OpenCores
URL https://opencores.org/ocsvn/or1k/or1k/trunk

Subversion Repositories or1k

[/] [or1k/] [tags/] [nog_patch_69/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 1044

Go to most recent revision | Details | Compare with Previous | View Log

Line No. Rev Author Line
1 879 markom 
/* bb.c -- OpenRISC Custom Unit Compiler, Basic Block handling
2 
 *    Copyright (C) 2002 Marko Mlinar, markom@opencores.org
3 
 *
4 
 *    This file is part of OpenRISC 1000 Architectural Simulator.
5 
 *
6 
 *    This program is free software; you can redistribute it and/or modify
7 
 *    it under the terms of the GNU General Public License as published by
8 
 *    the Free Software Foundation; either version 2 of the License, or
9 
 *    (at your option) any later version.
10 
 *
11 
 *    This program is distributed in the hope that it will be useful,
12 
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
13 
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14 
 *    GNU General Public License for more details.
15 
 *
16 
 *    You should have received a copy of the GNU General Public License
17 
 *    along with this program; if not, write to the Free Software
18 
 *    Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */
19 
 
20 
#include <stdio.h>
21 
#include <stdlib.h>
22 
#include <stdarg.h>
23 
#include <assert.h>
24 897 markom 
#include "sim-config.h"
25 
#include "abstract.h"
26 879 markom 
#include "cuc.h"
27 
#include "insn.h"
28 
#include "support/profile.h"
29 
 
30 925 markom 
/* prints out bb string */
31 
void print_bb_num (int num)
32 
{
33 997 markom 
  if (num < 0) PRINTF ("*");
34 
  else if (num == BBID_END) PRINTF ("END");
35 
  else if (num == BBID_START) PRINTF ("START");
36 
  else PRINTF ("%2x", num);
37 925 markom 
}
38 
 
39 879 markom 
/* Print out basic blocks */
40 
void print_cuc_bb (cuc_func *f, char *s)
41 
{
42 
  int i;
43 997 markom 
  PRINTF ("------- %s -------\n", s);
44 879 markom 
  for (i = 0; i < f->num_bb; i++) {
45 997 markom 
    if (f->bb[i].insn) PRINTF ("\n---- BB%-2x * %x ---- ", i, f->bb[i].cnt);
46 
    else PRINTF ("BB%-2x: %4x-%-4x", i, f->bb[i].first, f->bb[i].last);
47 
    PRINTF (" type %02x tmp %i ", f->bb[i].type, f->bb[i].tmp);
48 
    PRINTF ("next "); print_bb_num (f->bb[i].next[0]);
49 
    PRINTF (" "); print_bb_num (f->bb[i].next[1]);
50 
    PRINTF (" prev "); print_bb_num (f->bb[i].prev[0]);
51 
    PRINTF (" "); print_bb_num (f->bb[i].prev[1]);
52 
    PRINTF ("\n");
53 879 markom 
 
54 
    if (f->bb[i].insn) print_insns (f->bb[i].insn, f->bb[i].ninsn, 0);
55 
  }
56 953 markom 
  if (f->nmsched) {
57 997 markom 
    PRINTF ("\nmsched: ");
58 953 markom 
    for (i = 0; i < f->nmsched; i++)
59 997 markom 
      PRINTF ("%x ", f->msched[i]);
60 
    PRINTF ("\n\n\n");
61 
  } else PRINTF ("\n");
62 897 markom 
  fflush (stdout);
63 879 markom 
}
64 
 
65 
/* Copies src basic block into destination */
66 
cuc_bb *cpy_bb (cuc_bb *dest, cuc_bb *src)
67 
{
68 897 markom 
  int i, j;
69 941 markom 
  dep_list *d;
70 897 markom 
  assert (dest != src);
71 879 markom 
  *dest = *src;
72 
  assert (dest->insn = malloc (sizeof (cuc_insn) * src->ninsn));
73 941 markom 
  for (i = 0; i < src->ninsn; i++) {
74 
    d = src->insn[i].dep;
75 879 markom 
    dest->insn[i] = src->insn[i];
76 941 markom 
    dest->insn[i].dep = NULL;
77 
    while (d) {
78 
      add_dep (&dest->insn[i].dep, d->ref);
79 
      d = d->next;
80 
    }
81 
  }
82 
 
83 
  d = src->mdep;
84 
  dest->mdep = NULL;
85 
  while (d) {
86 
    add_dep (&dest->mdep, d->ref);
87 
    d = d->next;
88 
  }
89 879 markom 
  if (src->ntim) {
90 
    assert (dest->tim = malloc (sizeof (cuc_timings) * src->ntim));
91 897 markom 
    for (i = 0; i < src->ntim; i++) {
92 
      dest->tim[i] = src->tim[i];
93 
      if (src->tim[i].nshared) {
94 
        assert (dest->tim[i].shared = malloc (sizeof (int) * src->tim[i].nshared));
95 
        for (j = 0; j < src->tim[i].nshared; j++)
96 
          dest->tim[i].shared[j] = src->tim[i].shared[j];
97 
      }
98 
    }
99 879 markom 
  }
100 
}
101 
 
102 
/* Duplicates function */
103 
cuc_func *dup_func (cuc_func *f)
104 
{
105 
  cuc_func *n = (cuc_func *) malloc (sizeof (cuc_func));
106 
  int b, i;
107 
  for (b = 0; b < f->num_bb; b++) cpy_bb (&n->bb[b], &f->bb[b]);
108 
  n->num_bb = f->num_bb;
109 
  assert (n->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
110 
  for (b = 0; b < f->num_init_bb; b++) n->init_bb_reloc[b] = f->init_bb_reloc[b];
111 
  n->num_init_bb = f->num_init_bb;
112 915 markom 
  for (i = 0; i < MAX_REGS; i++) {
113 
    n->saved_regs[i] = f->saved_regs[i];
114 
    n->lur[i] = f->lur[i];
115 
    n->used_regs[i] = f->used_regs[i];
116 
  }
117 879 markom 
  n->start_addr = f->start_addr;
118 
  n->end_addr = f->end_addr;
119 
  n->orig_time = f->orig_time;
120 
  n->nmsched = f->nmsched;
121 1041 markom 
  n->num_runs = f->num_runs;
122 879 markom 
  for (i = 0; i < f->nmsched; i++) {
123 
    n->msched[i] = f->msched[i];
124 
    n->mtype[i] = f->mtype[i];
125 
  }
126 906 markom 
  n->nfdeps = f->nfdeps;
127 
  if (f->nfdeps) {
128 
    f->fdeps = (cuc_func **) malloc (sizeof (cuc_func *) * f->nfdeps);
129 
    for (i = 0; i < f->nfdeps; i++) n->fdeps[i] = f->fdeps[i];
130 
  }
131 879 markom 
  return n;
132 
}
133 
 
