[musl] [PATCH] aarch64: add optimized memcpy, memmove and memset

[musl] [PATCH] aarch64: add optimized memcpy, memmove and memset

[Szabolcs Nagy]

[-- Attachment #1: Type: text/plain, Size: 600 bytes --]

minimal edits to upstream version for easier updates
and because this code was benchmarked across many cores.

gcc generates slow code for the current c implementations.

the integer memcpy was chosen instead of the simd one,
this performs better on little cores, i think this is
the more conservative choice for now.

note: there are upcoming security architectures which
may mean updates to these functions (BTI - landing pads,
PAUTH - return address signing, MTE - 16byte tag granule
may affect optimized strcmp etc, not relevant yet), but
runtime support for these will need other libc changes.


[-- Attachment #2: 0001-aarch64-add-optimized-memcpy-memmove-and-memset.patch --]
[-- Type: text/x-diff, Size: 10107 bytes --]

From 633a481dbed9923cab1b3d5985eb4ebd8780f12d Mon Sep 17 00:00:00 2001
From: Szabolcs Nagy <nsz@port70.net>
Date: Wed, 25 Mar 2020 21:30:00 +0000
Subject: [PATCH] aarch64: add optimized memcpy, memmove and memset

The implementations are from

https://github.com/ARM-software/optimized-routines

at commit a0ad28dbcb9d52be85dfe622f4c43d9262641300
only the include path and function names are modified in the files

string/aarch64/memcpy.S
string/aarch64/memset.S

and memmove is an alias of memcpy.
---
 src/string/aarch64/asmdefs.h |  31 +++++
 src/string/aarch64/memcpy.S  | 239 +++++++++++++++++++++++++++++++++++
 src/string/aarch64/memmove.s |   1 +
 src/string/aarch64/memset.S  | 114 +++++++++++++++++
 4 files changed, 385 insertions(+)
 create mode 100644 src/string/aarch64/asmdefs.h
 create mode 100644 src/string/aarch64/memcpy.S
 create mode 100644 src/string/aarch64/memmove.s
 create mode 100644 src/string/aarch64/memset.S

