[PATCH 2/5] arm: Implement armv6 optimized strcpy

[Richard Henderson <rth@twiddle.net>]

Four times faster than the byte-by-byte default version.
---
 ports/sysdeps/arm/armv6/stpcpy.S |   1 +
 ports/sysdeps/arm/armv6/strcpy.S | 218 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 219 insertions(+)
 create mode 100644 ports/sysdeps/arm/armv6/stpcpy.S
 create mode 100644 ports/sysdeps/arm/armv6/strcpy.S

diff --git a/ports/sysdeps/arm/armv6/stpcpy.S b/ports/sysdeps/arm/armv6/stpcpy.S
new file mode 100644
index 0000000..21a4f38
--- /dev/null
+++ b/ports/sysdeps/arm/armv6/stpcpy.S
@@ -0,0 +1 @@
+/* Defined in strcpy.S.  */
diff --git a/ports/sysdeps/arm/armv6/strcpy.S b/ports/sysdeps/arm/armv6/strcpy.S
new file mode 100644
index 0000000..41f6443
--- /dev/null
+++ b/ports/sysdeps/arm/armv6/strcpy.S
@@ -0,0 +1,218 @@
+/* strcpy -- copy a nul-terminated string.
+   Copyright (C) 2013 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 <sysdep.h>
+
+/* Endian independent macros for shifting bytes within registers.  */
+#ifdef __ARMEB__
+#define lsh_gt		lsr
+#define lsh_ls		lsl
+#else
+#define lsh_gt		lsl
+#define lsh_ls		lsr
+#endif
+
+	.syntax unified
+	.text
+
+ENTRY (__stpcpy)
+	@ Signal stpcpy with NULL in IP.
+	mov	ip, #0
+	b	0f
+END (__stpcpy)
+
+weak_alias (__stpcpy, stpcpy)
+libc_hidden_def (__stpcpy)
+libc_hidden_builtin_def (stpcpy)
+
+ENTRY (strcpy)
+	@ Signal strcpy with DEST in IP.
+	mov	ip, r0
+0:
+	pld	[r0]
+	pld	[r1]
+
+	@ To cater to long strings, we want 8 byte alignment in the source.
+	@ To cater to small strings, we don't want to start that right away.
+	@ Loop up to 16 times, less whatever it takes to reach alignment.
+	and	r3, r1, #7
+	rsb	r3, r3, #16
+
+	@ Loop until we find ...
+1:	ldrb	r2, [r1], #1
+	subs	r3, r3, #1		@ ... the alignment point
+	strb	r2, [r0], #1
+	it	ne
+	cmpne	r2, #0			@ ... or EOS
+	bne	1b
+
+	@ Disambiguate the exit possibilites above
+	cmp	r2, #0			@ Found EOS
+	beq	.Lreturn
+
+	@ Load the next two words asap
+	ldrd	r2, r3, [r1], #8
+	pld	[r0, #64]
+	pld	[r1, #64]
+
+	@ For longer strings, we actaully need a stack frame.
+	push	{ r4, r5, r6, r7 }
+	cfi_adjust_cfa_offset (16)
+	cfi_rel_offset (r4, 0)
+	cfi_rel_offset (r5, 4)
+	cfi_rel_offset (r6, 8)
+	cfi_rel_offset (r7, 12)
+
+	@ Subtracting (unsigned saturating) from 1 for any byte means result
+	@ of 1 for any byte that was originally zero and 0 otherwise.
+	@ Therefore we consider the lsb of each byte the "found" bit.
+#ifdef ARCH_HAS_T2
+	movw	r7, #0x0101
+	tst	r0, #3			@ Test alignment of DEST
+	movt	r7, #0x0101
+#else
+	ldr	ip, =0x01010101
+	tst	r0, #3
+#endif
+	bne	.Lunaligned
+
+	@ So now source (r1) is aligned to 8, and dest (r0) is aligned to 4.
+	@ Loop, reading 8 bytes at a time, searching for EOS.
+	.balign	16
+2:	uqsub8	r4, r7, r2		@ Find EOS
+	uqsub8	r5, r7, r3
+	pld	[r1, #128]
+	cmp	r4, #0			@ EOS in first word?
