Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *
4 * Optimized version of the standard strlen() function
5 *
6 *
7 * Inputs:
8 * in0 address of string
9 *
10 * Outputs:
11 * ret0 the number of characters in the string (0 if empty string)
12 * does not count the \0
13 *
14 * Copyright (C) 1999, 2001 Hewlett-Packard Co
15 * Stephane Eranian <eranian@hpl.hp.com>
16 *
17 * 09/24/99 S.Eranian add speculation recovery code
18 */
19
20#include <asm/asmmacro.h>
21#include <asm/export.h>
22
23//
24//
25// This is an enhanced version of the basic strlen. it includes a combination
26// of compute zero index (czx), parallel comparisons, speculative loads and
27// loop unroll using rotating registers.
28//
29// General Ideas about the algorithm:
30// The goal is to look at the string in chunks of 8 bytes.
31// so we need to do a few extra checks at the beginning because the
32// string may not be 8-byte aligned. In this case we load the 8byte
33// quantity which includes the start of the string and mask the unused
34// bytes with 0xff to avoid confusing czx.
35// We use speculative loads and software pipelining to hide memory
36// latency and do read ahead safely. This way we defer any exception.
37//
38// Because we don't want the kernel to be relying on particular
39// settings of the DCR register, we provide recovery code in case
40// speculation fails. The recovery code is going to "redo" the work using
41// only normal loads. If we still get a fault then we generate a
42// kernel panic. Otherwise we return the strlen as usual.
43//
44// The fact that speculation may fail can be caused, for instance, by
45// the DCR.dm bit being set. In this case TLB misses are deferred, i.e.,
46// a NaT bit will be set if the translation is not present. The normal
47// load, on the other hand, will cause the translation to be inserted
48// if the mapping exists.
49//
50// It should be noted that we execute recovery code only when we need
51// to use the data that has been speculatively loaded: we don't execute
52// recovery code on pure read ahead data.
53//
54// Remarks:
55// - the cmp r0,r0 is used as a fast way to initialize a predicate
56// register to 1. This is required to make sure that we get the parallel
57// compare correct.
58//
59// - we don't use the epilogue counter to exit the loop but we need to set
60// it to zero beforehand.
61//
62// - after the loop we must test for Nat values because neither the
63// czx nor cmp instruction raise a NaT consumption fault. We must be
64// careful not to look too far for a Nat for which we don't care.
65// For instance we don't need to look at a NaT in val2 if the zero byte
66// was in val1.
67//
68// - Clearly performance tuning is required.
69//
70//
71//
72#define saved_pfs r11
73#define tmp r10
74#define base r16
75#define orig r17
76#define saved_pr r18
77#define src r19
78#define mask r20
79#define val r21
80#define val1 r22
81#define val2 r23
82
83GLOBAL_ENTRY(strlen)
84 .prologue
85 .save ar.pfs, saved_pfs
86 alloc saved_pfs=ar.pfs,11,0,0,8 // rotating must be multiple of 8
87
88 .rotr v[2], w[2] // declares our 4 aliases
89
90 extr.u tmp=in0,0,3 // tmp=least significant 3 bits
91 mov orig=in0 // keep trackof initial byte address
92 dep src=0,in0,0,3 // src=8byte-aligned in0 address
93 .save pr, saved_pr
94 mov saved_pr=pr // preserve predicates (rotation)
95 ;;
96
97 .body
98
99 ld8 v[1]=[src],8 // must not speculate: can fail here
100 shl tmp=tmp,3 // multiply by 8bits/byte
101 mov mask=-1 // our mask
102 ;;
103 ld8.s w[1]=[src],8 // speculatively load next
104 cmp.eq p6,p0=r0,r0 // sets p6 to true for cmp.and
105 sub tmp=64,tmp // how many bits to shift our mask on the right
106 ;;
107 shr.u mask=mask,tmp // zero enough bits to hold v[1] valuable part
108 mov ar.ec=r0 // clear epilogue counter (saved in ar.pfs)
109 ;;
110 add base=-16,src // keep track of aligned base
111 or v[1]=v[1],mask // now we have a safe initial byte pattern
112 ;;
1131:
114 ld8.s v[0]=[src],8 // speculatively load next
115 czx1.r val1=v[1] // search 0 byte from right
116 czx1.r val2=w[1] // search 0 byte from right following 8bytes
117 ;;
118 ld8.s w[0]=[src],8 // speculatively load next to next
119 cmp.eq.and p6,p0=8,val1 // p6 = p6 and val1==8
120 cmp.eq.and p6,p0=8,val2 // p6 = p6 and mask==8
121(p6) br.wtop.dptk 1b // loop until p6 == 0
122 ;;
123 //
124 // We must return try the recovery code iff
125 // val1_is_nat || (val1==8 && val2_is_nat)
126 //
127 // XXX Fixme
128 // - there must be a better way of doing the test
129 //
130 cmp.eq p8,p9=8,val1 // p6 = val1 had zero (disambiguate)
131 tnat.nz p6,p7=val1 // test NaT on val1
132(p6) br.cond.spnt .recover // jump to recovery if val1 is NaT
133 ;;
134 //
135 // if we come here p7 is true, i.e., initialized for // cmp
136 //
137 cmp.eq.and p7,p0=8,val1// val1==8?
