[thirdparty/glibc.git] / sysdeps / aarch64 / strncmp.S

/* Copyright (C) 2013-2015 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>

/* Assumptions:
 *
 * ARMv8-a, AArch64
 */

#define REP8_01 0x0101010101010101
#define REP8_7f 0x7f7f7f7f7f7f7f7f
#define REP8_80 0x8080808080808080

/* Parameters and result.  */
#define src1		x0
#define src2		x1
#define limit		x2
#define result		x0

/* Internal variables.  */
#define data1		x3
#define data1w		w3
#define data2		x4
#define data2w		w4
#define has_nul		x5
#define diff		x6
#define syndrome	x7
#define tmp1		x8
#define tmp2		x9
#define tmp3		x10
#define zeroones	x11
#define pos		x12
#define limit_wd	x13
#define mask		x14
#define endloop		x15

ENTRY_ALIGN_AND_PAD (strncmp, 6, 7)
	cbz	limit, L(ret0)
	eor	tmp1, src1, src2
	mov	zeroones, #REP8_01
	tst	tmp1, #7
	b.ne	L(misaligned8)
	ands	tmp1, src1, #7
	b.ne	L(mutual_align)
	/* Calculate the number of full and partial words -1.  */
	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
	lsr	limit_wd, limit_wd, #3	/* Convert to Dwords.  */

	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
	   (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
	   can be done in parallel across the entire word.  */
	/* Start of performance-critical section  -- one 64B cache line.  */
L(loop_aligned):
	ldr	data1, [src1], #8
	ldr	data2, [src2], #8
L(start_realigned):
	subs	limit_wd, limit_wd, #1
	sub	tmp1, data1, zeroones
	orr	tmp2, data1, #REP8_7f
	eor	diff, data1, data2	/* Non-zero if differences found.  */
	csinv	endloop, diff, xzr, pl	/* Last Dword or differences.  */
	bics	has_nul, tmp1, tmp2	/* Non-zero if NUL terminator.  */
	ccmp	endloop, #0, #0, eq
	b.eq	L(loop_aligned)
	/* End of performance-critical section  -- one 64B cache line.  */

	/* Not reached the limit, must have found the end or a diff.  */
	tbz	limit_wd, #63, L(not_limit)

	/* Limit % 8 == 0 => all bytes significant.  */
	ands	limit, limit, #7
	b.eq	L(not_limit)

	lsl	limit, limit, #3	/* Bits -> bytes.  */
	mov	mask, #~0
#ifdef __AARCH64EB__
	lsr	mask, mask, limit
#else
	lsl	mask, mask, limit
#endif
	bic	data1, data1, mask
	bic	data2, data2, mask

	/* Make sure that the NUL byte is marked in the syndrome.  */
	orr	has_nul, has_nul, mask

L(not_limit):
	orr	syndrome, diff, has_nul

#ifndef	__AARCH64EB__
	rev	syndrome, syndrome
	rev	data1, data1
	/* The MS-non-zero bit of the syndrome marks either the first bit
	   that is different, or the top bit of the first zero byte.
	   Shifting left now will bring the critical information into the
	   top bits.  */
	clz	pos, syndrome
	rev	data2, data2
	lsl	data1, data1, pos
	lsl	data2, data2, pos
	/* But we need to zero-extend (char is unsigned) the value and then
	   perform a signed 32-bit subtraction.  */
	lsr	data1, data1, #56
	sub	result, data1, data2, lsr #56
	RET
#else
	/* For big-endian we cannot use the trick with the syndrome value
	   as carry-propagation can corrupt the upper bits if the trailing
	   bytes in the string contain 0x01.  */
	/* However, if there is no NUL byte in the dword, we can generate
	   the result directly.  We can't just subtract the bytes as the
	   MSB might be significant.  */
	cbnz	has_nul, 1f
	cmp	data1, data2
	cset	result, ne
	cneg	result, result, lo
	RET
1:
	/* Re-compute the NUL-byte detection, using a byte-reversed value.  */
	rev	tmp3, data1
	sub	tmp1, tmp3, zeroones
	orr	tmp2, tmp3, #REP8_7f
	bic	has_nul, tmp1, tmp2
	rev	has_nul, has_nul
	orr	syndrome, diff, has_nul
	clz	pos, syndrome
	/* The MS-non-zero bit of the syndrome marks either the first bit
	   that is different, or the top bit of the first zero byte.
	   Shifting left now will bring the critical information into the
	   top bits.  */
	lsl	data1, data1, pos
	lsl	data2, data2, pos
	/* But we need to zero-extend (char is unsigned) the value and then
	   perform a signed 32-bit subtraction.  */
	lsr	data1, data1, #56
	sub	result, data1, data2, lsr #56
	RET
#endif

