1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | /* |
3 | * arch/alpha/lib/ev67-strchr.S |
4 | * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com> |
5 | * |
6 | * Return the address of a given character within a null-terminated |
7 | * string, or null if it is not found. |
8 | * |
9 | * Much of the information about 21264 scheduling/coding comes from: |
10 | * Compiler Writer's Guide for the Alpha 21264 |
11 | * abbreviated as 'CWG' in other comments here |
12 | * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html |
13 | * Scheduling notation: |
14 | * E - either cluster |
15 | * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1 |
16 | * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1 |
17 | * Try not to change the actual algorithm if possible for consistency. |
18 | */ |
19 | #include <linux/export.h> |
20 | #include <asm/regdef.h> |
21 | |
22 | .set noreorder |
23 | .set noat |
24 | |
25 | .align 4 |
26 | .globl strchr |
27 | .ent strchr |
28 | strchr: |
29 | .frame sp, 0, ra |
30 | .prologue 0 |
31 | |
32 | ldq_u t0, 0(a0) # L : load first quadword Latency=3 |
33 | and a1, 0xff, t3 # E : 00000000000000ch |
34 | insbl a1, 1, t5 # U : 000000000000ch00 |
35 | insbl a1, 7, a2 # U : ch00000000000000 |
36 | |
37 | insbl t3, 6, a3 # U : 00ch000000000000 |
38 | or t5, t3, a1 # E : 000000000000chch |
39 | andnot a0, 7, v0 # E : align our loop pointer |
40 | lda t4, -1 # E : build garbage mask |
41 | |
42 | mskqh t4, a0, t4 # U : only want relevant part of first quad |
43 | or a2, a3, a2 # E : chch000000000000 |
44 | inswl a1, 2, t5 # E : 00000000chch0000 |
45 | inswl a1, 4, a3 # E : 0000chch00000000 |
46 | |
47 | or a1, a2, a1 # E : chch00000000chch |
48 | or a3, t5, t5 # E : 0000chchchch0000 |
49 | cmpbge zero, t0, t2 # E : bits set iff byte == zero |
50 | cmpbge zero, t4, t4 # E : bits set iff byte is garbage |
51 | |
52 | /* This quad is _very_ serialized. Lots of stalling happens */ |
53 | or t5, a1, a1 # E : chchchchchchchch |
54 | xor t0, a1, t1 # E : make bytes == c zero |
55 | cmpbge zero, t1, t3 # E : bits set iff byte == c |
56 | or t2, t3, t0 # E : bits set iff char match or zero match |
57 | |
58 | andnot t0, t4, t0 # E : clear garbage bits |
59 | cttz t0, a2 # U0 : speculative (in case we get a match) |
60 | nop # E : |
61 | bne t0, $found # U : |
62 | |
63 | /* |
64 | * Yuk. This loop is going to stall like crazy waiting for the |
65 | * data to be loaded. Not much can be done about it unless it's |
66 | * unrolled multiple times - is that safe to do in kernel space? |
67 | * Or would exception handling recovery code do the trick here? |
68 | */ |
69 | $loop: ldq t0, 8(v0) # L : Latency=3 |
70 | addq v0, 8, v0 # E : |
71 | xor t0, a1, t1 # E : |
72 | cmpbge zero, t0, t2 # E : bits set iff byte == 0 |
73 | |
74 | cmpbge zero, t1, t3 # E : bits set iff byte == c |
75 | or t2, t3, t0 # E : |
76 | cttz t3, a2 # U0 : speculative (in case we get a match) |
77 | beq t0, $loop # U : |
78 | |
79 | $found: negq t0, t1 # E : clear all but least set bit |
80 | and t0, t1, t0 # E : |
81 | and t0, t3, t1 # E : bit set iff byte was the char |
82 | addq v0, a2, v0 # E : Add in the bit number from above |
83 | |
84 | cmoveq t1, $31, v0 # E : Two mapping slots, latency = 2 |
85 | nop |
86 | nop |
87 | ret # L0 : |
88 | |
89 | .end strchr |
90 | EXPORT_SYMBOL(strchr) |
91 | |