[thirdparty/glibc.git] / sysdeps / powerpc / powerpc64 / power8 / memrchr.S

/* Optimized memrchr implementation for PowerPC64/POWER8.
   Copyright (C) 2017-2019 Free Software Foundation, Inc.
   Contributed by Luis Machado <luisgpm@br.ibm.com>.
   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
   <https://www.gnu.org/licenses/>.  */

#include <sysdep.h>

/* int [r3] memrchr (char *s [r3], int byte [r4], int size [r5])  */

#ifndef MEMRCHR
# define MEMRCHR __memrchr
#endif
	.machine  power8
ENTRY_TOCLESS (MEMRCHR)
	CALL_MCOUNT 3
	add	r7, r3, r5      /* Calculate the last acceptable address.  */
	neg	r0, r7
	addi	r7, r7, -1
	mr	r10, r3
	clrrdi	r6, r7, 7
	li	r9, 3<<5
	dcbt	r9, r6, 8       /* Stream hint, decreasing addresses.  */

	/* Replicate BYTE to doubleword.  */
	insrdi	r4, r4, 8, 48
	insrdi	r4, r4, 16, 32
	insrdi	r4, r4, 32, 0
	li	r6, -8
	li	r9, -1
	rlwinm	r0, r0, 3, 26, 28 /* Calculate padding.  */
	clrrdi	r8, r7, 3
	srd	r9, r9, r0
	cmpldi	r5, 32
	clrrdi	r0, r10, 3
	ble	L(small_range)

#ifdef __LITTLE_ENDIAN__
	ldx	r12, 0, r8
#else
	ldbrx	r12, 0, r8      /* Load reversed doubleword from memory.  */
#endif
	cmpb	r3, r12, r4     /* Check for BYTE in DWORD1.  */
	and	r3, r3, r9
	cmpldi	cr7, r3, 0      /* If r3 == 0, no BYTEs have been found.  */
	bne	cr7, L(done)

	/* Are we now aligned to a quadword boundary?  If so, skip to
	   the main loop.  Otherwise, go through the alignment code.  */
	andi.	r12, r8, 15
	beq	cr0, L(align_qw)

	/* Handle DWORD2 of pair.  */
#ifdef __LITTLE_ENDIAN__
	ldx	r12, r8, r6
#else
	ldbrx	r12, r8, r6
#endif
	addi	r8, r8, -8
	cmpb	r3, r12, r4
	cmpldi	cr7, r3, 0
	bne	cr7, L(done)

	.align	4
	/* At this point, r8 is 16B aligned.  */
L(align_qw):
	sub	r5, r8, r0
	vspltisb	v0, 0
	/* Precompute vbpermq constant.  */
	vspltisb	v10, 3
	li	r0, 0
	lvsl	v11, r0, r0
	vslb	v10, v11, v10
	mtvrd	v1, r4
	vspltb	v1, v1, 7
	cmpldi	r5, 64
	ble	L(tail64)
	/* Are we 64-byte aligned? If so, jump to the vectorized loop.
	   Note: aligning to 64-byte will necessarily slow down performance for
	   strings around 64 bytes in length due to the extra comparisons
	   required to check alignment for the vectorized loop.  This is a
	   necessary tradeoff we are willing to take in order to speed up the
	   calculation for larger strings.  */
	andi.	r11, r8, 63
	beq	cr0, L(preloop_64B)
	/* In order to begin the 64B loop, it needs to be 64
	   bytes aligned.  So read until it is 64B aligned.  */
	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	addi	r5, r5, -16

	andi.	r11, r8, 63
	beq	cr0, L(preloop_64B)
	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	addi	r5, r5, -16

	andi.	r11, r8, 63
	beq	cr0, L(preloop_64B)
	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	addi	r5, r5, -16
	/* At this point it should be 64B aligned.
	   Prepare for the 64B loop.  */
L(preloop_64B):
	cmpldi	r5, 64		/* Check if r5 < 64.  */
	ble	L(tail64)
	srdi	r9, r5, 6	/* Number of loop iterations.  */
	mtctr	r9		/* Setup the counter.  */
	li	r11, 16		/* Load required offsets.  */
	li	r9, 32
	li	r7, 48

