Advertisement
Guest User

Sin Optimization icpc_O3_mavx

a guest
Feb 4th, 2014
106
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 12.15 KB | None | 0 0
  1. icpc_O3_mavx: file format elf64-x86-64
  2.  
  3. 0000000000402450 <sin1(double)>:
  4. 402450: c5 fb 59 c8 vmulsd %xmm0,%xmm0,%xmm1
  5. 402454: c5 f3 59 15 44 24 00 vmulsd 0x2444(%rip),%xmm1,%xmm2 # 4048a0 <.L_2il0floatpacket.442+0x80>
  6. 40245b: 00
  7. 40245c: c5 eb 58 1d 34 24 00 vaddsd 0x2434(%rip),%xmm2,%xmm3 # 404898 <.L_2il0floatpacket.442+0x78>
  8. 402463: 00
  9. 402464: c5 e3 59 e0 vmulsd %xmm0,%xmm3,%xmm4
  10. 402468: c5 db 59 e8 vmulsd %xmm0,%xmm4,%xmm5
  11. 40246c: c5 d3 58 35 34 24 00 vaddsd 0x2434(%rip),%xmm5,%xmm6 # 4048a8 <.L_2il0floatpacket.442+0x88>
  12. 402473: 00
  13. 402474: c5 cb 59 f8 vmulsd %xmm0,%xmm6,%xmm7
  14. 402478: c5 43 59 c0 vmulsd %xmm0,%xmm7,%xmm8
  15. 40247c: c5 3b 58 0d 2c 24 00 vaddsd 0x242c(%rip),%xmm8,%xmm9 # 4048b0 <.L_2il0floatpacket.442+0x90>
  16. 402483: 00
  17. 402484: c5 33 59 d0 vmulsd %xmm0,%xmm9,%xmm10
  18. 402488: c5 2b 59 d8 vmulsd %xmm0,%xmm10,%xmm11
  19. 40248c: c5 23 58 25 24 24 00 vaddsd 0x2424(%rip),%xmm11,%xmm12 # 4048b8 <.L_2il0floatpacket.442+0x98>
  20. 402493: 00
  21. 402494: c5 1b 59 e8 vmulsd %xmm0,%xmm12,%xmm13
  22. 402498: c5 13 59 f0 vmulsd %xmm0,%xmm13,%xmm14
  23. 40249c: c5 0b 58 3d 1c 24 00 vaddsd 0x241c(%rip),%xmm14,%xmm15 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  24. 4024a3: 00
  25. 4024a4: c5 83 59 c8 vmulsd %xmm0,%xmm15,%xmm1
  26. 4024a8: c5 f3 59 d0 vmulsd %xmm0,%xmm1,%xmm2
  27. 4024ac: c5 eb 58 1d 14 24 00 vaddsd 0x2414(%rip),%xmm2,%xmm3 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  28. 4024b3: 00
  29. 4024b4: c5 e3 59 e0 vmulsd %xmm0,%xmm3,%xmm4
  30. 4024b8: c5 db 59 e8 vmulsd %xmm0,%xmm4,%xmm5
  31. 4024bc: c5 d3 59 f0 vmulsd %xmm0,%xmm5,%xmm6
  32. 4024c0: c5 cb 58 c0 vaddsd %xmm0,%xmm6,%xmm0
  33. 4024c4: c3 retq
  34. 4024c5: 0f 1f 40 00 nopl 0x0(%rax)
  35. 4024c9: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  36.  
