2 * arch/alpha/lib/ev6-copy_user.S
4 * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
6 * Copy to/from user space, handling exceptions as we go.. This
7 * isn't exactly pretty.
9 * This is essentially the same as "memcpy()", but with a few twists.
10 * Notably, we have to make sure that $0 is always up-to-date and
11 * contains the right "bytes left to copy" value (and that it is updated
12 * only _after_ a successful copy). There is also some rather minor
13 * exception setup stuff..
15 * NOTE! This is not directly C-callable, because the calling semantics are
20 * destination address in $6
21 * source address in $7
22 * return address in $28
25 * bytes left to copy in $0
28 * $1,$2,$3,$4,$5,$6,$7
30 * Much of the information about 21264 scheduling/coding comes from:
31 * Compiler Writer's Guide for the Alpha 21264
32 * abbreviated as 'CWG' in other comments here
33 * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
34 * Scheduling notation:
36 * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1
37 * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1
40 /* Allow an exception for an insn; exit if we get one. */
43 .section __ex_table,"a"; \
45 lda $31, $exitin-99b($31); \
50 .section __ex_table,"a"; \
52 lda $31, $exitout-99b($31); \
59 # Pipeline info: Slotting & Comments
61 ldgp $29,0($27) # we do exceptions -- we need the gp.
62 # Macro instruction becomes ldah/lda
65 subq $0, 32, $1 # .. E .. .. : Is this going to be a small copy?
66 beq $0, $zerolength # U .. .. .. : U L U L
68 and $6,7,$3 # .. .. .. E : is leading dest misalignment
69 ble $1, $onebyteloop # .. .. U .. : 1st branch : small amount of data
70 beq $3, $destaligned # .. U .. .. : 2nd (one cycle fetcher stall)
71 subq $3, 8, $3 # E .. .. .. : L U U L : trip counter
73 * The fetcher stall also hides the 1 cycle cross-cluster stall for $3 (L --> U)
74 * This loop aligns the destination a byte at a time
75 * We know we have at least one trip through this loop
78 EXI( ldbu $1,0($7) ) # .. .. .. L : Keep loads separate from stores
79 addq $6,1,$6 # .. .. E .. : Section 3.8 in the CWG
80 addq $3,1,$3 # .. E .. .. :
81 nop # E .. .. .. : U L U L
84 * the -1 is to compensate for the inc($6) done in a previous quadpack
85 * which allows us zero dependencies within either quadpack in the loop
87 EXO( stb $1,-1($6) ) # .. .. .. L :
88 addq $7,1,$7 # .. .. E .. : Section 3.8 in the CWG
89 subq $0,1,$0 # .. E .. .. :
90 bne $3, $aligndest # U .. .. .. : U L U L
93 * If we fell through into here, we have a minimum of 33 - 7 bytes
94 * If we arrived via branch, we have a minimum of 32 bytes
97 and $7,7,$1 # .. .. .. E : Check _current_ source alignment
98 bic $0,7,$4 # .. .. E .. : number bytes as a quadword loop
99 EXI( ldq_u $3,0($7) ) # .. L .. .. : Forward fetch for fallthrough code
100 beq $1,$quadaligned # U .. .. .. : U L U L
103 * In the worst case, we've just executed an ldq_u here from 0($7)
104 * and we'll repeat it once if we take the branch
107 /* Misaligned quadword loop - not unrolled. Leave it that way. */
109 EXI( ldq_u $2,8($7) ) # .. .. .. L :
110 subq $4,8,$4 # .. .. E .. :
111 extql $3,$7,$3 # .. U .. .. :
112 extqh $2,$7,$1 # U .. .. .. : U U L L
114 bis $3,$1,$1 # .. .. .. E :
115 EXO( stq $1,0($6) ) # .. .. L .. :
116 addq $7,8,$7 # .. E .. .. :
117 subq $0,8,$0 # E .. .. .. : U L L U
119 addq $6,8,$6 # .. .. .. E :
120 bis $2,$2,$3 # .. .. E .. :
122 bne $4,$misquad # U .. .. .. : U L U L
127 beq $0,$zerolength # U .. .. .. : U L U L
129 /* We know we have at least one trip through the byte loop */
130 EXI ( ldbu $2,0($7) ) # .. .. .. L : No loads in the same quad
131 addq $6,1,$6 # .. .. E .. : as the store (Section 3.8 in CWG)
133 br $31, $dirtyentry # L0 .. .. .. : L U U L
134 /* Do the trailing byte loop load, then hop into the store part of the loop */
137 * A minimum of (33 - 7) bytes to do a quad at a time.
