graph.cpp 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400
  1. /*****************************************************************************
  2. * dcc project CFG related functions
  3. * (C) Cristina Cifuentes
  4. ****************************************************************************/
  5. #include "dcc.h"
  6. #include <string.h>
  7. #if __BORLAND__
  8. #include <alloc.h>
  9. #else
  10. #include <malloc.h> /* For free() */
  11. #endif
  12. #include "graph.h"
  13. //static BB * rmJMP(Function * pProc, int marker, BB * pBB);
  14. static void mergeFallThrough(Function * pProc, BB * pBB);
  15. static void dfsNumbering(BB * pBB, std::vector<BB*> &dfsLast, int *first, int *last);
  16. /*****************************************************************************
  17. * createCFG - Create the basic control flow graph
  18. ****************************************************************************/
  19. void Function::createCFG()
  20. {
  21. /* Splits Icode associated with the procedure into Basic Blocks.
  22. * The links between BBs represent the control flow graph of the
  23. * procedure.
  24. * A Basic Block is defined to end on one of the following instructions:
  25. * 1) Conditional and unconditional jumps
  26. * 2) CALL(F)
  27. * 3) RET(F)
  28. * 4) On the instruction before a join (a flagged TARGET)
  29. * 5) Repeated string instructions
  30. * 6) End of procedure
  31. */
  32. int i;
  33. int ip, start;
  34. BB * psBB;
  35. BB * pBB;
  36. iICODE pIcode = Icode.begin();
  37. stats.numBBbef = stats.numBBaft = 0;
  38. for (ip = start = 0; pIcode!=Icode.end(); ip++, pIcode++)
  39. {
  40. LLInst *ll = pIcode->ll();
  41. /* Stick a NOWHERE_NODE on the end if we terminate
  42. * with anything other than a ret, jump or terminate */
  43. if (ip + 1 == Icode.size() &&
  44. (not ll->testFlags(TERMINATES)) &&
  45. ll->opcode != iJMP && ll->opcode != iJMPF &&
  46. ll->opcode != iRET && ll->opcode != iRETF)
  47. {
  48. pBB=BB::Create(start, ip, NOWHERE_NODE, 0, this);
  49. }
  50. /* Only process icodes that have valid instructions */
  51. else if (not ll->testFlags(NO_CODE) )
  52. {
  53. switch (ll->opcode) {
  54. case iJB: case iJBE: case iJAE: case iJA:
  55. case iJL: case iJLE: case iJGE: case iJG:
  56. case iJE: case iJNE: case iJS: case iJNS:
  57. case iJO: case iJNO: case iJP: case iJNP:
  58. case iJCXZ:
  59. pBB = BB::Create(start, ip, TWO_BRANCH, 2, this);
  60. CondJumps:
  61. start = ip + 1;
  62. pBB->edges[0].ip = (uint32_t)start;
  63. /* This is for jumps off into nowhere */
  64. if ( ll->testFlags(NO_LABEL) )
  65. {
  66. pBB->edges.pop_back();
  67. }
  68. else
  69. pBB->edges[1].ip = ll->src.op();
  70. break;
  71. case iLOOP: case iLOOPE: case iLOOPNE:
  72. pBB = BB::Create(start, ip, LOOP_NODE, 2, this);
  73. goto CondJumps;
  74. case iJMPF: case iJMP:
  75. if (ll->testFlags(SWITCH))
  76. {
  77. pBB = BB::Create(start, ip, MULTI_BRANCH, ll->caseTbl.numEntries, this);
  78. for (i = 0; i < ll->caseTbl.numEntries; i++)
  79. pBB->edges[i].ip = ll->caseTbl.entries[i];
  80. hasCase = TRUE;
  81. }
  82. else if ((ll->getFlag() & (I | NO_LABEL)) == I) //TODO: WHY NO_LABEL TESTIT
  83. {
  84. pBB = BB::Create(start, ip, ONE_BRANCH, 1, this);
  85. pBB->edges[0].ip = ll->src.op();
  86. }
  87. else
  88. BB::Create(start, ip, NOWHERE_NODE, 0, this);
  89. start = ip + 1;
  90. break;
  91. case iCALLF: case iCALL:
  92. {
  93. Function * p = ll->src.proc.proc;
  94. if (p)
  95. i = ((p->flg) & TERMINATES) ? 0 : 1;
  96. else
  97. i = 1;
  98. pBB = BB::Create(start, ip, CALL_NODE, i, this);
  99. start = ip + 1;
  100. if (i)
  101. pBB->edges[0].ip = (uint32_t)start;
  102. }
  103. break;
  104. case iRET: case iRETF:
  105. BB::Create(start, ip, RETURN_NODE, 0, this);
  106. start = ip + 1;
  107. break;
  108. default:
  109. /* Check for exit to DOS */
  110. iICODE next1=++iICODE(pIcode);
  111. if ( ll->testFlags(TERMINATES) )
  112. {
  113. pBB = BB::Create(start, ip, TERMINATE_NODE, 0, this);
  114. start = ip + 1;
  115. }
  116. /* Check for a fall through */
