Advertisement
Bisqwit

Bubsy's code with __restrict__

Oct 20th, 2014
367
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 1.63 KB | None | 0 0
  1. .file "tmp1.cc"
  2. .section .text.unlikely,"ax",@progbits
  3. .LCOLDB0:
  4. .text
  5. .LHOTB0:
  6. .p2align 4,,15
  7. .globl _Z7upscalePKjPj
  8. .type _Z7upscalePKjPj, @function
  9. _Z7upscalePKjPj:
  10. .LFB0:
  11. .cfi_startproc
  12. leaq 1228800(%rsi), %r8
  13. .p2align 4,,10
  14. .p2align 3
  15. .L3:
  16. leaq 2560(%rsi), %rcx
  17. movq %rsi, %rax
  18. movq %rdi, %rdx
  19. .p2align 4,,10
  20. .p2align 3
  21. .L2:
  22. movdqu (%rdx), %xmm0
  23. addq $32, %rax
  24. addq $16, %rdx
  25. pshufd $80, %xmm0, %xmm1
  26. pshufd $250, %xmm0, %xmm0
  27. movups %xmm1, -32(%rax)
  28. movups %xmm0, -16(%rax)
  29. movups %xmm1, 2528(%rax)
  30. movups %xmm0, 2544(%rax)
  31. cmpq %rax, %rcx
  32. jne .L2
  33. addq $5120, %rsi
  34. addq $1280, %rdi
  35. cmpq %r8, %rsi
  36. jne .L3
  37. rep ret
  38. .cfi_endproc
  39. .LFE0:
  40. .size _Z7upscalePKjPj, .-_Z7upscalePKjPj
  41. .section .text.unlikely
  42. .LCOLDE0:
  43. .text
  44. .LHOTE0:
  45. .section .text.unlikely
  46. .LCOLDB1:
  47. .text
  48. .LHOTB1:
  49. .p2align 4,,15
  50. .globl _Z8upscale2PKjPj
  51. .type _Z8upscale2PKjPj, @function
  52. _Z8upscale2PKjPj:
  53. .LFB1:
  54. .cfi_startproc
  55. leaq 1228800(%rsi), %r8
  56. .p2align 4,,10
  57. .p2align 3
  58. .L9:
  59. leaq 2560(%rsi), %rcx
  60. movq %rsi, %rax
  61. movq %rdi, %rdx
  62. .p2align 4,,10
  63. .p2align 3
  64. .L8:
  65. movdqu (%rdx), %xmm0
  66. addq $32, %rax
  67. addq $16, %rdx
  68. pshufd $80, %xmm0, %xmm1
  69. pshufd $250, %xmm0, %xmm0
  70. movups %xmm1, -32(%rax)
  71. movups %xmm0, -16(%rax)
  72. movups %xmm1, 2528(%rax)
  73. movups %xmm0, 2544(%rax)
  74. cmpq %rax, %rcx
  75. jne .L8
  76. addq $5120, %rsi
  77. addq $1280, %rdi
  78. cmpq %r8, %rsi
  79. jne .L9
  80. rep ret
  81. .cfi_endproc
  82. .LFE1:
  83. .size _Z8upscale2PKjPj, .-_Z8upscale2PKjPj
  84. .section .text.unlikely
  85. .LCOLDE1:
  86. .text
  87. .LHOTE1:
  88. .ident "GCC: (Debian 4.9.1-16) 4.9.1"
  89. .section .note.GNU-stack,"",@progbits
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement