fp2e_sub2.s 7.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310
  1. # File: dclxvi-20130329/fp2e_sub2.s
  2. # Author: Ruben Niederhagen, Peter Schwabe
  3. # Public Domain
  4. # qhasm: enter fp2e_sub2_qhasm
  5. .text
  6. .p2align 5
  7. .globl _fp2e_sub2_qhasm
  8. .globl fp2e_sub2_qhasm
  9. _fp2e_sub2_qhasm:
  10. fp2e_sub2_qhasm:
  11. mov %rsp,%r11
  12. and $31,%r11
  13. add $0,%r11
  14. sub %r11,%rsp
  15. # qhasm: int64 0rop
  16. # qhasm: int64 0op1
  17. # qhasm: input 0rop
  18. # qhasm: input 0op1
  19. # qhasm: int6464 0r0
  20. # qhasm: int6464 0r1
  21. # qhasm: int6464 0r2
  22. # qhasm: int6464 0r3
  23. # qhasm: int6464 0r4
  24. # qhasm: int6464 0r5
  25. # qhasm: int6464 0r6
  26. # qhasm: int6464 0r7
  27. # qhasm: int6464 0r8
  28. # qhasm: int6464 0r9
  29. # qhasm: int6464 0r10
  30. # qhasm: int6464 0r11
  31. # qhasm: int6464 0t0
  32. # qhasm: int6464 0t1
  33. # qhasm: int6464 0t2
  34. # qhasm: int6464 0t3
  35. # qhasm: 0r0 = *(int128 *)(0rop + 0)
  36. # asm 1: movdqa 0(<0rop=int64#1),>0r0=int6464#1
  37. # asm 2: movdqa 0(<0rop=%rdi),>0r0=%xmm0
  38. movdqa 0(%rdi),%xmm0
  39. # qhasm: 0r1 = *(int128 *)(0rop + 16)
  40. # asm 1: movdqa 16(<0rop=int64#1),>0r1=int6464#2
  41. # asm 2: movdqa 16(<0rop=%rdi),>0r1=%xmm1
  42. movdqa 16(%rdi),%xmm1
  43. # qhasm: 0r2 = *(int128 *)(0rop + 32)
  44. # asm 1: movdqa 32(<0rop=int64#1),>0r2=int6464#3
  45. # asm 2: movdqa 32(<0rop=%rdi),>0r2=%xmm2
  46. movdqa 32(%rdi),%xmm2
  47. # qhasm: 0r3 = *(int128 *)(0rop + 48)
  48. # asm 1: movdqa 48(<0rop=int64#1),>0r3=int6464#4
  49. # asm 2: movdqa 48(<0rop=%rdi),>0r3=%xmm3
  50. movdqa 48(%rdi),%xmm3
  51. # qhasm: 0r4 = *(int128 *)(0rop + 64)
  52. # asm 1: movdqa 64(<0rop=int64#1),>0r4=int6464#5
  53. # asm 2: movdqa 64(<0rop=%rdi),>0r4=%xmm4
  54. movdqa 64(%rdi),%xmm4
  55. # qhasm: 0r5 = *(int128 *)(0rop + 80)
  56. # asm 1: movdqa 80(<0rop=int64#1),>0r5=int6464#6
  57. # asm 2: movdqa 80(<0rop=%rdi),>0r5=%xmm5
  58. movdqa 80(%rdi),%xmm5
  59. # qhasm: 0r6 = *(int128 *)(0rop + 96)
  60. # asm 1: movdqa 96(<0rop=int64#1),>0r6=int6464#7
  61. # asm 2: movdqa 96(<0rop=%rdi),>0r6=%xmm6
  62. movdqa 96(%rdi),%xmm6
  63. # qhasm: 0r7 = *(int128 *)(0rop + 112)
  64. # asm 1: movdqa 112(<0rop=int64#1),>0r7=int6464#8
  65. # asm 2: movdqa 112(<0rop=%rdi),>0r7=%xmm7
  66. movdqa 112(%rdi),%xmm7
  67. # qhasm: 0r8 = *(int128 *)(0rop + 128)