	/* Handle r5 > 64.  Loop over the bytes in strides of 64B.  */
	.align 4
L(loop):
	addi	r8, r8, -64	/* Adjust address for the next iteration.  */
	lvx	v2, 0, r8	/* Load 4 quadwords.  */
	lvx	v3, r8, r11
	lvx	v4, v8, r9
	lvx	v5, v8, r7
	vcmpequb	v6, v1, v2
	vcmpequb	v7, v1, v3
	vcmpequb	v8, v1, v4
	vcmpequb	v9, v1, v5
	vor	v11, v6, v7
	vor	v12, v8, v9
	vor	v11, v11, v12	/* Compare and merge into one VR for speed.  */
	vcmpequb.	v11, v0, v11
	bnl	cr6, L(found)
	bdnz	L(loop)
	clrldi	r5, r5, 58

	/* Handle remainder of 64B loop or r5 > 64.  */
	.align	4
L(tail64):
	cmpldi	r5, 0
	beq	L(null)
	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	cmpldi	cr6, r5, 16
	ble	cr6, L(null)
	addi	r5, r5, -16

	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	cmpldi	cr6, r5, 16
	ble	cr6, L(null)
	addi	r5, r5, -16

	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	cmpldi	cr6, r5, 16
	ble	cr6, L(null)
	addi	r5, r5, -16

	addi	r8, r8, -16
	lvx	v4, 0, r8
	vcmpequb	v6, v1, v4
	vcmpequb.	v11, v0, v6
	bnl	cr6, L(found_16B)
	li	r3, 0
	blr

	/* Found a match in 64B loop.  */
	.align	4
L(found):
	/* Permute the first bit of each byte into bits 48-63.  */
	vbpermq	v6, v6, v10
	vbpermq	v7, v7, v10
	vbpermq	v8, v8, v10
	vbpermq	v9, v9, v10
	/* Shift each component into its correct position for merging.  */
#ifdef __LITTLE_ENDIAN__
	vsldoi	v7, v7, v7, 2
	vsldoi	v8, v8, v8, 4
	vsldoi	v9, v9, v9, 6
#else
	vsldoi	v6, v6, v6, 6
	vsldoi	v7, v7, v7, 4
	vsldoi	v8, v8, v8, 2
#endif
	/* Merge the results and move to a GPR.  */
	vor	v11, v6, v7
	vor	v4, v9, v8
	vor	v4, v11, v4
	mfvrd	r5, v4
#ifdef __LITTLE_ENDIAN__
	cntlzd	r6, r5	/* Count leading zeros before the match.  */
#else
	addi	r6, r5, -1
	andc	r6, r6, r5
	popcntd	r6, r6
#endif
	addi	r8, r8, 63
	sub	r3, r8, r6	/* Compute final address.  */
	cmpld	cr7, r3, r10
	bgelr	cr7
	li	r3, 0
	blr

	/* Found a match in last 16 bytes.  */
	.align	4
L(found_16B):
	cmpld	r8, r10		/* Are we on the last QW?  */
	bge	L(last)
	/* Now discard bytes before starting address.  */
	sub	r9, r10, r8
	mtvrd	v9, r9
	vspltisb	v8, 3
	/* Mask unwanted bytes.  */
#ifdef __LITTLE_ENDIAN__
	lvsr	v7, 0, r10
	vperm   v6, v0, v6, v7
	vsldoi	v9, v0, v9, 8
	vsl	v9, v9, v8
	vslo	v6, v6, v9
#else
	lvsl	v7, 0, r10
	vperm   v6, v6, v0, v7
	vsldoi	v9, v0, v9, 8
	vsl	v9, v9, v8
	vsro	v6, v6, v9
#endif
L(last):
	/* Permute the first bit of each byte into bits 48-63.  */
	vbpermq	v6, v6, v10
	/* Shift each component into its correct position for merging.  */
#ifdef __LITTLE_ENDIAN__
	vsldoi	v6, v6, v6, 6
	mfvrd	r7, v6
	cntlzd	r6, r7	/* Count leading zeros before the match.  */
#else
	mfvrd	r7, v6
	addi	r6, r7, -1
	andc	r6, r6, r7
	popcntd	r6, r6
#endif
	addi	r8, r8, 15
	sub	r3, r8, r6	/* Compute final address.  */
	cmpld	r6, r5
	bltlr
	li	r3, 0
	blr