138 * Based upon the usage context, it's worth the effort to unroll this loop
139 * $0 - number of bytes to be moved
140 * $4 - number of bytes to move as quadwords
141 * $6 is current destination address
142 * $7 is current source address
145 subq $4, 32, $2 # .. .. .. E : do not unroll for small stuff
148 blt $2, $onequad # U .. .. .. : U L U L
151 * There is a significant assumption here that the source and destination
152 * addresses differ by more than 32 bytes. In this particular case, a
153 * sparsity of registers further bounds this to be a minimum of 8 bytes.
154 * But if this isn't met, then the output result will be incorrect.
155 * Furthermore, due to a lack of available registers, we really can't
156 * unroll this to be an 8x loop (which would enable us to use the wh64
157 * instruction memory hint instruction).
160 EXI( ldq $1,0($7) ) # .. .. .. L
161 EXI( ldq $2,8($7) ) # .. .. L ..
162 subq $4,32,$4 # .. E .. ..
163 nop # E .. .. .. : U U L L
165 addq $7,16,$7 # .. .. .. E
166 EXO( stq $1,0($6) ) # .. .. L ..
167 EXO( stq $2,8($6) ) # .. L .. ..
168 subq $0,16,$0 # E .. .. .. : U L L U
170 addq $6,16,$6 # .. .. .. E
171 EXI( ldq $1,0($7) ) # .. .. L ..
172 EXI( ldq $2,8($7) ) # .. L .. ..
173 subq $4, 32, $3 # E .. .. .. : U U L L : is there enough for another trip?
175 EXO( stq $1,0($6) ) # .. .. .. L
176 EXO( stq $2,8($6) ) # .. .. L ..
177 subq $0,16,$0 # .. E .. ..
178 addq $7,16,$7 # E .. .. .. : U L L U
182 addq $6,16,$6 # .. E .. ..
183 bgt $3,$unroll4 # U .. .. .. : U L U L
208 * For small copies (or the tail of a larger copy), do a very simple byte loop.
209 * There's no point in doing a lot of complex alignment calculations to try to
210 * to quadword stuff for a small amount of data.
211 * $0 - remaining number of bytes left to copy
212 * $6 - current dest addr
213 * $7 - current source addr
217 EXI ( ldbu $2,0($7) ) # .. .. .. L : No loads in the same quad
218 addq $6,1,$6 # .. .. E .. : as the store (Section 3.8 in CWG)
220 nop # E .. .. .. : U L U L
224 * the -1 is to compensate for the inc($6) done in a previous quadpack
225 * which allows us zero dependencies within either quadpack in the loop
227 EXO ( stb $2,-1($6) ) # .. .. .. L :
228 addq $7,1,$7 # .. .. E .. : quadpack as the load
229 subq $0,1,$0 # .. E .. .. : change count _after_ copy
230 bgt $0,$onebyteloop # U .. .. .. : U L U L
233 $exitout: # Destination for exception recovery(?)
237 ret $31,($28),1 # L0 .. .. .. : L U L U
241 /* A stupid byte-by-byte zeroing of the rest of the output
242 buffer. This cures security holes by never leaving
243 random kernel data around to be copied elsewhere. */
251 EXO ( stb $31,0($6) ) # L