  37. 00000000004024d0 <sin2(double)>:
  38. 4024d0: c5 f9 28 c8 vmovapd %xmm0,%xmm1
  39. 4024d4: c5 f3 59 c1 vmulsd %xmm1,%xmm1,%xmm0
  40. 4024d8: c5 f3 59 d8 vmulsd %xmm0,%xmm1,%xmm3
  41. 4024dc: c5 7b 59 2d bc 23 00 vmulsd 0x23bc(%rip),%xmm0,%xmm13 # 4048a0 <.L_2il0floatpacket.442+0x80>
  42. 4024e3: 00
  43. 4024e4: c5 e3 59 15 dc 23 00 vmulsd 0x23dc(%rip),%xmm3,%xmm2 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  44. 4024eb: 00
  45. 4024ec: c5 fb 59 f3 vmulsd %xmm3,%xmm0,%xmm6
  46. 4024f0: c5 13 58 35 a0 23 00 vaddsd 0x23a0(%rip),%xmm13,%xmm14 # 404898 <.L_2il0floatpacket.442+0x78>
  47. 4024f7: 00
  48. 4024f8: c5 f3 58 e2 vaddsd %xmm2,%xmm1,%xmm4
  49. 4024fc: c5 cb 59 2d bc 23 00 vmulsd 0x23bc(%rip),%xmm6,%xmm5 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  50. 402503: 00
  51. 402504: c5 7b 59 ce vmulsd %xmm6,%xmm0,%xmm9
  52. 402508: c5 0b 59 f8 vmulsd %xmm0,%xmm14,%xmm15
  53. 40250c: c5 db 58 fd vaddsd %xmm5,%xmm4,%xmm7
  54. 402510: c5 33 59 05 a0 23 00 vmulsd 0x23a0(%rip),%xmm9,%xmm8 # 4048b8 <.L_2il0floatpacket.442+0x98>
  55. 402517: 00
  56. 402518: c4 41 7b 59 e1 vmulsd %xmm9,%xmm0,%xmm12
  57. 40251d: c4 41 43 58 d0 vaddsd %xmm8,%xmm7,%xmm10
  58. 402522: c5 1b 59 1d 86 23 00 vmulsd 0x2386(%rip),%xmm12,%xmm11 # 4048b0 <.L_2il0floatpacket.442+0x90>
  59. 402529: 00
  60. 40252a: c4 41 7b 59 e4 vmulsd %xmm12,%xmm0,%xmm12
  61. 40252f: c5 83 58 05 71 23 00 vaddsd 0x2371(%rip),%xmm15,%xmm0 # 4048a8 <.L_2il0floatpacket.442+0x88>
  62. 402536: 00
  63. 402537: c4 c1 2b 58 cb vaddsd %xmm11,%xmm10,%xmm1
  64. 40253c: c4 41 7b 59 ec vmulsd %xmm12,%xmm0,%xmm13
  65. 402541: c5 93 58 c1 vaddsd %xmm1,%xmm13,%xmm0
  66. 402545: c3 retq
  67. 402546: 0f 1f 00 nopl (%rax)
  68. 402549: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  69.  
  70. 0000000000402550 <sin3(double)>:
  71. 402550: c5 7b 59 e0 vmulsd %xmm0,%xmm0,%xmm12
  72. 402554: c5 9b 59 0d 44 23 00 vmulsd 0x2344(%rip),%xmm12,%xmm1 # 4048a0 <.L_2il0floatpacket.442+0x80>
  73. 40255b: 00
  74. 40255c: c5 f3 58 15 34 23 00 vaddsd 0x2334(%rip),%xmm1,%xmm2 # 404898 <.L_2il0floatpacket.442+0x78>
  75. 402563: 00
  76. 402564: c5 9b 59 da vmulsd %xmm2,%xmm12,%xmm3
  77. 402568: c5 e3 58 25 38 23 00 vaddsd 0x2338(%rip),%xmm3,%xmm4 # 4048a8 <.L_2il0floatpacket.442+0x88>
  78. 40256f: 00
  79. 402570: c5 9b 59 ec vmulsd %xmm4,%xmm12,%xmm5
  80. 402574: c5 d3 58 35 34 23 00 vaddsd 0x2334(%rip),%xmm5,%xmm6 # 4048b0 <.L_2il0floatpacket.442+0x90>
  81. 40257b: 00
  82. 40257c: c5 9b 59 fe vmulsd %xmm6,%xmm12,%xmm7
  83. 402580: c5 43 58 05 30 23 00 vaddsd 0x2330(%rip),%xmm7,%xmm8 # 4048b8 <.L_2il0floatpacket.442+0x98>
  84. 402587: 00
  85. 402588: c4 41 1b 59 c8 vmulsd %xmm8,%xmm12,%xmm9
  86. 40258d: c5 33 58 15 2b 23 00 vaddsd 0x232b(%rip),%xmm9,%xmm10 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  87. 402594: 00
  88. 402595: c4 41 1b 59 da vmulsd %xmm10,%xmm12,%xmm11
  89. 40259a: c5 23 58 2d 26 23 00 vaddsd 0x2326(%rip),%xmm11,%xmm13 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  90. 4025a1: 00
  91. 4025a2: c4 41 1b 59 f5 vmulsd %xmm13,%xmm12,%xmm14
  92. 4025a7: c5 0b 58 3d 21 23 00 vaddsd 0x2321(%rip),%xmm14,%xmm15 # 4048d0 <.L_2il0floatpacket.442+0xb0>
  93. 4025ae: 00
  94. 4025af: c4 c1 7b 59 c7 vmulsd %xmm15,%xmm0,%xmm0
  95. 4025b4: c3 retq
  96. 4025b5: 0f 1f 40 00 nopl 0x0(%rax)
  97. 4025b9: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  98.  