	/* r3 has the output of the cmpb instruction, that is, it contains
	   0xff in the same position as BYTE in the original
	   word from the string.  Use that to calculate the pointer.
	   We need to make sure BYTE is *before* the end of the
	   range.  */
L(done):
	cntlzd	r9, r3	      /* Count leading zeros before the match.  */
	cmpld	r8, r0         /* Are we on the last word?  */
	srdi	r6, r9, 3	      /* Convert leading zeros to bytes.  */
	addi	r0, r6, -7
	sub	r3, r8, r0
	cmpld	cr7, r3, r10
	bnelr
	bgelr	cr7
	li	r3, 0
	blr

	.align	4
L(null):
	li	r3, 0
	blr

/* Deals with size <= 32.  */
	.align	4
L(small_range):
	cmpldi	r5, 0
	beq	L(null)

#ifdef __LITTLE_ENDIAN__
	ldx	r12, 0, r8
#else
	ldbrx	r12, 0, r8      /* Load reversed doubleword from memory.  */
#endif
	cmpb	r3, r12, r4     /* Check for BYTE in DWORD1.  */
	and	r3, r3, r9
	cmpldi	cr7, r3, 0
	bne	cr7, L(done)

	/* Are we done already?  */
	cmpld	r8, r0
	addi	r8, r8, -8
	beqlr

	.align	5
L(loop_small):
#ifdef __LITTLE_ENDIAN__
	ldx	r12, 0, r8
#else
	ldbrx	r12, 0, r8
#endif
	cmpb	r3, r12, r4
	cmpld	r8, r0
	cmpldi	cr7, r3, 0
	bne	cr7, L(done)
	addi	r8, r8, -8
	bne	L(loop_small)
	blr