134 
/* Releases memory allocated by function */
135 
void free_func (cuc_func *f)
136 
{
137 
  int b, i;
138 
  for (b = 0; b < f->num_bb; b++) {
139 
    for (i = 0; i < f->bb[b].ninsn; i++)
140 
      dispose_list (&f->bb[b].insn[i].dep);
141 
    if (f->bb[b].insn) free (f->bb[b].insn);
142 897 markom 
    for (i = 0; i < f->bb[b].ntim; i++)
143 
      if (f->bb[b].tim[i].nshared && f->bb[b].tim[i].shared)
144 
        free (f->bb[b].tim[i].shared);
145 879 markom 
    if (f->bb[b].tim && f->bb[b].ntim) free (f->bb[b].tim);
146 
  }
147 
  free (f);
148 
}
149 
 
150 
/* Recalculates last_used_reg */
151 
void recalc_last_used_reg (cuc_func *f, int b)
152 
{
153 
  int i;
154 
  cuc_bb *bb = &f->bb[b];
155 
 
156 
  /* rebuild last used reg array */
157 
  if (bb->insn[0].index == II_LRBB) bb->last_used_reg[LRBB_REG] = 0;
158 
  else bb->last_used_reg[LRBB_REG] = -1;
159 
 
160 
  for (i = 1; i < MAX_REGS - 1; i++) bb->last_used_reg[i] = -1;
161 
 
162 
    /* Create references */
163 
  for (i = 0; i < bb->ninsn; i++) {
164 
    int k;
165 
    /* Now check for destination operand(s) */
166 
    for (k = 0; k < MAX_OPERANDS; k++) if (bb->insn[i].opt[k] & OPT_DEST)
167 
      if ((bb->insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
168 
        && (int)bb->insn[i].op[k] >= 0) {
169 
        bb->last_used_reg[bb->insn[i].op[k]] = REF (b, i);
170 
      }
171 
  }
172 
}
173 
 
174 
/* Set the BB limits */
175 
void detect_bb (cuc_func *f)
176 
{
177 
  int i, j, end_bb = 0, eb = 0;
178 
 
179 
  /* Mark block starts/ends */
180 
  for (i = 0; i < num_insn; i++) {
181 
    if (end_bb) insn[i].type |= IT_BBSTART;
182 
    end_bb = 0;
183 
    if (insn[i].type & IT_BRANCH) {
184 
      int jt = insn[i].op[0];
185 
      insn[i].type |= IT_BBEND;
186 
      end_bb = 1;
187 
      if (jt < 0 || jt >= num_insn) {
188 
        fprintf (stderr, "Instruction #%i:Jump out of function '%s'.\n", i, insn[i].disasm);
189 
        exit (1);
190 
      }
191 
      if (jt > 0) insn[jt - 1].type |= IT_BBEND;
192 
      insn[jt].type |= IT_BBSTART;
193 
    }
194 
  }
195 
 
196 
  /* Initialize bb array */
197 
  insn[0].type |= IT_BBSTART;
198 
  insn[num_insn - 1].type |= IT_BBEND;
199 
  f->num_bb = 0;
200 
  for (i = 0; i < num_insn; i++) {
201 
    if (insn[i].type & IT_BBSTART) {
202 
      f->bb[f->num_bb].first = i;
203 
      f->bb[f->num_bb].cnt = 0;
204 
    }
205 
    /* Determine repetitions of a loop */
206 
    if (insn[i].type & IT_BBEND) {
207 
      f->bb[f->num_bb].type = 0;
208 
      f->bb[f->num_bb].last = i;
209 
      f->bb[f->num_bb].next[0] = f->bb[f->num_bb].next[1] = -1;
210 
      f->bb[f->num_bb].tmp = 0;
211 
      f->bb[f->num_bb].ntim = 0;
212 
      f->num_bb++;
213 
      assert (f->num_bb < MAX_BB);
214 
    }
215 
  }
216 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INIT");
217 879 markom 
 
218 
  /* Build forward connections between BBs */
219 
  for (i = 0; i < f->num_bb; i++)
220 
    if (insn[f->bb[i].last].type & IT_BRANCH) {
221 
      int j;
222 
      assert (insn[f->bb[i].last].index == II_BF);
223 
      /* Find block this instruction jumps to */
224 
      for (j = 0; j < f->num_bb; j++)
225 
        if (f->bb[j].first == insn[f->bb[i].last].op[0]) break;
226 
      assert (j < f->num_bb);
227 
 
228 
      /* Convert the jump address to BB link */
229 
      insn[f->bb[i].last].op[0] = j; insn[f->bb[i].last].opt[0] = OPT_BB;
230 
 
231 
      /* Make a link */
232 
      f->bb[i].next[0] = j;
233 
      if (++f->bb[j].tmp > 2) eb++;
234 
      f->bb[i].next[1] = i + 1;
235 
      if (++f->bb[i + 1].tmp > 2) eb++;
236 
    } else if (f->bb[i].last == num_insn - 1) { /* Last instruction doesn't have to do anything */
237 
      f->bb[i].type |= BB_END;
238 
    } else {
239 
      f->bb[i].next[0] = i + 1;
240 
      if (++f->bb[i + 1].tmp > 2) eb++;
241 
    }
242 
 
243 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_NEXT");
244 879 markom 
 
245 
  /* Build backward connections, but first insert artificial blocks
246 
   * to handle more than 2 connections */
247 883 markom 
  cucdebug (6, "artificial %i %i\n", f->num_bb, eb);
248 879 markom 
  end_bb = f->num_bb + eb;
249 
  for (i = f->num_bb - 1; i >= 0; i--) {
250 
    j = f->bb[i].tmp;
251 
    if (f->bb[i].tmp > 2) f->bb[i].tmp = -f->bb[i].tmp;
252 
    f->bb[--end_bb] = f->bb[i];
253 
    reloc[i] = end_bb;
254 
    while (j-- > 2) {
255 
      f->bb[--end_bb].first = f->bb[i].first;
256 
      f->bb[end_bb].last = -1;
257 925 markom 
      f->bb[end_bb].type &= ~BB_END;
258 879 markom 
      f->bb[end_bb].next[0] = -1;
259 
      f->bb[end_bb].next[1] = -1;
260 
      f->bb[end_bb].tmp = 0;
261 
      f->bb[end_bb].cnt = f->bb[i].cnt;
262 
      f->bb[end_bb].ntim = 0;
263 
    }
264 
  }
265 
  f->num_bb += eb;
266 
 