  68. # asm 1: movdqa 128(<0rop=int64#1),>0r8=int6464#9
  69. # asm 2: movdqa 128(<0rop=%rdi),>0r8=%xmm8
  70. movdqa 128(%rdi),%xmm8
  71. # qhasm: 0r9 = *(int128 *)(0rop + 144)
  72. # asm 1: movdqa 144(<0rop=int64#1),>0r9=int6464#10
  73. # asm 2: movdqa 144(<0rop=%rdi),>0r9=%xmm9
  74. movdqa 144(%rdi),%xmm9
  75. # qhasm: 0r10 = *(int128 *)(0rop + 160)
  76. # asm 1: movdqa 160(<0rop=int64#1),>0r10=int6464#11
  77. # asm 2: movdqa 160(<0rop=%rdi),>0r10=%xmm10
  78. movdqa 160(%rdi),%xmm10
  79. # qhasm: 0r11 = *(int128 *)(0rop + 176)
  80. # asm 1: movdqa 176(<0rop=int64#1),>0r11=int6464#12
  81. # asm 2: movdqa 176(<0rop=%rdi),>0r11=%xmm11
  82. movdqa 176(%rdi),%xmm11
  83. # qhasm: int6464 1t0
  84. # qhasm: int6464 1t1
  85. # qhasm: int6464 1t2
  86. # qhasm: int6464 1t3
  87. # qhasm: 1t0 = *(int128 *)(0op1 + 0)
  88. # asm 1: movdqa 0(<0op1=int64#2),>1t0=int6464#13
  89. # asm 2: movdqa 0(<0op1=%rsi),>1t0=%xmm12
  90. movdqa 0(%rsi),%xmm12
  91. # qhasm: 1t1 = *(int128 *)(0op1 + 16)
  92. # asm 1: movdqa 16(<0op1=int64#2),>1t1=int6464#14
  93. # asm 2: movdqa 16(<0op1=%rsi),>1t1=%xmm13
  94. movdqa 16(%rsi),%xmm13
  95. # qhasm: 1t2 = *(int128 *)(0op1 + 32)
  96. # asm 1: movdqa 32(<0op1=int64#2),>1t2=int6464#15
  97. # asm 2: movdqa 32(<0op1=%rsi),>1t2=%xmm14
  98. movdqa 32(%rsi),%xmm14
  99. # qhasm: 1t3 = *(int128 *)(0op1 + 48)
  100. # asm 1: movdqa 48(<0op1=int64#2),>1t3=int6464#16
  101. # asm 2: movdqa 48(<0op1=%rsi),>1t3=%xmm15
  102. movdqa 48(%rsi),%xmm15
  103. # qhasm: float6464 0r0 -= 1t0
  104. # asm 1: subpd <1t0=int6464#13,<0r0=int6464#1
  105. # asm 2: subpd <1t0=%xmm12,<0r0=%xmm0
  106. subpd %xmm12,%xmm0
  107. # qhasm: float6464 0r1 -= 1t1
  108. # asm 1: subpd <1t1=int6464#14,<0r1=int6464#2
  109. # asm 2: subpd <1t1=%xmm13,<0r1=%xmm1
  110. subpd %xmm13,%xmm1
  111. # qhasm: float6464 0r2 -= 1t2
  112. # asm 1: subpd <1t2=int6464#15,<0r2=int6464#3
  113. # asm 2: subpd <1t2=%xmm14,<0r2=%xmm2
  114. subpd %xmm14,%xmm2
  115. # qhasm: float6464 0r3 -= 1t3
  116. # asm 1: subpd <1t3=int6464#16,<0r3=int6464#4
  117. # asm 2: subpd <1t3=%xmm15,<0r3=%xmm3
  118. subpd %xmm15,%xmm3
  119. # qhasm: 1t0 = *(int128 *)(0op1 + 64)
  120. # asm 1: movdqa 64(<0op1=int64#2),>1t0=int6464#13
  121. # asm 2: movdqa 64(<0op1=%rsi),>1t0=%xmm12
  122. movdqa 64(%rsi),%xmm12
  123. # qhasm: 1t1 = *(int128 *)(0op1 + 80)