END (MEMRCHR)
weak_alias (__memrchr, memrchr)
libc_hidden_builtin_def (memrchr)
Commit	Line	Data
59ba2d2b	1	/* Optimized memrchr implementation for PowerPC64/POWER8.
04277e02	2	Copyright (C) 2017-2019 Free Software Foundation, Inc.
59ba2d2b RS	3	Contributed by Luis Machado <luisgpm@br.ibm.com>.
	4	This file is part of the GNU C Library.
	5
	6	The GNU C Library is free software; you can redistribute it and/or
	7	modify it under the terms of the GNU Lesser General Public
	8	License as published by the Free Software Foundation; either
	9	version 2.1 of the License, or (at your option) any later version.
	10
	11	The GNU C Library is distributed in the hope that it will be useful,
	12	but WITHOUT ANY WARRANTY; without even the implied warranty of
	13	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	14	Lesser General Public License for more details.
	15
	16	You should have received a copy of the GNU Lesser General Public
	17	License along with the GNU C Library; if not, see
5a82c748	18	<https://www.gnu.org/licenses/>. */
59ba2d2b RS	19
	20	#include <sysdep.h>
	21
	22	/* int [r3] memrchr (char s [r3], int byte [r4], int size [r5]) /
	23
59ba2d2b RS	24	#ifndef MEMRCHR
	25	# define MEMRCHR __memrchr
	26	#endif
066020c5	27	.machine power8
59ba2d2b RS	28	ENTRY_TOCLESS (MEMRCHR)
	29	CALL_MCOUNT 3
	30	add r7, r3, r5 /* Calculate the last acceptable address. */
	31	neg r0, r7
	32	addi r7, r7, -1
	33	mr r10, r3
	34	clrrdi r6, r7, 7
	35	li r9, 3<<5
	36	dcbt r9, r6, 8 /* Stream hint, decreasing addresses. */
	37
	38	/* Replicate BYTE to doubleword. */
	39	insrdi r4, r4, 8, 48
	40	insrdi r4, r4, 16, 32
	41	insrdi r4, r4, 32, 0
	42	li r6, -8
	43	li r9, -1
	44	rlwinm r0, r0, 3, 26, 28 /* Calculate padding. */
	45	clrrdi r8, r7, 3
	46	srd r9, r9, r0
	47	cmpldi r5, 32
	48	clrrdi r0, r10, 3
	49	ble L(small_range)
	50
	51	#ifdef __LITTLE_ENDIAN__
	52	ldx r12, 0, r8
	53	#else
	54	ldbrx r12, 0, r8 /* Load reversed doubleword from memory. */
	55	#endif
	56	cmpb r3, r12, r4 /* Check for BYTE in DWORD1. */
	57	and r3, r3, r9
	58	cmpldi cr7, r3, 0 /* If r3 == 0, no BYTEs have been found. */
	59	bne cr7, L(done)
	60
	61	/* Are we now aligned to a quadword boundary? If so, skip to
	62	the main loop. Otherwise, go through the alignment code. */
	63	andi. r12, r8, 15
	64	beq cr0, L(align_qw)
	65
	66	/* Handle DWORD2 of pair. */
	67	#ifdef __LITTLE_ENDIAN__
	68	ldx r12, r8, r6
	69	#else
	70	ldbrx r12, r8, r6
	71	#endif
	72	addi r8, r8, -8
	73	cmpb r3, r12, r4
	74	cmpldi cr7, r3, 0
	75	bne cr7, L(done)
	76
	77	.align 4
	78	/* At this point, r8 is 16B aligned. */
	79	L(align_qw):
	80	sub r5, r8, r0
	81	vspltisb v0, 0
	82	/* Precompute vbpermq constant. */
	83	vspltisb v10, 3
	84	li r0, 0
	85	lvsl v11, r0, r0
	86	vslb v10, v11, v10
066020c5	87	mtvrd v1, r4
59ba2d2b RS	88	vspltb v1, v1, 7
	89	cmpldi r5, 64
	90	ble L(tail64)
	91	/* Are we 64-byte aligned? If so, jump to the vectorized loop.
	92	Note: aligning to 64-byte will necessarily slow down performance for
	93	strings around 64 bytes in length due to the extra comparisons
	94	required to check alignment for the vectorized loop. This is a