267 
  /* relocate jump instructions */
268 
  for (i = 0; i < num_insn; i++)
269 
    for (j = 0; j < MAX_OPERANDS; j++)
270 
      if (insn[i].opt[j] & OPT_BB)
271 
        insn[i].op[j] = reloc[insn[i].op[j]];
272 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT-reloc");
273 879 markom 
  for (i = 0; i < f->num_bb; i++) {
274 
    if (f->bb[i].next[0] >= 0) {
275 
      int t = reloc[f->bb[i].next[0]];
276 
      if (f->bb[t].tmp < 0) {
277 
        f->bb[t].tmp = -f->bb[t].tmp;
278 
        t -= f->bb[t].tmp - 2;
279 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
280 
      f->bb[i].next[0] = t;
281 
    }
282 
    if (f->bb[i].next[1] >= 0) {
283 
      int t = reloc[f->bb[i].next[1]];
284 
      if (f->bb[t].tmp < 0) {
285 
        f->bb[t].tmp = -f->bb[t].tmp;
286 
        t -= f->bb[t].tmp - 2;
287 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
288 
      f->bb[i].next[1] = t;
289 
    }
290 
    /* artificial blocks do not have relocations, hardcode them */
291 
    if (f->bb[i].last < 0) f->bb[i].next[0] = i + 1;
292 
  }
293 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT");
294 879 markom 
 
295 
  /* Uncoditional branched do not continue to next block */
296 
  for (i = 0; i < f->num_bb; i++) {
297 
    cuc_insn *ii;
298 
    if (f->bb[i].last < 0) continue;
299 
    ii = &insn[f->bb[i].last];
300 
    /* Unconditional branch? */
301 
    if (ii->type & IT_BRANCH && ii->opt[1] & OPT_CONST) {
302 
      change_insn_type (ii, II_NOP);
303 1044 markom 
#if 0
304 879 markom 
      if (f->bb[i].next[1] == i + 1) f->bb[i].next[0] = f->bb[i].next[1];
305 1044 markom 
#endif
306 879 markom 
      f->bb[i].next[1] = -1;
307 
    }
308 1044 markom 
  }
309 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_UNCOND_JUMP");
310 879 markom 
 
311 
  /* Add backward connections */
312 
  for (i = 0; i < f->num_bb; i++)
313 
    f->bb[i].prev[0] = f->bb[i].prev[1] = -1;
314 
 
315 
  for (i = 0; i < f->num_bb; i++) {
316 
    if (f->bb[i].next[0] >= 0) {
317 
      int t = f->bb[i].next[0];
318 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
319 
      else {
320 
        assert (f->bb[t].prev[1] < 0);
321 
        f->bb[t].prev[1] = i;
322 
      }
323 
    }
324 
    if (f->bb[i].next[1] >= 0) {
325 
      int t = f->bb[i].next[1];
326 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
327 
      else {
328 
        assert (f->bb[t].prev[1] < 0);
329 
        f->bb[t].prev[1] = i;
330 
      }
331 
    }
332 
  }
333 932 markom 
  /* Add START marker */
334 
  assert (f->bb[0].prev[0] < 0);
335 
  f->bb[0].prev[0]= BBID_START;
336 
 
337 925 markom 
  /* Add END marker */
338 
  for (i = 0; i < f->num_bb; i++)
339 
    if (f->bb[i].type & BB_END) {
340 
      assert (f->bb[i].next[0] < 0);
341 
      f->bb[i].next[0] = BBID_END;
342 
    }
343 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_PREV");
344 879 markom 
}
345 
 
346 905 markom 
/* We do a quick check if there are some anomalies with references */
347 
void cuc_check (cuc_func *f)
348 
{
349 
  int i, j, k;
350 973 markom 
  cucdebug (1, "cuc_check\n");
351 905 markom 
  for (i = 0; i < f->num_bb; i++) {
352 930 markom 
    if (!f->bb[i].insn && f->bb[i].ninsn) goto err;
353 
    for (j = 0; j < f->bb[i].ninsn; j++) {
354 
      cuc_insn *ii = &f->bb[i].insn[j];
355 934 markom 
      if ((ii->index == II_CMOV || ii->index == II_ADD) && ii->type & IT_COND) {
356 930 markom 
        k = 0;
357 
        assert (ii->opt[k] & OPT_REGISTER);
358 
        if ((signed)ii->op[k] >= 0 && ii->op[k] != FLAG_REG && ii->op[k] != LRBB_REG) {
359 973 markom 
          cucdebug (1, "%x %x\n", ii->opt[0], ii->op[0]);
360 930 markom 
          goto err;
361 
        }
362 
      }
363 905 markom 
      for (k = 0; k < MAX_OPERANDS; k++)
364 930 markom 
        if (ii->opt[k] & OPT_REF) {
365 
          int t = ii->op[k];
366 
          if (REF_BB(t) >= f->num_bb || REF_I (t) >= f->bb[REF_BB(t)].ninsn
367 934 markom 
           || (ii->index == II_CMOV || ii->index == II_ADD) && (
368 930 markom 
                (f->INSN(t).type & IT_COND) != (ii->type & IT_COND) && k < 3
369 
              || !(f->INSN(t).type & IT_COND) && k == 3)) goto err;
370 905 markom 
        }
371 930 markom 
    }
372 905 markom 
  }
373 930 markom 
  return;
374 
err:
375 973 markom 
  cucdebug (1, "Anomaly detected at %x.%x[%i]\n", i, j, k);
376 930 markom 
  print_cuc_bb (f, "ANOMALY");
377 
  exit (1);
378 905 markom 
}
379 
 
380 879 markom 
/* Build basic blocks */
381 
void build_bb (cuc_func *f)
382 
{
383 
  int i, j, k;
384 
  for (i = 0; i < f->num_bb; i++) {
385 
    if (f->bb[i].last < 0) f->bb[i].ninsn = MAX_REGS - 1;
386 
    else f->bb[i].ninsn = f->bb[i].last - f->bb[i].first + 1 + MAX_REGS - 1;
387 
    assert (f->bb[i].ninsn >= MAX_REGS - 1);
388 
    f->bb[i].insn = (cuc_insn *) malloc (sizeof (cuc_insn) * f->bb[i].ninsn);
389 
    assert (f->bb[i].insn);
390 
    f->bb[i].nmemory = 0;
391 
    f->bb[i].unrolled = 1;
392 
 
