long)
 
-	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):
-	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 (MEMMOVE)
 libc_hidden_builtin_def (MEMMOVE)
 
@@ -293,18 +117,15 @@  libc_hidden_builtin_def (MEMMOVE)
    medium copies of 17..96 bytes which are fully unrolled. Large copies
    of more than 96 bytes align the destination and use load-and-merge
    approach in the case src and dst addresses are unaligned not evenly,
-   so that, loads and stores are always aligned.
-   Large copies use an unrolled loop processing 64 bytes per iteration.
-   The current optimized memcpy implementation is not compatible with
-   memmove and is separated from it completely.
-
-   memcpy implementation below is not compatible with memmove
-   because of pipelined loads/stores, which are faster, but they
-   can't be used in the case of overlapping memmove arrays */
+   so that, actual loads and stores are always aligned.
+   Large copies use the loops processing 64 bytes per iteration for
+   unaligned case and 128 bytes per iteration for aligned ones.
+*/
 
 #define MEMCPY_PREFETCH_LDR 640
 
 ENTRY (MEMCPY)
+
 	DELOUSE (0)
 	DELOUSE (1)
 	DELOUSE (2)
@@ -373,29 +194,26 @@  L(bytes_0_to_8):
 	/* 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.  */
 L(bytes_0_to_3):
-	cbz     count, L(end)
+	cbz	count, 1f
 	lsr	tmp1, count, 1
 	ldrb	A_lw, [src]
 	ldrb	A_hw, [srcend, -1]
 	add	dstend, dstin, count
 	ldrb	B_lw, [src, tmp1]
-	strb    A_lw, [dstin]
 	strb	B_lw, [dstin, tmp1]
 	strb	A_hw, [dstend, -1]
-L(end):
+	strb	A_lw, [dstin]
+1:
 	ret
 
 	.p2align 4
 
 L(memcpy_copy96):
 	/* Copying 65..96 bytes. A_q (first 16 bytes) and
-	   E_q(last 16 bytes) are already loaded.
-
-	   The size is large enough to benefit from aligned
-	   loads */
+	   E_q(last 16 bytes) are already loaded. The size
+	   is large enough to benefit from aligned loads */
 	bic	src, src, 15
 	ldp	B_q, C_q, [src]
-	str     A_q, [dstin]
 	/* Loaded 64 bytes, second 16-bytes chunk can be
 	   overlapping with the first chunk by tmp1 bytes.
 	   Stored 16 bytes. */
@@ -408,44 +226,47 @@  L(memcpy_copy96):
 	b.gt	L(copy96_medium)
 	ldr	D_q, [src, 32]
 	stp	B_q, C_q, [dst, 16]
-	str     E_q, [dstend, -16]
 	str	D_q, [dst, 48]
+	str	A_q, [dstin]
+	str	E_q, [dstend, -16]
 	ret
 
 	.p2align 4
 L(copy96_medium):
-	ldp     D_q, A_q, [src, 32]
-	str     B_q, [dst, 16]
+	ldp	D_q, G_q, [src, 32]
 	cmp	count, 96
 	b.gt	L(copy96_large)
-	str     E_q, [dstend, -16]
+	str	B_q, [dst, 16]
 	stp	C_q, D_q, [dst, 32]
-	str     A_q, [dst, 64]
+	str	G_q, [dst, 64]
+	str	A_q, [dstin]
+	str	E_q, [dstend, -16]
 	ret
 
 L(copy96_large):
 	ldr	F_q, [src, 64]
+	str	B_q, [dst, 16]
 	stp	C_q, D_q, [dst, 32]
+	stp	G_q, F_q, [dst, 64]
+	str	A_q, [dstin]
 	str	E_q, [dstend, -16]
-	stp     A_q, F_q, [dst, 64]
 	ret
 
 	.p2align 4
 L(memcopy_long):
 	bic	src, src, 15
 	ldp	B_q, C_q, [src], #32
-	str     A_q, [dstin]
 	sub	dst, dstin, tmp1
 	add	count, count, tmp1
 	add	dst, dst, 16
 	and	tmp1, dst, 15
 	ldp	D_q, E_q, [src], #32
-	str     B_q, [dst], #16
+	str	A_q, [dstin]
 
