1/*
2 * arch/alpha/lib/ev67-strrchr.S
3 * 21264 version by Rick Gorton <rick.gorton@alpha-processor.com>
4 *
5 * Finds length of a 0-terminated string.  Optimized for the
6 * Alpha architecture:
7 *
8 *	- memory accessed as aligned quadwords only
9 *	- uses bcmpge to compare 8 bytes in parallel
10 *
11 * Much of the information about 21264 scheduling/coding comes from:
12 *	Compiler Writer's Guide for the Alpha 21264
13 *	abbreviated as 'CWG' in other comments here
14 *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
15 * Scheduling notation:
16 *	E	- either cluster
17 *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
18 *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
19 */
20
21
22#include <asm/regdef.h>
23
24	.set noreorder
25	.set noat
26
27	.align 4
28	.ent strrchr
29	.globl strrchr
30strrchr:
31	.frame sp, 0, ra
32	.prologue 0
33
34	and	a1, 0xff, t2	# E : 00000000000000ch
35	insbl	a1, 1, t4	# U : 000000000000ch00
36	insbl	a1, 2, t5	# U : 0000000000ch0000
37	ldq_u   t0, 0(a0)	# L : load first quadword Latency=3
38
39	mov	zero, t6	# E : t6 is last match aligned addr
40	or	t2, t4, a1	# E : 000000000000chch
41	sll	t5, 8, t3	# U : 00000000ch000000
42	mov	zero, t8	# E : t8 is last match byte compare mask
43
44	andnot  a0, 7, v0	# E : align source addr
45	or	t5, t3, t3	# E : 00000000chch0000
46	sll	a1, 32, t2	# U : 0000chch00000000
47	sll	a1, 48, t4	# U : chch000000000000
48
49	or	t4, a1, a1	# E : chch00000000chch
50	or	t2, t3, t2	# E : 0000chchchch0000
51	or	a1, t2, a1	# E : chchchchchchchch
52	lda	t5, -1		# E : build garbage mask
53
54	cmpbge  zero, t0, t1	# E : bits set iff byte == zero
55	mskqh	t5, a0, t4	# E : Complete garbage mask
56	xor	t0, a1, t2	# E : make bytes == c zero
57	cmpbge	zero, t4, t4	# E : bits set iff byte is garbage
58
59	cmpbge  zero, t2, t3	# E : bits set iff byte == c
60	andnot	t1, t4, t1	# E : clear garbage from null test
61	andnot	t3, t4, t3	# E : clear garbage from char test
62	bne	t1, $eos	# U : did we already hit the terminator?
63
64	/* Character search main loop */
65$loop:
66	ldq	t0, 8(v0)	# L : load next quadword
67	cmovne	t3, v0, t6	# E : save previous comparisons match
68	nop			#   : Latency=2, extra map slot (keep nop with cmov)
69	nop
70
71	cmovne	t3, t3, t8	# E : Latency=2, extra map slot
72	nop			#   : keep with cmovne
73	addq	v0, 8, v0	# E :
74	xor	t0, a1, t2	# E :
75
76	cmpbge	zero, t0, t1	# E : bits set iff byte == zero
77	cmpbge	zero, t2, t3	# E : bits set iff byte == c
78	beq	t1, $loop	# U : if we havnt seen a null, loop
79	nop
80
81	/* Mask out character matches after terminator */
82$eos:
83	negq	t1, t4		# E : isolate first null byte match
84	and	t1, t4, t4	# E :
85	subq	t4, 1, t5	# E : build a mask of the bytes upto...
86	or	t4, t5, t4	# E : ... and including the null
87
88	and	t3, t4, t3	# E : mask out char matches after null
89	cmovne	t3, t3, t8	# E : save it, if match found Latency=2, extra map slot
90	nop			#   : Keep with cmovne
91	nop
92
93	cmovne	t3, v0, t6	# E :
94	nop			#   : Keep with cmovne
95	/* Locate the address of the last matched character */
96	ctlz	t8, t2		# U0 : Latency=3 (0x40 for t8=0)
97	nop
98
99	cmoveq	t8, 0x3f, t2	# E : Compensate for case when no match is seen
100	nop			# E : hide the cmov latency (2) behind ctlz latency
101	lda	t5, 0x3f($31)	# E :
102	subq	t5, t2, t5	# E : Normalize leading zero count
103
104	addq	t6, t5, v0	# E : and add to quadword address
105	ret			# L0 : Latency=3
106	nop
107	nop
108
109	.end strrchr
110