xref: /openbmc/linux/arch/alpha/lib/strrchr.S (revision f3c78e94)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 * arch/alpha/lib/strrchr.S
4 * Contributed by Richard Henderson (rth@tamu.edu)
5 *
6 * Return the address of the last occurrence of a given character
7 * within a null-terminated string, or null if it is not found.
8 */
9#include <linux/export.h>
10#include <asm/regdef.h>
11
12	.set noreorder
13	.set noat
14
15	.align 3
16	.ent strrchr
17	.globl strrchr
18strrchr:
19	.frame sp, 0, ra
20	.prologue 0
21
22	zapnot	a1, 1, a1	# e0    : zero extend our test character
23	mov	zero, t6	# .. e1 : t6 is last match aligned addr
24	sll	a1, 8, t5	# e0    : replicate our test character
25	mov	zero, t8	# .. e1 : t8 is last match byte compare mask
26	or	t5, a1, a1	# e0    :
27	ldq_u   t0, 0(a0)	# .. e1 : load first quadword
28	sll	a1, 16, t5	# e0    :
29	andnot  a0, 7, v0	# .. e1 : align source addr
30	or	t5, a1, a1	# e0    :
31	lda	t4, -1		# .. e1 : build garbage mask
32	sll	a1, 32, t5	# e0    :
33	cmpbge  zero, t0, t1	# .. e1 : bits set iff byte == zero
34	mskqh	t4, a0, t4	# e0    :
35	or	t5, a1, a1	# .. e1 : character replication complete
36	xor	t0, a1, t2	# e0    : make bytes == c zero
37	cmpbge	zero, t4, t4	# .. e1 : bits set iff byte is garbage
38	cmpbge  zero, t2, t3	# e0    : bits set iff byte == c
39	andnot	t1, t4, t1	# .. e1 : clear garbage from null test
40	andnot	t3, t4, t3	# e0    : clear garbage from char test
41	bne	t1, $eos	# .. e1 : did we already hit the terminator?
42
43	/* Character search main loop */
44$loop:
45	ldq	t0, 8(v0)	# e0    : load next quadword
46	cmovne	t3, v0, t6	# .. e1 : save previous comparisons match
47	cmovne	t3, t3, t8	# e0    :
48	addq	v0, 8, v0	# .. e1 :
49	xor	t0, a1, t2	# e0    :
50	cmpbge	zero, t0, t1	# .. e1 : bits set iff byte == zero
51	cmpbge	zero, t2, t3	# e0    : bits set iff byte == c
52	beq	t1, $loop	# .. e1 : if we havnt seen a null, loop
53
54	/* Mask out character matches after terminator */
55$eos:
56	negq	t1, t4		# e0    : isolate first null byte match
57	and	t1, t4, t4	# e1    :
58	subq	t4, 1, t5	# e0    : build a mask of the bytes up to...
59	or	t4, t5, t4	# e1    : ... and including the null
60
61	and	t3, t4, t3	# e0    : mask out char matches after null
62	cmovne	t3, t3, t8	# .. e1 : save it, if match found
63	cmovne	t3, v0, t6	# e0    :
64
65	/* Locate the address of the last matched character */
66
67	/* Retain the early exit for the ev4 -- the ev5 mispredict penalty
68	   is 5 cycles -- the same as just falling through.  */
69	beq	t8, $retnull	# .. e1 :
70
71	and	t8, 0xf0, t2	# e0    : binary search for the high bit set
72	cmovne	t2, t2, t8	# .. e1 (zdb)
73	cmovne	t2, 4, t2	# e0    :
74	and	t8, 0xcc, t1	# .. e1 :
75	cmovne	t1, t1, t8	# e0    :
76	cmovne	t1, 2, t1	# .. e1 :
77	and	t8, 0xaa, t0	# e0    :
78	cmovne	t0, 1, t0	# .. e1 (zdb)
79	addq	t2, t1, t1	# e0    :
80	addq	t6, t0, v0	# .. e1 : add our aligned base ptr to the mix
81	addq	v0, t1, v0	# e0    :
82	ret			# .. e1 :
83
84$retnull:
85	mov	zero, v0	# e0    :
86	ret			# .. e1 :
87
88	.end strrchr
89	EXPORT_SYMBOL(strrchr)
90