Guest User

icc -fno-alias -Wall -std=c99 -finline-functions -Ofast -ma

a guest
Dec 16th, 2013
149
0
Never
Not a member of Pastebin yet? Sign Up, it unlocks many cool features!
text 18.85 KB | None | 0 0
  1. 0000000000400970 <main>:
  2. 400970: 55 push %rbp
  3. 400971: 48 89 e5 mov %rsp,%rbp
  4. 400974: 48 83 e4 80 and $0xffffffffffffff80,%rsp
  5. 400978: 41 56 push %r14
  6. 40097a: 41 57 push %r15
  7. 40097c: 48 83 ec 70 sub $0x70,%rsp
  8. 400980: 41 89 fe mov %edi,%r14d
  9. 400983: bf 03 00 00 00 mov $0x3,%edi
  10. 400988: 49 89 f7 mov %rsi,%r15
  11. 40098b: e8 e0 07 00 00 callq 401170 <__intel_new_proc_init_G>
  12. 400990: c5 f8 ae 5c 24 08 vstmxcsr 0x8(%rsp)
  13. 400996: 81 4c 24 08 40 80 00 orl $0x8040,0x8(%rsp)
  14. 40099d: 00
  15. 40099e: c5 f8 ae 54 24 08 vldmxcsr 0x8(%rsp)
  16. 4009a4: 41 83 fe 01 cmp $0x1,%r14d
  17. 4009a8: 0f 8e fa 03 00 00 jle 400da8 <main+0x438>
  18. 4009ae: 49 8b 77 08 mov 0x8(%r15),%rsi
  19. 4009b2: 8a 06 mov (%rsi),%al
  20. 4009b4: 3c 30 cmp $0x30,%al
  21. 4009b6: 0f 84 dc 03 00 00 je 400d98 <main+0x428>
  22. 4009bc: 3c 31 cmp $0x31,%al
  23. 4009be: 0f 85 b6 03 00 00 jne 400d7a <main+0x40a>
  24. 4009c4: b8 b0 0e 40 00 mov $0x400eb0,%eax
  25. 4009c9: 49 89 c6 mov %rax,%r14
  26. 4009cc: ba 04 00 00 00 mov $0x4,%edx
  27. 4009d1: 33 c9 xor %ecx,%ecx
  28. 4009d3: c5 fa 6f 15 35 2c 00 vmovdqu 0x2c35(%rip),%xmm2 # 403610 <.L_2il0floatpacket.14>
  29. 4009da: 00
  30. 4009db: c5 f9 6e da vmovd %edx,%xmm3
  31. 4009df: 33 d2 xor %edx,%edx
  32. 4009e1: c5 f9 70 e3 00 vpshufd $0x0,%xmm3,%xmm4
  33. 4009e6: c5 e9 fe cc vpaddd %xmm4,%xmm2,%xmm1
  34. 4009ea: c5 f1 fe c4 vpaddd %xmm4,%xmm1,%xmm0
  35. 4009ee: c5 f9 fe ec vpaddd %xmm4,%xmm0,%xmm5
  36. 4009f2: c5 f8 5b da vcvtdq2ps %xmm2,%xmm3
  37. 4009f6: c5 f8 5b d1 vcvtdq2ps %xmm1,%xmm2
  38. 4009fa: c5 f8 5b c8 vcvtdq2ps %xmm0,%xmm1
  39. 4009fe: c5 f8 5b c5 vcvtdq2ps %xmm5,%xmm0
  40. 400a02: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  41. 400a09: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  42. 400a10: 48 ff c1 inc %rcx
  43. 400a13: c5 f8 11 9a e0 72 60 vmovups %xmm3,0x6072e0(%rdx)
  44. 400a1a: 00
  45. 400a1b: c5 f8 11 92 f0 72 60 vmovups %xmm2,0x6072f0(%rdx)
