1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
|
/*
* arch/alpha/lib/ev6-copy_user.S
*
* 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
*
* Copy to/from user space, handling exceptions as we go.. This
* isn't exactly pretty.
*
* This is essentially the same as "memcpy()", but with a few twists.
* Notably, we have to make sure that $0 is always up-to-date and
* contains the right "bytes left to copy" value (and that it is updated
* only _after_ a successful copy). There is also some rather minor
* exception setup stuff..
*
* NOTE! This is not directly C-callable, because the calling semantics are
* different:
*
* Inputs:
* length in $0
* destination address in $6
* source address in $7
* return address in $28
*
* Outputs:
* bytes left to copy in $0
*
* Clobbers:
* $1,$2,$3,$4,$5,$6,$7
*
* Much of the information about 21264 scheduling/coding comes from:
* Compiler Writer's Guide for the Alpha 21264
* abbreviated as 'CWG' in other comments here
* ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
* Scheduling notation:
* E - either cluster
* U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
* L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
*/
/* Allow an exception for an insn; exit if we get one. */
#define EXI(x,y...) \
99: x,##y; \
.section __ex_table,"a"; \
.long 99b - .; \
lda $31, $exitin-99b($31); \
.previous
#define EXO(x,y...) \
99: x,##y; \
.section __ex_table,"a"; \
.long 99b - .; \
lda $31, $exitout-99b($31); \
.previous
.set noat
.align 4
.globl __copy_user
.ent __copy_user
# Pipeline info: Slotting & Comments
__copy_user:
.prologue 0
subq $0, 32, $1 # .. E .. .. : Is this going to be a small copy?
beq $0, $zerolength # U .. .. .. : U L U L
and $6,7,$3 # .. .. .. E : is leading dest misalignment
ble $1, $onebyteloop # .. .. U .. : 1st branch : small amount of data
beq $3, $destaligned # .. U .. .. : 2nd (one cycle fetcher stall)
subq $3, 8, $3 # E .. .. .. : L U U L : trip counter
/*
* The fetcher stall also hides the 1 cycle cross-cluster stall for $3 (L --> U)
* This loop aligns the destination a byte at a time
* We know we have at least one trip through this loop
*/
$aligndest:
EXI( ldbu $1,0($7) ) # .. .. .. L : Keep loads separate from stores
addq $6,1,$6 # .. .. E .. : Section 3.8 in the CWG
addq $3,1,$3 # .. E .. .. :
nop # E .. .. .. : U L U L
/*
* the -1 is to compensate for the inc($6) done in a previous quadpack
* which allows us zero dependencies within either quadpack in the loop
*/
EXO( stb $1,-1($6) ) # .. .. .. L :
addq $7,1,$7 # .. .. E .. : Section 3.8 in the CWG
subq $0,1,$0 # .. E .. .. :
bne $3, $aligndest # U .. .. .. : U L U L
/*
* If we fell through into here, we have a minimum of 33 - 7 bytes
* If we arrived via branch, we have a minimum of 32 bytes
*/
$destaligned:
and $7,7,$1 # .. .. .. E : Check _current_ source alignment
bic $0,7,$4 # .. .. E .. : number bytes as a quadword loop
EXI( ldq_u $3,0($7) ) # .. L .. .. : Forward fetch for fallthrough code
beq $1,$quadaligned # U .. .. .. : U L U L
/*
* In the worst case, we've just executed an ldq_u here from 0($7)
* and we'll repeat it once if we take the branch
*/
/* Misaligned quadword loop - not unrolled. Leave it that way. */
$misquad:
EXI( ldq_u $2,8($7) ) # .. .. .. L :
subq $4,8,$4 # .. .. E .. :
extql $3,$7,$3 # .. U .. .. :
extqh $2,$7,$1 # U .. .. .. : U U L L
bis $3,$1,$1 # .. .. .. E :
EXO( stq $1,0($6) ) # .. .. L .. :
addq $7,8,$7 # .. E .. .. :
subq $0,8,$0 # E .. .. .. : U L L U
addq $6,8,$6 # .. .. .. E :
bis $2,$2,$3 # .. .. E .. :
nop # .. E .. .. :
bne $4,$misquad # U .. .. .. : U L U L
nop # .. .. .. E
nop # .. .. E ..