diff --git a/src/string/aarch64/asmdefs.h b/src/string/aarch64/asmdefs.h
new file mode 100644
index 00000000..7d143a96
--- /dev/null
+++ b/src/string/aarch64/asmdefs.h
@@ -0,0 +1,31 @@
+/*
+ * Macros for asm code.
+ *
+ * Copyright (c) 2019, Arm Limited.
+ * SPDX-License-Identifier: MIT
+ */
+
+#ifndef _ASMDEFS_H
+#define _ASMDEFS_H
+
+#define ENTRY_ALIGN(name, alignment)	\
+  .global name;		\
+  .type name,%function;	\
+  .align alignment;		\
+  name:			\
+  .cfi_startproc;
+
+#define ENTRY(name)	ENTRY_ALIGN(name, 6)
+
+#define ENTRY_ALIAS(name)	\
+  .global name;		\
+  .type name,%function;	\
+  name:
+
+#define END(name)	\
+  .cfi_endproc;		\
+  .size name, .-name;
+
+#define L(l) .L ## l
+
+#endif
diff --git a/src/string/aarch64/memcpy.S b/src/string/aarch64/memcpy.S
new file mode 100644
index 00000000..5e76e13a
--- /dev/null
+++ b/src/string/aarch64/memcpy.S
@@ -0,0 +1,239 @@
+/*
+ * memcpy - copy memory area
+ *
+ * Copyright (c) 2012-2020, Arm Limited.
+ * SPDX-License-Identifier: MIT
+ */
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64, unaligned accesses.
+ *
+ */
+
+#include "asmdefs.h"
+
+#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 B_l	x8
+#define B_lw	w8
+#define B_h	x9
+#define C_l	x10
+#define C_lw	w10
+#define C_h	x11
+#define D_l	x12
+#define D_h	x13
+#define E_l	x14
+#define E_h	x15
+#define F_l	x16
+#define F_h	x17
+#define G_l	count
+#define G_h	dst
+#define H_l	src
+#define H_h	srcend
+#define tmp1	x14
+
+/* This implementation handles overlaps and supports both memcpy and memmove
+   from a single entry point.  It uses unaligned accesses and branchless
+   sequences to keep the code small, simple and improve performance.
+
+   Copies are split into 3 main cases: small copies of up to 32 bytes, medium
+   copies of up to 128 bytes, and large copies.  The overhead of the overlap
+   check is negligible since it is only required for large copies.
+
+   Large copies use a software pipelined loop processing 64 bytes per iteration.
+   The destination pointer is 16-byte aligned to minimize unaligned accesses.
+   The loop tail is handled by always copying 64 bytes from the end.
+*/
+
+ENTRY (memcpy)
+ENTRY_ALIAS (memmove)
+	add	srcend, src, count
+	add	dstend, dstin, count
+	cmp	count, 128
+	b.hi	L(copy_long)
+	cmp	count, 32
+	b.hi	L(copy32_128)
+
+	/* Small copies: 0..32 bytes.  */
+	cmp	count, 16
+	b.lo	L(copy16)
+	ldp	A_l, A_h, [src]
+	ldp	D_l, D_h, [srcend, -16]
+	stp	A_l, A_h, [dstin]
+	stp	D_l, D_h, [dstend, -16]
+	ret
+
+	/* Copy 8-15 bytes.  */
+L(copy16):
+	tbz	count, 3, L(copy8)
+	ldr	A_l, [src]
+	ldr	A_h, [srcend, -8]
+	str	A_l, [dstin]
+	str	A_h, [dstend, -8]
+	ret
+
+	.p2align 3
+	/* Copy 4-7 bytes.  */
+L(copy8):
+	tbz	count, 2, L(copy4)
+	ldr	A_lw, [src]
+	ldr	B_lw, [srcend, -4]
+	str	A_lw, [dstin]
+	str	B_lw, [dstend, -4]
+	ret
+
+	/* Copy 0..3 bytes using a branchless sequence.  */
+L(copy4):
+	cbz	count, L(copy0)
+	lsr	tmp1, count, 1
+	ldrb	A_lw, [src]
+	ldrb	C_lw, [srcend, -1]
+	ldrb	B_lw, [src, tmp1]
+	strb	A_lw, [dstin]
+	strb	B_lw, [dstin, tmp1]
+	strb	C_lw, [dstend, -1]
+L(copy0):
+	ret
+
+	.p2align 4
+	/* Medium copies: 33..128 bytes.  */
+L(copy32_128):
+	ldp	A_l, A_h, [src]
+	ldp	B_l, B_h, [src, 16]
+	ldp	C_l, C_h, [srcend, -32]
+	ldp	D_l, D_h, [srcend, -16]
+	cmp	count, 64
+	b.hi	L(copy128)
+	stp	A_l, A_h, [dstin]
+	stp	B_l, B_h, [dstin, 16]
+	stp	C_l, C_h, [dstend, -32]
+	stp	D_l, D_h, [dstend, -16]
+	ret
+
+	.p2align 4
+	/* Copy 65..128 bytes.  */
+L(copy128):
+	ldp	E_l, E_h, [src, 32]
+	ldp	F_l, F_h, [src, 48]
+	cmp	count, 96
+	b.ls	L(copy96)
+	ldp	G_l, G_h, [srcend, -64]
+	ldp	H_l, H_h, [srcend, -48]
+	stp	G_l, G_h, [dstend, -64]
+	stp	H_l, H_h, [dstend, -48]
+L(copy96):
+	stp	A_l, A_h, [dstin]
+	stp	B_l, B_h, [dstin, 16]
+	stp	E_l, E_h, [dstin, 32]
+	stp	F_l, F_h, [dstin, 48]
+	stp	C_l, C_h, [dstend, -32]
+	stp	D_l, D_h, [dstend, -16]
+	ret
+
+	.p2align 4
+	/* Copy more than 128 bytes.  */
+L(copy_long):
+	/* Use backwards copy if there is an overlap.  */
+	sub	tmp1, dstin, src
+	cbz	tmp1, L(copy0)
+	cmp	tmp1, count
+	b.lo	L(copy_long_backwards)
+
+	/* Copy 16 bytes and then align dst to 16-byte alignment.  */
+
+	ldp	D_l, D_h, [src]
+	and	tmp1, dstin, 15
+	bic	dst, dstin, 15
+	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(copy64_from_end)
+
+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 iteration and copy 64 bytes from the end.  */
+L(copy64_from_end):
+	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
+
+	/* Large backwards copy for overlapping copies.
+	   Copy 16 bytes and then align dst to 16-byte alignment.  */
+L(copy_long_backwards):
+	ldp	D_l, D_h, [srcend, -16]
+	and	tmp1, dstend, 15
+	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	L(copy64_from_start)
+
+L(loop64_backwards):
+	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	L(loop64_backwards)
+
+	/* Write the last iteration and copy 64 bytes from the start.  */
+L(copy64_from_start):
+	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]
+	ret
+
+END (memcpy)
diff --git a/src/string/aarch64/memmove.s b/src/string/aarch64/memmove.s
new file mode 100644
index 00000000..0dc2941f
--- /dev/null
+++ b/src/string/aarch64/memmove.s
@@ -0,0 +1 @@
+// see memcpy.S
diff --git a/src/string/aarch64/memset.S b/src/string/aarch64/memset.S
new file mode 100644
index 00000000..7aca4a1e
--- /dev/null
+++ b/src/string/aarch64/memset.S
@@ -0,0 +1,114 @@
+/*
+ * memset - fill memory with a constant byte
+ *
+ * Copyright (c) 2012-2020, Arm Limited.
+ * SPDX-License-Identifier: MIT
+ */
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64, Advanced SIMD, unaligned accesses.
+ *
+ */
+
+#include "asmdefs.h"
+
+#define dstin	x0
+#define val	x1
+#define valw	w1
+#define count	x2
+#define dst	x3
+#define dstend	x4
+#define zva_val	x5
+
+ENTRY (memset)
+
+	dup	v0.16B, valw
+	add	dstend, dstin, count
+
+	cmp	count, 96
+	b.hi	L(set_long)
+	cmp	count, 16
+	b.hs	L(set_medium)
+	mov	val, v0.D[0]
+
+	/* Set 0..15 bytes.  */
+	tbz	count, 3, 1f
+	str	val, [dstin]
+	str	val, [dstend, -8]
+	ret
+	nop
+1:	tbz	count, 2, 2f
+	str	valw, [dstin]
+	str	valw, [dstend, -4]
+	ret
+2:	cbz	count, 3f
+	strb	valw, [dstin]
+	tbz	count, 1, 3f
+	strh	valw, [dstend, -2]
+3:	ret
+
+	/* Set 17..96 bytes.  */
+L(set_medium):
+	str	q0, [dstin]
+	tbnz	count, 6, L(set96)
+	str	q0, [dstend, -16]
+	tbz	count, 5, 1f
+	str	q0, [dstin, 16]
+	str	q0, [dstend, -32]
+1:	ret
+
+	.p2align 4
+	/* Set 64..96 bytes.  Write 64 bytes from the start and
+	   32 bytes from the end.  */
+L(set96):
+	str	q0, [dstin, 16]
+	stp	q0, q0, [dstin, 32]
+	stp	q0, q0, [dstend, -32]
+	ret
+
+	.p2align 4
+L(set_long):
+	and	valw, valw, 255
+	bic	dst, dstin, 15
+	str	q0, [dstin]
+	cmp	count, 160
+	ccmp	valw, 0, 0, hs
+	b.ne	L(no_zva)
+
+#ifndef SKIP_ZVA_CHECK
+	mrs	zva_val, dczid_el0
+	and	zva_val, zva_val, 31
+	cmp	zva_val, 4		/* ZVA size is 64 bytes.  */
+	b.ne	L(no_zva)
+#endif
+	str	q0, [dst, 16]
+	stp	q0, q0, [dst, 32]
+	bic	dst, dst, 63
+	sub	count, dstend, dst	/* Count is now 64 too large.  */
+	sub	count, count, 128	/* Adjust count and bias for loop.  */
+
+	.p2align 4
+L(zva_loop):
+	add	dst, dst, 64
+	dc	zva, dst
+	subs	count, count, 64
+	b.hi	L(zva_loop)
+	stp	q0, q0, [dstend, -64]
+	stp	q0, q0, [dstend, -32]
+	ret
+
+L(no_zva):
+	sub	count, dstend, dst	/* Count is 16 too large.  */
+	sub	dst, dst, 16		/* Dst is biased by -32.  */
+	sub	count, count, 64 + 16	/* Adjust count and bias for loop.  */
+L(no_zva_loop):
+	stp	q0, q0, [dst, 32]
+	stp	q0, q0, [dst, 64]!
+	subs	count, count, 64
+	b.hi	L(no_zva_loop)
+	stp	q0, q0, [dstend, -64]
+	stp	q0, q0, [dstend, -32]
+	ret
+
+END (memset)
-- 
2.24.1

Re: [musl] [PATCH] aarch64: add optimized memcpy, memmove and memset

[Rich Felker]

On Wed, Mar 25, 2020 at 10:45:45PM +0100, Szabolcs Nagy wrote:
> minimal edits to upstream version for easier updates
> and because this code was benchmarked across many cores.
> 
> gcc generates slow code for the current c implementations.
> 
> the integer memcpy was chosen instead of the simd one,
> this performs better on little cores, i think this is
> the more conservative choice for now.

I think this was discussed before on IRC, and I'm not particularly
opposed to these especially since aarch64 is one of the most important
archs these days. However I would really like to avoid adding more asm
source files with the function flow written in asm when the only thing
that really needs to benefit from asm is the inner loop body. I know
nothing has happened on this front since we last talked about it, so
it's very possible that the answer is just "we need something with
decent performance in the short term and nobody has cycles to spend on
doing it better right now and so we we should just use the asm
files"...

> note: there are upcoming security architectures which
> may mean updates to these functions (BTI - landing pads,
> PAUTH - return address signing, MTE - 16byte tag granule
> may affect optimized strcmp etc, not relevant yet), but
> runtime support for these will need other libc changes.

If these mattered they'd be another reason to prefer having the
function in C with minimal inline asm or just extensions for unaligned
loads/stores, but MTE is the only one of these that's interesting and
it doesn't conflict with any current code in musl at all (nothing does
unaligned overreads; they have to be assumed to be able to fault
anyway).

Rich