| Paul Mundt | 4466b20 | 2008-12-12 16:34:44 +0900 | [diff] [blame^] | 1 | /* Cloned and hacked for uClibc by Paul Mundt, December 2003 */ |
| 2 | /* Modified by SuperH, Inc. September 2003 */ |
| 3 | ! |
| 4 | ! Fast SH memcpy |
| 5 | ! |
| 6 | ! by Toshiyasu Morita (tm@netcom.com) |
| 7 | ! hacked by J"orn Rernnecke (joern.rennecke@superh.com) ("o for o-umlaut) |
| 8 | ! SH5 code Copyright 2002 SuperH Ltd. |
| 9 | ! |
| 10 | ! Entry: ARG0: destination pointer |
| 11 | ! ARG1: source pointer |
| 12 | ! ARG2: byte count |
| 13 | ! |
| 14 | ! Exit: RESULT: destination pointer |
| 15 | ! any other registers in the range r0-r7: trashed |
| 16 | ! |
| 17 | ! Notes: Usually one wants to do small reads and write a longword, but |
| 18 | ! unfortunately it is difficult in some cases to concatanate bytes |
| 19 | ! into a longword on the SH, so this does a longword read and small |
| 20 | ! writes. |
| 21 | ! |
| 22 | ! This implementation makes two assumptions about how it is called: |
| 23 | ! |
| 24 | ! 1.: If the byte count is nonzero, the address of the last byte to be |
| 25 | ! copied is unsigned greater than the address of the first byte to |
| 26 | ! be copied. This could be easily swapped for a signed comparison, |
| 27 | ! but the algorithm used needs some comparison. |
| 28 | ! |
| 29 | ! 2.: When there are two or three bytes in the last word of an 11-or-more |
| 30 | ! bytes memory chunk to b copied, the rest of the word can be read |
| 31 | ! without side effects. |
| 32 | ! This could be easily changed by increasing the minumum size of |
| 33 | ! a fast memcpy and the amount subtracted from r7 before L_2l_loop be 2, |
| 34 | ! however, this would cost a few extra cyles on average. |
| 35 | ! For SHmedia, the assumption is that any quadword can be read in its |
| 36 | ! enirety if at least one byte is included in the copy. |
| 37 | ! |
| 38 | |
| 39 | .section .text..SHmedia32,"ax" |
| 40 | .globl memcpy |
| 41 | .type memcpy, @function |
| 42 | .align 5 |
| 43 | |
| 44 | memcpy: |
| 45 | |
| 46 | #define LDUAQ(P,O,D0,D1) ldlo.q P,O,D0; ldhi.q P,O+7,D1 |
| 47 | #define STUAQ(P,O,D0,D1) stlo.q P,O,D0; sthi.q P,O+7,D1 |
| 48 | #define LDUAL(P,O,D0,D1) ldlo.l P,O,D0; ldhi.l P,O+3,D1 |
| 49 | #define STUAL(P,O,D0,D1) stlo.l P,O,D0; sthi.l P,O+3,D1 |
| 50 | |
| 51 | ld.b r3,0,r63 |
| 52 | pta/l Large,tr0 |
| 53 | movi 25,r0 |
| 54 | bgeu/u r4,r0,tr0 |
| 55 | nsb r4,r0 |
| 56 | shlli r0,5,r0 |
| 57 | movi (L1-L0+63*32 + 1) & 0xffff,r1 |
| 58 | sub r1, r0, r0 |
| 59 | L0: ptrel r0,tr0 |
| 60 | add r2,r4,r5 |
| 61 | ptabs r18,tr1 |
| 62 | add r3,r4,r6 |
| 63 | blink tr0,r63 |
| 64 | |
| 65 | /* Rearranged to make cut2 safe */ |
| 66 | .balign 8 |
| 67 | L4_7: /* 4..7 byte memcpy cntd. */ |
| 68 | stlo.l r2, 0, r0 |
| 69 | or r6, r7, r6 |
| 70 | sthi.l r5, -1, r6 |
| 71 | stlo.l r5, -4, r6 |
| 72 | blink tr1,r63 |
| 73 | |
| 74 | .balign 8 |
| 75 | L1: /* 0 byte memcpy */ |
| 76 | nop |
| 77 | blink tr1,r63 |
| 78 | nop |
| 79 | nop |
| 80 | nop |
| 81 | nop |
| 82 | |
| 83 | L2_3: /* 2 or 3 byte memcpy cntd. */ |
| 84 | st.b r5,-1,r6 |
| 85 | blink tr1,r63 |
| 86 | |
| 87 | /* 1 byte memcpy */ |
| 88 | ld.b r3,0,r0 |
| 89 | st.b r2,0,r0 |
| 90 | blink tr1,r63 |
| 91 | |
| 92 | L8_15: /* 8..15 byte memcpy cntd. */ |
| 93 | stlo.q r2, 0, r0 |
| 94 | or r6, r7, r6 |
| 95 | sthi.q r5, -1, r6 |
| 96 | stlo.q r5, -8, r6 |
| 97 | blink tr1,r63 |
| 98 | |
| 99 | /* 2 or 3 byte memcpy */ |
| 100 | ld.b r3,0,r0 |
| 101 | ld.b r2,0,r63 |
| 102 | ld.b r3,1,r1 |
| 103 | st.b r2,0,r0 |
| 104 | pta/l L2_3,tr0 |
| 105 | ld.b r6,-1,r6 |
| 106 | st.b r2,1,r1 |
| 107 | blink tr0, r63 |
| 108 | |
| 109 | /* 4 .. 7 byte memcpy */ |
| 110 | LDUAL (r3, 0, r0, r1) |
| 111 | pta L4_7, tr0 |
| 112 | ldlo.l r6, -4, r7 |
| 113 | or r0, r1, r0 |
| 114 | sthi.l r2, 3, r0 |
| 115 | ldhi.l r6, -1, r6 |
| 116 | blink tr0, r63 |
| 117 | |
| 118 | /* 8 .. 15 byte memcpy */ |
| 119 | LDUAQ (r3, 0, r0, r1) |
| 120 | pta L8_15, tr0 |
| 121 | ldlo.q r6, -8, r7 |
| 122 | or r0, r1, r0 |
| 123 | sthi.q r2, 7, r0 |
| 124 | ldhi.q r6, -1, r6 |
| 125 | blink tr0, r63 |
| 126 | |
| 127 | /* 16 .. 24 byte memcpy */ |
| 128 | LDUAQ (r3, 0, r0, r1) |
| 129 | LDUAQ (r3, 8, r8, r9) |
| 130 | or r0, r1, r0 |
| 131 | sthi.q r2, 7, r0 |
| 132 | or r8, r9, r8 |
| 133 | sthi.q r2, 15, r8 |
| 134 | ldlo.q r6, -8, r7 |
| 135 | ldhi.q r6, -1, r6 |
| 136 | stlo.q r2, 8, r8 |
| 137 | stlo.q r2, 0, r0 |
| 138 | or r6, r7, r6 |
| 139 | sthi.q r5, -1, r6 |
| 140 | stlo.q r5, -8, r6 |
| 141 | blink tr1,r63 |
| 142 | |
| 143 | Large: |
| 144 | ld.b r2, 0, r63 |
| 145 | pta/l Loop_ua, tr1 |
| 146 | ori r3, -8, r7 |
| 147 | sub r2, r7, r22 |
| 148 | sub r3, r2, r6 |
| 149 | add r2, r4, r5 |
| 150 | ldlo.q r3, 0, r0 |
| 151 | addi r5, -16, r5 |
| 152 | movi 64+8, r27 // could subtract r7 from that. |
| 153 | stlo.q r2, 0, r0 |
| 154 | sthi.q r2, 7, r0 |
| 155 | ldx.q r22, r6, r0 |
| 156 | bgtu/l r27, r4, tr1 |
| 157 | |
| 158 | addi r5, -48, r27 |
| 159 | pta/l Loop_line, tr0 |
| 160 | addi r6, 64, r36 |
| 161 | addi r6, -24, r19 |
| 162 | addi r6, -16, r20 |
| 163 | addi r6, -8, r21 |
| 164 | |
| 165 | Loop_line: |
| 166 | ldx.q r22, r36, r63 |
| 167 | alloco r22, 32 |
| 168 | addi r22, 32, r22 |
| 169 | ldx.q r22, r19, r23 |
| 170 | sthi.q r22, -25, r0 |
| 171 | ldx.q r22, r20, r24 |
| 172 | ldx.q r22, r21, r25 |
| 173 | stlo.q r22, -32, r0 |
| 174 | ldx.q r22, r6, r0 |
| 175 | sthi.q r22, -17, r23 |
| 176 | sthi.q r22, -9, r24 |
| 177 | sthi.q r22, -1, r25 |
| 178 | stlo.q r22, -24, r23 |
| 179 | stlo.q r22, -16, r24 |
| 180 | stlo.q r22, -8, r25 |
| 181 | bgeu r27, r22, tr0 |
| 182 | |
| 183 | Loop_ua: |
| 184 | addi r22, 8, r22 |
| 185 | sthi.q r22, -1, r0 |
| 186 | stlo.q r22, -8, r0 |
| 187 | ldx.q r22, r6, r0 |
| 188 | bgtu/l r5, r22, tr1 |
| 189 | |
| 190 | add r3, r4, r7 |
| 191 | ldlo.q r7, -8, r1 |
| 192 | sthi.q r22, 7, r0 |
| 193 | ldhi.q r7, -1, r7 |
| 194 | ptabs r18,tr1 |
| 195 | stlo.q r22, 0, r0 |
| 196 | or r1, r7, r1 |
| 197 | sthi.q r5, 15, r1 |
| 198 | stlo.q r5, 8, r1 |
| 199 | blink tr1, r63 |
| 200 | |
| 201 | .size memcpy,.-memcpy |