nop # .. E .. ..
beq $0,$zerolength # U .. .. .. : U L U L
/* We know we have at least one trip through the byte loop */
EXI ( ldbu $2,0($7) ) # .. .. .. L : No loads in the same quad
addq $6,1,$6 # .. .. E .. : as the store (Section 3.8 in CWG)
nop # .. E .. .. :
br $31, $dirtyentry # L0 .. .. .. : L U U L
/* Do the trailing byte loop load, then hop into the store part of the loop */
/*
* A minimum of (33 - 7) bytes to do a quad at a time.
* Based upon the usage context, it's worth the effort to unroll this loop
* $0 - number of bytes to be moved
* $4 - number of bytes to move as quadwords
* $6 is current destination address
* $7 is current source address
*/
$quadaligned:
subq $4, 32, $2 # .. .. .. E : do not unroll for small stuff
nop # .. .. E ..
nop # .. E .. ..
blt $2, $onequad # U .. .. .. : U L U L
/*
* There is a significant assumption here that the source and destination
* addresses differ by more than 32 bytes. In this particular case, a
* sparsity of registers further bounds this to be a minimum of 8 bytes.
* But if this isn't met, then the output result will be incorrect.
* Furthermore, due to a lack of available registers, we really can't
* unroll this to be an 8x loop (which would enable us to use the wh64
* instruction memory hint instruction).
*/
$unroll4:
EXI( ldq $1,0($7) ) # .. .. .. L
EXI( ldq $2,8($7) ) # .. .. L ..
subq $4,32,$4 # .. E .. ..
nop # E .. .. .. : U U L L
addq $7,16,$7 # .. .. .. E
EXO( stq $1,0($6) ) # .. .. L ..
EXO( stq $2,8($6) ) # .. L .. ..
subq $0,16,$0 # E .. .. .. : U L L U
addq $6,16,$6 # .. .. .. E
EXI( ldq $1,0($7) ) # .. .. L ..
EXI( ldq $2,8($7) ) # .. L .. ..
subq $4, 32, $3 # E .. .. .. : U U L L : is there enough for another trip?
EXO( stq $1,0($6) ) # .. .. .. L
EXO( stq $2,8($6) ) # .. .. L ..
subq $0,16,$0 # .. E .. ..
addq $7,16,$7 # E .. .. .. : U L L U
nop # .. .. .. E
nop # .. .. E ..
addq $6,16,$6 # .. E .. ..
bgt $3,$unroll4 # U .. .. .. : U L U L
nop
nop
nop
beq $4, $noquads
$onequad:
EXI( ldq $1,0($7) )
subq $4,8,$4
addq $7,8,$7
nop
EXO( stq $1,0($6) )
subq $0,8,$0
addq $6,8,$6
bne $4,$onequad
$noquads:
nop
nop
nop
beq $0,$zerolength
/*
* For small copies (or the tail of a larger copy), do a very simple byte loop.
* There's no point in doing a lot of complex alignment calculations to try to
* to quadword stuff for a small amount of data.
* $0 - remaining number of bytes left to copy
* $6 - current dest addr
* $7 - current source addr
*/
$onebyteloop:
EXI ( ldbu $2,0($7) ) # .. .. .. L : No loads in the same quad
addq $6,1,$6 # .. .. E .. : as the store (Section 3.8 in CWG)
nop # .. E .. .. :
nop # E .. .. .. : U L U L
$dirtyentry:
/*
* the -1 is to compensate for the inc($6) done in a previous quadpack
* which allows us zero dependencies within either quadpack in the loop
*/
EXO ( stb $2,-1($6) ) # .. .. .. L :
addq $7,1,$7 # .. .. E .. : quadpack as the load
subq $0,1,$0 # .. E .. .. : change count _after_ copy
bgt $0,$onebyteloop # U .. .. .. : U L U L
$zerolength:
$exitin:
$exitout: # Destination for exception recovery(?)
nop # .. .. .. E
nop # .. .. E ..
nop # .. E .. ..
ret $31,($28),1 # L0 .. .. .. : L U L U
.end __copy_user
|