	95	necessary tradeoff we are willing to take in order to speed up the
	96	calculation for larger strings. */
	97	andi. r11, r8, 63
	98	beq cr0, L(preloop_64B)
	99	/* In order to begin the 64B loop, it needs to be 64
	100	bytes aligned. So read until it is 64B aligned. */
	101	addi r8, r8, -16
	102	lvx v4, 0, r8
	103	vcmpequb v6, v1, v4
	104	vcmpequb. v11, v0, v6
	105	bnl cr6, L(found_16B)
	106	addi r5, r5, -16
	107
	108	andi. r11, r8, 63
	109	beq cr0, L(preloop_64B)
	110	addi r8, r8, -16
	111	lvx v4, 0, r8
	112	vcmpequb v6, v1, v4
	113	vcmpequb. v11, v0, v6
	114	bnl cr6, L(found_16B)
	115	addi r5, r5, -16
	116
	117	andi. r11, r8, 63
	118	beq cr0, L(preloop_64B)
	119	addi r8, r8, -16
	120	lvx v4, 0, r8
	121	vcmpequb v6, v1, v4
	122	vcmpequb. v11, v0, v6
	123	bnl cr6, L(found_16B)
	124	addi r5, r5, -16
	125	/* At this point it should be 64B aligned.
	126	Prepare for the 64B loop. */
	127	L(preloop_64B):
	128	cmpldi r5, 64 /* Check if r5 < 64. */
	129	ble L(tail64)
	130	srdi r9, r5, 6 /* Number of loop iterations. */
	131	mtctr r9 /* Setup the counter. */
	132	li r11, 16 /* Load required offsets. */
	133	li r9, 32
	134	li r7, 48
	135
	136	/* Handle r5 > 64. Loop over the bytes in strides of 64B. */
	137	.align 4
	138	L(loop):
	139	addi r8, r8, -64 /* Adjust address for the next iteration. */
	140	lvx v2, 0, r8 /* Load 4 quadwords. */
	141	lvx v3, r8, r11
	142	lvx v4, v8, r9
	143	lvx v5, v8, r7
	144	vcmpequb v6, v1, v2
	145	vcmpequb v7, v1, v3
	146	vcmpequb v8, v1, v4
	147	vcmpequb v9, v1, v5
	148	vor v11, v6, v7
	149	vor v12, v8, v9
	150	vor v11, v11, v12 /* Compare and merge into one VR for speed. */
	151	vcmpequb. v11, v0, v11
152	bnl cr6, L(found)
153	bdnz L(loop)
154	clrldi r5, r5, 58
155
156	/* Handle remainder of 64B loop or r5 > 64. */
157	.align 4
158	L(tail64):
159	cmpldi r5, 0
160	beq L(null)
161	addi r8, r8, -16
162	lvx v4, 0, r8
163	vcmpequb v6, v1, v4
164	vcmpequb. v11, v0, v6
165	bnl cr6, L(found_16B)
166	cmpldi cr6, r5, 16
167	ble cr6, L(null)
168	addi r5, r5, -16
169
170	addi r8, r8, -16
171	lvx v4, 0, r8
172	vcmpequb v6, v1, v4
173	vcmpequb. v11, v0, v6
174	bnl cr6, L(found_16B)
175	cmpldi cr6, r5, 16
176	ble cr6, L(null)
177	addi r5, r5, -16
178
179	addi r8, r8, -16
180	lvx v4, 0, r8
181	vcmpequb v6, v1, v4
182	vcmpequb. v11, v0, v6
183	bnl cr6, L(found_16B)
184	cmpldi cr6, r5, 16
185	ble cr6, L(null)
186	addi r5, r5, -16
187
188	addi r8, r8, -16
189	lvx v4, 0, r8
190	vcmpequb v6, v1, v4
191	vcmpequb. v11, v0, v6
192	bnl cr6, L(found_16B)
193	li r3, 0
194	blr
195
196	/* Found a match in 64B loop. */
197	.align 4
198	L(found):
199	/* Permute the first bit of each byte into bits 48-63. */
066020c5 RFF	200	vbpermq v6, v6, v10
	201	vbpermq v7, v7, v10
	202	vbpermq v8, v8, v10
	203	vbpermq v9, v9, v10
59ba2d2b RS	204	/* Shift each component into its correct position for merging. */
	205	#ifdef __LITTLE_ENDIAN__
	206	vsldoi v7, v7, v7, 2
	207	vsldoi v8, v8, v8, 4
	208	vsldoi v9, v9, v9, 6
	209	#else
	210	vsldoi v6, v6, v6, 6
	211	vsldoi v7, v7, v7, 4
	212	vsldoi v8, v8, v8, 2
	213	#endif
	214	/* Merge the results and move to a GPR. */