  46. 400a22: 00
  47. 400a23: c5 f8 11 8a 00 73 60 vmovups %xmm1,0x607300(%rdx)
  48. 400a2a: 00
  49. 400a2b: c5 f8 11 82 10 73 60 vmovups %xmm0,0x607310(%rdx)
  50. 400a32: 00
  51. 400a33: c5 f8 11 9a 20 73 60 vmovups %xmm3,0x607320(%rdx)
  52. 400a3a: 00
  53. 400a3b: c5 f8 11 92 30 73 60 vmovups %xmm2,0x607330(%rdx)
  54. 400a42: 00
  55. 400a43: c5 f8 11 8a 40 73 60 vmovups %xmm1,0x607340(%rdx)
  56. 400a4a: 00
  57. 400a4b: c5 f8 11 82 50 73 60 vmovups %xmm0,0x607350(%rdx)
  58. 400a52: 00
  59. 400a53: 48 81 c2 80 00 00 00 add $0x80,%rdx
  60. 400a5a: 48 81 f9 f4 01 00 00 cmp $0x1f4,%rcx
  61. 400a61: 72 ad jb 400a10 <main+0xa0>
  62. 400a63: 48 89 1c 24 mov %rbx,(%rsp)
  63. 400a67: 45 33 ff xor %r15d,%r15d
  64. 400a6a: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
  65. 400a6e: 48 89 c3 mov %rax,%rbx
  66. 400a71: bf 40 63 60 00 mov $0x606340,%edi
  67. 400a76: 33 f6 xor %esi,%esi
  68. 400a78: ba a0 0f 00 00 mov $0xfa0,%edx
  69. 400a7d: e8 be 06 00 00 callq 401140 <_intel_fast_memset>
  70. 400a82: 33 c9 xor %ecx,%ecx
  71. 400a84: 33 d2 xor %edx,%edx
  72. 400a86: 49 3b de cmp %r14,%rbx
  73. 400a89: 0f 84 d0 00 00 00 je 400b5f <main+0x1ef>
  74. 400a8f: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
  75. 400a93: c5 fc 10 92 e0 72 60 vmovups 0x6072e0(%rdx),%ymm2
  76. 400a9a: 00
  77. 400a9b: c5 fc 10 aa 00 73 60 vmovups 0x607300(%rdx),%ymm5
  78. 400aa2: 00
  79. 400aa3: c5 fa 10 0c cd 40 63 vmovss 0x606340(,%rcx,8),%xmm1
  80. 400aaa: 60 00
  81. 400aac: c5 7a 10 3c cd 44 63 vmovss 0x606344(,%rcx,8),%xmm15
  82. 400ab3: 60 00
  83. 400ab5: c5 ec 59 e2 vmulps %ymm2,%ymm2,%ymm4
  84. 400ab9: c5 d4 59 fd vmulps %ymm5,%ymm5,%ymm7
  85. 400abd: c4 e3 75 18 d8 01 vinsertf128 $0x1,%xmm0,%ymm1,%ymm3
  86. 400ac3: c5 fc 10 8a 20 73 60 vmovups 0x607320(%rdx),%ymm1
  87. 400aca: 00
  88. 400acb: c5 e4 58 f4 vaddps %ymm4,%ymm3,%ymm6
  89. 400acf: c5 fc 10 a2 40 73 60 vmovups 0x607340(%rdx),%ymm4
  90. 400ad6: 00
  91. 400ad7: c5 f4 59 d9 vmulps %ymm1,%ymm1,%ymm3
  92. 400adb: c5 4c 58 c7 vaddps %ymm7,%ymm6,%ymm8
  93. 400adf: c5 dc 59 f4 vmulps %ymm4,%ymm4,%ymm6
  94. 400ae3: 48 81 c2 80 00 00 00 add $0x80,%rdx
  95. 400aea: c4 e3 05 18 d0 01 vinsertf128 $0x1,%xmm0,%ymm15,%ymm2
  96. 400af0: c5 ec 58 eb vaddps %ymm3,%ymm2,%ymm5
  97. 400af4: c5 d4 58 fe vaddps %ymm6,%ymm5,%ymm7
  98. 400af8: c4 43 7d 19 c1 01 vextractf128 $0x1,%ymm8,%xmm9
  99. 400afe: c4 41 38 58 d1 vaddps %xmm9,%xmm8,%xmm10
  100. 400b03: c4 41 28 12 da vmovhlps %xmm10,%xmm10,%xmm11
  101. 400b08: c4 41 28 58 e3 vaddps %xmm11,%xmm10,%xmm12
  102. 400b0d: c4 41 18 c6 ec f5 vshufps $0xf5,%xmm12,%xmm12,%xmm13
  103. 400b13: c4 41 1a 58 f5 vaddss %xmm13,%xmm12,%xmm14
  104. 400b18: c5 7a 11 34 cd 40 63 vmovss %xmm14,0x606340(,%rcx,8)
  105. 400b1f: 60 00
  106. 400b21: c4 c3 7d 19 f8 01 vextractf128 $0x1,%ymm7,%xmm8
  107. 400b27: c4 41 40 58 c8 vaddps %xmm8,%xmm7,%xmm9
  108. 400b2c: c4 41 30 12 d1 vmovhlps %xmm9,%xmm9,%xmm10
  109. 400b31: c4 41 30 58 da vaddps %xmm10,%xmm9,%xmm11
  110. 400b36: c4 41 20 c6 e3 f5 vshufps $0xf5,%xmm11,%xmm11,%xmm12
  111. 400b3c: c4 41 22 58 ec vaddss %xmm12,%xmm11,%xmm13
  112. 400b41: c5 7a 11 2c cd 44 63 vmovss %xmm13,0x606344(,%rcx,8)
  113. 400b48: 60 00
  114. 400b4a: 48 ff c1 inc %rcx
  115. 400b4d: 48 81 f9 f4 01 00 00 cmp $0x1f4,%rcx
  116. 400b54: 0f 82 39 ff ff ff jb 400a93 <main+0x123>
  117. 400b5a: e9 cb 00 00 00 jmpq 400c2a <main+0x2ba>
  118. 400b5f: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
  119. 400b63: c5 fc 10 92 e0 72 60 vmovups 0x6072e0(%rdx),%ymm2
  120. 400b6a: 00
  121. 400b6b: c5 fc 10 aa 00 73 60 vmovups 0x607300(%rdx),%ymm5
  122. 400b72: 00
  123. 400b73: c5 fa 10 0c cd 40 63 vmovss 0x606340(,%rcx,8),%xmm1
  124. 400b7a: 60 00
  125. 400b7c: c5 7a 10 3c cd 44 63 vmovss 0x606344(,%rcx,8),%xmm15
  126. 400b83: 60 00
  127. 400b85: c5 ec 59 e2 vmulps %ymm2,%ymm2,%ymm4
  128. 400b89: c5 d4 59 fd vmulps %ymm5,%ymm5,%ymm7
  129. 400b8d: c4 e3 75 18 d8 01 vinsertf128 $0x1,%xmm0,%ymm1,%ymm3
  130. 400b93: c5 fc 10 8a 20 73 60 vmovups 0x607320(%rdx),%ymm1
  131. 400b9a: 00
  132. 400b9b: c5 e4 58 f4 vaddps %ymm4,%ymm3,%ymm6
  133. 400b9f: c5 fc 10 a2 40 73 60 vmovups 0x607340(%rdx),%ymm4
  134. 400ba6: 00
  135. 400ba7: c5 f4 59 d9 vmulps %ymm1,%ymm1,%ymm3
  136. 400bab: c5 4c 58 c7 vaddps %ymm7,%ymm6,%ymm8
  137. 400baf: c5 dc 59 f4 vmulps %ymm4,%ymm4,%ymm6
  138. 400bb3: 48 81 c2 80 00 00 00 add $0x80,%rdx
  139. 400bba: c4 e3 05 18 d0 01 vinsertf128 $0x1,%xmm0,%ymm15,%ymm2
  140. 400bc0: c5 ec 58 eb vaddps %ymm3,%ymm2,%ymm5
  141. 400bc4: c5 d4 58 fe vaddps %ymm6,%ymm5,%ymm7
  142. 400bc8: c4 43 7d 19 c1 01 vextractf128 $0x1,%ymm8,%xmm9
  143. 400bce: c4 41 38 58 d1 vaddps %xmm9,%xmm8,%xmm10
  144. 400bd3: c4 41 28 12 da vmovhlps %xmm10,%xmm10,%xmm11
  145. 400bd8: c4 41 28 58 e3 vaddps %xmm11,%xmm10,%xmm12
  146. 400bdd: c4 41 18 c6 ec f5 vshufps $0xf5,%xmm12,%xmm12,%xmm13
  147. 400be3: c4 41 1a 58 f5 vaddss %xmm13,%xmm12,%xmm14
  148. 400be8: c5 7a 11 34 cd 40 63 vmovss %xmm14,0x606340(,%rcx,8)
  149. 400bef: 60 00
  150. 400bf1: c4 c3 7d 19 f8 01 vextractf128 $0x1,%ymm7,%xmm8
  151. 400bf7: c4 41 40 58 c8 vaddps %xmm8,%xmm7,%xmm9
  152. 400bfc: c4 41 30 12 d1 vmovhlps %xmm9,%xmm9,%xmm10
  153. 400c01: c4 41 30 58 da vaddps %xmm10,%xmm9,%xmm11
  154. 400c06: c4 41 20 c6 e3 f5 vshufps $0xf5,%xmm11,%xmm11,%xmm12
  155. 400c0c: c4 41 22 58 ec vaddss %xmm12,%xmm11,%xmm13
  156. 400c11: c5 7a 11 2c cd 44 63 vmovss %xmm13,0x606344(,%rcx,8)
  157. 400c18: 60 00
  158. 400c1a: 48 ff c1 inc %rcx
  159. 400c1d: 48 81 f9 f4 01 00 00 cmp $0x1f4,%rcx
  160. 400c24: 0f 82 39 ff ff ff jb 400b63 <main+0x1f3>
  161. 400c2a: 49 ff c7 inc %r15
  162. 400c2d: 49 81 ff a0 86 01 00 cmp $0x186a0,%r15
  163. 400c34: 0f 82 37 fe ff ff jb 400a71 <main+0x101>
  164. 400c3a: 48 8b 1c 24 mov (%rsp),%rbx
  165. 400c3e: 33 c0 xor %eax,%eax
  166. 400c40: c5 f5 57 c9 vxorpd %ymm1,%ymm1,%ymm1
  167. 400c44: c5 fd 28 c1 vmovapd %ymm1,%ymm0
  168. 400c48: c5 fc 10 14 85 40 63 vmovups 0x606340(,%rax,4),%ymm2
  169. 400c4f: 60 00
  170. 400c51: c5 7c 10 14 85 80 63 vmovups 0x606380(,%rax,4),%ymm10
  171. 400c58: 60 00
  172. 400c5a: c5 fc 5a e2 vcvtps2pd %xmm2,%ymm4
  173. 400c5e: c4 e3 7d 19 d3 01 vextractf128 $0x1,%ymm2,%xmm3
  174. 400c64: c5 f5 58 f4 vaddpd %ymm4,%ymm1,%ymm6
  175. 400c68: c5 fc 10 0c 85 60 63 vmovups 0x606360(,%rax,4),%ymm1
  176. 400c6f: 60 00
  177. 400c71: c5 fc 5a eb vcvtps2pd %xmm3,%ymm5
  178. 400c75: c5 7d 58 c5 vaddpd %ymm5,%ymm0,%ymm8
  179. 400c79: c4 e3 7d 19 c8 01 vextractf128 $0x1,%ymm1,%xmm0
  180. 400c7f: c5 7c 5a c8 vcvtps2pd %xmm0,%ymm9
  181. 400c83: c5 fc 5a f9 vcvtps2pd %xmm1,%ymm7
  182. 400c87: c5 fc 10 0c 85 a0 63 vmovups 0x6063a0(,%rax,4),%ymm1
  183. 400c8e: 60 00
  184. 400c90: c4 41 3d 58 f1 vaddpd %ymm9,%ymm8,%ymm14
  185. 400c95: c5 4d 58 e7 vaddpd %ymm7,%ymm6,%ymm12
  186. 400c99: 48 83 c0 20 add $0x20,%rax
  187. 400c9d: c4 43 7d 19 d3 01 vextractf128 $0x1,%ymm10,%xmm11
  188. 400ca3: c4 41 7c 5a fb vcvtps2pd %xmm11,%ymm15
  189. 400ca8: c4 41 7c 5a ea vcvtps2pd %xmm10,%ymm13
  190. 400cad: c4 c1 0d 58 df vaddpd %ymm15,%ymm14,%ymm3
  191. 400cb2: c4 c1 1d 58 c5 vaddpd %ymm13,%ymm12,%ymm0
  192. 400cb7: c4 c3 7d 19 ce 01 vextractf128 $0x1,%ymm1,%xmm14
  193. 400cbd: c5 fc 5a d1 vcvtps2pd %xmm1,%ymm2
  194. 400cc1: c4 c1 7c 5a e6 vcvtps2pd %xmm14,%ymm4
  195. 400cc6: c5 fd 58 ca vaddpd %ymm2,%ymm0,%ymm1
  196. 400cca: c5 e5 58 c4 vaddpd %ymm4,%ymm3,%ymm0
  197. 400cce: 48 3d e0 03 00 00 cmp $0x3e0,%rax
  198. 400cd4: 0f 82 6e ff ff ff jb 400c48 <main+0x2d8>
  199. 400cda: c5 f5 58 c8 vaddpd %ymm0,%ymm1,%ymm1
  200. 400cde: bf 60 36 40 00 mov $0x403660,%edi
  201. 400ce3: c5 ca 5a 35 d5 65 20 vcvtss2sd 0x2065d5(%rip),%xmm6,%xmm6 # 6072c0 <globalDst+0xf80>
  202. 400cea: 00
  203. 400ceb: c5 c2 5a 3d d1 65 20 vcvtss2sd 0x2065d1(%rip),%xmm7,%xmm7 # 6072c4 <globalDst+0xf84>
  204. 400cf2: 00
  205. 400cf3: c5 32 5a 0d cd 65 20 vcvtss2sd 0x2065cd(%rip),%xmm9,%xmm9 # 6072c8 <globalDst+0xf88>
  206. 400cfa: 00
  207. 400cfb: c5 22 5a 1d c9 65 20 vcvtss2sd 0x2065c9(%rip),%xmm11,%xmm11 # 6072cc <globalDst+0xf8c>
  208. 400d02: 00
  209. 400d03: c5 12 5a 2d c5 65 20 vcvtss2sd 0x2065c5(%rip),%xmm13,%xmm13 # 6072d0 <globalDst+0xf90>
  210. 400d0a: 00
  211. 400d0b: c5 02 5a 3d c1 65 20 vcvtss2sd 0x2065c1(%rip),%xmm15,%xmm15 # 6072d4 <globalDst+0xf94>
  212. 400d12: 00
  213. 400d13: be a0 86 01 00 mov $0x186a0,%esi
  214. 400d18: b8 01 00 00 00 mov $0x1,%eax
  215. 400d1d: c4 e3 7d 19 ca 01 vextractf128 $0x1,%ymm1,%xmm2
  216. 400d23: c5 f1 58 da vaddpd %xmm2,%xmm1,%xmm3
  217. 400d27: c5 f2 5a 0d a9 65 20 vcvtss2sd 0x2065a9(%rip),%xmm1,%xmm1 # 6072d8 <globalDst+0xf98>
  218. 400d2e: 00
  219. 400d2f: c5 ea 5a 15 a5 65 20 vcvtss2sd 0x2065a5(%rip),%xmm2,%xmm2 # 6072dc <globalDst+0xf9c>
  220. 400d36: 00
  221. 400d37: c5 e1 15 e3 vunpckhpd %xmm3,%xmm3,%xmm4
  222. 400d3b: c5 e3 58 ec vaddsd %xmm4,%xmm3,%xmm5
  223. 400d3f: c5 4b 58 c5 vaddsd %xmm5,%xmm6,%xmm8
  224. 400d43: c4 41 43 58 d0 vaddsd %xmm8,%xmm7,%xmm10
  225. 400d48: c4 41 33 58 e2 vaddsd %xmm10,%xmm9,%xmm12
  226. 400d4d: c4 41 23 58 f4 vaddsd %xmm12,%xmm11,%xmm14
  227. 400d52: c4 c1 13 58 c6 vaddsd %xmm14,%xmm13,%xmm0
  228. 400d57: c5 83 58 e8 vaddsd %xmm0,%xmm15,%xmm5
  229. 400d5b: c5 f3 58 dd vaddsd %xmm5,%xmm1,%xmm3
  230. 400d5f: c5 eb 58 c3 vaddsd %xmm3,%xmm2,%xmm0
  231. 400d63: c5 f8 77 vzeroupper
  232. 400d66: e8 75 fa ff ff callq 4007e0 <printf@plt>
  233. 400d6b: 33 c0 xor %eax,%eax
  234. 400d6d: 48 83 c4 70 add $0x70,%rsp
  235. 400d71: 41 5f pop %r15
  236. 400d73: 41 5e pop %r14
  237. 400d75: 48 89 ec mov %rbp,%rsp
  238. 400d78: 5d pop %rbp
  239. 400d79: c3 retq
  240. 400d7a: bf 40 36 40 00 mov $0x403640,%edi
  241. 400d7f: 33 c0 xor %eax,%eax
  242. 400d81: e8 5a fa ff ff callq 4007e0 <printf@plt>
  243. 400d86: b8 01 00 00 00 mov $0x1,%eax
  244. 400d8b: 48 83 c4 70 add $0x70,%rsp
  245. 400d8f: 41 5f pop %r15
  246. 400d91: 41 5e pop %r14
  247. 400d93: 48 89 ec mov %rbp,%rsp
  248. 400d96: 5d pop %rbp
  249. 400d97: c3 retq
  250. 400d98: b8 d0 0d 40 00 mov $0x400dd0,%eax
  251. 400d9d: 41 be b0 0e 40 00 mov $0x400eb0,%r14d
  252. 400da3: e9 24 fc ff ff jmpq 4009cc <main+0x5c>
  253. 400da8: bf 30 36 40 00 mov $0x403630,%edi
  254. 400dad: 33 c0 xor %eax,%eax
  255. 400daf: 49 8b 37 mov (%r15),%rsi
  256. 400db2: e8 29 fa ff ff callq 4007e0 <printf@plt>
  257. 400db7: b8 01 00 00 00 mov $0x1,%eax
  258. 400dbc: 48 83 c4 70 add $0x70,%rsp
  259. 400dc0: 41 5f pop %r15
  260. 400dc2: 41 5e pop %r14
  261. 400dc4: 48 89 ec mov %rbp,%rsp
  262. 400dc7: 5d pop %rbp
  263. 400dc8: c3 retq
  264. 400dc9: 0f 1f 80 00 00 00 00 nopl 0x0(%rax)
  265.  
  266. 0000000000400dd0 <sum_of_squares_0>:
  267. 400dd0: 33 d2 xor %edx,%edx
  268. 400dd2: 33 c0 xor %eax,%eax
  269. 400dd4: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
  270. 400dd8: c5 f8 10 14 30 vmovups (%rax,%rsi,1),%xmm2
  271. 400ddd: c5 f8 10 74 30 20 vmovups 0x20(%rax,%rsi,1),%xmm6
  272. 400de3: c5 fa 10 0c d7 vmovss (%rdi,%rdx,8),%xmm1
  273. 400de8: c4 e3 6d 18 5c 30 10 vinsertf128 $0x1,0x10(%rax,%rsi,1),%ymm2,%ymm3
  274. 400def: 01
  275. 400df0: c5 e4 59 eb vmulps %ymm3,%ymm3,%ymm5
  276. 400df4: c5 f8 10 5c 30 40 vmovups 0x40(%rax,%rsi,1),%xmm3
  277. 400dfa: c5 fa 10 54 d7 04 vmovss 0x4(%rdi,%rdx,8),%xmm2
  278. 400e00: c4 e3 4d 18 7c 30 30 vinsertf128 $0x1,0x30(%rax,%rsi,1),%ymm6,%ymm7
  279. 400e07: 01
  280. 400e08: c5 44 59 cf vmulps %ymm7,%ymm7,%ymm9
  281. 400e0c: c5 f8 10 7c 30 60 vmovups 0x60(%rax,%rsi,1),%xmm7
  282. 400e12: c4 e3 75 18 e0 01 vinsertf128 $0x1,%xmm0,%ymm1,%ymm4
  283. 400e18: c5 5c 58 c5 vaddps %ymm5,%ymm4,%ymm8
  284. 400e1c: c4 41 3c 58 d1 vaddps %ymm9,%ymm8,%ymm10
  285. 400e21: c4 e3 65 18 64 30 50 vinsertf128 $0x1,0x50(%rax,%rsi,1),%ymm3,%ymm4
  286. 400e28: 01
  287. 400e29: c5 dc 59 f4 vmulps %ymm4,%ymm4,%ymm6
  288. 400e2d: c4 63 45 18 44 30 70 vinsertf128 $0x1,0x70(%rax,%rsi,1),%ymm7,%ymm8
  289. 400e34: 01
  290. 400e35: 48 05 80 00 00 00 add $0x80,%rax
  291. 400e3b: c4 43 7d 19 d3 01 vextractf128 $0x1,%ymm10,%xmm11
  292. 400e41: c4 41 28 58 e3 vaddps %xmm11,%xmm10,%xmm12
  293. 400e46: c4 41 3c 59 d0 vmulps %ymm8,%ymm8,%ymm10
  294. 400e4b: c4 41 18 12 ec vmovhlps %xmm12,%xmm12,%xmm13
  295. 400e50: c4 41 18 58 f5 vaddps %xmm13,%xmm12,%xmm14
  296. 400e55: c4 41 08 c6 fe f5 vshufps $0xf5,%xmm14,%xmm14,%xmm15
  297. 400e5b: c4 c1 0a 58 cf vaddss %xmm15,%xmm14,%xmm1
  298. 400e60: c5 fa 11 0c d7 vmovss %xmm1,(%rdi,%rdx,8)
  299. 400e65: c4 e3 6d 18 e8 01 vinsertf128 $0x1,%xmm0,%ymm2,%ymm5
  300. 400e6b: c5 54 58 ce vaddps %ymm6,%ymm5,%ymm9
  301. 400e6f: c4 41 34 58 da vaddps %ymm10,%ymm9,%ymm11
  302. 400e74: c4 43 7d 19 dc 01 vextractf128 $0x1,%ymm11,%xmm12
  303. 400e7a: c4 41 20 58 ec vaddps %xmm12,%xmm11,%xmm13
  304. 400e7f: c4 41 10 12 f5 vmovhlps %xmm13,%xmm13,%xmm14
  305. 400e84: c4 41 10 58 fe vaddps %xmm14,%xmm13,%xmm15
  306. 400e89: c4 c1 00 c6 cf f5 vshufps $0xf5,%xmm15,%xmm15,%xmm1
  307. 400e8f: c5 82 58 d1 vaddss %xmm1,%xmm15,%xmm2
  308. 400e93: c5 fa 11 54 d7 04 vmovss %xmm2,0x4(%rdi,%rdx,8)
  309. 400e99: 48 ff c2 inc %rdx
  310. 400e9c: 48 81 fa f4 01 00 00 cmp $0x1f4,%rdx
  311. 400ea3: 0f 82 2f ff ff ff jb 400dd8 <sum_of_squares_0+0x8>
  312. 400ea9: c5 f8 77 vzeroupper
  313. 400eac: c3 retq
  314. 400ead: 0f 1f 00 nopl (%rax)
  315.  
