Bcj2Enc.c 7.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312
  1. /* Bcj2Enc.c -- BCJ2 Encoder (Converter for x86 code)
  2. 2014-11-10 : Igor Pavlov : Public domain */
  3. #include "Precomp.h"
  4. /* #define SHOW_STAT */
  5. #ifdef SHOW_STAT
  6. #include <stdio.h>
  7. #define PRF(x) x
  8. #else
  9. #define PRF(x)
  10. #endif
  11. #include <windows.h>
  12. #include <string.h>
  13. #include "Bcj2.h"
  14. #include "CpuArch.h"
  15. #define CProb UInt16
  16. #define kTopValue ((UInt32)1 << 24)
  17. #define kNumModelBits 11
  18. #define kBitModelTotal (1 << kNumModelBits)
  19. #define kNumMoveBits 5
  20. void Bcj2Enc_Init(CBcj2Enc *p)
  21. {
  22. unsigned i;
  23. p->state = BCJ2_ENC_STATE_OK;
  24. p->finishMode = BCJ2_ENC_FINISH_MODE_CONTINUE;
  25. p->prevByte = 0;
  26. p->cache = 0;
  27. p->range = 0xFFFFFFFF;
  28. p->low = 0;
  29. p->cacheSize = 1;
  30. p->ip = 0;
  31. p->fileIp = 0;
  32. p->fileSize = 0;
  33. p->relatLimit = BCJ2_RELAT_LIMIT;
  34. p->tempPos = 0;
  35. p->flushPos = 0;
  36. for (i = 0; i < sizeof(p->probs) / sizeof(p->probs[0]); i++)
  37. p->probs[i] = kBitModelTotal >> 1;
  38. }
  39. static Bool MY_FAST_CALL RangeEnc_ShiftLow(CBcj2Enc *p)
  40. {
  41. if ((UInt32)p->low < (UInt32)0xFF000000 || (UInt32)(p->low >> 32) != 0)
  42. {
  43. Byte *buf = p->bufs[BCJ2_STREAM_RC];
  44. do
  45. {
  46. if (buf == p->lims[BCJ2_STREAM_RC])
  47. {
  48. p->state = BCJ2_STREAM_RC;
  49. p->bufs[BCJ2_STREAM_RC] = buf;
  50. return True;
  51. }
  52. *buf++ = (Byte)(p->cache + (Byte)(p->low >> 32));
  53. p->cache = 0xFF;
  54. }
  55. while (--p->cacheSize);
  56. p->bufs[BCJ2_STREAM_RC] = buf;
  57. p->cache = (Byte)((UInt32)p->low >> 24);
  58. }
  59. p->cacheSize++;
  60. p->low = (UInt32)p->low << 8;
  61. return False;
  62. }
  63. static void Bcj2Enc_Encode_2(CBcj2Enc *p)
  64. {
  65. if (BCJ2_IS_32BIT_STREAM(p->state))
  66. {
  67. Byte *cur = p->bufs[p->state];
  68. if (cur == p->lims[p->state])
  69. return;
  70. SetBe32(cur, p->tempTarget);
  71. p->bufs[p->state] = cur + 4;
  72. }
  73. p->state = BCJ2_ENC_STATE_ORIG;
  74. for (;;)
  75. {
  76. if (p->range < kTopValue)
  77. {
  78. if (RangeEnc_ShiftLow(p))
  79. return;
  80. p->range <<= 8;
  81. }
  82. {
  83. {
  84. const Byte *src = p->src;
  85. const Byte *srcLim;
  86. Byte *dest;
  87. SizeT num = p->srcLim - src;
  88. if (p->finishMode == BCJ2_ENC_FINISH_MODE_CONTINUE)
  89. {
  90. if (num <= 4)
  91. return;
  92. num -= 4;
  93. }
  94. else if (num == 0)
  95. break;
  96. dest = p->bufs[BCJ2_STREAM_MAIN];
  97. if (num > (SizeT)(p->lims[BCJ2_STREAM_MAIN] - dest))
  98. {
  99. num = p->lims[BCJ2_STREAM_MAIN] - dest;
  100. if (num == 0)
  101. {
  102. p->state = BCJ2_STREAM_MAIN;
  103. return;
  104. }
  105. }
  106. srcLim = src + num;
  107. if (p->prevByte == 0x0F && (src[0] & 0xF0) == 0x80)
  108. *dest = src[0];
  109. else for (;;)
  110. {
  111. Byte b = *src;
  112. *dest = b;
  113. if (b != 0x0F)
  114. {
  115. if ((b & 0xFE) == 0xE8)
  116. break;
  117. dest++;
  118. if (++src != srcLim)
  119. continue;
  120. break;
  121. }
  122. dest++;
  123. if (++src == srcLim)
  124. break;
  125. if ((*src & 0xF0) != 0x80)
  126. continue;
  127. *dest = *src;
  128. break;
  129. }
  130. num = src - p->src;
  131. if (src == srcLim)
  132. {
  133. p->prevByte = src[-1];
  134. p->bufs[BCJ2_STREAM_MAIN] = dest;
  135. p->src = src;
  136. p->ip += (UInt32)num;
  137. continue;
  138. }
  139. {
