1/* 2 * arch/alpha/lib/ev6-memcpy.S 3 * 21264 version by Rick Gorton <rick.gorton@alpha-processor.com> 4 * 5 * Reasonably optimized memcpy() routine for the Alpha 21264 6 * 7 * - memory accessed as aligned quadwords only 8 * - uses bcmpge to compare 8 bytes in parallel 9 * 10 * Much of the information about 21264 scheduling/coding comes from: 11 * Compiler Writer's Guide for the Alpha 21264 12 * abbreviated as 'CWG' in other comments here 13 * ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html 14 * Scheduling notation: 15 * E - either cluster 16 * U - upper subcluster; U0 - subcluster U0; U1 - subcluster U1 17 * L - lower subcluster; L0 - subcluster L0; L1 - subcluster L1 18 * 19 * Temp usage notes: 20 * $1,$2, - scratch 21 */ 22 23 .set noreorder 24 .set noat 25 26 .align 4 27 .globl memcpy 28 .ent memcpy 29memcpy: 30 .frame $30,0,$26,0 31 .prologue 0 32 33 mov $16, $0 # E : copy dest to return 34 ble $18, $nomoredata # U : done with the copy? 35 xor $16, $17, $1 # E : are source and dest alignments the same? 36 and $1, 7, $1 # E : are they the same mod 8? 37 38 bne $1, $misaligned # U : Nope - gotta do this the slow way 39 /* source and dest are same mod 8 address */ 40 and $16, 7, $1 # E : Are both 0mod8? 41 beq $1, $both_0mod8 # U : Yes 42 nop # E : 43 44 /* 45 * source and dest are same misalignment. move a byte at a time 46 * until a 0mod8 alignment for both is reached. 47 * At least one byte more to move 48 */ 49 50$head_align: 51 ldbu $1, 0($17) # L : grab a byte 52 subq $18, 1, $18 # E : count-- 53 addq $17, 1, $17 # E : src++ 54 stb $1, 0($16) # L : 55 addq $16, 1, $16 # E : dest++ 56 and $16, 7, $1 # E : Are we at 0mod8 yet? 57 ble $18, $nomoredata # U : done with the copy? 58 bne $1, $head_align # U : 59 60$both_0mod8: 61 cmple $18, 127, $1 # E : Can we unroll the loop? 62 bne $1, $no_unroll # U : 63 and $16, 63, $1 # E : get mod64 alignment 64 beq $1, $do_unroll # U : no single quads to fiddle 65 66$single_head_quad: 67 ldq $1, 0($17) # L : get 8 bytes 68 subq $18, 8, $18 # E : count -= 8 69 addq $17, 8, $17 # E : src += 8 70 nop # E : 71 72 stq $1, 0($16) # L : store 73 addq $16, 8, $16 # E : dest += 8 74 and $16, 63, $1 # E : get mod64 alignment 75 bne $1, $single_head_quad # U : still not fully aligned 76 77$do_unroll: 78 addq $16, 64, $7 # E : Initial (+1 trip) wh64 address 79 cmple $18, 127, $1 # E : Can we go through the unrolled loop? 80 bne $1, $tail_quads # U : Nope 81 nop # E : 82 83$unroll_body: 84 wh64 ($7) # L1 : memory subsystem hint: 64 bytes at 85 # ($7) are about to be over-written 86 ldq $6, 0($17) # L0 : bytes 0..7 87 nop # E : 88 nop # E : 89 90 ldq $4, 8($17) # L : bytes 8..15 91 ldq $5, 16($17) # L : bytes 16..23 92 addq $7, 64, $7 # E : Update next wh64 address 93 nop # E : 94 95 ldq $3, 24($17) # L : bytes 24..31 96 addq $16, 64, $1 # E : fallback value for wh64 97 nop # E : 98 nop # E : 99 100 addq $17, 32, $17 # E : src += 32 bytes 101 stq $6, 0($16) # L : bytes 0..7 102 nop # E : 103 nop # E : 104 105 stq $4, 8($16) # L : bytes 8..15 106 stq $5, 16($16) # L : bytes 16..23 107 subq $18, 192, $2 # E : At least two more trips to go? 108 nop # E : 109 110 stq $3, 24($16) # L : bytes 24..31 111 addq $16, 32, $16 # E : dest += 32 bytes 112 nop # E : 113 nop # E : 114 115 ldq $6, 0($17) # L : bytes 0..7 116 ldq $4, 8($17) # L : bytes 8..15 117 cmovlt $2, $1, $7 # E : Latency 2, extra map slot - Use 118 # fallback wh64 address if < 2 more trips 119 nop # E : 120 121 ldq $5, 16($17) # L : bytes 16..23 122 ldq $3, 24($17) # L : bytes 24..31 123 addq $16, 32, $16 # E : dest += 32 124 subq $18, 64, $18 # E : count -= 64 125 126 addq $17, 32, $17 # E : src += 32 127 stq $6, -32($16) # L : bytes 0..7 128 stq $4, -24($16) # L : bytes 8..15 129 cmple $18, 63, $1 # E : At least one more trip? 130 131 stq $5, -16($16) # L : bytes 16..23 132 stq $3, -8($16) # L : bytes 24..31 133 nop # E : 134 beq $1, $unroll_body 135 136$tail_quads: 137$no_unroll: 138 .align 4 139 subq $18, 8, $18 # E : At least a quad left? 140 blt $18, $less_than_8 # U : Nope 141 nop # E : 142 nop # E : 143 144$move_a_quad: 145 ldq $1, 0($17) # L : fetch 8 146 subq $18, 8, $18 # E : count -= 8 147 addq $17, 8, $17 # E : src += 8 148 nop # E : 149 150 stq $1, 0($16) # L : store 8 151 addq $16, 8, $16 # E : dest += 8 152 bge $18, $move_a_quad # U : 153 nop # E : 154 155$less_than_8: 156 .align 4 157 addq $18, 8, $18 # E : add back for trailing bytes 158 ble $18, $nomoredata # U : All-done 159 nop # E : 160 nop # E : 161 162 /* Trailing bytes */ 163$tail_bytes: 164 subq $18, 1, $18 # E : count-- 165 ldbu $1, 0($17) # L : fetch a byte 166 addq $17, 1, $17 # E : src++ 167 nop # E : 168 169 stb $1, 0($16) # L : store a byte 170 addq $16, 1, $16 # E : dest++ 171 bgt $18, $tail_bytes # U : more to be done? 172 nop # E : 173 174 /* branching to exit takes 3 extra cycles, so replicate exit here */ 175 ret $31, ($26), 1 # L0 : 176 nop # E : 177 nop # E : 178 nop # E : 179 180$misaligned: 181 mov $0, $4 # E : dest temp 182 and $0, 7, $1 # E : dest alignment mod8 183 beq $1, $dest_0mod8 # U : life doesnt totally suck 184 nop 185 186$aligndest: 187 ble $18, $nomoredata # U : 188 ldbu $1, 0($17) # L : fetch a byte 189 subq $18, 1, $18 # E : count-- 190 addq $17, 1, $17 # E : src++ 191 192 stb $1, 0($4) # L : store it 193 addq $4, 1, $4 # E : dest++ 194 and $4, 7, $1 # E : dest 0mod8 yet? 195 bne $1, $aligndest # U : go until we are aligned. 196 197 /* Source has unknown alignment, but dest is known to be 0mod8 */ 198$dest_0mod8: 199 subq $18, 8, $18 # E : At least a quad left? 200 blt $18, $misalign_tail # U : Nope 201 ldq_u $3, 0($17) # L : seed (rotating load) of 8 bytes 202 nop # E : 203 204$mis_quad: 205 ldq_u $16, 8($17) # L : Fetch next 8 206 extql $3, $17, $3 # U : masking 207 extqh $16, $17, $1 # U : masking 208 bis $3, $1, $1 # E : merged bytes to store 209 210 subq $18, 8, $18 # E : count -= 8 211 addq $17, 8, $17 # E : src += 8 212 stq $1, 0($4) # L : store 8 (aligned) 213 mov $16, $3 # E : "rotate" source data 214 215 addq $4, 8, $4 # E : dest += 8 216 bge $18, $mis_quad # U : More quads to move 217 nop 218 nop 219 220$misalign_tail: 221 addq $18, 8, $18 # E : account for tail stuff 222 ble $18, $nomoredata # U : 223 nop 224 nop 225 226$misalign_byte: 227 ldbu $1, 0($17) # L : fetch 1 228 subq $18, 1, $18 # E : count-- 229 addq $17, 1, $17 # E : src++ 230 nop # E : 231 232 stb $1, 0($4) # L : store 233 addq $4, 1, $4 # E : dest++ 234 bgt $18, $misalign_byte # U : more to go? 235 nop 236 237 238$nomoredata: 239 ret $31, ($26), 1 # L0 : 240 nop # E : 241 nop # E : 242 nop # E : 243 244 .end memcpy 245 246/* For backwards module compatibility. */ 247__memcpy = memcpy 248.globl __memcpy 249