  99. 00000000004025c0 <sin4(double)>:
  100. 4025c0: c5 7b 59 e0 vmulsd %xmm0,%xmm0,%xmm12
  101. 4025c4: c4 41 1b 59 cc vmulsd %xmm12,%xmm12,%xmm9
  102. 4025c9: c5 b3 59 35 cf 22 00 vmulsd 0x22cf(%rip),%xmm9,%xmm6 # 4048a0 <.L_2il0floatpacket.442+0x80>
  103. 4025d0: 00
  104. 4025d1: c5 b3 59 0d bf 22 00 vmulsd 0x22bf(%rip),%xmm9,%xmm1 # 404898 <.L_2il0floatpacket.442+0x78>
  105. 4025d8: 00
  106. 4025d9: c5 cb 58 3d c7 22 00 vaddsd 0x22c7(%rip),%xmm6,%xmm7 # 4048a8 <.L_2il0floatpacket.442+0x88>
  107. 4025e0: 00
  108. 4025e1: c5 f3 58 15 c7 22 00 vaddsd 0x22c7(%rip),%xmm1,%xmm2 # 4048b0 <.L_2il0floatpacket.442+0x90>
  109. 4025e8: 00
  110. 4025e9: c5 33 59 c7 vmulsd %xmm7,%xmm9,%xmm8
  111. 4025ed: c5 b3 59 da vmulsd %xmm2,%xmm9,%xmm3
  112. 4025f1: c5 3b 58 15 bf 22 00 vaddsd 0x22bf(%rip),%xmm8,%xmm10 # 4048b8 <.L_2il0floatpacket.442+0x98>
  113. 4025f8: 00
  114. 4025f9: c5 e3 58 25 bf 22 00 vaddsd 0x22bf(%rip),%xmm3,%xmm4 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  115. 402600: 00
  116. 402601: c4 41 33 59 da vmulsd %xmm10,%xmm9,%xmm11
  117. 402606: c5 b3 59 ec vmulsd %xmm4,%xmm9,%xmm5
  118. 40260a: c5 23 58 2d b6 22 00 vaddsd 0x22b6(%rip),%xmm11,%xmm13 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  119. 402611: 00
  120. 402612: c5 53 58 35 b6 22 00 vaddsd 0x22b6(%rip),%xmm5,%xmm14 # 4048d0 <.L_2il0floatpacket.442+0xb0>
  121. 402619: 00
  122. 40261a: c4 41 1b 59 fd vmulsd %xmm13,%xmm12,%xmm15
  123. 40261f: c4 c1 0b 58 cf vaddsd %xmm15,%xmm14,%xmm1
  124. 402624: c5 fb 59 c1 vmulsd %xmm1,%xmm0,%xmm0
  125. 402628: c3 retq
  126. 402629: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  127.  