  124. # asm 1: movdqa 80(<0op1=int64#2),>1t1=int6464#14
  125. # asm 2: movdqa 80(<0op1=%rsi),>1t1=%xmm13
  126. movdqa 80(%rsi),%xmm13
  127. # qhasm: 1t2 = *(int128 *)(0op1 + 96)
  128. # asm 1: movdqa 96(<0op1=int64#2),>1t2=int6464#15
  129. # asm 2: movdqa 96(<0op1=%rsi),>1t2=%xmm14
  130. movdqa 96(%rsi),%xmm14
  131. # qhasm: 1t3 = *(int128 *)(0op1 + 112)
  132. # asm 1: movdqa 112(<0op1=int64#2),>1t3=int6464#16
  133. # asm 2: movdqa 112(<0op1=%rsi),>1t3=%xmm15
  134. movdqa 112(%rsi),%xmm15
  135. # qhasm: float6464 0r4 -= 1t0
  136. # asm 1: subpd <1t0=int6464#13,<0r4=int6464#5
  137. # asm 2: subpd <1t0=%xmm12,<0r4=%xmm4
  138. subpd %xmm12,%xmm4
  139. # qhasm: float6464 0r5 -= 1t1
  140. # asm 1: subpd <1t1=int6464#14,<0r5=int6464#6
  141. # asm 2: subpd <1t1=%xmm13,<0r5=%xmm5
  142. subpd %xmm13,%xmm5
  143. # qhasm: float6464 0r6 -= 1t2
  144. # asm 1: subpd <1t2=int6464#15,<0r6=int6464#7
  145. # asm 2: subpd <1t2=%xmm14,<0r6=%xmm6
  146. subpd %xmm14,%xmm6
  147. # qhasm: float6464 0r7 -= 1t3
  148. # asm 1: subpd <1t3=int6464#16,<0r7=int6464#8
  149. # asm 2: subpd <1t3=%xmm15,<0r7=%xmm7
  150. subpd %xmm15,%xmm7
  151. # qhasm: 1t0 = *(int128 *)(0op1 + 128)
  152. # asm 1: movdqa 128(<0op1=int64#2),>1t0=int6464#13
  153. # asm 2: movdqa 128(<0op1=%rsi),>1t0=%xmm12
  154. movdqa 128(%rsi),%xmm12
  155. # qhasm: 1t1 = *(int128 *)(0op1 + 144)
  156. # asm 1: movdqa 144(<0op1=int64#2),>1t1=int6464#14
  157. # asm 2: movdqa 144(<0op1=%rsi),>1t1=%xmm13
  158. movdqa 144(%rsi),%xmm13
  159. # qhasm: 1t2 = *(int128 *)(0op1 + 160)
  160. # asm 1: movdqa 160(<0op1=int64#2),>1t2=int6464#15
  161. # asm 2: movdqa 160(<0op1=%rsi),>1t2=%xmm14
  162. movdqa 160(%rsi),%xmm14
  163. # qhasm: 1t3 = *(int128 *)(0op1 + 176)
  164. # asm 1: movdqa 176(<0op1=int64#2),>1t3=int6464#16
  165. # asm 2: movdqa 176(<0op1=%rsi),>1t3=%xmm15
  166. movdqa 176(%rsi),%xmm15
  167. # qhasm: float6464 0r8 -= 1t0
  168. # asm 1: subpd <1t0=int6464#13,<0r8=int6464#9
  169. # asm 2: subpd <1t0=%xmm12,<0r8=%xmm8
  170. subpd %xmm12,%xmm8
  171. # qhasm: float6464 0r9 -= 1t1
  172. # asm 1: subpd <1t1=int6464#14,<0r9=int6464#10
  173. # asm 2: subpd <1t1=%xmm13,<0r9=%xmm9
  174. subpd %xmm13,%xmm9
  175. # qhasm: float6464 0r10 -= 1t2
  176. # asm 1: subpd <1t2=int6464#15,<0r10=int6464#11
  177. # asm 2: subpd <1t2=%xmm14,<0r10=%xmm10
  178. subpd %xmm14,%xmm10
  179. # qhasm: float6464 0r11 -= 1t3