L(mutual_align):
	/* Sources are mutually aligned, but are not currently at an
	   alignment boundary.  Round down the addresses and then mask off
	   the bytes that precede the start point.
	   We also need to adjust the limit calculations, but without
	   overflowing if the limit is near ULONG_MAX.  */
	bic	src1, src1, #7
	bic	src2, src2, #7
	ldr	data1, [src1], #8
	neg	tmp3, tmp1, lsl #3	/* 64 - bits(bytes beyond align). */
	ldr	data2, [src2], #8
	mov	tmp2, #~0
	sub	limit_wd, limit, #1	/* limit != 0, so no underflow.  */
#ifdef __AARCH64EB__
	/* Big-endian.  Early bytes are at MSB.  */
	lsl	tmp2, tmp2, tmp3	/* Shift (tmp1 & 63).  */
#else
	/* Little-endian.  Early bytes are at LSB.  */
	lsr	tmp2, tmp2, tmp3	/* Shift (tmp1 & 63).  */
#endif
	and	tmp3, limit_wd, #7
	lsr	limit_wd, limit_wd, #3
	/* Adjust the limit. Only low 3 bits used, so overflow irrelevant.  */
	add	limit, limit, tmp1
	add	tmp3, tmp3, tmp1
	orr	data1, data1, tmp2
	orr	data2, data2, tmp2
	add	limit_wd, limit_wd, tmp3, lsr #3
	b	L(start_realigned)

