misc_amips.s 3.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178
  1. #*
  2. #* Some misc routines for Allegrex MIPS
  3. #* (C) notaz, 2007-2008
  4. #*
  5. #* This work is licensed under the terms of MAME license.
  6. #* See COPYING file in the top-level directory.
  7. #*
  8. .set noreorder
  9. .set noat
  10. .text
  11. .align 4
  12. .globl memset32 # int *dest, int c, int count
  13. memset32:
  14. ms32_aloop:
  15. andi $t0, $a0, 0x3f
  16. beqz $t0, ms32_bloop_prep
  17. nop
  18. sw $a1, 0($a0)
  19. addiu $a2, -1
  20. beqz $a2, ms32_return
  21. addiu $a0, 4
  22. j ms32_aloop
  23. nop
  24. ms32_bloop_prep:
  25. srl $t0, $a2, 4 # we will do 64 bytes per iteration (cache line)
  26. beqz $t0, ms32_bloop_end
  27. ms32_bloop:
  28. addiu $t0, -1
  29. cache 0x18, ($a0) # create dirty exclusive
  30. sw $a1, 0x00($a0)
  31. sw $a1, 0x04($a0)
  32. sw $a1, 0x08($a0)
  33. sw $a1, 0x0c($a0)
  34. sw $a1, 0x10($a0)
  35. sw $a1, 0x14($a0)
  36. sw $a1, 0x18($a0)
  37. sw $a1, 0x1c($a0)
  38. sw $a1, 0x20($a0)
  39. sw $a1, 0x24($a0)
  40. sw $a1, 0x28($a0)
  41. sw $a1, 0x2c($a0)
  42. sw $a1, 0x30($a0)
  43. sw $a1, 0x34($a0)
  44. sw $a1, 0x38($a0)
  45. sw $a1, 0x3c($a0)
  46. bnez $t0, ms32_bloop
  47. addiu $a0, 0x40
  48. ms32_bloop_end:
  49. andi $a2, $a2, 0x0f
  50. beqz $a2, ms32_return
  51. ms32_cloop:
  52. addiu $a2, -1
  53. sw $a1, 0($a0)
  54. bnez $a2, ms32_cloop
  55. addiu $a0, 4
  56. ms32_return:
  57. jr $ra
  58. nop
  59. .globl memset32_uncached # int *dest, int c, int count
  60. memset32_uncached:
  61. srl $t0, $a2, 3 # we will do 32 bytes per iteration
  62. beqz $t0, ms32u_bloop_end
  63. ms32u_bloop:
  64. addiu $t0, -1
  65. sw $a1, 0x00($a0)
  66. sw $a1, 0x04($a0)
  67. sw $a1, 0x08($a0)
  68. sw $a1, 0x0c($a0)
  69. sw $a1, 0x10($a0)
  70. sw $a1, 0x14($a0)
  71. sw $a1, 0x18($a0)
  72. sw $a1, 0x1c($a0)
  73. bnez $t0, ms32u_bloop
  74. addiu $a0, 0x20
  75. ms32u_bloop_end:
  76. andi $a2, $a2, 0x0f
  77. beqz $a2, ms32u_return
  78. ms32u_cloop:
  79. addiu $a2, -1
  80. sw $a1, 0($a0)
  81. bnez $a2, ms32u_cloop
  82. addiu $a0, 4
  83. ms32u_return:
  84. jr $ra
  85. nop
  86. .globl memcpy32 # int *dest, int *src, int count
  87. memcpy32:
  88. mc32_aloop:
  89. andi $t0, $a0, 0x3f
  90. beqz $t0, mc32_bloop_prep
  91. nop
  92. lw $t1, 0($a1)
  93. addiu $a2, -1
  94. sw $t1, 0($a0)
  95. beqz $a2, mc32_return
  96. addiu $a0, 4
  97. j mc32_aloop
  98. addiu $a1, 4
  99. mc32_bloop_prep:
  100. srl $t0, $a2, 4 # we will do 64 bytes per iteration (cache line)
  101. beqz $t0, mc32_bloop_end
  102. mc32_bloop:
  103. addiu $t0, -1
  104. cache 0x18, ($a0) # create dirty exclusive
  105. lw $t2, 0x00($a1)
  106. lw $t3, 0x04($a1)
  107. lw $t4, 0x08($a1)
  108. lw $t5, 0x0c($a1)
  109. lw $t6, 0x10($a1)
  110. lw $t7, 0x14($a1)
  111. lw $t8, 0x18($a1)
  112. lw $t9, 0x1c($a1)
  113. sw $t2, 0x00($a0)
  114. sw $t3, 0x04($a0)
  115. sw $t4, 0x08($a0)
  116. sw $t5, 0x0c($a0)
  117. sw $t6, 0x10($a0)
  118. sw $t7, 0x14($a0)
  119. sw $t8, 0x18($a0)
  120. sw $t9, 0x1c($a0)
  121. lw $t2, 0x20($a1)
  122. lw $t3, 0x24($a1)
  123. lw $t4, 0x28($a1)
  124. lw $t5, 0x2c($a1)
  125. lw $t6, 0x30($a1)
  126. lw $t7, 0x34($a1)
  127. lw $t8, 0x38($a1)
  128. lw $t9, 0x3c($a1)
  129. sw $t2, 0x20($a0)
  130. sw $t3, 0x24($a0)
  131. sw $t4, 0x28($a0)
  132. sw $t5, 0x2c($a0)
  133. sw $t6, 0x30($a0)
  134. sw $t7, 0x34($a0)
  135. sw $t8, 0x38($a0)
  136. sw $t9, 0x3c($a0)
  137. addiu $a0, 0x40
  138. bnez $t0, mc32_bloop
  139. addiu $a1, 0x40
  140. mc32_bloop_end:
  141. andi $a2, $a2, 0x0f
  142. beqz $a2, mc32_return
  143. mc32_cloop:
  144. lw $t1, 0($a1)
  145. addiu $a2, -1
  146. addiu $a1, 4
  147. sw $t1, 0($a0)
  148. bnez $a2, mc32_cloop
  149. addiu $a0, 4
  150. mc32_return:
  151. jr $ra
  152. nop
  153. # vim:filetype=mips