393 
    /* Save space for conditional moves, exclude r0, place lrbb instead */
394 
    change_insn_type (&f->bb[i].insn[0], II_LRBB);
395 
    strcpy (f->bb[i].insn[0].disasm, "lrbb");
396 930 markom 
    f->bb[i].insn[0].type = IT_UNUSED | IT_COND;
397 879 markom 
    f->bb[i].insn[0].dep = NULL;
398 
    f->bb[i].insn[0].op[0] = LRBB_REG; f->bb[i].insn[0].opt[0] = OPT_REGISTER | OPT_DEST;
399 
    f->bb[i].insn[0].opt[1] = OPT_LRBB;
400 
    f->bb[i].insn[0].opt[2] = f->bb[i].insn[0].opt[3] = OPT_NONE;
401 
    for (j = 1; j < MAX_REGS - 1; j++) {
402 
      change_insn_type (&f->bb[i].insn[j], II_CMOV);
403 
      strcpy (f->bb[i].insn[j].disasm, "cmov");
404 930 markom 
      f->bb[i].insn[j].type = j == FLAG_REG || j == LRBB_REG ? IT_COND : 0;
405 879 markom 
      f->bb[i].insn[j].dep = NULL;
406 
      f->bb[i].insn[j].opt[0] = f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_REGISTER;
407 
      f->bb[i].insn[j].opt[0] |= OPT_DEST;
408 
      f->bb[i].insn[j].op[0] = f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = j;
409 
      f->bb[i].insn[j].op[3] = LRBB_REG; f->bb[i].insn[j].opt[3] = OPT_REGISTER;
410 
    }
411 897 markom 
 
412 
    /* Relocate instructions */
413 879 markom 
    for (j = MAX_REGS - 1; j < f->bb[i].ninsn; j++) {
414 
      f->bb[i].insn[j] = insn[f->bb[i].first + j - (MAX_REGS - 1)];
415 
      for (k = 0; k < MAX_OPERANDS; k++)
416 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
417 
          int b1;
418 
          for (b1 = 0; b1 < i; b1++)
419 898 markom 
            if (f->bb[b1].first <= (signed) f->bb[i].insn[j].op[k]
420 
              && (signed)f->bb[i].insn[j].op[k] <= f->bb[b1].last) break;
421 879 markom 
          assert (b1 < f->num_bb);
422 
          f->bb[i].insn[j].op[k] = REF (b1, f->bb[i].insn[j].op[k] - f->bb[b1].first + MAX_REGS - 1);
423 
        }
424 
      if (f->bb[i].insn[j].type & IT_MEMORY) f->bb[i].nmemory++;
425 
    }
426 
  }
427 905 markom 
  cuc_check (f);
428 879 markom 
}
429 
 
430 1042 markom 
/* Does simplification on blocks A, B, C:
431 
   A->B->C, A->C to just A->B->C */
432 
static void simplify_bb (cuc_func *f, int pred, int s1, int s2, int neg)
433 
{
434 
  cuc_insn *last;
435 
  int i;
436 
  if (cuc_debug >= 3) print_cuc_bb (f, "BEFORE_SIMPLIFY");
437 
  cucdebug (3, "simplify %x->%x->%x (%i)\n", pred, s1, s2, neg);
438 
  assert (s2 != pred); /* Shouldn't occur => stupid */
439 
  f->bb[pred].next[1] = -1;
440 
  f->bb[pred].next[0] = s1;
441 
 
442 
  if (f->bb[s2].prev[0] == pred) {
443 
    f->bb[s2].prev[0] = f->bb[s2].prev[1];
444 
    f->bb[s2].prev[1] = -1;
445 
  } else if (f->bb[s2].prev[1] == pred) {
446 
    f->bb[s2].prev[1] = -1;
447 
  } else assert (0);
448 
 
449 
  last = &f->bb[pred].insn[f->bb[pred].ninsn - 1];
450 
  assert (last->type & IT_BRANCH);
451 
  for (i = 0; i < f->bb[s2].ninsn; i++) {
452 
    cuc_insn *ii= &f->bb[s2].insn[i];
453 
    if (ii->index == II_LRBB) {
454 
      change_insn_type (ii, II_CMOV);
455 
      ii->type = IT_COND;
456 
      ii->op[1] = neg ? 0 : 1; ii->opt[1] = OPT_CONST;
457 
      ii->op[2] = neg ? 1 : 0; ii->opt[2] = OPT_CONST;
458 
      ii->op[3] = last->op[1]; ii->opt[3] = last->opt[1];
459 
    }
460 
  }
461 
  change_insn_type (last, II_NOP);
462 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_SIMPLIFY");
463 
}
464 
 
465 879 markom 
/* type == 0; keep predecessor condition
466 
 * type == 1; keep successor condition
467 
 * type == 2; join loop unrolled blocks */
468 
static void join_bb (cuc_func *f, int pred, int succ, int type)
469 
{
470 905 markom 
  int i, j, k, n1, n2, ninsn, add_cond = 0;
471 879 markom 
  unsigned long cond_op, cond_opt;
472 
  cuc_insn *insn;
473 
 
474 905 markom 
  if (cuc_debug) cuc_check (f);
475 897 markom 
  cucdebug (3, "%x <= %x+%x (%i)\n", pred, pred, succ, type);
476 
  cucdebug (3, "%x %x\n", f->bb[pred].ninsn, f->bb[succ].ninsn);
477 
  if (cuc_debug >= 3) fflush (stdout);
478 879 markom 
 
479 905 markom 
  n1 = f->bb[pred].ninsn;
480 
  n2 = f->bb[succ].ninsn;
481 
  if (n1 <= 0
482 
   || !(f->bb[pred].insn[n1 - 1].type & IT_BRANCH)) type = 1;
483 879 markom 
  if (type == 0 && f->bb[succ].prev[0] == f->bb[succ].next[0]) add_cond = 1;
484 905 markom 
  if (type == 2) add_cond = 1;
485 
 
486 
  assert (f->bb[pred].next[0] == f->bb[succ].next[0] || type != 2); /* not supported */
487 879 markom 
 
