memcpy.S 4.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205
  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. #include <features.h>
  39. .section .text..SHmedia32,"ax"
  40. .globl memcpy
  41. .type memcpy, @function
  42. .align 5
  43. memcpy:
  44. #define LDUAQ(P,O,D0,D1) ldlo.q P,O,D0; ldhi.q P,O+7,D1
  45. #define STUAQ(P,O,D0,D1) stlo.q P,O,D0; sthi.q P,O+7,D1
  46. #define LDUAL(P,O,D0,D1) ldlo.l P,O,D0; ldhi.l P,O+3,D1
  47. #define STUAL(P,O,D0,D1) stlo.l P,O,D0; sthi.l P,O+3,D1
  48. ld.b r3,0,r63
  49. pta/l Large,tr0
  50. movi 25,r0
  51. bgeu/u r4,r0,tr0
  52. nsb r4,r0
  53. shlli r0,5,r0
  54. movi (L1-L0+63*32 + 1) & 0xffff,r1
  55. sub r1, r0, r0
  56. L0: ptrel r0,tr0
  57. add r2,r4,r5
  58. ptabs r18,tr1
  59. add r3,r4,r6
  60. blink tr0,r63
  61. /* Rearranged to make cut2 safe */
  62. .balign 8
  63. L4_7: /* 4..7 byte memcpy cntd. */
  64. stlo.l r2, 0, r0
  65. or r6, r7, r6
  66. sthi.l r5, -1, r6
  67. stlo.l r5, -4, r6
  68. blink tr1,r63
  69. .balign 8
  70. L1: /* 0 byte memcpy */
  71. nop
  72. blink tr1,r63
  73. nop
  74. nop
  75. nop
  76. nop
  77. L2_3: /* 2 or 3 byte memcpy cntd. */
  78. st.b r5,-1,r6
  79. blink tr1,r63
  80. /* 1 byte memcpy */
  81. ld.b r3,0,r0
  82. st.b r2,0,r0
  83. blink tr1,r63
  84. L8_15: /* 8..15 byte memcpy cntd. */
  85. stlo.q r2, 0, r0
  86. or r6, r7, r6
  87. sthi.q r5, -1, r6
  88. stlo.q r5, -8, r6
  89. blink tr1,r63
  90. /* 2 or 3 byte memcpy */
  91. ld.b r3,0,r0
  92. ld.b r2,0,r63
  93. ld.b r3,1,r1
  94. st.b r2,0,r0
  95. pta/l L2_3,tr0
  96. ld.b r6,-1,r6
  97. st.b r2,1,r1
  98. blink tr0, r63
  99. /* 4 .. 7 byte memcpy */
  100. LDUAL (r3, 0, r0, r1)
  101. pta L4_7, tr0
  102. ldlo.l r6, -4, r7
  103. or r0, r1, r0
  104. sthi.l r2, 3, r0
  105. ldhi.l r6, -1, r6
  106. blink tr0, r63
  107. /* 8 .. 15 byte memcpy */
  108. LDUAQ (r3, 0, r0, r1)
  109. pta L8_15, tr0
  110. ldlo.q r6, -8, r7
  111. or r0, r1, r0
  112. sthi.q r2, 7, r0
  113. ldhi.q r6, -1, r6
  114. blink tr0, r63
  115. /* 16 .. 24 byte memcpy */
  116. LDUAQ (r3, 0, r0, r1)
  117. LDUAQ (r3, 8, r8, r9)
  118. or r0, r1, r0
  119. sthi.q r2, 7, r0
  120. or r8, r9, r8
  121. sthi.q r2, 15, r8
  122. ldlo.q r6, -8, r7
  123. ldhi.q r6, -1, r6
  124. stlo.q r2, 8, r8
  125. stlo.q r2, 0, r0
  126. or r6, r7, r6
  127. sthi.q r5, -1, r6
  128. stlo.q r5, -8, r6
  129. blink tr1,r63
  130. Large:
  131. ld.b r2, 0, r63
  132. pta/l Loop_ua, tr1
  133. ori r3, -8, r7
  134. sub r2, r7, r22
  135. sub r3, r2, r6
  136. add r2, r4, r5
  137. ldlo.q r3, 0, r0
  138. addi r5, -16, r5
  139. movi 64+8, r27 // could subtract r7 from that.
  140. stlo.q r2, 0, r0
  141. sthi.q r2, 7, r0
  142. ldx.q r22, r6, r0
  143. bgtu/l r27, r4, tr1
  144. addi r5, -48, r27
  145. pta/l Loop_line, tr0
  146. addi r6, 64, r36
  147. addi r6, -24, r19
  148. addi r6, -16, r20
  149. addi r6, -8, r21
  150. Loop_line:
  151. ldx.q r22, r36, r63
  152. alloco r22, 32
  153. addi r22, 32, r22
  154. ldx.q r22, r19, r23
  155. sthi.q r22, -25, r0
  156. ldx.q r22, r20, r24
  157. ldx.q r22, r21, r25
  158. stlo.q r22, -32, r0
  159. ldx.q r22, r6, r0
  160. sthi.q r22, -17, r23
  161. sthi.q r22, -9, r24
  162. sthi.q r22, -1, r25
  163. stlo.q r22, -24, r23
  164. stlo.q r22, -16, r24
  165. stlo.q r22, -8, r25
  166. bgeu r27, r22, tr0
  167. Loop_ua:
  168. addi r22, 8, r22
  169. sthi.q r22, -1, r0
  170. stlo.q r22, -8, r0
  171. ldx.q r22, r6, r0
  172. bgtu/l r5, r22, tr1
  173. add r3, r4, r7
  174. ldlo.q r7, -8, r1
  175. sthi.q r22, 7, r0
  176. ldhi.q r7, -1, r7
  177. ptabs r18,tr1
  178. stlo.q r22, 0, r0
  179. or r1, r7, r1
  180. sthi.q r5, 15, r1
  181. stlo.q r5, 8, r1
  182. blink tr1, r63
  183. .size memcpy,.-memcpy
  184. libc_hidden_def(memcpy)