  180. # asm 1: subpd <1t3=int6464#16,<0r11=int6464#12
  181. # asm 2: subpd <1t3=%xmm15,<0r11=%xmm11
  182. subpd %xmm15,%xmm11
  183. # qhasm: *(int128 *)(0rop + 0) = 0r0
  184. # asm 1: movdqa <0r0=int6464#1,0(<0rop=int64#1)
  185. # asm 2: movdqa <0r0=%xmm0,0(<0rop=%rdi)
  186. movdqa %xmm0,0(%rdi)
  187. # qhasm: *(int128 *)(0rop + 16) = 0r1
  188. # asm 1: movdqa <0r1=int6464#2,16(<0rop=int64#1)
  189. # asm 2: movdqa <0r1=%xmm1,16(<0rop=%rdi)
  190. movdqa %xmm1,16(%rdi)
  191. # qhasm: *(int128 *)(0rop + 32) = 0r2
  192. # asm 1: movdqa <0r2=int6464#3,32(<0rop=int64#1)
  193. # asm 2: movdqa <0r2=%xmm2,32(<0rop=%rdi)
  194. movdqa %xmm2,32(%rdi)
  195. # qhasm: *(int128 *)(0rop + 48) = 0r3
  196. # asm 1: movdqa <0r3=int6464#4,48(<0rop=int64#1)
  197. # asm 2: movdqa <0r3=%xmm3,48(<0rop=%rdi)
  198. movdqa %xmm3,48(%rdi)
  199. # qhasm: *(int128 *)(0rop + 64) = 0r4
  200. # asm 1: movdqa <0r4=int6464#5,64(<0rop=int64#1)
  201. # asm 2: movdqa <0r4=%xmm4,64(<0rop=%rdi)
  202. movdqa %xmm4,64(%rdi)
  203. # qhasm: *(int128 *)(0rop + 80) = 0r5
  204. # asm 1: movdqa <0r5=int6464#6,80(<0rop=int64#1)
  205. # asm 2: movdqa <0r5=%xmm5,80(<0rop=%rdi)
  206. movdqa %xmm5,80(%rdi)
  207. # qhasm: *(int128 *)(0rop + 96) = 0r6
  208. # asm 1: movdqa <0r6=int6464#7,96(<0rop=int64#1)
  209. # asm 2: movdqa <0r6=%xmm6,96(<0rop=%rdi)
  210. movdqa %xmm6,96(%rdi)
  211. # qhasm: *(int128 *)(0rop + 112) = 0r7
  212. # asm 1: movdqa <0r7=int6464#8,112(<0rop=int64#1)
  213. # asm 2: movdqa <0r7=%xmm7,112(<0rop=%rdi)
  214. movdqa %xmm7,112(%rdi)
  215. # qhasm: *(int128 *)(0rop + 128) = 0r8
  216. # asm 1: movdqa <0r8=int6464#9,128(<0rop=int64#1)
  217. # asm 2: movdqa <0r8=%xmm8,128(<0rop=%rdi)
  218. movdqa %xmm8,128(%rdi)
  219. # qhasm: *(int128 *)(0rop + 144) = 0r9
  220. # asm 1: movdqa <0r9=int6464#10,144(<0rop=int64#1)
  221. # asm 2: movdqa <0r9=%xmm9,144(<0rop=%rdi)
  222. movdqa %xmm9,144(%rdi)
  223. # qhasm: *(int128 *)(0rop + 160) = 0r10
  224. # asm 1: movdqa <0r10=int6464#11,160(<0rop=int64#1)
  225. # asm 2: movdqa <0r10=%xmm10,160(<0rop=%rdi)
  226. movdqa %xmm10,160(%rdi)
  227. # qhasm: *(int128 *)(0rop + 176) = 0r11
  228. # asm 1: movdqa <0r11=int6464#12,176(<0rop=int64#1)
  229. # asm 2: movdqa <0r11=%xmm11,176(<0rop=%rdi)
  230. movdqa %xmm11,176(%rdi)
  231. # qhasm: leave
  232. add %r11,%rsp
  233. mov %rdi,%rax
  234. mov %rsi,%rdx
  235. ret