138 tnat.nz.and p7,p0=val2 // test NaT if val2
139(p7) br.cond.spnt .recover // jump to recovery if val2 is NaT
140 ;;
141(p8) mov val1=val2 // the other test got us out of the loop
142(p8) adds src=-16,src // correct position when 3 ahead
143(p9) adds src=-24,src // correct position when 4 ahead
144 ;;
145 sub ret0=src,orig // distance from base
146 sub tmp=8,val1 // which byte in word
147 mov pr=saved_pr,0xffffffffffff0000
148 ;;
149 sub ret0=ret0,tmp // adjust
150 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
151 br.ret.sptk.many rp // end of normal execution
152
153 //
154 // Outlined recovery code when speculation failed
155 //
156 // This time we don't use speculation and rely on the normal exception
157 // mechanism. that's why the loop is not as good as the previous one
158 // because read ahead is not possible
159 //
160 // IMPORTANT:
161 // Please note that in the case of strlen() as opposed to strlen_user()
162 // we don't use the exception mechanism, as this function is not
163 // supposed to fail. If that happens it means we have a bug and the
164 // code will cause of kernel fault.
165 //
166 // XXX Fixme
167 // - today we restart from the beginning of the string instead
168 // of trying to continue where we left off.
169 //
170.recover:
171 ld8 val=[base],8 // will fail if unrecoverable fault
172 ;;
173 or val=val,mask // remask first bytes
174 cmp.eq p0,p6=r0,r0 // nullify first ld8 in loop
175 ;;
176 //
177 // ar.ec is still zero here
178 //
1792:
180(p6) ld8 val=[base],8 // will fail if unrecoverable fault
181 ;;
182 czx1.r val1=val // search 0 byte from right
183 ;;
184 cmp.eq p6,p0=8,val1 // val1==8 ?
185(p6) br.wtop.dptk 2b // loop until p6 == 0
186 ;; // (avoid WAW on p63)
187 sub ret0=base,orig // distance from base
188 sub tmp=8,val1
189 mov pr=saved_pr,0xffffffffffff0000
190 ;;
191 sub ret0=ret0,tmp // length=now - back -1
192 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
193 br.ret.sptk.many rp // end of successful recovery code
194END(strlen)
195EXPORT_SYMBOL(strlen)
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>
20
21//
22//
23// This is an enhanced version of the basic strlen. it includes a combination
24// of compute zero index (czx), parallel comparisons, speculative loads and
25// loop unroll using rotating registers.
26//
27// General Ideas about the algorithm:
28// The goal is to look at the string in chunks of 8 bytes.
29// so we need to do a few extra checks at the beginning because the
30// string may not be 8-byte aligned. In this case we load the 8byte
31// quantity which includes the start of the string and mask the unused
32// bytes with 0xff to avoid confusing czx.
33// We use speculative loads and software pipelining to hide memory
34// latency and do read ahead safely. This way we defer any exception.
35//
36// Because we don't want the kernel to be relying on particular
37// settings of the DCR register, we provide recovery code in case
38// speculation fails. The recovery code is going to "redo" the work using
39// only normal loads. If we still get a fault then we generate a
40// kernel panic. Otherwise we return the strlen as usual.
41//
42// The fact that speculation may fail can be caused, for instance, by
43// the DCR.dm bit being set. In this case TLB misses are deferred, i.e.,
44// a NaT bit will be set if the translation is not present. The normal
45// load, on the other hand, will cause the translation to be inserted
46// if the mapping exists.
47//
48// It should be noted that we execute recovery code only when we need
49// to use the data that has been speculatively loaded: we don't execute
50// recovery code on pure read ahead data.
51//
52// Remarks:
53// - the cmp r0,r0 is used as a fast way to initialize a predicate
54// register to 1. This is required to make sure that we get the parallel
55// compare correct.
56//
57// - we don't use the epilogue counter to exit the loop but we need to set
58// it to zero beforehand.
59//
60// - after the loop we must test for Nat values because neither the
61// czx nor cmp instruction raise a NaT consumption fault. We must be
62// careful not to look too far for a Nat for which we don't care.
63// For instance we don't need to look at a NaT in val2 if the zero byte
64// was in val1.
65//
66// - Clearly performance tuning is required.