488 905 markom 
  ninsn = n1 + n2 + (type == 1 ? 0 : 1) + (add_cond ? MAX_REGS : 0);
489 879 markom 
 
490 
  insn = (cuc_insn *) malloc (ninsn * sizeof (cuc_insn));
491 905 markom 
  for (i = 0; i < n1; i++) insn[i] = f->bb[pred].insn[i];
492 
  /* when type == 0, we move the last (jump) instruction to the end */
493 879 markom 
  if (type == 0 || type == 2) {
494 905 markom 
    /* Move first branch instruction to the end */
495 
    assert (insn[n1 - 1].type & IT_BRANCH);
496 
    insn[ninsn - 1] = insn[n1 - 1];
497 
    cond_op = insn[n1 - 1].op[1];
498 
    cond_opt = insn[n1 - 1].opt[1];
499 
 
500 
    /* Remove old branch */
501 
    change_insn_type (&insn[n1 - 1], II_NOP);
502 879 markom 
  }
503 
  /* Copy second block */
504 905 markom 
  for (i = 0; i < n2; i++) insn[i + n1] = f->bb[succ].insn[i];
505 902 markom 
 
506 
  /* and when type == 2, we may need to add sfor instruction, to quit when either is true */
507 
  if (type == 2) {
508 905 markom 
    /* Move second branch instruction to the end */
509 
    if (insn[n1 + n2 - 1].type & IT_BRANCH) {
510 
      insn[ninsn - 1] = insn[n1 + n2 - 1];
511 
 
512 
      /* Use conditional from cmov FLAG_REG, c_p, c_s, c_p */
513 
      insn[ninsn - 1].op[1] = REF (pred, n1 + n2 + FLAG_REG); insn[ninsn - 1].opt[1] = OPT_REF;
514 
 
515 
      /* Remove old one */
516 
      change_insn_type (&insn[n1 + n2 - 1], II_NOP);
517 
    } else change_insn_type (&insn[ninsn - 1], II_NOP); /* do not use branch slot */
518 902 markom 
  }
519 
 
520 905 markom 
#if 1
521 927 markom 
  /* LRBB at start of succ BB is not valid anymore */
522 933 markom 
  if (n1 > 0 && insn[n1].index == II_LRBB) {
523 925 markom 
    if (type == 1) {
524 
      /* We have two possibilities, how this could have happened:
525 
         1. we just moved second predecessor of succ to pred,
526 
            pred now having two predecessors => everything is ok
527 
         2. we just moved second predecessor of succ to pred,
528 
            now, having just one predecessor => LRBB is not needed anymore */
529 
      if (f->bb[pred].prev[1] < 0) { /* handle second option */
530 
        change_insn_type (&insn[n1], II_ADD);
531 
        insn[n1].op[1] = 1; insn[n1].opt[1] = OPT_CONST;
532 
        insn[n1].op[2] = 0; insn[n1].opt[2] = OPT_CONST;
533 
        insn[n1].opt[3] = OPT_NONE;
534 
      }
535 
    } else {
536 
      assert (0); /* not tested yet */
537 
      change_insn_type (&insn[n1], II_NOP);
538 
      for (i = n1; i < ninsn; i++)
539 
        if (insn[i].index == II_CMOV && insn[i].op[3] == REF (pred, n1)) {
540 
          assert (insn[i].opt[3] == OPT_REF);
541 
          insn[i].op[3] = cond_op;
542 
          insn[i].opt[3] = cond_opt;
543 
          if (f->bb[pred].next[0] != succ) {
544 
            unsigned long t; /* negate conditional -- exchange */
545 
            assert (f->bb[pred].next[1] == succ);
546 
            t = insn[i].op[1];
547 
            insn[i].op[1] = insn[i].op[2];
548 
            insn[i].op[2] = t;
549 
            t = insn[i].opt[1];
550 
            insn[i].opt[1] = insn[i].opt[2];
551 
            insn[i].opt[2] = t;
552 
          }
553 902 markom 
        }
554 925 markom 
    }
555 902 markom 
  }
556 905 markom 
#endif
557 879 markom 
 
558 
  for (i = 0; i < ninsn; i++) reloc[i] = -1;
559 
 
560 
  /* Add conditional instructions if required */
561 
  if (add_cond) {
562 
    recalc_last_used_reg (f, pred);
563 
    recalc_last_used_reg (f, succ);
564 
 
565 
    /* r0 -- add nop for it */
566 905 markom 
    change_insn_type (&insn[n1 + n2], II_NOP);
567 879 markom 
    for (i = 1; i < MAX_REGS; i++) {
568 905 markom 
      cuc_insn *ii = &insn[n1 + n2 + i];
569 879 markom 
      int a = f->bb[pred].last_used_reg[i];
570 
      int b = f->bb[succ].last_used_reg[i];
571 
 
572 
      if (b < 0) change_insn_type (ii, II_NOP);
573 
      else if (a < 0) {
574 
        change_insn_type (ii, II_ADD);
575 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
576 879 markom 
        ii->dep = NULL;
577 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
578 
        ii->op[1] = b; ii->opt[1] = OPT_REF;
579 
        ii->op[2] = 0; ii->opt[2] = OPT_CONST;
580 905 markom 
        ii->opt[3] = OPT_NONE;
581 879 markom 
      } else if (b >= 0) {
582 
        change_insn_type (ii, II_CMOV);
583 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
584 879 markom 
        ii->dep = NULL;
585 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
586 
        ii->op[1] = a; ii->opt[1] = OPT_REF;
587 
        ii->op[2] = b; ii->opt[2] = OPT_REF;
588 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
589 905 markom 
        reloc[REF_I(a)] = REF (pred, n1 + n2 + i);
590 879 markom 
      }
591 
      sprintf (ii->disasm, "cmov (join BB)");
592 
    }
593 
  }
594 
 