  117. else if (next1 != Icode.end())
  118. {
  119. assert(next1->loc_ip==ip+1);
  120. if (next1->ll()->testFlags(TARGET | CASE))
  121. {
  122. pBB = BB::Create(start, ip, FALL_NODE, 1, this);
  123. start = ip + 1;
  124. pBB->edges[0].ip = (uint32_t)start;
  125. }
  126. }
  127. break;
  128. }
  129. }
  130. }
  131. auto iter=heldBBs.begin();
  132. /* Convert list of BBs into a graph */
  133. for (; iter!=heldBBs.end(); ++iter)
  134. {
  135. pBB = *iter;
  136. for (i = 0; i < pBB->edges.size(); i++)
  137. {
  138. ip = pBB->edges[i].ip;
  139. if (ip >= SYNTHESIZED_MIN)
  140. {
  141. fatalError (INVALID_SYNTHETIC_BB);
  142. return ;
  143. }
  144. auto iter2=std::find_if(heldBBs.begin(),heldBBs.end(),
  145. [ip](BB *psBB)->bool {return psBB->begin()==ip;});
  146. if(iter2==heldBBs.end())
  147. fatalError(NO_BB, ip, name.c_str());
  148. psBB = *iter2;
  149. pBB->edges[i].BBptr = psBB;
  150. psBB->inEdges.push_back(0);
  151. }
  152. }
  153. }
  154. void Function::markImpure()
  155. {
  156. SYM * psym;
  157. for(ICODE &icod : Icode)
  158. {
  159. if ( not icod.ll()->testFlags(SYM_USE | SYM_DEF))
  160. continue;
  161. psym = &symtab[icod.ll()->caseTbl.numEntries];
  162. for (int c = (int)psym->label; c < (int)psym->label+psym->size; c++)
  163. {
  164. if (BITMAP(c, BM_CODE))
  165. {
  166. icod.ll()->setFlags(IMPURE);
  167. flg |= IMPURE;
  168. break;
  169. }
  170. }
  171. }
  172. }
  173. /*****************************************************************************
  174. * newBB - Allocate new BB and link to end of list
  175. *****************************************************************************/
  176. /*****************************************************************************
  177. * freeCFG - Deallocates a cfg
  178. ****************************************************************************/
  179. void Function::freeCFG()
  180. {
  181. for(BB *p : heldBBs)
  182. delete p;
  183. }
  184. /*****************************************************************************
  185. * compressCFG - Remove redundancies and add in-edge information
  186. ****************************************************************************/
  187. void Function::compressCFG()
  188. {
  189. BB * pBB, *pNxt;
  190. int ip, first=0, last, i;
  191. /* First pass over BB list removes redundant jumps of the form
  192. * (Un)Conditional -> Unconditional jump */
  193. auto iter=m_cfg.begin();
  194. for (;iter!=m_cfg.end(); ++iter)
  195. {
  196. pBB = *iter;
  197. if(pBB->inEdges.empty() || (pBB->nodeType != ONE_BRANCH && pBB->nodeType != TWO_BRANCH))
  198. continue;
  199. for (i = 0; i < pBB->edges.size(); i++)
  200. {
  201. ip = pBB->rbegin();
  202. pNxt = pBB->edges[i].BBptr->rmJMP(ip, pBB->edges[i].BBptr);
  203. if (not pBB->edges.empty()) /* Might have been clobbered */
  204. {
  205. pBB->edges[i].BBptr = pNxt;
  206. assert(pBB->back().loc_ip==ip);
  207. pBB->back().ll()->SetImmediateOp((uint32_t)pNxt->begin());
  208. //Icode[ip].SetImmediateOp((uint32_t)pNxt->begin());
  209. }
  210. }
  211. }
  212. /* Next is a depth-first traversal merging any FALL_NODE or
  213. * ONE_BRANCH that fall through to a node with that as their only
  214. * in-edge. */
  215. m_cfg.front()->mergeFallThrough(Icode);
  216. /* Remove redundant BBs created by the above compressions
  217. * and allocate in-edge arrays as required. */
  218. stats.numBBaft = stats.numBBbef;
  219. for(auto iter=m_cfg.begin(); iter!=m_cfg.end(); ++iter)
  220. {
  221. pBB = *iter;
  222. if (pBB->inEdges.empty())
  223. {
  224. if (iter == m_cfg.begin()) /* Init it misses out on */
  225. pBB->index = UN_INIT;
  226. else
  227. {
  228. delete pBB;
  229. stats.numBBaft--;
  230. }
  231. }
  232. else
  233. {
  234. pBB->inEdgeCount = pBB->inEdges.size();
  235. }
  236. }
  237. /* Allocate storage for dfsLast[] array */
  238. numBBs = stats.numBBaft;
  239. m_dfsLast.resize(numBBs,0); // = (BB **)allocMem(numBBs * sizeof(BB *))
  240. /* Now do a dfs numbering traversal and fill in the inEdges[] array */
