Página Principal Explorar Ajuda
Iniciar Sessão
oss
/
uclibc-ng
4
1
Fork 0
Ficheiros Problemas 3 Pull Requests 0 Wiki
Árvore: 0c033f3d8a
Ramos Etiquetas
uclibc-ng
/
libc
/
string
/
sh64
/
memset.S

memset.S 2.0 KB
Histórico Em bruto

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697 
  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. 	.hidden	__memset
    26. 

  26. 	.type	__memset, @function
    27. 

  27. 

  28. 	.align 5
    29. 

  29. 

  30. __memset:
    31. 

  31. 	pta/l multiquad, tr0
    32. 

  32. 	andi r2, 7, r22
    33. 

  33. 	ptabs r18, tr2
    34. 

  34. 	mshflo.b r3,r3,r3
    35. 

  35. 	add r4, r22, r23
    36. 

  36. 	mperm.w r3, r63, r3	// Fill pattern now in every byte of r3
    37. 

  37. 

  38. 	movi 8, r9
    39. 

  39. 	bgtu/u r23, r9, tr0 // multiquad
    40. 

  40. 

  41. 	beqi/u r4, 0, tr2       // Return with size 0 - ensures no mem accesses
    42. 

  42. 	ldlo.q r2, 0, r7
    43. 

  43. 	shlli r4, 2, r4
    44. 

  44. 	movi -1, r8
    45. 

  45. 	SHHI r8, r4, r8
    46. 

  46. 	SHHI r8, r4, r8
    47. 

  47. 	mcmv r7, r8, r3
    48. 

  48. 	stlo.q r2, 0, r3
    49. 

  49. 	blink tr2, r63
    50. 

  50. 

  51. multiquad:
    52. 

  52. 	pta/l lastquad, tr0
    53. 

  53. 	stlo.q r2, 0, r3
    54. 

  54. 	shlri r23, 3, r24
    55. 

  55. 	add r2, r4, r5
    56. 

  56. 	beqi/u r24, 1, tr0 // lastquad
    57. 

  57. 	pta/l loop, tr1
    58. 

  58. 	sub r2, r22, r25
    59. 

  59. 	andi r5, -8, r20   // calculate end address and
    60. 

  60. 	addi r20, -7*8, r8 // loop end address; This might overflow, so we need
    61. 

  61. 	                   // to use a different test before we start the loop
    62. 

  62. 	bge/u r24, r9, tr1 // loop
    63. 

  63. 	st.q r25, 8, r3
    64. 

  64. 	st.q r20, -8, r3
    65. 

  65. 	shlri r24, 1, r24
    66. 

  66. 	beqi/u r24, 1, tr0 // lastquad
    67. 

  67. 	st.q r25, 16, r3
    68. 

  68. 	st.q r20, -16, r3
    69. 

  69. 	beqi/u r24, 2, tr0 // lastquad
    70. 

  70. 	st.q r25, 24, r3
    71. 

  71. 	st.q r20, -24, r3
    72. 

  72. lastquad:
    73. 

  73. 	sthi.q r5, -1, r3
    74. 

  74. 	blink tr2,r63
    75. 

  75. 

  76. loop:
    77. 

  77. !!!	alloco r25, 32	// QQQ comment out for short-term fix to SHUK #3895.
    78. 

  78. 			// QQQ commenting out is locically correct, but sub-optimal
    79. 

  79. 			// QQQ Sean McGoogan - 4th April 2003.
    80. 

  80. 	st.q r25, 8, r3
    81. 

  81. 	st.q r25, 16, r3
    82. 

  82. 	st.q r25, 24, r3
    83. 

  83. 	st.q r25, 32, r3
    84. 

  84. 	addi r25, 32, r25
    85. 

  85. 	bgeu/l r8, r25, tr1 // loop
    86. 

  86. 

  87. 	st.q r20, -40, r3
    88. 

  88. 	st.q r20, -32, r3
    89. 

  89. 	st.q r20, -24, r3
    90. 

  90. 	st.q r20, -16, r3
    91. 

  91. 	st.q r20, -8, r3
    92. 

  92. 	sthi.q r5, -1, r3
    93. 

  93. 	blink tr2,r63
    94. 

  94. 

  95. 	.size	__memset,.-__memset
    96. 

  96. 

  97. strong_alias(__memset,memset)
    98.