67//
68//
69//
70#define saved_pfs r11
71#define tmp r10
72#define base r16
73#define orig r17
74#define saved_pr r18
75#define src r19
76#define mask r20
77#define val r21
78#define val1 r22
79#define val2 r23
80
81GLOBAL_ENTRY(strlen)
82 .prologue
83 .save ar.pfs, saved_pfs
84 alloc saved_pfs=ar.pfs,11,0,0,8 // rotating must be multiple of 8
85
86 .rotr v[2], w[2] // declares our 4 aliases
87
88 extr.u tmp=in0,0,3 // tmp=least significant 3 bits
89 mov orig=in0 // keep trackof initial byte address
90 dep src=0,in0,0,3 // src=8byte-aligned in0 address
91 .save pr, saved_pr
92 mov saved_pr=pr // preserve predicates (rotation)
93 ;;
94
95 .body
96
97 ld8 v[1]=[src],8 // must not speculate: can fail here
98 shl tmp=tmp,3 // multiply by 8bits/byte
99 mov mask=-1 // our mask
100 ;;
101 ld8.s w[1]=[src],8 // speculatively load next
102 cmp.eq p6,p0=r0,r0 // sets p6 to true for cmp.and
103 sub tmp=64,tmp // how many bits to shift our mask on the right
104 ;;
105 shr.u mask=mask,tmp // zero enough bits to hold v[1] valuable part
106 mov ar.ec=r0 // clear epilogue counter (saved in ar.pfs)
107 ;;
108 add base=-16,src // keep track of aligned base
109 or v[1]=v[1],mask // now we have a safe initial byte pattern
110 ;;
1111:
112 ld8.s v[0]=[src],8 // speculatively load next
113 czx1.r val1=v[1] // search 0 byte from right
114 czx1.r val2=w[1] // search 0 byte from right following 8bytes
115 ;;
116 ld8.s w[0]=[src],8 // speculatively load next to next
117 cmp.eq.and p6,p0=8,val1 // p6 = p6 and val1==8
118 cmp.eq.and p6,p0=8,val2 // p6 = p6 and mask==8
119(p6) br.wtop.dptk 1b // loop until p6 == 0
120 ;;
121 //
122 // We must return try the recovery code iff
123 // val1_is_nat || (val1==8 && val2_is_nat)
124 //
125 // XXX Fixme
126 // - there must be a better way of doing the test
127 //
128 cmp.eq p8,p9=8,val1 // p6 = val1 had zero (disambiguate)
129 tnat.nz p6,p7=val1 // test NaT on val1
130(p6) br.cond.spnt .recover // jump to recovery if val1 is NaT
131 ;;
132 //
133 // if we come here p7 is true, i.e., initialized for // cmp
134 //
135 cmp.eq.and p7,p0=8,val1// val1==8?
136 tnat.nz.and p7,p0=val2 // test NaT if val2
137(p7) br.cond.spnt .recover // jump to recovery if val2 is NaT
138 ;;
139(p8) mov val1=val2 // the other test got us out of the loop
140(p8) adds src=-16,src // correct position when 3 ahead
141(p9) adds src=-24,src // correct position when 4 ahead
142 ;;
143 sub ret0=src,orig // distance from base
144 sub tmp=8,val1 // which byte in word
145 mov pr=saved_pr,0xffffffffffff0000
146 ;;
147 sub ret0=ret0,tmp // adjust
148 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
149 br.ret.sptk.many rp // end of normal execution
150
151 //
152 // Outlined recovery code when speculation failed
153 //
154 // This time we don't use speculation and rely on the normal exception
155 // mechanism. that's why the loop is not as good as the previous one
156 // because read ahead is not possible
157 //
158 // IMPORTANT:
159 // Please note that in the case of strlen() as opposed to strlen_user()
160 // we don't use the exception mechanism, as this function is not
161 // supposed to fail. If that happens it means we have a bug and the
162 // code will cause of kernel fault.
163 //
164 // XXX Fixme
165 // - today we restart from the beginning of the string instead
166 // of trying to continue where we left off.
167 //
168.recover:
169 ld8 val=[base],8 // will fail if unrecoverable fault
170 ;;
171 or val=val,mask // remask first bytes
172 cmp.eq p0,p6=r0,r0 // nullify first ld8 in loop
173 ;;
174 //
175 // ar.ec is still zero here
176 //
1772:
178(p6) ld8 val=[base],8 // will fail if unrecoverable fault
179 ;;
180 czx1.r val1=val // search 0 byte from right
181 ;;
182 cmp.eq p6,p0=8,val1 // val1==8 ?
183(p6) br.wtop.dptk 2b // loop until p6 == 0
184 ;; // (avoid WAW on p63)
185 sub ret0=base,orig // distance from base
186 sub tmp=8,val1
187 mov pr=saved_pr,0xffffffffffff0000
188 ;;
189 sub ret0=ret0,tmp // length=now - back -1
190 mov ar.pfs=saved_pfs // because of ar.ec, restore no matter what
191 br.ret.sptk.many rp // end of successful recovery code
192END(strlen)