  128. 0000000000402630 <sin5(double)>:
  129. 402630: c5 fb 59 f8 vmulsd %xmm0,%xmm0,%xmm7
  130. 402634: c5 43 59 ef vmulsd %xmm7,%xmm7,%xmm13
  131. 402638: c4 41 43 59 dd vmulsd %xmm13,%xmm7,%xmm11
  132. 40263d: c5 a3 59 25 5b 22 00 vmulsd 0x225b(%rip),%xmm11,%xmm4 # 4048a0 <.L_2il0floatpacket.442+0x80>
  133. 402644: 00
  134. 402645: c5 a3 59 0d 4b 22 00 vmulsd 0x224b(%rip),%xmm11,%xmm1 # 404898 <.L_2il0floatpacket.442+0x78>
  135. 40264c: 00
  136. 40264d: c5 23 59 25 53 22 00 vmulsd 0x2253(%rip),%xmm11,%xmm12 # 4048a8 <.L_2il0floatpacket.442+0x88>
  137. 402654: 00
  138. 402655: c5 db 58 2d 53 22 00 vaddsd 0x2253(%rip),%xmm4,%xmm5 # 4048b0 <.L_2il0floatpacket.442+0x90>
  139. 40265c: 00
  140. 40265d: c5 f3 58 15 53 22 00 vaddsd 0x2253(%rip),%xmm1,%xmm2 # 4048b8 <.L_2il0floatpacket.442+0x98>
  141. 402664: 00
  142. 402665: c5 1b 58 35 53 22 00 vaddsd 0x2253(%rip),%xmm12,%xmm14 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  143. 40266c: 00
  144. 40266d: c5 a3 59 f5 vmulsd %xmm5,%xmm11,%xmm6
  145. 402671: c5 a3 59 da vmulsd %xmm2,%xmm11,%xmm3
  146. 402675: c4 c1 13 59 ce vmulsd %xmm14,%xmm13,%xmm1
  147. 40267a: c5 4b 58 05 46 22 00 vaddsd 0x2246(%rip),%xmm6,%xmm8 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  148. 402681: 00
  149. 402682: c5 63 58 0d 46 22 00 vaddsd 0x2246(%rip),%xmm3,%xmm9 # 4048d0 <.L_2il0floatpacket.442+0xb0>
  150. 402689: 00
  151. 40268a: c4 41 43 59 d0 vmulsd %xmm8,%xmm7,%xmm10
  152. 40268f: c4 41 33 58 fa vaddsd %xmm10,%xmm9,%xmm15
  153. 402694: c5 83 58 d1 vaddsd %xmm1,%xmm15,%xmm2
  154. 402698: c5 fb 59 c2 vmulsd %xmm2,%xmm0,%xmm0
  155. 40269c: c3 retq
  156. 40269d: 0f 1f 00 nopl (%rax)
  157.  
  158. 00000000004026a0 <sin6(double)>:
  159. 4026a0: c5 7b 59 d0 vmulsd %xmm0,%xmm0,%xmm10
  160. 4026a4: c5 ab 59 3d f4 21 00 vmulsd 0x21f4(%rip),%xmm10,%xmm7 # 4048a0 <.L_2il0floatpacket.442+0x80>
  161. 4026ab: 00
  162. 4026ac: c5 ab 59 0d 04 22 00 vmulsd 0x2204(%rip),%xmm10,%xmm1 # 4048b8 <.L_2il0floatpacket.442+0x98>
  163. 4026b3: 00
  164. 4026b4: c4 c1 2b 59 f2 vmulsd %xmm10,%xmm10,%xmm6
  165. 4026b9: c5 43 58 05 d7 21 00 vaddsd 0x21d7(%rip),%xmm7,%xmm8 # 404898 <.L_2il0floatpacket.442+0x78>
  166. 4026c0: 00
  167. 4026c1: c5 f3 58 15 f7 21 00 vaddsd 0x21f7(%rip),%xmm1,%xmm2 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  168. 4026c8: 00
  169. 4026c9: c5 4b 59 ee vmulsd %xmm6,%xmm6,%xmm13
  170. 4026cd: c4 41 2b 59 c8 vmulsd %xmm8,%xmm10,%xmm9
  171. 4026d2: c5 ab 59 da vmulsd %xmm2,%xmm10,%xmm3
  172. 4026d6: c5 33 58 1d ca 21 00 vaddsd 0x21ca(%rip),%xmm9,%xmm11 # 4048a8 <.L_2il0floatpacket.442+0x88>
  173. 4026dd: 00
  174. 4026de: c5 e3 58 25 e2 21 00 vaddsd 0x21e2(%rip),%xmm3,%xmm4 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  175. 4026e5: 00
  176. 4026e6: c4 41 2b 59 e3 vmulsd %xmm11,%xmm10,%xmm12
  177. 4026eb: c5 ab 59 ec vmulsd %xmm4,%xmm10,%xmm5
  178. 4026ef: c5 1b 58 35 b9 21 00 vaddsd 0x21b9(%rip),%xmm12,%xmm14 # 4048b0 <.L_2il0floatpacket.442+0x90>
  179. 4026f6: 00
  180. 4026f7: c5 53 58 3d d1 21 00 vaddsd 0x21d1(%rip),%xmm5,%xmm15 # 4048d0 <.L_2il0floatpacket.442+0xb0>
  181. 4026fe: 00
  182. 4026ff: c4 c1 13 59 ce vmulsd %xmm14,%xmm13,%xmm1
  183. 402704: c5 83 58 d1 vaddsd %xmm1,%xmm15,%xmm2
  184. 402708: c5 fb 59 c2 vmulsd %xmm2,%xmm0,%xmm0
  185. 40270c: c3 retq
  186. 40270d: 0f 1f 00 nopl (%rax)
  187.  
