mul_amd64.s 2.8 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189
  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 mul(dest, a, b *[5]uint64)
  7. TEXT ·mul(SB),0,$128-24
  8. MOVQ dest+0(FP), DI
  9. MOVQ a+8(FP), SI
  10. MOVQ b+16(FP), DX
  11. MOVQ SP,R11
  12. MOVQ $31,CX
  13. NOTQ CX
  14. ANDQ CX,SP
  15. ADDQ $32,SP
  16. MOVQ R11,0(SP)
  17. MOVQ R12,8(SP)
  18. MOVQ R13,16(SP)
  19. MOVQ R14,24(SP)
  20. MOVQ R15,32(SP)
  21. MOVQ BX,40(SP)
  22. MOVQ BP,48(SP)
  23. MOVQ DI,56(SP)
  24. MOVQ DX,CX
  25. MOVQ 24(SI),DX
  26. IMUL3Q $19,DX,AX
  27. MOVQ AX,64(SP)
  28. MULQ 16(CX)
  29. MOVQ AX,R8
  30. MOVQ DX,R9
  31. MOVQ 32(SI),DX
  32. IMUL3Q $19,DX,AX
  33. MOVQ AX,72(SP)
  34. MULQ 8(CX)
  35. ADDQ AX,R8
  36. ADCQ DX,R9
  37. MOVQ 0(SI),AX
  38. MULQ 0(CX)
  39. ADDQ AX,R8
  40. ADCQ DX,R9
  41. MOVQ 0(SI),AX
  42. MULQ 8(CX)
  43. MOVQ AX,R10
  44. MOVQ DX,R11
  45. MOVQ 0(SI),AX
  46. MULQ 16(CX)
  47. MOVQ AX,R12
  48. MOVQ DX,R13
  49. MOVQ 0(SI),AX
  50. MULQ 24(CX)
  51. MOVQ AX,R14
  52. MOVQ DX,R15
  53. MOVQ 0(SI),AX
  54. MULQ 32(CX)
  55. MOVQ AX,BX
  56. MOVQ DX,BP
  57. MOVQ 8(SI),AX
  58. MULQ 0(CX)
  59. ADDQ AX,R10
  60. ADCQ DX,R11
  61. MOVQ 8(SI),AX
  62. MULQ 8(CX)
  63. ADDQ AX,R12
  64. ADCQ DX,R13
  65. MOVQ 8(SI),AX
  66. MULQ 16(CX)
  67. ADDQ AX,R14
  68. ADCQ DX,R15
  69. MOVQ 8(SI),AX
  70. MULQ 24(CX)
  71. ADDQ AX,BX
  72. ADCQ DX,BP
  73. MOVQ 8(SI),DX
  74. IMUL3Q $19,DX,AX
  75. MULQ 32(CX)
  76. ADDQ AX,R8
  77. ADCQ DX,R9
  78. MOVQ 16(SI),AX
  79. MULQ 0(CX)
  80. ADDQ AX,R12
  81. ADCQ DX,R13
  82. MOVQ 16(SI),AX
  83. MULQ 8(CX)
  84. ADDQ AX,R14
  85. ADCQ DX,R15
  86. MOVQ 16(SI),AX
  87. MULQ 16(CX)
  88. ADDQ AX,BX
  89. ADCQ DX,BP
  90. MOVQ 16(SI),DX
  91. IMUL3Q $19,DX,AX
  92. MULQ 24(CX)
  93. ADDQ AX,R8
  94. ADCQ DX,R9
  95. MOVQ 16(SI),DX
  96. IMUL3Q $19,DX,AX
  97. MULQ 32(CX)
  98. ADDQ AX,R10
  99. ADCQ DX,R11
  100. MOVQ 24(SI),AX
  101. MULQ 0(CX)
  102. ADDQ AX,R14
  103. ADCQ DX,R15
  104. MOVQ 24(SI),AX
  105. MULQ 8(CX)
  106. ADDQ AX,BX
  107. ADCQ DX,BP
  108. MOVQ 64(SP),AX
  109. MULQ 24(CX)
  110. ADDQ AX,R10
  111. ADCQ DX,R11
  112. MOVQ 64(SP),AX
  113. MULQ 32(CX)
  114. ADDQ AX,R12
  115. ADCQ DX,R13
  116. MOVQ 32(SI),AX
  117. MULQ 0(CX)
  118. ADDQ AX,BX
  119. ADCQ DX,BP
  120. MOVQ 72(SP),AX
  121. MULQ 16(CX)
  122. ADDQ AX,R10
  123. ADCQ DX,R11
  124. MOVQ 72(SP),AX
  125. MULQ 24(CX)
  126. ADDQ AX,R12
  127. ADCQ DX,R13
  128. MOVQ 72(SP),AX
  129. MULQ 32(CX)
  130. ADDQ AX,R14
  131. ADCQ DX,R15
  132. MOVQ ·REDMASK51(SB),SI
  133. SHLQ $13,R9:R8
  134. ANDQ SI,R8
  135. SHLQ $13,R11:R10
  136. ANDQ SI,R10
  137. ADDQ R9,R10
  138. SHLQ $13,R13:R12
  139. ANDQ SI,R12
  140. ADDQ R11,R12
  141. SHLQ $13,R15:R14
  142. ANDQ SI,R14
  143. ADDQ R13,R14
  144. SHLQ $13,BP:BX
  145. ANDQ SI,BX
  146. ADDQ R15,BX
  147. IMUL3Q $19,BP,DX
  148. ADDQ DX,R8
  149. MOVQ R8,DX
  150. SHRQ $51,DX
  151. ADDQ R10,DX
  152. MOVQ DX,CX
  153. SHRQ $51,DX
  154. ANDQ SI,R8
  155. ADDQ R12,DX
  156. MOVQ DX,R9
  157. SHRQ $51,DX
  158. ANDQ SI,CX
  159. ADDQ R14,DX
  160. MOVQ DX,AX
  161. SHRQ $51,DX
  162. ANDQ SI,R9
  163. ADDQ BX,DX
  164. MOVQ DX,R10
  165. SHRQ $51,DX
  166. ANDQ SI,AX
  167. IMUL3Q $19,DX,DX
  168. ADDQ DX,R8
  169. ANDQ SI,R10
  170. MOVQ R8,0(DI)
  171. MOVQ CX,8(DI)
  172. MOVQ R9,16(DI)
  173. MOVQ AX,24(DI)
  174. MOVQ R10,32(DI)
  175. MOVQ 0(SP),R11
  176. MOVQ 8(SP),R12
  177. MOVQ 16(SP),R13
  178. MOVQ 24(SP),R14
  179. MOVQ 32(SP),R15
  180. MOVQ 40(SP),BX
  181. MOVQ 48(SP),BP
  182. MOVQ R11,SP
  183. MOVQ DI,AX
  184. MOVQ SI,DX
  185. RET