square_amd64.s 2.3 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151
  1. // Copyright 2012 The Go Authors. All rights reserved.
  2. // Use of this source code is governed by a BSD-style
  3. // license that can be found in the LICENSE file.
  4. // This code was translated into a form compatible with 6a from the public
  5. // domain sources in SUPERCOP: http://bench.cr.yp.to/supercop.html
  6. // func square(dest, src *[5]uint64)
  7. TEXT ·square(SB),7,$96-16
  8. MOVQ dest+0(FP), DI
  9. MOVQ src+8(FP), SI
  10. MOVQ SP,R11
  11. MOVQ $31,CX
  12. NOTQ CX
  13. ANDQ CX,SP
  14. ADDQ $32, SP
  15. MOVQ R11,0(SP)
  16. MOVQ R12,8(SP)
  17. MOVQ R13,16(SP)
  18. MOVQ R14,24(SP)
  19. MOVQ R15,32(SP)
  20. MOVQ BX,40(SP)
  21. MOVQ BP,48(SP)
  22. MOVQ 0(SI),AX
  23. MULQ 0(SI)
  24. MOVQ AX,CX
  25. MOVQ DX,R8
  26. MOVQ 0(SI),AX
  27. SHLQ $1,AX
  28. MULQ 8(SI)
  29. MOVQ AX,R9
  30. MOVQ DX,R10
  31. MOVQ 0(SI),AX
  32. SHLQ $1,AX
  33. MULQ 16(SI)
  34. MOVQ AX,R11
  35. MOVQ DX,R12
  36. MOVQ 0(SI),AX
  37. SHLQ $1,AX
  38. MULQ 24(SI)
  39. MOVQ AX,R13
  40. MOVQ DX,R14
  41. MOVQ 0(SI),AX
  42. SHLQ $1,AX
  43. MULQ 32(SI)
  44. MOVQ AX,R15
  45. MOVQ DX,BX
  46. MOVQ 8(SI),AX
  47. MULQ 8(SI)
  48. ADDQ AX,R11
  49. ADCQ DX,R12
  50. MOVQ 8(SI),AX
  51. SHLQ $1,AX
  52. MULQ 16(SI)
  53. ADDQ AX,R13
  54. ADCQ DX,R14
  55. MOVQ 8(SI),AX
  56. SHLQ $1,AX
  57. MULQ 24(SI)
  58. ADDQ AX,R15
  59. ADCQ DX,BX
  60. MOVQ 8(SI),DX
  61. IMUL3Q $38,DX,AX
  62. MULQ 32(SI)
  63. ADDQ AX,CX
  64. ADCQ DX,R8
  65. MOVQ 16(SI),AX
  66. MULQ 16(SI)
  67. ADDQ AX,R15
  68. ADCQ DX,BX
  69. MOVQ 16(SI),DX
  70. IMUL3Q $38,DX,AX
  71. MULQ 24(SI)
  72. ADDQ AX,CX
  73. ADCQ DX,R8
  74. MOVQ 16(SI),DX
  75. IMUL3Q $38,DX,AX
  76. MULQ 32(SI)
  77. ADDQ AX,R9
  78. ADCQ DX,R10
  79. MOVQ 24(SI),DX
  80. IMUL3Q $19,DX,AX
  81. MULQ 24(SI)
  82. ADDQ AX,R9
  83. ADCQ DX,R10
  84. MOVQ 24(SI),DX
  85. IMUL3Q $38,DX,AX
  86. MULQ 32(SI)
  87. ADDQ AX,R11
  88. ADCQ DX,R12
  89. MOVQ 32(SI),DX
  90. IMUL3Q $19,DX,AX
  91. MULQ 32(SI)
  92. ADDQ AX,R13
  93. ADCQ DX,R14
  94. MOVQ ·REDMASK51(SB),SI
  95. SHLQ $13,R8:CX
  96. ANDQ SI,CX
  97. SHLQ $13,R10:R9
  98. ANDQ SI,R9
  99. ADDQ R8,R9
  100. SHLQ $13,R12:R11
  101. ANDQ SI,R11
  102. ADDQ R10,R11
  103. SHLQ $13,R14:R13
  104. ANDQ SI,R13
  105. ADDQ R12,R13
  106. SHLQ $13,BX:R15
  107. ANDQ SI,R15
  108. ADDQ R14,R15
  109. IMUL3Q $19,BX,DX
  110. ADDQ DX,CX
  111. MOVQ CX,DX
  112. SHRQ $51,DX
  113. ADDQ R9,DX
  114. ANDQ SI,CX
  115. MOVQ DX,R8
  116. SHRQ $51,DX
  117. ADDQ R11,DX
  118. ANDQ SI,R8
  119. MOVQ DX,R9
  120. SHRQ $51,DX
  121. ADDQ R13,DX
  122. ANDQ SI,R9
  123. MOVQ DX,AX
  124. SHRQ $51,DX
  125. ADDQ R15,DX
  126. ANDQ SI,AX
  127. MOVQ DX,R10
  128. SHRQ $51,DX
  129. IMUL3Q $19,DX,DX
  130. ADDQ DX,CX
  131. ANDQ SI,R10
  132. MOVQ CX,0(DI)
  133. MOVQ R8,8(DI)
  134. MOVQ R9,16(DI)
  135. MOVQ AX,24(DI)
  136. MOVQ R10,32(DI)
  137. MOVQ 0(SP),R11
  138. MOVQ 8(SP),R12
  139. MOVQ 16(SP),R13
  140. MOVQ 24(SP),R14
  141. MOVQ 32(SP),R15
  142. MOVQ 40(SP),BX
  143. MOVQ 48(SP),BP
  144. MOVQ R11,SP
  145. MOVQ DI,AX
  146. MOVQ SI,DX
  147. RET