Home Explore Help
Sign In
oss
/
uclibc-ng
4
1
Fork 0
Files Issues 3 Pull Requests 0 Wiki
Tree: a29eb4f241
Branches Tags
uclibc-ng
/
libc
/
string
/
sh64
/
memset.S

memset.S 2.0 KB
History Raw

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596 
  1. /* Cloned and hacked for uClibc by Paul Mundt, December 2003 */
    2. 

  2. /* Modified by SuperH, Inc. September 2003 */
    3. 

  3. !
    4. 

  4. ! Fast SH memset
    5. 

  5. !
    6. 

  6. ! by Toshiyasu Morita (tm@netcom.com)
    7. 

  7. !
    8. 

  8. ! SH5 code by J"orn Rennecke (joern.rennecke@superh.com)
    9. 

  9. ! Copyright 2002 SuperH Ltd.
    10. 

  10. !
    11. 

  11. 

  12. #include <features.h>
    13. 

  13. #include <endian.h>
    14. 

  14. 

  15. #if __BYTE_ORDER == __LITTLE_ENDIAN
    16. 

  16. #define SHHI shlld
    17. 

  17. #define SHLO shlrd
    18. 

  18. #else
    19. 

  19. #define SHHI shlrd
    20. 

  20. #define SHLO shlld
    21. 

  21. #endif
    22. 

  22. 

  23. 	.section .text..SHmedia32,"ax"
    24. 

  24. 	.globl	memset
    25. 

  25. 	.type	memset, @function
    26. 

  26. 

  27. 	.align 5
    28. 

  28. 

  29. memset:
    30. 

  30. 	pta/l multiquad, tr0
    31. 

  31. 	andi r2, 7, r22
    32. 

  32. 	ptabs r18, tr2
    33. 

  33. 	mshflo.b r3,r3,r3
    34. 

  34. 	add r4, r22, r23
    35. 

  35. 	mperm.w r3, r63, r3	// Fill pattern now in every byte of r3
    36. 

  36. 

  37. 	movi 8, r9
    38. 

  38. 	bgtu/u r23, r9, tr0 // multiquad
    39. 

  39. 

  40. 	beqi/u r4, 0, tr2       // Return with size 0 - ensures no mem accesses
    41. 

  41. 	ldlo.q r2, 0, r7
    42. 

  42. 	shlli r4, 2, r4
    43. 

  43. 	movi -1, r8
    44. 

  44. 	SHHI r8, r4, r8
    45. 

  45. 	SHHI r8, r4, r8
    46. 

  46. 	mcmv r7, r8, r3
    47. 

  47. 	stlo.q r2, 0, r3
    48. 

  48. 	blink tr2, r63
    49. 

  49. 

  50. multiquad:
    51. 

  51. 	pta/l lastquad, tr0
    52. 

  52. 	stlo.q r2, 0, r3
    53. 

  53. 	shlri r23, 3, r24
    54. 

  54. 	add r2, r4, r5
    55. 

  55. 	beqi/u r24, 1, tr0 // lastquad
    56. 

  56. 	pta/l loop, tr1
    57. 

  57. 	sub r2, r22, r25
    58. 

  58. 	andi r5, -8, r20   // calculate end address and
    59. 

  59. 	addi r20, -7*8, r8 // loop end address; This might overflow, so we need
    60. 

  60. 	                   // to use a different test before we start the loop
    61. 

  61. 	bge/u r24, r9, tr1 // loop
    62. 

  62. 	st.q r25, 8, r3
    63. 

  63. 	st.q r20, -8, r3
    64. 

  64. 	shlri r24, 1, r24
    65. 

  65. 	beqi/u r24, 1, tr0 // lastquad
    66. 

  66. 	st.q r25, 16, r3
    67. 

  67. 	st.q r20, -16, r3
    68. 

  68. 	beqi/u r24, 2, tr0 // lastquad
    69. 

  69. 	st.q r25, 24, r3
    70. 

  70. 	st.q r20, -24, r3
    71. 

  71. lastquad:
    72. 

  72. 	sthi.q r5, -1, r3
    73. 

  73. 	blink tr2,r63
    74. 

  74. 

  75. loop:
    76. 

  76. !!!	alloco r25, 32	// QQQ comment out for short-term fix to SHUK #3895.
    77. 

  77. 			// QQQ commenting out is locically correct, but sub-optimal
    78. 

  78. 			// QQQ Sean McGoogan - 4th April 2003.
    79. 

  79. 	st.q r25, 8, r3
    80. 

  80. 	st.q r25, 16, r3
    81. 

  81. 	st.q r25, 24, r3
    82. 

  82. 	st.q r25, 32, r3
    83. 

  83. 	addi r25, 32, r25
    84. 

  84. 	bgeu/l r8, r25, tr1 // loop
    85. 

  85. 

  86. 	st.q r20, -40, r3
    87. 

  87. 	st.q r20, -32, r3
    88. 

  88. 	st.q r20, -24, r3
    89. 

  89. 	st.q r20, -16, r3
    90. 

  90. 	st.q r20, -8, r3
    91. 

  91. 	sthi.q r5, -1, r3
    92. 

  92. 	blink tr2,r63
    93. 

  93. 

  94. 	.size	memset,.-memset
    95. 

  95. 

  96. libc_hidden_def(memset)
    97.