  188. 0000000000402710 <sin7(double)>:
  189. 402710: c5 f9 28 c8 vmovapd %xmm0,%xmm1
  190. 402714: c5 73 59 e1 vmulsd %xmm1,%xmm1,%xmm12
  191. 402718: c4 c1 1b 59 dc vmulsd %xmm12,%xmm12,%xmm3
  192. 40271d: c4 c1 73 59 fc vmulsd %xmm12,%xmm1,%xmm7
  193. 402722: c5 e3 59 d3 vmulsd %xmm3,%xmm3,%xmm2
  194. 402726: c5 f3 59 c2 vmulsd %xmm2,%xmm1,%xmm0
  195. 40272a: c5 1b 59 0d 6e 21 00 vmulsd 0x216e(%rip),%xmm12,%xmm9 # 4048a0 <.L_2il0floatpacket.442+0x80>
  196. 402731: 00
  197. 402732: c5 9b 59 25 7e 21 00 vmulsd 0x217e(%rip),%xmm12,%xmm4 # 4048b8 <.L_2il0floatpacket.442+0x98>
  198. 402739: 00
  199. 40273a: c5 33 58 15 56 21 00 vaddsd 0x2156(%rip),%xmm9,%xmm10 # 404898 <.L_2il0floatpacket.442+0x78>
  200. 402741: 00
  201. 402742: c5 db 58 2d 76 21 00 vaddsd 0x2176(%rip),%xmm4,%xmm5 # 4048c0 <.L_2il0floatpacket.442+0xa0>
  202. 402749: 00
  203. 40274a: c4 41 1b 59 da vmulsd %xmm10,%xmm12,%xmm11
  204. 40274f: c5 9b 59 f5 vmulsd %xmm5,%xmm12,%xmm6
  205. 402753: c5 23 58 2d 4d 21 00 vaddsd 0x214d(%rip),%xmm11,%xmm13 # 4048a8 <.L_2il0floatpacket.442+0x88>
  206. 40275a: 00
  207. 40275b: c5 4b 58 05 65 21 00 vaddsd 0x2165(%rip),%xmm6,%xmm8 # 4048c8 <.L_2il0floatpacket.442+0xa8>
  208. 402762: 00
  209. 402763: c4 41 1b 59 f5 vmulsd %xmm13,%xmm12,%xmm14
  210. 402768: c4 c1 43 59 d0 vmulsd %xmm8,%xmm7,%xmm2
  211. 40276d: c5 0b 58 3d 3b 21 00 vaddsd 0x213b(%rip),%xmm14,%xmm15 # 4048b0 <.L_2il0floatpacket.442+0x90>
  212. 402774: 00
  213. 402775: c4 c1 7b 59 c7 vmulsd %xmm15,%xmm0,%xmm0
  214. 40277a: c5 eb 58 d8 vaddsd %xmm0,%xmm2,%xmm3
  215. 40277e: c5 e3 58 c1 vaddsd %xmm1,%xmm3,%xmm0
  216. 402782: c3 retq
  217. 402783: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1)
  218. 402788: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1)
  219. 40278f: 00
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement