d4dag.s 8.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251
  1. # mark_description "Intel(R) C Intel(R) 64 Compiler Classic for applications running on Intel(R) 64, Version 2021.3.0 Build 2021";
  2. # mark_description "0609_000000";
  3. # mark_description "-m64 -c -Wall -S -fverbose-asm -qopt-report-embed";
  4. .file "d4dag.c"
  5. .text
  6. ..TXTST0:
  7. .L_2__routine_start_d4d_version_0:
  8. # -- Begin d4d_version
  9. .text
  10. # mark_begin;
  11. .align 16,0x90
  12. .globl d4d_version
  13. # --- d4d_version(void)
  14. d4d_version:
  15. ..B1.1: # Preds ..B1.0
  16. # Execution count [1.00e+00]
  17. .cfi_startproc
  18. ..___tag_value_d4d_version.1:
  19. ..L2:
  20. #86.1
  21. movl $10, %eax #87.9
  22. ret #87.9
  23. .align 16,0x90
  24. # LOE
  25. .cfi_endproc
  26. # mark_end;
  27. .type d4d_version,@function
  28. .size d4d_version,.-d4d_version
  29. ..LNd4d_version.0:
  30. .data
  31. # -- End d4d_version
  32. .text
  33. .L_2__routine_start_d4d_deinit_1:
  34. # -- Begin d4d_deinit
  35. .text
  36. # mark_begin;
  37. .align 16,0x90
  38. .globl d4d_deinit
  39. # --- d4d_deinit(void)
  40. d4d_deinit:
  41. ..B2.1: # Preds ..B2.0
  42. # Execution count [1.00e+00]
  43. .cfi_startproc
  44. ..___tag_value_d4d_deinit.4:
  45. ..L5:
  46. #124.1
  47. pushq %rsi #124.1
  48. .cfi_def_cfa_offset 16
  49. movq d4d__main(%rip), %rdi #126.7
  50. testq %rdi, %rdi #126.7
  51. jne ..B2.3 # Prob 67% #126.7
  52. # LOE rbx rbp rdi r12 r13 r14 r15
  53. ..B2.2: # Preds ..B2.1
  54. # Execution count [3.26e-01]
  55. xorl %eax, %eax #127.10
  56. popq %rcx #127.10
  57. .cfi_def_cfa_offset 8
  58. ret #127.10
  59. .cfi_def_cfa_offset 16
  60. # LOE
  61. ..B2.3: # Preds ..B2.1
  62. # Execution count [6.74e-01]
  63. ..___tag_value_d4d_deinit.9:
  64. call *8(%rdi) #131.2
  65. ..___tag_value_d4d_deinit.10:
  66. # LOE rbx rbp r12 r13 r14 r15
  67. ..B2.4: # Preds ..B2.3
  68. # Execution count [6.74e-01]
  69. movq $0, d4d__main(%rip) #132.2
  70. xorl %eax, %eax #133.9
  71. popq %rcx #133.9
  72. .cfi_def_cfa_offset 8
  73. ret #133.9
  74. .align 16,0x90
  75. # LOE
  76. .cfi_endproc
  77. # mark_end;
  78. .type d4d_deinit,@function
  79. .size d4d_deinit,.-d4d_deinit
  80. ..LNd4d_deinit.1:
  81. .data
  82. # -- End d4d_deinit
  83. .text
  84. .L_2__routine_start_d4d_init_2:
  85. # -- Begin d4d_init
  86. .text
  87. # mark_begin;
  88. .align 16,0x90
  89. .globl d4d_init
  90. # --- d4d_init(void *(*)(unsigned int), void (*)(void *))
  91. d4d_init:
  92. # parameter 1: %rdi
  93. # parameter 2: %rsi
  94. ..B3.1: # Preds ..B3.0
  95. # Execution count [1.00e+00]
  96. .cfi_startproc
  97. ..___tag_value_d4d_init.13:
  98. ..L14:
  99. #101.1
  100. pushq %r14 #101.1
  101. .cfi_def_cfa_offset 16
  102. .cfi_offset 14, -16
  103. pushq %r15 #101.1
  104. .cfi_def_cfa_offset 24
  105. .cfi_offset 15, -24
  106. pushq %rsi #101.1
  107. .cfi_def_cfa_offset 32
  108. movq %rdi, %r14 #101.1
  109. movq %rsi, %r15 #101.1
  110. testq %r14, %r14 #103.6
  111. je ..B3.8 # Prob 5% #103.6
  112. # LOE rbx rbp r12 r13 r14 r15
  113. ..B3.2: # Preds ..B3.1
  114. # Execution count [9.50e-01]
  115. testq %r15, %r15 #106.7
  116. je ..B3.8 # Prob 15% #106.7
  117. # LOE rbx rbp r12 r13 r14 r15
  118. ..B3.3: # Preds ..B3.2
  119. # Execution count [7.99e-01]
  120. movl $16, %edi #109.35
  121. ..___tag_value_d4d_init.20:
