xref: /linux/arch/alpha/lib/strrchr.S (revision f3a8b6645dc2e60d11f20c1c23afd964ff4e55ae)
1/*
2 * arch/alpha/lib/strrchr.S
3 * Contributed by Richard Henderson (rth@tamu.edu)
4 *
5 * Return the address of the last occurrence of a given character
6 * within a null-terminated string, or null if it is not found.
7 */
8#include <asm/export.h>
9#include <asm/regdef.h>
10
11	.set noreorder
12	.set noat
13
14	.align 3
15	.ent strrchr
16	.globl strrchr
17strrchr:
18	.frame sp, 0, ra
19	.prologue 0
20
21	zapnot	a1, 1, a1	# e0    : zero extend our test character
22	mov	zero, t6	# .. e1 : t6 is last match aligned addr
23	sll	a1, 8, t5	# e0    : replicate our test character
24	mov	zero, t8	# .. e1 : t8 is last match byte compare mask
25	or	t5, a1, a1	# e0    :
26	ldq_u   t0, 0(a0)	# .. e1 : load first quadword
27	sll	a1, 16, t5	# e0    :
28	andnot  a0, 7, v0	# .. e1 : align source addr
29	or	t5, a1, a1	# e0    :
30	lda	t4, -1		# .. e1 : build garbage mask
31	sll	a1, 32, t5	# e0    :
32	cmpbge  zero, t0, t1	# .. e1 : bits set iff byte == zero
33	mskqh	t4, a0, t4	# e0    :
34	or	t5, a1, a1	# .. e1 : character replication complete
35	xor	t0, a1, t2	# e0    : make bytes == c zero
36	cmpbge	zero, t4, t4	# .. e1 : bits set iff byte is garbage
37	cmpbge  zero, t2, t3	# e0    : bits set iff byte == c
38	andnot	t1, t4, t1	# .. e1 : clear garbage from null test
39	andnot	t3, t4, t3	# e0    : clear garbage from char test
40	bne	t1, $eos	# .. e1 : did we already hit the terminator?
41
42	/* Character search main loop */
43$loop:
44	ldq	t0, 8(v0)	# e0    : load next quadword
45	cmovne	t3, v0, t6	# .. e1 : save previous comparisons match
46	cmovne	t3, t3, t8	# e0    :
47	addq	v0, 8, v0	# .. e1 :
48	xor	t0, a1, t2	# e0    :
49	cmpbge	zero, t0, t1	# .. e1 : bits set iff byte == zero
50	cmpbge	zero, t2, t3	# e0    : bits set iff byte == c
51	beq	t1, $loop	# .. e1 : if we havnt seen a null, loop
52
53	/* Mask out character matches after terminator */
54$eos:
55	negq	t1, t4		# e0    : isolate first null byte match
56	and	t1, t4, t4	# e1    :
57	subq	t4, 1, t5	# e0    : build a mask of the bytes up to...
58	or	t4, t5, t4	# e1    : ... and including the null
59
60	and	t3, t4, t3	# e0    : mask out char matches after null
61	cmovne	t3, t3, t8	# .. e1 : save it, if match found
62	cmovne	t3, v0, t6	# e0    :
63
64	/* Locate the address of the last matched character */
65
66	/* Retain the early exit for the ev4 -- the ev5 mispredict penalty
67	   is 5 cycles -- the same as just falling through.  */
68	beq	t8, $retnull	# .. e1 :
69
70	and	t8, 0xf0, t2	# e0    : binary search for the high bit set
71	cmovne	t2, t2, t8	# .. e1 (zdb)
72	cmovne	t2, 4, t2	# e0    :
73	and	t8, 0xcc, t1	# .. e1 :
74	cmovne	t1, t1, t8	# e0    :
75	cmovne	t1, 2, t1	# .. e1 :
76	and	t8, 0xaa, t0	# e0    :
77	cmovne	t0, 1, t0	# .. e1 (zdb)
78	addq	t2, t1, t1	# e0    :
79	addq	t6, t0, v0	# .. e1 : add our aligned base ptr to the mix
80	addq	v0, t1, v0	# e0    :
81	ret			# .. e1 :
82
83$retnull:
84	mov	zero, v0	# e0    :
85	ret			# .. e1 :
86
87	.end strrchr
88	EXPORT_SYMBOL(strrchr)
89