blob: b789db19275443d092494ea80a41be772dedb96a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * arch/alpha/lib/ev6-copy_page.S
3 *
4 * Copy an entire page.
5 */
6
7/* The following comparison of this routine vs the normal copy_page.S
8 was written by an unnamed ev6 hardware designer and forwarded to me
9 via Steven Hobbs <hobbs@steven.zko.dec.com>.
10
11 First Problem: STQ overflows.
12 -----------------------------
13
14 It would be nice if EV6 handled every resource overflow efficiently,
15 but for some it doesn't. Including store queue overflows. It causes
16 a trap and a restart of the pipe.
17
18 To get around this we sometimes use (to borrow a term from a VSSAD
19 researcher) "aeration". The idea is to slow the rate at which the
20 processor receives valid instructions by inserting nops in the fetch
21 path. In doing so, you can prevent the overflow and actually make
22 the code run faster. You can, of course, take advantage of the fact
23 that the processor can fetch at most 4 aligned instructions per cycle.
24
25 I inserted enough nops to force it to take 10 cycles to fetch the
26 loop code. In theory, EV6 should be able to execute this loop in
27 9 cycles but I was not able to get it to run that fast -- the initial
28 conditions were such that I could not reach this optimum rate on
29 (chaotic) EV6. I wrote the code such that everything would issue
30 in order.
31
32 Second Problem: Dcache index matches.
33 -------------------------------------
34
35 If you are going to use this routine on random aligned pages, there
36 is a 25% chance that the pages will be at the same dcache indices.
37 This results in many nasty memory traps without care.
38
39 The solution is to schedule the prefetches to avoid the memory
40 conflicts. I schedule the wh64 prefetches farther ahead of the
41 read prefetches to avoid this problem.
42
43 Third Problem: Needs more prefetching.
44 --------------------------------------
45
46 In order to improve the code I added deeper prefetching to take the
47 most advantage of EV6's bandwidth.
48
49 I also prefetched the read stream. Note that adding the read prefetch
50 forced me to add another cycle to the inner-most kernel - up to 11
51 from the original 8 cycles per iteration. We could improve performance
52 further by unrolling the loop and doing multiple prefetches per cycle.
53
54 I think that the code below will be very robust and fast code for the
55 purposes of copying aligned pages. It is slower when both source and
56 destination pages are in the dcache, but it is my guess that this is
57 less important than the dcache miss case. */
58
59
60 .text
61 .align 4
62 .global copy_page
63 .ent copy_page
64copy_page:
65 .prologue 0
66
67 /* Prefetch 5 read cachelines; write-hint 10 cache lines. */
68 wh64 ($16)
69 ldl $31,0($17)
70 ldl $31,64($17)
71 lda $1,1*64($16)
72
73 wh64 ($1)
74 ldl $31,128($17)
75 ldl $31,192($17)
76 lda $1,2*64($16)
77
78 wh64 ($1)
79 ldl $31,256($17)
80 lda $18,118
81 lda $1,3*64($16)
82
83 wh64 ($1)
84 nop
85 lda $1,4*64($16)
86 lda $2,5*64($16)
87
88 wh64 ($1)
89 wh64 ($2)
90 lda $1,6*64($16)
91 lda $2,7*64($16)
92
93 wh64 ($1)
94 wh64 ($2)
95 lda $1,8*64($16)
96 lda $2,9*64($16)
97
98 wh64 ($1)
99 wh64 ($2)
100 lda $19,10*64($16)
101 nop
102
103 /* Main prefetching/write-hinting loop. */
1041: ldq $0,0($17)
105 ldq $1,8($17)
106 unop
107 unop
108
109 unop
110 unop
111 ldq $2,16($17)
112 ldq $3,24($17)
113
114 ldq $4,32($17)
115 ldq $5,40($17)
116 unop
117 unop
118
119 unop
120 unop
121 ldq $6,48($17)
122 ldq $7,56($17)
123
124 ldl $31,320($17)
125 unop
126 unop
127 unop
128
129 /* This gives the extra cycle of aeration above the minimum. */
130 unop
131 unop
132 unop
133 unop
134
135 wh64 ($19)
136 unop
137 unop
138 unop
139
140 stq $0,0($16)
141 subq $18,1,$18
142 stq $1,8($16)
143 unop
144
145 unop
146 stq $2,16($16)
147 addq $17,64,$17
148 stq $3,24($16)
149
150 stq $4,32($16)
151 stq $5,40($16)
152 addq $19,64,$19
153 unop
154
155 stq $6,48($16)
156 stq $7,56($16)
157 addq $16,64,$16
158 bne $18, 1b
159
160 /* Prefetch the final 5 cache lines of the read stream. */
161 lda $18,10
162 ldl $31,320($17)
163 ldl $31,384($17)
164 ldl $31,448($17)
165
166 ldl $31,512($17)
167 ldl $31,576($17)
168 nop
169 nop
170
171 /* Non-prefetching, non-write-hinting cleanup loop for the
172 final 10 cache lines. */
1732: ldq $0,0($17)
174 ldq $1,8($17)
175 ldq $2,16($17)
176 ldq $3,24($17)
177
178 ldq $4,32($17)
179 ldq $5,40($17)
180 ldq $6,48($17)
181 ldq $7,56($17)
182
183 stq $0,0($16)
184 subq $18,1,$18
185 stq $1,8($16)
186 addq $17,64,$17
187
188 stq $2,16($16)
189 stq $3,24($16)
190 stq $4,32($16)
191 stq $5,40($16)
192
193 stq $6,48($16)
194 stq $7,56($16)
195 addq $16,64,$16
196 bne $18, 2b
197
198 ret
199 nop
200 unop
201 nop
202
203 .end copy_page