1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3 * arch/alpha/lib/ev67-strrchr.S
4 * 21264 version by Rick Gorton <rick.gorton@al
5 *
6 * Finds length of a 0-terminated string. Opt
7 * Alpha architecture:
8 *
9 * - memory accessed as aligned quadwords
10 * - uses bcmpge to compare 8 bytes in pa
11 *
12 * Much of the information about 21264 schedul
13 * Compiler Writer's Guide for the Alpha
14 * abbreviated as 'CWG' in other comments
15 * ftp.digital.com/pub/Digital/info/semic
16 * Scheduling notation:
17 * E - either cluster
18 * U - upper subcluster; U0 - subcl
19 * L - lower subcluster; L0 - subcl
20 */
21
22 #include <linux/export.h>
23 #include <asm/regdef.h>
24
25 .set noreorder
26 .set noat
27
28 .align 4
29 .ent strrchr
30 .globl strrchr
31 strrchr:
32 .frame sp, 0, ra
33 .prologue 0
34
35 and a1, 0xff, t2 # E : 00000000
36 insbl a1, 1, t4 # U : 00000000
37 insbl a1, 2, t5 # U : 00000000
38 ldq_u t0, 0(a0) # L : load fir
39
40 mov zero, t6 # E : t6 is la
41 or t2, t4, a1 # E : 00000000
42 sll t5, 8, t3 # U : 00000000
43 mov zero, t8 # E : t8 is la
44
45 andnot a0, 7, v0 # E : align so
46 or t5, t3, t3 # E : 00000000
47 sll a1, 32, t2 # U : 0000chch
48 sll a1, 48, t4 # U : chch0000
49
50 or t4, a1, a1 # E : chch0000
51 or t2, t3, t2 # E : 0000chch
52 or a1, t2, a1 # E : chchchch
53 lda t5, -1 # E : build ga
54
55 cmpbge zero, t0, t1 # E : bits set
56 mskqh t5, a0, t4 # E : Complete
57 xor t0, a1, t2 # E : make byt
58 cmpbge zero, t4, t4 # E : bits set
59
60 cmpbge zero, t2, t3 # E : bits set
61 andnot t1, t4, t1 # E : clear ga
62 andnot t3, t4, t3 # E : clear ga
63 bne t1, $eos # U : did we a
64
65 /* Character search main loop */
66 $loop:
67 ldq t0, 8(v0) # L : load nex
68 cmovne t3, v0, t6 # E : save pre
69 nop # : Latency=
70 nop
71
72 cmovne t3, t3, t8 # E : Latency=
73 nop # : keep wit
74 addq v0, 8, v0 # E :
75 xor t0, a1, t2 # E :
76
77 cmpbge zero, t0, t1 # E : bits set
78 cmpbge zero, t2, t3 # E : bits set
79 beq t1, $loop # U : if we ha
80 nop
81
82 /* Mask out character matches after te
83 $eos:
84 negq t1, t4 # E : isolate
85 and t1, t4, t4 # E :
86 subq t4, 1, t5 # E : build a
87 or t4, t5, t4 # E : ... and
88
89 and t3, t4, t3 # E : mask out
90 cmovne t3, t3, t8 # E : save it,
91 nop # : Keep wit
92 nop
93
94 cmovne t3, v0, t6 # E :
95 nop # : Keep wit
96 /* Locate the address of the last matc
97 ctlz t8, t2 # U0 : Latency
98 nop
99
100 cmoveq t8, 0x3f, t2 # E : Compensa
101 nop # E : hide the
102 lda t5, 0x3f($31) # E :
103 subq t5, t2, t5 # E : Normaliz
104
105 addq t6, t5, v0 # E : and add
106 ret # L0 : Latency
107 nop
108 nop
109
110 .end strrchr
111 EXPORT_SYMBOL(strrchr)
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.