595 905 markom 
  if (cuc_debug) cuc_check (f);
596 925 markom 
  if (f->bb[succ].type & BB_END) {
597 
    f->bb[pred].type |= BB_END;
598 933 markom 
    if (ninsn > 0 && insn[ninsn - 1].type & IT_BRANCH && insn[ninsn - 1].op[0] == succ) {
599 925 markom 
      assert (insn[ninsn - 1].opt[0] & OPT_BB);
600 
      insn[ninsn - 1].op[0] = BBID_END;
601 
    }
602 
  }
603 879 markom 
  i = 0;
604 925 markom 
  assert (f->bb[pred].next[0] >= 0 && f->bb[pred].next[0] != BBID_END);
605 879 markom 
  switch (type) {
606 
  case 0:
607 
    if (f->bb[pred].next[0] == succ) f->bb[pred].next[0] = f->bb[succ].next[0];
608 
    if (f->bb[pred].next[1] == succ) f->bb[pred].next[1] = f->bb[succ].next[0];
609 
    assert (f->bb[succ].next[1] < 0);
610 
    break;
611 
  case 1:
612 
    f->bb[pred].next[0] = f->bb[succ].next[0];
613 
    f->bb[pred].next[1] = f->bb[succ].next[1];
614 
    break;
615 905 markom 
  case 2:
616 
    f->bb[pred].next[0] = f->bb[succ].next[0];
617 
    f->bb[pred].next[1] = f->bb[succ].next[1];
618 
    break;
619 879 markom 
  }
620 924 markom 
  if (f->bb[pred].next[0] < 0) f->bb[pred].next[0] = f->bb[pred].next[1];
621 879 markom 
  if (f->bb[pred].next[0] == f->bb[pred].next[1]) f->bb[pred].next[1] = -1;
622 927 markom 
 
623 
  /* We just did something stupid -- we joined two predecessors into one;
624 
     succ may need the information from which block we came.  We will repair
625 
     this by converting LRBB to CMOV */
626 930 markom 
  for (j = 0; j < 2; j++) {
627 
    int nb = f->bb[pred].next[j];
628 927 markom 
    int t;
629 
 
630 
    /* check just valid connections */
631 
    if (nb < 0 || nb == BBID_END) continue;
632 
 
633 
    /* check type */
634 
    if (f->bb[nb].prev[0] == pred && f->bb[nb].prev[1] == succ) t = 1;
635 
    else if (f->bb[nb].prev[1] == pred && f->bb[nb].prev[0] == succ) t = 0;
636 
    else continue;
637 
 
638 930 markom 
    /* check all LRBB instructions.  */
639 
    for (i = 0; i < f->bb[nb].ninsn; i++)
640 
      if (f->bb[nb].insn[i].index == II_LRBB) {
641 
        cuc_insn *lrbb =&f->bb[nb].insn[i];
642 
        change_insn_type (lrbb, II_CMOV);
643 
        lrbb->op[1] = t; lrbb->opt[1] = OPT_CONST;
644 
        lrbb->op[2] = 1 - t; lrbb->opt[2] = OPT_CONST;
645 
        lrbb->op[3] = cond_op; lrbb->opt[3] = cond_opt;
646 
        lrbb->type |= IT_COND;
647 
      }
648 927 markom 
  }
649 
 
650 879 markom 
  f->bb[succ].type = BB_DEAD;
651 997 markom 
  //PRINTF (" %x %x %x %x %x\n", f->bb[pred].next[0], f->bb[pred].next[1], f->bb[succ].next[0], f->bb[succ].next[1], insn[ninsn - 1].type);
652 905 markom 
  /* remove branch instruction, if there is only one successor */
653 933 markom 
  if (f->bb[pred].next[1] < 0 && ninsn > 0 && insn[ninsn - 1].type & IT_BRANCH) {
654 925 markom 
    assert (f->bb[pred].next[0] != pred); /* end BB, loop should not be possible */
655 905 markom 
    change_insn_type (&insn[ninsn - 1], II_NOP);
656 925 markom 
  }
657 879 markom 
 
658 
  /* Set max count */
659 
  if (f->bb[pred].cnt < f->bb[succ].cnt) f->bb[pred].cnt = f->bb[succ].cnt;
660 
  f->bb[pred].ninsn = ninsn;
661 905 markom 
  f->bb[succ].ninsn = 0;
662 879 markom 
  free (f->bb[pred].insn); f->bb[pred].insn = NULL;
663 
  free (f->bb[succ].insn); f->bb[succ].insn = NULL;
664 
  f->bb[pred].insn = insn;
665 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
666 
    if (f->bb[i].prev[0] == succ) f->bb[i].prev[0] = pred;
667 
    if (f->bb[i].prev[1] == succ) f->bb[i].prev[1] = pred;
668 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
669 
    for (j = 0; j < f->bb[i].ninsn; j++)
670 
      for (k = 0; k < MAX_OPERANDS; k++)
671 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
672 
          /* Check if we are referencing successor BB -> relocate to second part of
673 
             the new block */
674 
          if (REF_BB (f->bb[i].insn[j].op[k]) == succ) {
675 905 markom 
            int t = f->bb[i].insn[j].op[k];
676 
            int ndest = REF (pred, REF_I (t) + n1);
677 997 markom 
            //PRINTF ("%x: %x %x\n", REF(i, j), t, ndest);
678 879 markom 
 
679 905 markom 
            /* We've found a reference to succ. block, being removed, relocate */
680 
            f->bb[i].insn[j].op[k] = ndest;
681 879 markom 
          } else if (REF_BB(f->bb[i].insn[j].op[k]) == pred) {
682 
            if (i != pred && reloc[REF_I(f->bb[i].insn[j].op[k])] >= 0) {
683 
              f->bb[i].insn[j].op[k] = reloc[REF_I(f->bb[i].insn[j].op[k])];
684 
            }
685 
          }
686 
        }
687 
  }
688 
 
689 905 markom 
  if (cuc_debug) cuc_check (f);
690 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "join");
691 879 markom 
}
692 
 
693 
/* Optimize basic blocks */
694 931 markom 
int optimize_bb (cuc_func *f)
695 879 markom 
{
696 931 markom 
  int modified = 0;
697 879 markom 
  int i, j;
698 
remove_lrbb:
699 
  /* we can remove lrbb instructions from blocks with just one predecessor */
700 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
701 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* exactly one predecessor */
702 
      for (j = 0; j < f->bb[i].ninsn; j++)
703 
        if (f->bb[i].insn[j].index == II_LRBB) {
704 
          cuc_insn *t;
705 925 markom 
          cucdebug (4, "-lrbb %x.%x\n", i, j);
706 879 markom 
 
