Merge remote-tracking branch 'lightnvm/for-next'
[deliverable/linux.git] / arch / ia64 / lib / strlen.S
CommitLineData
1da177e4
LT
1/*
2 *
3 * Optimized version of the standard strlen() function
4 *
5 *
6 * Inputs:
7 * in0 address of string
8 *
9 * Outputs:
10 * ret0 the number of characters in the string (0 if empty string)
11 * does not count the \0
12 *
13 * Copyright (C) 1999, 2001 Hewlett-Packard Co
14 * Stephane Eranian <eranian@hpl.hp.com>
15 *
16 * 09/24/99 S.Eranian add speculation recovery code
17 */
18
19#include <asm/asmmacro.h>
e007c533 20#include <asm/export.h>
1da177e4
LT
21
22//
23//
24// This is an enhanced version of the basic strlen. it includes a combination
25// of compute zero index (czx), parallel comparisons, speculative loads and
26// loop unroll using rotating registers.
27//
28// General Ideas about the algorithm:
29// The goal is to look at the string in chunks of 8 bytes.
30// so we need to do a few extra checks at the beginning because the
31// string may not be 8-byte aligned. In this case we load the 8byte
32// quantity which includes the start of the string and mask the unused
33// bytes with 0xff to avoid confusing czx.
34// We use speculative loads and software pipelining to hide memory
35// latency and do read ahead safely. This way we defer any exception.
36//
37// Because we don't want the kernel to be relying on particular
38// settings of the DCR register, we provide recovery code in case
39// speculation fails. The recovery code is going to "redo" the work using
40// only normal loads. If we still get a fault then we generate a
41// kernel panic. Otherwise we return the strlen as usual.
42//
43// The fact that speculation may fail can be caused, for instance, by
44// the DCR.dm bit being set. In this case TLB misses are deferred, i.e.,
45// a NaT bit will be set if the translation is not present. The normal
46// load, on the other hand, will cause the translation to be inserted
47// if the mapping exists.
48//
49// It should be noted that we execute recovery code only when we need
50// to use the data that has been speculatively loaded: we don't execute
51// recovery code on pure read ahead data.
52//
53// Remarks:
54// - the cmp r0,r0 is used as a fast way to initialize a predicate
55// register to 1. This is required to make sure that we get the parallel
56// compare correct.
57//
58// - we don't use the epilogue counter to exit the loop but we need to set
59// it to zero beforehand.
60//
61// - after the loop we must test for Nat values because neither the
62// czx nor cmp instruction raise a NaT consumption fault. We must be
63// careful not to look too far for a Nat for which we don't care.
64// For instance we don't need to look at a NaT in val2 if the zero byte
65// was in val1.
66//
67// - Clearly performance tuning is required.
68//
69//
70//
71#define saved_pfs r11
72#define tmp r10
73#define base r16
74#define orig r17
75#define saved_pr r18
76#define src r19
77#define mask r20
78#define val r21
79#define val1 r22
80#define val2 r23
81
82GLOBAL_ENTRY(strlen)
83 .prologue
84 .save ar.pfs, saved_pfs
85 alloc saved_pfs=ar.pfs,11,0,0,8 // rotating must be multiple of 8
86
87 .rotr v[2], w[2] // declares our 4 aliases
88
89 extr.u tmp=in0,0,3 // tmp=least significant 3 bits
90 mov orig=in0 // keep trackof initial byte address
91 dep src=0,in0,0,3 // src=8byte-aligned in0 address
92 .save pr, saved_pr
93 mov saved_pr=pr // preserve predicates (rotation)
94 ;;
95
96 .body
97
98 ld8 v[1]=[src],8 // must not speculate: can fail here
99 shl tmp=tmp,3 // multiply by 8bits/byte
100 mov mask=-1 // our mask
101 ;;
102 ld8.s w[1]=[src],8 // speculatively load next
103 cmp.eq p6,p0=r0,r0 // sets p6 to true for cmp.and
104 sub tmp=64,tmp // how many bits to shift our mask on the right
105 ;;
106 shr.u mask=mask,tmp // zero enough bits to hold v[1] valuable part
107 mov ar.ec=r0 // clear epilogue counter (saved in ar.pfs)
108 ;;
109 add base=-16,src // keep track of aligned base
110 or v[1]=v[1],mask // now we have a safe initial byte pattern
111 ;;
1121:
113 ld8.s v[0]=[src],8 // speculatively load next
114 czx1.r val1=v[1] // search 0 byte from right
115 czx1.r val2=w[1] // search 0 byte from right following 8bytes
116 ;;
117 ld8.s w[0]=[src],8 // speculatively load next to next
118 cmp.eq.and p6,p0=8,val1 // p6 = p6 and val1==8
119 cmp.eq.and p6,p0=8,val2 // p6 = p6 and mask==8
120(p6) br.wtop.dptk 1b // loop until p6 == 0
121 ;;
122 //
123 // We must return try the recovery code iff
124 // val1_is_nat || (val1==8 && val2_is_nat)
125 //
126 // XXX Fixme
127 // - there must be a better way of doing the test
128 //
129 cmp.eq p8,p9=8,val1 // p6 = val1 had zero (disambiguate)
130 tnat.nz p6,p7=val1 // test NaT on val1
131(p6) br.cond.spnt .recover // jump to recovery if val1 is NaT
132 ;;
133 //
134 // if we come here p7 is true, i.e., initialized for // cmp
135 //
136 cmp.eq.and p7,p0=8,val1// val1==8?
137 tnat.nz.and p7,p0=val2 // test NaT if val2
138(p7) br.cond.spnt .recover // jump to recovery if val2 is NaT
139 ;;
140(p8) mov val1=val2 // the other test got us out of the loop
141(p8) adds src=-16,src // correct position when 3 ahead
142(p9) adds src=-24,src // correct position when 4 ahead
143 ;;
144 sub ret0=src,orig // distance from base
145 sub tmp=8,val1 // which byte in word
146 mov pr=saved_pr,0xffffffffffff0000
147 ;;
148 sub ret0=ret0,tmp // adjust
149 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
150 br.ret.sptk.many rp // end of normal execution
151
152 //
153 // Outlined recovery code when speculation failed
154 //
155 // This time we don't use speculation and rely on the normal exception
156 // mechanism. that's why the loop is not as good as the previous one
157 // because read ahead is not possible
158 //
159 // IMPORTANT:
160 // Please note that in the case of strlen() as opposed to strlen_user()
161 // we don't use the exception mechanism, as this function is not
162 // supposed to fail. If that happens it means we have a bug and the
163 // code will cause of kernel fault.
164 //
165 // XXX Fixme
166 // - today we restart from the beginning of the string instead
167 // of trying to continue where we left off.
168 //
169.recover:
170 ld8 val=[base],8 // will fail if unrecoverable fault
171 ;;
172 or val=val,mask // remask first bytes
173 cmp.eq p0,p6=r0,r0 // nullify first ld8 in loop
174 ;;
175 //
176 // ar.ec is still zero here
177 //
1782:
179(p6) ld8 val=[base],8 // will fail if unrecoverable fault
180 ;;
181 czx1.r val1=val // search 0 byte from right
182 ;;
183 cmp.eq p6,p0=8,val1 // val1==8 ?
184(p6) br.wtop.dptk 2b // loop until p6 == 0
185 ;; // (avoid WAW on p63)
186 sub ret0=base,orig // distance from base
187 sub tmp=8,val1
188 mov pr=saved_pr,0xffffffffffff0000
189 ;;
190 sub ret0=ret0,tmp // length=now - back -1
191 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
192 br.ret.sptk.many rp // end of successful recovery code
193END(strlen)
e007c533 194EXPORT_SYMBOL(strlen)
This page took 0.7419 seconds and 5 git commands to generate.