  316. 0000000000400eb0 <sum_of_squares_1>:
  317. 400eb0: 33 d2 xor %edx,%edx
  318. 400eb2: 33 c0 xor %eax,%eax
  319. 400eb4: c5 f8 57 c0 vxorps %xmm0,%xmm0,%xmm0
  320. 400eb8: c5 f8 10 14 30 vmovups (%rax,%rsi,1),%xmm2
  321. 400ebd: c5 f8 10 74 30 20 vmovups 0x20(%rax,%rsi,1),%xmm6
  322. 400ec3: c5 fa 10 0c d7 vmovss (%rdi,%rdx,8),%xmm1
  323. 400ec8: c4 e3 6d 18 5c 30 10 vinsertf128 $0x1,0x10(%rax,%rsi,1),%ymm2,%ymm3
  324. 400ecf: 01
  325. 400ed0: c5 e4 59 eb vmulps %ymm3,%ymm3,%ymm5
  326. 400ed4: c5 f8 10 5c 30 40 vmovups 0x40(%rax,%rsi,1),%xmm3
  327. 400eda: c5 fa 10 54 d7 04 vmovss 0x4(%rdi,%rdx,8),%xmm2
  328. 400ee0: c4 e3 4d 18 7c 30 30 vinsertf128 $0x1,0x30(%rax,%rsi,1),%ymm6,%ymm7
  329. 400ee7: 01
  330. 400ee8: c5 44 59 cf vmulps %ymm7,%ymm7,%ymm9
  331. 400eec: c5 f8 10 7c 30 60 vmovups 0x60(%rax,%rsi,1),%xmm7
  332. 400ef2: c4 e3 75 18 e0 01 vinsertf128 $0x1,%xmm0,%ymm1,%ymm4
  333. 400ef8: c5 5c 58 c5 vaddps %ymm5,%ymm4,%ymm8
  334. 400efc: c4 41 3c 58 d1 vaddps %ymm9,%ymm8,%ymm10
  335. 400f01: c4 e3 65 18 64 30 50 vinsertf128 $0x1,0x50(%rax,%rsi,1),%ymm3,%ymm4
  336. 400f08: 01
  337. 400f09: c5 dc 59 f4 vmulps %ymm4,%ymm4,%ymm6
  338. 400f0d: c4 63 45 18 44 30 70 vinsertf128 $0x1,0x70(%rax,%rsi,1),%ymm7,%ymm8
  339. 400f14: 01
  340. 400f15: 48 05 80 00 00 00 add $0x80,%rax
  341. 400f1b: c4 43 7d 19 d3 01 vextractf128 $0x1,%ymm10,%xmm11
  342. 400f21: c4 41 28 58 e3 vaddps %xmm11,%xmm10,%xmm12
  343. 400f26: c4 41 3c 59 d0 vmulps %ymm8,%ymm8,%ymm10
  344. 400f2b: c4 41 18 12 ec vmovhlps %xmm12,%xmm12,%xmm13
  345. 400f30: c4 41 18 58 f5 vaddps %xmm13,%xmm12,%xmm14
  346. 400f35: c4 41 08 c6 fe f5 vshufps $0xf5,%xmm14,%xmm14,%xmm15
  347. 400f3b: c4 c1 0a 58 cf vaddss %xmm15,%xmm14,%xmm1
  348. 400f40: c5 fa 11 0c d7 vmovss %xmm1,(%rdi,%rdx,8)
  349. 400f45: c4 e3 6d 18 e8 01 vinsertf128 $0x1,%xmm0,%ymm2,%ymm5
  350. 400f4b: c5 54 58 ce vaddps %ymm6,%ymm5,%ymm9
  351. 400f4f: c4 41 34 58 da vaddps %ymm10,%ymm9,%ymm11
  352. 400f54: c4 43 7d 19 dc 01 vextractf128 $0x1,%ymm11,%xmm12
  353. 400f5a: c4 41 20 58 ec vaddps %xmm12,%xmm11,%xmm13
  354. 400f5f: c4 41 10 12 f5 vmovhlps %xmm13,%xmm13,%xmm14
  355. 400f64: c4 41 10 58 fe vaddps %xmm14,%xmm13,%xmm15
  356. 400f69: c4 c1 00 c6 cf f5 vshufps $0xf5,%xmm15,%xmm15,%xmm1
  357. 400f6f: c5 82 58 d1 vaddss %xmm1,%xmm15,%xmm2
  358. 400f73: c5 fa 11 54 d7 04 vmovss %xmm2,0x4(%rdi,%rdx,8)
  359. 400f79: 48 ff c2 inc %rdx
  360. 400f7c: 48 81 fa f4 01 00 00 cmp $0x1f4,%rdx
  361. 400f83: 0f 82 2f ff ff ff jb 400eb8 <sum_of_squares_1+0x8>
  362. 400f89: c5 f8 77 vzeroupper
  363. 400f8c: c3 retq
  364. 400f8d: 0f 1f 00 nopl (%rax)
Advertisement
Add Comment
Please, Sign In to add comment