Category

Development Platform
Home > SourceCode/Document > Linux-Unix program > View Code
memcpy.S
Package: linux-2.4.20.tar.gz　[view]
Upload User: jlfgdled
Upload Date: 2013-04-10
Package Size: 33168k
Code Size: 2k
Category:

Linux-Unix program

Development Platform:

Unix_Linux

memcpy.S：Code Content
  1. /* Copyright 2002 Andi Kleen */
  3. /*
  4.  * memcpy - Copy a memory block.
  5.  *
  6.  * Input:
  7.  * rdi destination
  8.  * rsi source
  9.  * rdx count
  10.  * 
  11.  * Output:
  12.  * rax original destination
  13.  */
  15.  // #define FIX_ALIGNMENT
  16.   .globl __memcpy
  17. .globl memcpy
  18. .p2align
  19. __memcpy:
  20. memcpy:
  21. pushq %rbx
  22. movq %rdi,%rax
  24. #ifdef FIX_ALIGNMENT
  25. movl %edi,%ecx
  26. andl $7,%ecx
  27. jnz  bad_alignment
  28. after_bad_alignment:
  29. #endif
  31. movq %rdx,%rcx
  32. movl $64,%ebx
  33. shrq $6,%rcx
  34. jz handle_tail
  36. loop_64:
  37. movq (%rsi),%r11
  38. movq 8(%rsi),%r8
  39. movq 2*8(%rsi),%r9
  40. movq 3*8(%rsi),%r10
  41. movq %r11,(%rdi)
  42. movq %r8,1*8(%rdi)
  43. movq %r9,2*8(%rdi)
  44. movq %r10,3*8(%rdi)
  46. movq 4*8(%rsi),%r11
  47. movq 5*8(%rsi),%r8
  48. movq 6*8(%rsi),%r9
  49. movq 7*8(%rsi),%r10
  50. movq %r11,4*8(%rdi)
  51. movq %r8,5*8(%rdi)
  52. movq %r9,6*8(%rdi)
  53. movq %r10,7*8(%rdi)
  55. addq %rbx,%rsi
  56. addq %rbx,%rdi
  57. decl %ecx
  58. jnz  loop_64
  60. handle_tail:
  61. movl %edx,%ecx
  62. andl $63,%ecx
  63. shrl $3,%ecx
  64. jz   handle_7
  65. movl $8,%ebx
  66. loop_8: 
  67. movq (%rsi),%r8
  68. movq %r8,(%rdi) 
  69. addq %rbx,%rdi
  70. addq %rbx,%rsi
  71. decl %ecx
  72. jnz  loop_8
  74. handle_7:
  75. movl %edx,%ecx
  76. andl $7,%ecx
  77. jz ende
  78. loop_1:
  79. movb (%rsi),%r8b
  80. movb %r8b,(%rdi) 
  81. incq %rdi
  82. incq %rsi
  83. decl %ecx
  84. jnz loop_1
  86. ende: 
  87. sfence
  88. popq %rbx
  89. ret
  92. #ifdef FIX_ALIGNMENT
  93. /* align destination */
  94. /* This is simpleminded. For bigger blocks it may make sense to align
  95.    src and dst to their aligned subset and handle the rest separately */
  96. bad_alignment:
  97. movl $8,%r9d
  98. subl %ecx,%r9d
  99. movl %r9d,%ecx
  100. subq %r9,%rdx
  101. js   small_alignment
  102. jz   small_alignment
  103. align_1:
  104. movb (%rsi),%r8b
  105. movb %r8b,(%rdi) 
  106. incq %rdi
  107. incq %rsi
  108. decl %ecx
  109. jnz  align_1
  110. jmp after_bad_alignment
  111. small_alignment:
  112. addq %r9,%rdx
  113. jmp handle_7
  114. #endif