memset-sh4.S 1.6 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108
  1. /* SPDX-License-Identifier: GPL-2.0 */
  2. /*
  3. * "memset" implementation for SH4
  4. *
  5. * Copyright (C) 1999 Niibe Yutaka
  6. * Copyright (c) 2009 STMicroelectronics Limited
  7. * Author: Stuart Menefy <stuart.menefy:st.com>
  8. */
  9. /*
  10. * void *memset(void *s, int c, size_t n);
  11. */
  12. #include <linux/linkage.h>
  13. ENTRY(memset)
  14. mov #12,r0
  15. add r6,r4
  16. cmp/gt r6,r0
  17. bt/s 40f ! if it's too small, set a byte at once
  18. mov r4,r0
  19. and #3,r0
  20. cmp/eq #0,r0
  21. bt/s 2f ! It's aligned
  22. sub r0,r6
  23. 1:
  24. dt r0
  25. bf/s 1b
  26. mov.b r5,@-r4
  27. 2: ! make VVVV
  28. extu.b r5,r5
  29. swap.b r5,r0 ! V0
  30. or r0,r5 ! VV
  31. swap.w r5,r0 ! VV00
  32. or r0,r5 ! VVVV
  33. ! Check if enough bytes need to be copied to be worth the big loop
  34. mov #0x40, r0 ! (MT)
  35. cmp/gt r6,r0 ! (MT) 64 > len => slow loop
  36. bt/s 22f
  37. mov r6,r0
  38. ! align the dst to the cache block size if necessary
  39. mov r4, r3
  40. mov #~(0x1f), r1
  41. and r3, r1
  42. cmp/eq r3, r1
  43. bt/s 11f ! dst is already aligned
  44. sub r1, r3 ! r3-r1 -> r3
  45. shlr2 r3 ! number of loops
  46. 10: mov.l r5,@-r4
  47. dt r3
  48. bf/s 10b
  49. add #-4, r6
  50. 11: ! dst is 32byte aligned
  51. mov r6,r2
  52. mov #-5,r0
  53. shld r0,r2 ! number of loops
  54. add #-32, r4
  55. mov r5, r0
  56. 12:
  57. movca.l r0,@r4
  58. mov.l r5,@(4, r4)
  59. mov.l r5,@(8, r4)
  60. mov.l r5,@(12,r4)
  61. mov.l r5,@(16,r4)
  62. mov.l r5,@(20,r4)
  63. add #-0x20, r6
  64. mov.l r5,@(24,r4)
  65. dt r2
  66. mov.l r5,@(28,r4)
  67. bf/s 12b
  68. add #-32, r4
  69. add #32, r4
  70. mov #8, r0
  71. cmp/ge r0, r6
  72. bf 40f
  73. mov r6,r0
  74. 22:
  75. shlr2 r0
  76. shlr r0 ! r0 = r6 >> 3
  77. 3:
  78. dt r0
  79. mov.l r5,@-r4 ! set 8-byte at once
  80. bf/s 3b
  81. mov.l r5,@-r4
  82. !
  83. mov #7,r0
  84. and r0,r6
  85. ! fill bytes (length may be zero)
  86. 40: tst r6,r6
  87. bt 5f
  88. 4:
  89. dt r6
  90. bf/s 4b
  91. mov.b r5,@-r4
  92. 5:
  93. rts
  94. mov r4,r0