+	pld	[r0, #128]
+	bne	3f
+	str	r2, [r0], #4
+	cmp	r5, #0			@ EOS in second word?
+	bne	4f
+	str	r3, [r0], #4
+	ldrd	r2, r3, [r1], #8
+	b	2b
+
+3:	sub	r1, r1, #4		@ backup to first word
+4:	sub	r1, r1, #4		@ backup to second word
+
+	@ ... then finish up any tail a byte at a time.
+	@ Note that we generally back up and re-read source bytes,
+	@ but we'll not re-write dest bytes.
+.Lbyte_loop:
+	ldrb	r2, [r1], #1
+	cmp	r2, #0
+	strb	r2, [r0], #1
+	bne	.Lbyte_loop
+
+	pop	{ r4, r5, r6, r7 }
+	cfi_remember_state
+	cfi_adjust_cfa_offset (-16)
+	cfi_restore (r4)
+	cfi_restore (r5)
+	cfi_restore (r6)
+	cfi_restore (r7)
+
+.Lreturn:
+	cmp	ip, #0			@ Was this strcpy or stpcpy?
+	ite	eq
+	subeq	r0, r0, #1		@ stpcpy: undo post-inc from store
+	movne	r0, ip			@ strcpy: return original dest
+	bx	lr
+
+.Lunaligned:
+	cfi_restore_state
+	@ Here, source is aligned to 8, but the destination is not word
+	@ aligned.  Therefore we have to shift the data in order to be
+	@ able to perform aligned word stores.
+
+	@ Find out which misalignment we're dealing with.
+	tst	r0, #1
+	beq	.Lunaligned2
+	tst	r0, #2
+	bne	.Lunaligned3
+	@ Fallthru to .Lunaligned1.
+
+.macro unaligned_copy	unalign
+	@ Prologue to unaligned loop.  Seed shifted non-zero bytes.
+	uqsub8	r4, r7, r2		@ Find EOS
+	uqsub8	r5, r7, r3
+	mvns	r4, r4			@ EOS in first word?
+	it	ne
+	subne	r1, r1, #8
+	bne	.Lbyte_loop
+#ifdef __ARMEB__
+	rev	r2, r2			@ Byte stores below need LE data
+#endif
+	@ Store a few bytes from the first word.
+	@ At the same time we align r0 and shift out bytes from r2.
+.rept	4-\unalign
+	strb	r2, [r0], #1
+	lsr	r2, r2, #8
+.endr
+#ifdef __ARMEB__
+	rev	r2, r2			@ Undo previous rev
+#endif
+	@ Rotated unaligned copy loop.  The tail of the prologue is
+	@ shared with the loop itself.
+	.balign 8
+1:	mvns	r5, r5			@ EOS in second word?
+	bne	4f
+	@ Combine first and second words
+	orr	r2, r2, r3, lsh_gt #(\unalign*8)
+	@ Save leftover bytes from the two words
+	lsh_ls	r6, r3, #((4-\unalign)*8)
+	str	r2, [r0], #4
+	@ The "real" start of the unaligned copy loop.
+	ldrd	r2, r3, [r1], #8	@ Load 8 more bytes
+	uqsub8	r4, r7, r2		@ Find EOS
+	pld	[r1, #128]
+	uqsub8	r5, r7, r3
+	pld	[r0, #128]
+	mvns	r4, r4			@ EOS in first word?
+	bne	3f
+	@ Combine the leftover and the first word
+	orr	r6, r6, r2, lsh_gt #(\unalign*8)
+	@ Discard used bytes from the first word.
+	lsh_ls	r2, r2, #((4-\unalign)*8)
+	str	r6, [r0], #4
+	b	1b
+	@ Found EOS in one of the words; adjust backward
+3:	sub	r1, r1, #4
+	mov	r2, r6
+4:	sub	r1, r1, #4
+	@ And store the remaining bytes from the leftover
+#ifdef __ARMEB__
+	rev	r2, r2
+#endif
+.rept	\unalign
+	strb	r2, [r0], #1
+	lsr	r2, r2, #8
+.endr
+	b	.Lbyte_loop
+.endm
+
+.Lunaligned1:
+	unaligned_copy	1
+.Lunaligned2:
+	unaligned_copy	2
+.Lunaligned3:
+	unaligned_copy	3
+
+END (strcpy)
+
+libc_hidden_builtin_def (strcpy)
-- 
1.8.1.2