  140. Byte context = (Byte)(num == 0 ? p->prevByte : src[-1]);
  141. Bool needConvert;
  142. p->bufs[BCJ2_STREAM_MAIN] = dest + 1;
  143. p->ip += (UInt32)num + 1;
  144. src++;
  145. needConvert = False;
  146. if ((SizeT)(p->srcLim - src) >= 4)
  147. {
  148. UInt32 relatVal = GetUi32(src);
  149. if ((p->fileSize == 0 || (UInt32)(p->ip + 4 + relatVal - p->fileIp) < p->fileSize)
  150. && ((relatVal + p->relatLimit) >> 1) < p->relatLimit)
  151. needConvert = True;
  152. }
  153. {
  154. UInt32 bound;
  155. unsigned ttt;
  156. Byte b = src[-1];
  157. CProb *prob = p->probs + (unsigned)(b == 0xE8 ? 2 + (unsigned)context : (b == 0xE9 ? 1 : 0));
  158. ttt = *prob;
  159. bound = (p->range >> kNumModelBits) * ttt;
  160. if (!needConvert)
  161. {
  162. p->range = bound;
  163. *prob = (CProb)(ttt + ((kBitModelTotal - ttt) >> kNumMoveBits));
  164. p->src = src;
  165. p->prevByte = b;
  166. continue;
  167. }
  168. p->low += bound;
  169. p->range -= bound;
  170. *prob = (CProb)(ttt - (ttt >> kNumMoveBits));
  171. {
  172. UInt32 relatVal = GetUi32(src);
  173. UInt32 absVal;
  174. p->ip += 4;
  175. absVal = p->ip + relatVal;
  176. p->prevByte = src[3];
  177. src += 4;
  178. p->src = src;
  179. {
  180. unsigned cj = (b == 0xE8) ? BCJ2_STREAM_CALL : BCJ2_STREAM_JUMP;
  181. Byte *cur = p->bufs[cj];
  182. if (cur == p->lims[cj])
  183. {
  184. p->state = cj;
  185. p->tempTarget = absVal;
  186. return;
  187. }
  188. SetBe32(cur, absVal);
  189. p->bufs[cj] = cur + 4;
  190. }
  191. }
  192. }
  193. }
  194. }
  195. }
  196. }
  197. if (p->finishMode != BCJ2_ENC_FINISH_MODE_END_STREAM)
  198. return;
  199. for (; p->flushPos < 5; p->flushPos++)
  200. if (RangeEnc_ShiftLow(p))
  201. return;
  202. p->state = BCJ2_ENC_STATE_OK;
  203. }
  204. void Bcj2Enc_Encode(CBcj2Enc *p)
  205. {
  206. PRF(printf("\n"));
  207. PRF(printf("---- ip = %8d tempPos = %8d src = %8d\n", p->ip, p->tempPos, p->srcLim - p->src));
  208. if (p->tempPos != 0)
  209. {
  210. unsigned extra = 0;
  211. for (;;)
  212. {
  213. const Byte *src = p->src;
  214. const Byte *srcLim = p->srcLim;
  215. unsigned finishMode = p->finishMode;
  216. p->src = p->temp;
  217. p->srcLim = p->temp + p->tempPos;
  218. if (src != srcLim)
  219. p->finishMode = BCJ2_ENC_FINISH_MODE_CONTINUE;
  220. PRF(printf(" ip = %8d tempPos = %8d src = %8d\n", p->ip, p->tempPos, p->srcLim - p->src));
  221. Bcj2Enc_Encode_2(p);
  222. {
  223. unsigned num = (unsigned)(p->src - p->temp);
  224. unsigned tempPos = p->tempPos - num;
  225. unsigned i;
  226. p->tempPos = tempPos;
  227. for (i = 0; i < tempPos; i++)
  228. p->temp[i] = p->temp[i + num];
  229. p->src = src;
  230. p->srcLim = srcLim;
  231. p->finishMode = finishMode;
  232. if (p->state != BCJ2_ENC_STATE_ORIG || src == srcLim)
  233. return;
  234. if (extra >= tempPos)
  235. {
  236. p->src = src - tempPos;
  237. p->tempPos = 0;
  238. break;
  239. }
  240. p->temp[tempPos] = src[0];
  241. p->tempPos = tempPos + 1;
  242. p->src = src + 1;
  243. extra++;
  244. }
  245. }
  246. }
  247. PRF(printf("++++ ip = %8d tempPos = %8d src = %8d\n", p->ip, p->tempPos, p->srcLim - p->src));
  248. Bcj2Enc_Encode_2(p);
  249. if (p->state == BCJ2_ENC_STATE_ORIG)
  250. {
  251. const Byte *src = p->src;
  252. unsigned rem = (unsigned)(p->srcLim - src);
  253. unsigned i;
  254. for (i = 0; i < rem; i++)
  255. p->temp[i] = src[i];
  256. p->tempPos = rem;
  257. p->src = src + rem;
  258. }
  259. }