707 
          /* Change to add LRBB, 0, 0 */
708 
          change_insn_type (&f->bb[i].insn[j], II_ADD);
709 
          f->bb[i].insn[j].type &= ~IT_VOLATILE;
710 
          f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_CONST;
711 
          f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = 0; /* always use left block */
712 
          f->bb[i].insn[j].opt[3] = OPT_NONE;
713 931 markom 
          modified = 1;
714 934 markom 
          if (f->bb[i].prev[0] != BBID_START && f->bb[f->bb[i].prev[0]].ninsn > 0) {
715 932 markom 
            t = &f->bb[f->bb[i].prev[0]].insn[f->bb[f->bb[i].prev[0]].ninsn - 1];
716 879 markom 
 
717 932 markom 
            /* If the predecessor still has a conditional jump instruction, we must be careful.
718 
               If next[0] == next[1] join them. Now we will link lrbb and correct the situation */
719 
            if (t->type & IT_BRANCH) { /* We must set a reference to branch result */
720 
              f->bb[i].insn[j].opt[1] = t->opt[1];
721 
              f->bb[i].insn[j].op[1] = t->op[1];
722 
              /* sometimes branch is not needed anymore */
723 
              if (f->bb[f->bb[i].prev[0]].next[1] < 0) change_insn_type (t, II_NOP);
724 
            }
725 879 markom 
          }
726 
        }
727 
    }
728 
  }
729 
 
730 927 markom 
  /* Ordering of joining types is cruical -- we should concat all directly connected BBs
731 
     together first, so when we do a type != 1 joining, we can remove LRBB, directly by
732 
     looking at number of its predeccessors */
733 879 markom 
 
734 
  /* Type 1 joining
735 
     1. link between pred & succ
736 
     2. no other pred's successors
737 925 markom 
     3. no other succ's predecessors, except if pred has max one */
738 932 markom 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
739 
    int p = f->bb[i].prev[0];
740 
    if (p < 0 || p == BBID_START) continue;
741 925 markom 
    /* one successor and max sum of 3 predecessors */
742 932 markom 
    if (f->bb[p].next[0] >= 0 && f->bb[p].next[1] < 0
743 
     && (f->bb[p].prev[1] < 0 || f->bb[i].prev[1] < 0)) {
744 925 markom 
      /* First we will move all predecessors from succ to pred, and then we will do
745 
         real type 1 joining */
746 932 markom 
      if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START) {
747 
        int p1 = f->bb[i].prev[1];
748 925 markom 
        /* joining is surely not worth another extra memory access */
749 932 markom 
        if (f->bb[p].nmemory) continue;
750 
        if (f->bb[p].prev[0] >= 0) {
751 
           assert (f->bb[p].prev[1] < 0);
752 
           f->bb[p].prev[1] = p1;
753 
        } else f->bb[p].prev[0] = p1;
754 
        if (f->bb[p1].next[0] == i) f->bb[p1].next[0] = p;
755 
        else if (f->bb[p1].next[1] == i) f->bb[p1].next[1] = p;
756 925 markom 
        else assert (0);
757 
        f->bb[i].prev[1] = -1;
758 
      }
759 932 markom 
      assert (p >= 0 && f->bb[i].prev[1] < 0); /* one predecessor */
760 
      join_bb (f, p, i, 1);
761 931 markom 
      modified = 1;
762 879 markom 
      goto remove_lrbb;
763 
    }
764 932 markom 
  }
765 927 markom 
 
766 
  /* Type 0 joining
767 
     1. link between pred & succ
768 
     2. no memory accesses in succ
769 
     3. optional pred's second successors
770 
     4. max. one succ's successors */
771 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
772 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START
773 
     && f->bb[i].prev[1] < 0 /* one predecessor */
774 927 markom 
     && f->bb[i].next[1] < 0 /* max. one successor */
775 
     && f->bb[i].nmemory == 0) {                  /* and no memory acceses */
776 
      join_bb (f, f->bb[i].prev[0], i, 0);
777 931 markom 
      modified = 1;
778 927 markom 
      goto remove_lrbb;
779 
    }
780 879 markom 
 
781 
  /* Type 2 joining
782 
     1. link between pred & succ
783 
     2. succ has exactly one predeccessor
784 
     3. pred & succ share common successor
785 
     4. optional succ's second successor */
786 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
787 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* one predecessor */
788 
      int p = f->bb[i].prev[0];
789 932 markom 
      if (p == BBID_START) continue;
790 905 markom 
#if 0 /* not yet supported */
791 
      if (f->bb[p].next[0] == i
792 
       && (f->bb[i].next[1] == f->bb[p].next[1]
793 
        || f->bb[i].next[1] == f->bb[p].next[0])) {
794 
        join_bb (f, p, i, 2);
795 897 markom 
        goto remove_lrbb;
796 
      }
797 905 markom 
#endif
798 
      if (f->bb[p].next[1] == i
799 
       && (f->bb[i].next[0] == f->bb[p].next[1]
800 
        || f->bb[i].next[0] == f->bb[p].next[0])) {
801 
        join_bb (f, p, i, 2);
802 931 markom 
        modified = 1;
803 905 markom 
        goto remove_lrbb;
804 
      }
805 879 markom 
    }
806 1042 markom 
 
807 
  /* BB simplify:
808 
     1. a block has exactly 2 successors A and B
809 
     2. A has exactly one successor -- B
810 
     3. A has no memory accesses
811 
     to:
812 
     flow always goes though A, LRBB is replaced by current block conditional
813 
    */
814 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
815 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END
816 
      && f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END) {
817 
      int a = f->bb[i].next[0];
818 
      int b = f->bb[i].next[1];
819 
      int neg = 0;
820 
      /* Exchange? */
821 
      if (f->bb[b].next[0] == a && f->bb[b].next[1] < 0) {
822 
        int t = a;
823 
        a = b;
824 
        b = t;
825 
        neg = 1;
826 
      }
827 
      /* Do the simplification if possible */
828 
      if (f->bb[a].next[0] == b && f->bb[a].next[1] < 0
829 
       && f->bb[a].nmemory == 0) {
830 
        simplify_bb (f, i, a, b, neg);
831 
        modified = 1;
832 
        goto remove_lrbb;
833 
      }
834 
    }
835 
 
836 931 markom 
  return modified;
837 879 markom 
}
838 
 
839 
/* Removes BBs marked as dead */
840 931 markom 
int remove_dead_bb (cuc_func *f)
841 879 markom 
{
842 
  int i, j, k, d = 0;
843 
 
844 
  for (i = 0; i < f->num_bb; i++) if (f->bb[i].type & BB_DEAD) {
845 
    if (f->bb[i].insn) free (f->bb[i].insn);
846 
    f->bb[i].insn = NULL;
847 
    reloc[i] = -1;
848 
  } else {
849 
    reloc[i] = d;
850 
    f->bb[d++] = f->bb[i];
851 
  }
852 931 markom 
  if (f->num_bb == d) return 0;
853 879 markom 
  f->num_bb = d;
854 
 