 	/* Already loaded 64+16 bytes. Check if at
 	   least 64 more bytes left */
 	subs	count, count, 64+64+16
-	b.lt    L(loop128_exit2)
+	b.lt	L(loop128_exit0)
 	cmp	count, MEMCPY_PREFETCH_LDR + 64 + 32
 	b.lt	L(loop128)
 	cbnz	tmp1, L(dst_unaligned)
@@ -454,66 +275,59 @@  L(memcopy_long):
 	.p2align 4
 
 L(loop128_prefetch):
-	str     C_q, [dst], #16
 	prfm	pldl1strm, [src, MEMCPY_PREFETCH_LDR]
-	str     D_q, [dst], #16
 	ldp	F_q, G_q, [src], #32
-	str	E_q, [dst], #16
-	ldp     H_q, A_q, [src], #32
-	str     F_q, [dst], #16
+	stp	B_q, C_q, [dst], #32
+	ldp	H_q, I_q, [src], #32
 	prfm	pldl1strm, [src, MEMCPY_PREFETCH_LDR]
-	str     G_q, [dst], #16
 	ldp	B_q, C_q, [src], #32
-	str	H_q, [dst], #16
+	stp	D_q, E_q, [dst], #32
 	ldp	D_q, E_q, [src], #32
-	stp	A_q, B_q, [dst], #32
+	stp	F_q, G_q, [dst], #32
+	stp	H_q, I_q, [dst], #32
 	subs	count, count, 128
 	b.ge	L(loop128_prefetch)
 
-L(preloop128):
 	add	count, count, MEMCPY_PREFETCH_LDR + 64 + 32
 	.p2align 4
 L(loop128):
 	ldp	F_q, G_q, [src], #32
-	str     C_q, [dst], #16
-	ldp     B_q, A_q, [src], #32
-	str     D_q, [dst], #16
-	stp     E_q, F_q, [dst], #32
-	stp     G_q, B_q, [dst], #32
+	ldp	H_q, I_q, [src], #32
+	stp	B_q, C_q, [dst], #32
+	stp	D_q, E_q, [dst], #32
 	subs	count, count, 64
 	b.lt	L(loop128_exit1)
-L(loop128_proceed):
 	ldp	B_q, C_q, [src], #32
-	str     A_q, [dst], #16
 	ldp	D_q, E_q, [src], #32
-	str     B_q, [dst], #16
+	stp	F_q, G_q, [dst], #32
+	stp	H_q, I_q, [dst], #32
 	subs	count, count, 64
 	b.ge	L(loop128)
-
-	.p2align 4
-L(loop128_exit2):
-	stp     C_q, D_q, [dst], #32
-	str     E_q, [dst], #16
-	b       L(copy_long_check32);
-
+L(loop128_exit0):
+	ldp	F_q, G_q, [srcend, -64]
+	ldp	H_q, I_q, [srcend, -32]
+	stp	B_q, C_q, [dst], #32
+	stp	D_q, E_q, [dst]
+	stp	F_q, G_q, [dstend, -64]
+	stp	H_q, I_q, [dstend, -32]
+	ret
 L(loop128_exit1):
-	/* A_q is still not stored and 0..63 bytes left,
-	   so, count is -64..-1.
-	   Check if less than 32 bytes left (count < -32) */
-	str     A_q, [dst], #16
-L(copy_long_check32):
-	cmn     count, 64
-	b.eq    L(copy_long_done)
-	cmn     count, 32
-	b.le    L(copy_long_last32)
-	ldp     B_q, C_q, [src]
-	stp     B_q, C_q, [dst]
-
-L(copy_long_last32):
-	ldp     F_q, G_q, [srcend, -32]
-	stp     F_q, G_q, [dstend, -32]
+	ldp	B_q, C_q, [srcend, -64]
+	ldp	D_q, E_q, [srcend, -32]
+	stp	F_q, G_q, [dst], #32
+	stp	H_q, I_q, [dst]
+	stp	B_q, C_q, [dstend, -64]
+	stp	D_q, E_q, [dstend, -32]
+	ret
 
-L(copy_long_done):
+L(dst_unaligned_tail):
+	ldp	C_q, D_q, [srcend, -64]
+	ldp	E_q, F_q, [srcend, -32]
+	stp	A_q, B_q, [dst], #32
+	stp	H_q, I_q, [dst], #16
+	str	G_q, [dst, tmp1]
+	stp	C_q, D_q, [dstend, -64]
+	stp	E_q, F_q, [dstend, -32]
 	ret
 