	215	vor v11, v6, v7
	216	vor v4, v9, v8
	217	vor v4, v11, v4
066020c5	218	mfvrd r5, v4
59ba2d2b RS	219	#ifdef __LITTLE_ENDIAN__
	220	cntlzd r6, r5 /* Count leading zeros before the match. */
	221	#else
	222	addi r6, r5, -1
	223	andc r6, r6, r5
	224	popcntd r6, r6
	225	#endif
	226	addi r8, r8, 63
	227	sub r3, r8, r6 /* Compute final address. */
5a907168 RS	228	cmpld cr7, r3, r10
	229	bgelr cr7
	230	li r3, 0
59ba2d2b RS	231	blr
	232
	233	/* Found a match in last 16 bytes. */
	234	.align 4
	235	L(found_16B):
5a907168 RS	236	cmpld r8, r10 /* Are we on the last QW? */
	237	bge L(last)
	238	/* Now discard bytes before starting address. */
	239	sub r9, r10, r8
066020c5	240	mtvrd v9, r9
5a907168 RS	241	vspltisb v8, 3
	242	/* Mask unwanted bytes. */
	243	#ifdef __LITTLE_ENDIAN__
	244	lvsr v7, 0, r10
	245	vperm v6, v0, v6, v7
	246	vsldoi v9, v0, v9, 8
	247	vsl v9, v9, v8
	248	vslo v6, v6, v9
	249	#else
	250	lvsl v7, 0, r10
	251	vperm v6, v6, v0, v7
	252	vsldoi v9, v0, v9, 8
	253	vsl v9, v9, v8
	254	vsro v6, v6, v9
	255	#endif
	256	L(last):
59ba2d2b	257	/* Permute the first bit of each byte into bits 48-63. */
066020c5	258	vbpermq v6, v6, v10
59ba2d2b RS	259	/* Shift each component into its correct position for merging. */
	260	#ifdef __LITTLE_ENDIAN__
	261	vsldoi v6, v6, v6, 6
066020c5	262	mfvrd r7, v6
59ba2d2b RS	263	cntlzd r6, r7 /* Count leading zeros before the match. */
59ba2d2b RS	264	#else
066020c5	265	mfvrd r7, v6
59ba2d2b RS	266	addi r6, r7, -1
	267	andc r6, r6, r7
	268	popcntd r6, r6
	269	#endif
	270	addi r8, r8, 15
	271	sub r3, r8, r6 /* Compute final address. */
	272	cmpld r6, r5
	273	bltlr
	274	li r3, 0
	275	blr
	276
	277	/* r3 has the output of the cmpb instruction, that is, it contains
	278	0xff in the same position as BYTE in the original
	279	word from the string. Use that to calculate the pointer.
	280	We need to make sure BYTE is before the end of the
	281	range. */
	282	L(done):
	283	cntlzd r9, r3 /* Count leading zeros before the match. */
	284	cmpld r8, r0 /* Are we on the last word? */
	285	srdi r6, r9, 3 /* Convert leading zeros to bytes. */
	286	addi r0, r6, -7
	287	sub r3, r8, r0
	288	cmpld cr7, r3, r10
	289	bnelr
	290	bgelr cr7
	291	li r3, 0
	292	blr
	293
	294	.align 4
	295	L(null):
	296	li r3, 0
	297	blr
	298
	299	/* Deals with size <= 32. */
	300	.align 4
	301	L(small_range):
	302	cmpldi r5, 0
	303	beq L(null)
	304
	305	#ifdef __LITTLE_ENDIAN__
	306	ldx r12, 0, r8
	307	#else
	308	ldbrx r12, 0, r8 /* Load reversed doubleword from memory. */
	309	#endif
	310	cmpb r3, r12, r4 /* Check for BYTE in DWORD1. */
	311	and r3, r3, r9
	312	cmpldi cr7, r3, 0
	313	bne cr7, L(done)
	314
	315	/* Are we done already? */
	316	cmpld r8, r0
	317	addi r8, r8, -8
	318	beqlr
	319
	320	.align 5
	321	L(loop_small):
	322	#ifdef __LITTLE_ENDIAN__
	323	ldx r12, 0, r8
	324	#else
	325	ldbrx r12, 0, r8
	326	#endif
	327	cmpb r3, r12, r4
	328	cmpld r8, r0
	329	cmpldi cr7, r3, 0
330	bne cr7, L(done)
331	addi r8, r8, -8
332	bne L(loop_small)
333	blr
334
335	END (MEMRCHR)
336	weak_alias (__memrchr, memrchr)
337	libc_hidden_builtin_def (memrchr)