  122. call *%r14 #109.35
  123. ..___tag_value_d4d_init.21:
  124. # LOE rax rbx rbp r12 r13 r14 r15
  125. ..B3.4: # Preds ..B3.3
  126. # Execution count [7.99e-01]
  127. movq %rax, d4d__main(%rip) #109.2
  128. testq %rax, %rax #110.7
  129. jne ..B3.6 # Prob 67% #110.7
  130. # LOE rax rbx rbp r12 r13 r14 r15
  131. ..B3.5: # Preds ..B3.4
  132. # Execution count [2.61e-01]
  133. movl $-2, %eax #111.10
  134. popq %rcx #111.10
  135. .cfi_def_cfa_offset 24
  136. .cfi_restore 15
  137. popq %r15 #111.10
  138. .cfi_def_cfa_offset 16
  139. .cfi_restore 14
  140. popq %r14 #111.10
  141. .cfi_def_cfa_offset 8
  142. ret #111.10
  143. .cfi_def_cfa_offset 32
  144. .cfi_offset 14, -16
  145. .cfi_offset 15, -24
  146. # LOE rbx rbp r12 r13 eax
  147. ..B3.6: # Preds ..B3.4
  148. # Execution count [5.39e-01]
  149. ..L30:
  150. # optimization report
  151. # LOOP WAS COMPLETELY UNROLLED
  152. # LOOP WAS VECTORIZED
  153. # VECTORIZATION HAS UNALIGNED MEMORY REFERENCES
  154. # VECTORIZATION SPEEDUP COEFFECIENT 2.667969
  155. # LOOP HAS ONE VECTOR ITERATION
  156. # VECTOR LENGTH 16
  157. # NORMALIZED VECTORIZATION OVERHEAD 0.812500
  158. # MAIN VECTOR TYPE: 8-bits integer
  159. # THE LOOP HAS FP INDUCTION
  160. pxor %xmm0, %xmm0 #144.4
  161. movdqu %xmm0, (%rax) #144.4
  162. movq %r14, (%rax) #115.2
  163. movq %r15, 8(%rax) #116.2
  164. xorl %eax, %eax #117.9
  165. # LOE rbx rbp r12 r13 eax
  166. ..B3.7: # Preds ..B3.6
  167. # Execution count [7.99e-01]
  168. popq %rcx #111.10
  169. .cfi_def_cfa_offset 24
  170. .cfi_restore 15
  171. popq %r15 #111.10
  172. .cfi_def_cfa_offset 16
  173. .cfi_restore 14
  174. popq %r14 #111.10
  175. .cfi_def_cfa_offset 8
  176. ret #111.10
  177. .cfi_def_cfa_offset 32
  178. .cfi_offset 14, -16
  179. .cfi_offset 15, -24
  180. # LOE
  181. ..B3.8: # Preds ..B3.2 ..B3.1
  182. # Execution count [5.04e-02]: Infreq
  183. movl $-1, %eax #104.9
  184. popq %rcx #104.9
  185. .cfi_def_cfa_offset 24
  186. .cfi_restore 15
  187. popq %r15 #104.9
  188. .cfi_def_cfa_offset 16
  189. .cfi_restore 14
  190. popq %r14 #104.9
  191. .cfi_def_cfa_offset 8
  192. ret #104.9
  193. .align 16,0x90
  194. # LOE
  195. .cfi_endproc
  196. # mark_end;
  197. .type d4d_init,@function
  198. .size d4d_init,.-d4d_init
  199. ..LNd4d_init.2:
  200. .data
  201. # -- End d4d_init
  202. .bss
  203. .align 8
  204. .align 8
  205. d4d__main:
  206. .type d4d__main,@object
  207. .size d4d__main,8
  208. .space 8 # pad
  209. .data
  210. .section .debug_opt_report, ""
  211. ..L45:
  212. .ascii ".itt_notify_tab\0"
  213. .word 258
  214. .word 48
  215. .long 2
  216. .long ..L46 - ..L45
  217. .long 48
  218. .long ..L47 - ..L45
  219. .long 25
  220. .long 0x00000008,0x00000000
  221. .long 0
  222. .long 0
  223. .long 0
  224. .long 1
  225. .quad ..L30
  226. .long 28
  227. .long 5
  228. ..L46:
  229. .long 1769238639
  230. .long 1635412333
  231. .long 1852795252
  232. .long 1885696607
  233. .long 1601466991
  234. .long 1936876918
  235. .long 7237481
  236. .long 1769238639
  237. .long 1635412333
  238. .long 1852795252
  239. .long 1885696607
  240. .long 7631471
  241. ..L47:
  242. .long -2045771008
  243. .long -2146430206
  244. .long -1065320320
  245. .long -2139062144
  246. .long -1783594366
  247. .long -2137151032
  248. .byte 16
  249. .section .note.GNU-stack, ""
  250. # End