  241. last = numBBs - 1;
  242. m_cfg.front()->dfsNumbering(m_dfsLast, &first, &last);
  243. }
  244. /****************************************************************************
  245. * rmJMP - If BB addressed is just a JMP it is replaced with its target
  246. ***************************************************************************/
  247. BB *BB::rmJMP(int marker, BB * pBB)
  248. {
  249. marker += DFS_JMP;
  250. while (pBB->nodeType == ONE_BRANCH && pBB->size() == 1)
  251. {
  252. if (pBB->traversed != marker)
  253. {
  254. pBB->traversed = marker;
  255. pBB->inEdges.pop_back();
  256. if (not pBB->inEdges.empty())
  257. {
  258. pBB->edges[0].BBptr->inEdges.push_back(0);
  259. }
  260. else
  261. {
  262. pBB->front().ll()->setFlags(NO_CODE);
  263. pBB->front().invalidate(); //pProc->Icode.SetLlInvalid(pBB->begin(), TRUE);
  264. }
  265. pBB = pBB->edges[0].BBptr;
  266. }
  267. else
  268. {
  269. /* We are going around in circles */
  270. pBB->nodeType = NOWHERE_NODE;
  271. pBB->front().ll()->src.SetImmediateOp(pBB->front().loc_ip);
  272. //pBB->front().ll()->src.immed.op = pBB->front().loc_ip;
  273. do {
  274. pBB = pBB->edges[0].BBptr;
  275. pBB->inEdges.pop_back(); // was --numInedges
  276. if (! pBB->inEdges.empty())
  277. {
  278. pBB->front().ll()->setFlags(NO_CODE);
  279. pBB->front().invalidate();
  280. // pProc->Icode.setFlags(pBB->start, NO_CODE);
  281. // pProc->Icode.SetLlInvalid(pBB->start, TRUE);
  282. }
  283. } while (pBB->nodeType != NOWHERE_NODE);
  284. pBB->edges.clear();
  285. }
  286. }
  287. return pBB;
  288. }
  289. /*****************************************************************************
  290. * mergeFallThrough
  291. ****************************************************************************/
  292. void BB::mergeFallThrough( CIcodeRec &Icode)
  293. {
  294. BB * pChild;
  295. int i;
  296. if (!this)
  297. {
  298. printf("mergeFallThrough on empty BB!\n");
  299. }
  300. while (nodeType == FALL_NODE || nodeType == ONE_BRANCH)
  301. {
  302. pChild = edges[0].BBptr;
  303. /* Jump to next instruction can always be removed */
  304. if (nodeType == ONE_BRANCH)
  305. {
  306. assert(Parent==pChild->Parent);
  307. if(back().loc_ip>pChild->front().loc_ip) // back edege
  308. break;
  309. auto iter=std::find_if(this->end2(),pChild->begin2(),[](ICODE &c)
  310. {return not c.ll()->testFlags(NO_CODE);});
  311. if (iter != pChild->begin2())
  312. break;
  313. back().ll()->setFlags(NO_CODE);
  314. back().invalidate();
  315. nodeType = FALL_NODE;
  316. length--;
  317. }
  318. /* If there's no other edges into child can merge */
  319. if (pChild->inEdges.size() != 1)
  320. break;
  321. nodeType = pChild->nodeType;
  322. length = (pChild->start - start) + pChild->length ;
  323. pChild->front().ll()->clrFlags(TARGET);
  324. edges.swap(pChild->edges);
  325. pChild->inEdges.clear();
  326. pChild->edges.clear();
  327. }
  328. traversed = DFS_MERGE;
  329. /* Process all out edges recursively */
  330. for (i = 0; i < edges.size(); i++)
  331. if (edges[i].BBptr->traversed != DFS_MERGE)
  332. edges[i].BBptr->mergeFallThrough(Icode);
  333. }
  334. /*****************************************************************************
  335. * dfsNumbering - Numbers nodes during first and last visits and determine
  336. * in-edges
  337. ****************************************************************************/
  338. void BB::dfsNumbering(std::vector<BB *> &dfsLast, int *first, int *last)
  339. {
  340. BB * pChild;
  341. traversed = DFS_NUM;
  342. dfsFirstNum = (*first)++;
  343. /* index is being used as an index to inEdges[]. */
  344. // for (i = 0; i < edges.size(); i++)
  345. for (auto edge : edges)
  346. {
  347. pChild = edge.BBptr;
  348. pChild->inEdges[pChild->index++] = this;
  349. /* Is this the last visit? */
  350. if (pChild->index == pChild->inEdges.size())
  351. pChild->index = UN_INIT;
  352. if (pChild->traversed != DFS_NUM)
  353. pChild->dfsNumbering(dfsLast, first, last);
  354. }
  355. dfsLastNum = *last;
  356. dfsLast[(*last)--] = this;
  357. }