 L(dst_unaligned):
@@ -542,17 +356,20 @@  L(dst_unaligned):
 	/* Store the 16 bytes to dst and align dst for further
 	   operations, several bytes will be stored at this
 	   address once more */
-	str     C_q, [dst], #16
+
 	ldp	F_q, G_q, [src], #32
+	stp	B_q, C_q, [dst], #32
 	bic	dst, dst, 15
-	subs    count, count, 32
+	sub	count, count, 32
 	adrp	tmp2, L(ext_table)
 	add	tmp2, tmp2, :lo12:L(ext_table)
 	add	tmp2, tmp2, tmp1, LSL #2
 	ldr	tmp3w, [tmp2]
 	add	tmp2, tmp2, tmp3w, SXTW
 	br	tmp2
-.p2align 4 ;\
+
+.p2align 4
+	/* to make the loop in each chunk 16-bytes aligned */
 	nop
 #define EXT_CHUNK(shft) \
 L(ext_size_ ## shft):;\
@@ -573,7 +390,7 @@  L(ext_size_ ## shft):;\
 	b.ge    1b;\
 2:;\
 	ext     I_v.16b, F_v.16b, G_v.16b, 16-shft;\
-	b	L(ext_tail);
+	b	L(dst_unaligned_tail);
 
 EXT_CHUNK(1)
 EXT_CHUNK(2)
@@ -591,12 +408,51 @@  EXT_CHUNK(13)
 EXT_CHUNK(14)
 EXT_CHUNK(15)
 
-L(ext_tail):
-	stp     A_q, B_q, [dst], #32
-	stp     H_q, I_q, [dst], #16
-	add     dst, dst, tmp1
-	str     G_q, [dst], #16
-	b       L(copy_long_check32)
+L(move_long):
+	.p2align 4
+1:
+	cbz	tmp1, 3f
+
+	add	srcend, src, count
+	add	dstend, dstin, count
+
+	and	tmp1, srcend, 15
+	ldr	D_q, [srcend, -16]
+	sub	srcend, srcend, tmp1
+	sub	count, count, tmp1
+	ldp	A_q, B_q, [srcend, -32]
+	str	D_q, [dstend, -16]
+	ldp	C_q, D_q, [srcend, -64]!
+	sub	dstend, dstend, tmp1
+	subs	count, count, 128
+	b.ls	2f
+
+	.p2align 4
+1:
+	subs	count, count, 64
+	stp	A_q, B_q, [dstend, -32]
+	ldp	A_q, B_q, [srcend, -32]
+	stp	C_q, D_q, [dstend, -64]!
+	ldp	C_q, D_q, [srcend, -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:
+	ldr	G_q, [src, 48]
+	str	B_q, [dstend, -16]
+	ldr	B_q, [src, 32]
+	str	A_q, [dstend, -32]
+	ldr	A_q, [src, 16]
+	str	D_q, [dstend, -48]
+	ldr	D_q, [src]
+	str	C_q, [dstend, -64]
+	str	G_q, [dstin, 48]
+	str	B_q, [dstin, 32]
+	str	A_q, [dstin, 16]
+	str	D_q, [dstin]
+3:	ret
 
 
 END (MEMCPY)
diff --git a/sysdeps/aarch64/multiarch/memmove.c b/sysdeps/aarch64/multiarch/memmove.c
index f58dde3..f3d341b 100644
--- a/sysdeps/aarch64/multiarch/memmove.c
+++ b/sysdeps/aarch64/multiarch/memmove.c
@@ -30,6 +30,7 @@  extern __typeof (__redirect_memmove) __libc_memmove;
 
 extern __typeof (__redirect_memmove) __memmove_generic attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_thunderx attribute_hidden;
+extern __typeof (__redirect_memmove) __memmove_thunderx2 attribute_hidden;
 extern __typeof (__redirect_memmove) __memmove_falkor attribute_hidden;
 
 libc_ifunc (__libc_memmove,
@@ -37,7 +38,9 @@  libc_ifunc (__libc_memmove,
 	     ? __memmove_thunderx
 	     : (IS_FALKOR (midr) || IS_PHECDA (midr)
 		? __memmove_falkor
-		: __memmove_generic)));
+		: (IS_THUNDERX2 (midr) || IS_THUNDERX2PA (midr)
+		  ? __memmove_thunderx2
+		  : __memmove_generic))));
 
 # undef memmove
 strong_alias (__libc_memmove, memmove);
-- 
2.7.4