L(ret0):
	mov	result, #0
	RET

	.p2align 6
L(misaligned8):
	sub	limit, limit, #1
1:
	/* Perhaps we can do better than this.  */
	ldrb	data1w, [src1], #1
	ldrb	data2w, [src2], #1
	subs	limit, limit, #1
	ccmp	data1w, #1, #0, cs	/* NZCV = 0b0000.  */
	ccmp	data1w, data2w, #0, cs	/* NZCV = 0b0000.  */
	b.eq	1b
	sub	result, data1, data2
	RET
END (strncmp)
libc_hidden_builtin_def (strncmp)
Commit	Line	Data
b168057a	1	/* Copyright (C) 2013-2015 Free Software Foundation, Inc.
a0b1cd88 MS	2
	3	This file is part of the GNU C Library.
	4
	5	The GNU C Library is free software; you can redistribute it and/or
	6	modify it under the terms of the GNU Lesser General Public
	7	License as published by the Free Software Foundation; either
	8	version 2.1 of the License, or (at your option) any later version.
	9
	10	The GNU C Library is distributed in the hope that it will be useful,
	11	but WITHOUT ANY WARRANTY; without even the implied warranty of
	12	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	13	Lesser General Public License for more details.
	14
	15	You should have received a copy of the GNU Lesser General Public
	16	License along with the GNU C Library. If not, see
	17	<http://www.gnu.org/licenses/>. */
	18
	19	#include <sysdep.h>
	20
	21	/* Assumptions:
	22	*
	23	* ARMv8-a, AArch64
	24	*/
	25
	26	#define REP8_01 0x0101010101010101
	27	#define REP8_7f 0x7f7f7f7f7f7f7f7f
	28	#define REP8_80 0x8080808080808080
	29
	30	/* Parameters and result. */
	31	#define src1 x0
	32	#define src2 x1
	33	#define limit x2
	34	#define result x0
	35
	36	/* Internal variables. */
	37	#define data1 x3
	38	#define data1w w3
	39	#define data2 x4
	40	#define data2w w4
	41	#define has_nul x5
	42	#define diff x6
	43	#define syndrome x7
	44	#define tmp1 x8
	45	#define tmp2 x9
	46	#define tmp3 x10
	47	#define zeroones x11
	48	#define pos x12
	49	#define limit_wd x13
	50	#define mask x14
	51	#define endloop x15
	52
	53	ENTRY_ALIGN_AND_PAD (strncmp, 6, 7)
	54	cbz limit, L(ret0)
	55	eor tmp1, src1, src2
	56	mov zeroones, #REP8_01
	57	tst tmp1, #7
	58	b.ne L(misaligned8)
	59	ands tmp1, src1, #7
	60	b.ne L(mutual_align)
	61	/* Calculate the number of full and partial words -1. */
	62	sub limit_wd, limit, #1 /* limit != 0, so no underflow. */
	63	lsr limit_wd, limit_wd, #3 /* Convert to Dwords. */
	64
	65	/* NUL detection works on the principle that (X - 1) & (~X) & 0x80
66	(=> (X - 1) & ~(X \| 0x7f)) is non-zero iff a byte is zero, and
67	can be done in parallel across the entire word. */
68	/* Start of performance-critical section -- one 64B cache line. */
69	L(loop_aligned):
70	ldr data1, [src1], #8
71	ldr data2, [src2], #8
72	L(start_realigned):
73	subs limit_wd, limit_wd, #1
74	sub tmp1, data1, zeroones
75	orr tmp2, data1, #REP8_7f
76	eor diff, data1, data2 /* Non-zero if differences found. */
77	csinv endloop, diff, xzr, pl /* Last Dword or differences. */
78	bics has_nul, tmp1, tmp2 /* Non-zero if NUL terminator. */
79	ccmp endloop, #0, #0, eq
80	b.eq L(loop_aligned)
81	/* End of performance-critical section -- one 64B cache line. */
82
83	/* Not reached the limit, must have found the end or a diff. */
84	tbz limit_wd, #63, L(not_limit)
85
86	/* Limit % 8 == 0 => all bytes significant. */
87	ands limit, limit, #7
88	b.eq L(not_limit)
89
90	lsl limit, limit, #3 /* Bits -> bytes. */
91	mov mask, #~0
92	#ifdef __AARCH64EB__
93	lsr mask, mask, limit
94	#else
95	lsl mask, mask, limit
96	#endif
97	bic data1, data1, mask
98	bic data2, data2, mask
99
100	/* Make sure that the NUL byte is marked in the syndrome. */
101	orr has_nul, has_nul, mask
102
103	L(not_limit):
104	orr syndrome, diff, has_nul
105
106	#ifndef __AARCH64EB__
107	rev syndrome, syndrome
108	rev data1, data1
109	/* The MS-non-zero bit of the syndrome marks either the first bit
110	that is different, or the top bit of the first zero byte.
111	Shifting left now will bring the critical information into the
112	top bits. */
113	clz pos, syndrome
114	rev data2, data2
115	lsl data1, data1, pos
116	lsl data2, data2, pos
117	/* But we need to zero-extend (char is unsigned) the value and then
118	perform a signed 32-bit subtraction. */
119	lsr data1, data1, #56
120	sub result, data1, data2, lsr #56
121	RET
122	#else
123	/* For big-endian we cannot use the trick with the syndrome value
124	as carry-propagation can corrupt the upper bits if the trailing
125	bytes in the string contain 0x01. */
126	/* However, if there is no NUL byte in the dword, we can generate
127	the result directly. We can't just subtract the bytes as the
128	MSB might be significant. */
129	cbnz has_nul, 1f
130	cmp data1, data2
131	cset result, ne
132	cneg result, result, lo
133	RET
134	1:
135	/* Re-compute the NUL-byte detection, using a byte-reversed value. */
136	rev tmp3, data1
137	sub tmp1, tmp3, zeroones
138	orr tmp2, tmp3, #REP8_7f
139	bic has_nul, tmp1, tmp2
140	rev has_nul, has_nul
141	orr syndrome, diff, has_nul
142	clz pos, syndrome
143	/* The MS-non-zero bit of the syndrome marks either the first bit
144	that is different, or the top bit of the first zero byte.
145	Shifting left now will bring the critical information into the
146	top bits. */
147	lsl data1, data1, pos
148	lsl data2, data2, pos
149	/* But we need to zero-extend (char is unsigned) the value and then
150	perform a signed 32-bit subtraction. */
151	lsr data1, data1, #56
152	sub result, data1, data2, lsr #56
153	RET
154	#endif
155
156	L(mutual_align):
157	/* Sources are mutually aligned, but are not currently at an
158	alignment boundary. Round down the addresses and then mask off
159	the bytes that precede the start point.
160	We also need to adjust the limit calculations, but without
161	overflowing if the limit is near ULONG_MAX. */
162	bic src1, src1, #7
163	bic src2, src2, #7
164	ldr data1, [src1], #8
165	neg tmp3, tmp1, lsl #3 /* 64 - bits(bytes beyond align). */
166	ldr data2, [src2], #8
167	mov tmp2, #~0
168	sub limit_wd, limit, #1 /* limit != 0, so no underflow. */
169	#ifdef __AARCH64EB__
170	/* Big-endian. Early bytes are at MSB. */
171	lsl tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */
172	#else
173	/* Little-endian. Early bytes are at LSB. */
174	lsr tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */
175	#endif
176	and tmp3, limit_wd, #7
177	lsr limit_wd, limit_wd, #3
178	/* Adjust the limit. Only low 3 bits used, so overflow irrelevant. */
179	add limit, limit, tmp1
180	add tmp3, tmp3, tmp1
181	orr data1, data1, tmp2
182	orr data2, data2, tmp2
183	add limit_wd, limit_wd, tmp3, lsr #3
184	b L(start_realigned)
185
186	L(ret0):
187	mov result, #0
188	RET
189
190	.p2align 6
191	L(misaligned8):
192	sub limit, limit, #1
193	1:
194	/* Perhaps we can do better than this. */
195	ldrb data1w, [src1], #1
196	ldrb data2w, [src2], #1
197	subs limit, limit, #1
198	ccmp data1w, #1, #0, cs /* NZCV = 0b0000. */
199	ccmp data1w, data2w, #0, cs /* NZCV = 0b0000. */
200	b.eq 1b
201	sub result, data1, data2
202	RET
203	END (strncmp)
204	libc_hidden_builtin_def (strncmp)