855 
  /* relocate initial blocks */
856 
  for (i = 0; i < f->num_init_bb; i++)
857 
    f->init_bb_reloc[i] = reloc[f->init_bb_reloc[i]];
858 
 
859 
  /* repair references */
860 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
861 973 markom 
          cucdebug (5, "%x %x %x %x %x\n", i, f->bb[i].prev[0], f->bb[i].prev[1], f->bb[i].next[0], f->bb[i].next[1]);
862 925 markom 
          fflush (stdout);
863 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START)
864 
      assert ((f->bb[i].prev[0] = reloc[f->bb[i].prev[0]]) >= 0);
865 
    if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START)
866 
      assert ((f->bb[i].prev[1] = reloc[f->bb[i].prev[1]]) >= 0);
867 925 markom 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END)
868 
      assert ((f->bb[i].next[0] = reloc[f->bb[i].next[0]]) >= 0);
869 
    if (f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END)
870 
      assert ((f->bb[i].next[1] = reloc[f->bb[i].next[1]]) >= 0);
871 879 markom 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
872 
    if (f->bb[i].next[0] == f->bb[i].next[1]) f->bb[i].next[1] = -1;
873 
 
874 
    for (j = 0; j < f->bb[i].ninsn; j++)
875 
      for (k = 0; k < MAX_OPERANDS; k++)
876 925 markom 
        if (f->bb[i].insn[j].opt[k] & OPT_BB && (signed)f->bb[i].insn[j].op[k] >= 0) {
877 
          if (f->bb[i].insn[j].op[k] != BBID_END)
878 
            assert ((f->bb[i].insn[j].op[k] = reloc[f->bb[i].insn[j].op[k]]) >= 0);
879 
        } else if (f->bb[i].insn[j].opt[k] & OPT_REF) {
880 879 markom 
          int t = f->bb[i].insn[j].op[k];
881 
          assert (reloc[REF_BB(t)] >= 0);
882 
          f->bb[i].insn[j].op[k] = REF (reloc[REF_BB(t)], REF_I (t));
883 
        }
884 
  }
885 931 markom 
  return 1;
886 879 markom 
}
887 
 
888 
/* Recursive calculation of dependencies */
889 
static int reg_dep_rec (cuc_func *f, int cur)
890 
{
891 
  int i, j;
892 
  cuc_insn *insn = f->bb[cur].insn;
893 
 
894 997 markom 
  //PRINTF ("\n %i", cur);
895 879 markom 
  /* Spread only, do not loop */
896 
  if (f->bb[cur].tmp) return;
897 
  f->bb[cur].tmp = 1;
898 997 markom 
  //PRINTF ("!   ");
899 879 markom 
 
900 
  for (i = 0; i < f->bb[cur].ninsn; i++) {
901 
    /* Check for destination operand(s) */
902 
    for (j = 0; j < MAX_OPERANDS; j++) if (insn[i].opt[j] & OPT_DEST)
903 
      if ((insn[i].opt[j] & ~OPT_DEST) == OPT_REGISTER && (signed)insn[i].op[j] >= 0) {
904 997 markom 
        //PRINTF ("%i:%i,%x ", insn[i].op[j], i, REF (cur, i));
905 879 markom 
        assert (insn[i].op[j] > 0 && insn[i].op[j] < MAX_REGS); /* r0 should never be dest */
906 
        f->bb[cur].last_used_reg[insn[i].op[j]] = REF (cur, i);
907 
      }
908 
  }
909 
 
910 925 markom 
  if (f->bb[cur].next[0] >= 0 && f->bb[cur].next[0] != BBID_END)
911 
    reg_dep_rec (f, f->bb[cur].next[0]);
912 
  if (f->bb[cur].next[1] >= 0 && f->bb[cur].next[1] != BBID_END)
913 
    reg_dep_rec (f, f->bb[cur].next[1]);
914 879 markom 
}
915 
 
916 
/* Detect register dependencies */
917 
void reg_dep (cuc_func *f)
918 
{
919 
  int i, b, c;
920 
 
921 
  /* Set dead blocks */
922 
  for (b = 0; b < f->num_bb; b++) {
923 
    f->bb[b].tmp = 0;
924 
    for (i = 0; i < MAX_REGS; i++) f->bb[b].last_used_reg[i] = -1;
925 
  }
926 
 
927 
  /* Start with first block and set dependecies of all reachable blocks */
928 
  /* At the same time set last_used_regs */
929 
  reg_dep_rec (f, 0);
930 
 
931 
  for (i = 0; i < f->num_bb; i++)
932 
    if (f->bb[i].tmp) f->bb[i].tmp = 0;
933 
    else f->bb[i].type |= BB_DEAD;
934 
 
935 
  /* Detect loops; mark BBs where loops must be broken */
936 
  for (c = 0; c < f->num_bb; c++) {
937 
    int min = 3, minb;
938 
 
939 
    /* search though all non-visited for minimum number of unvisited predecessors */
940 
    for (b = 0; b < f->num_bb; b++) if (!f->bb[b].tmp) {
941 
      int tmp = 0;
942 932 markom 
      if (f->bb[b].prev[0] >= 0 && f->bb[b].prev[0] != BBID_START
943 
       && !f->bb[f->bb[b].prev[0]].tmp) tmp++;
944 
      if (f->bb[b].prev[1] >= 0 && f->bb[b].prev[1] != BBID_START
945 
       && !f->bb[f->bb[b].prev[1]].tmp) tmp++;
946 879 markom 
      if (tmp < min) {
947 
        minb = b;
948 
        min = tmp;
949 
        if (tmp == 0) break; /* We already have the best one */
950 
      }
951 
    }
952 
    b = minb;
953 
    f->bb[b].tmp = 1; /* Mark visited */
954 883 markom 
    cucdebug (3, "minb %i min %i\n", minb, min);
955 879 markom 
    if (min) { /* We just broke the loop */
956 
      f->bb[b].type |= BB_INLOOP;
957 
    }
958 
  }
959 
 
960 
  /* Set real predecessors in cmov instructions to previous blocks */
961 
  for (b = 0; b < f->num_bb; b++)
962 
    for (i = 1; i < MAX_REGS - 1; i++) {
963 
      int pa, pb;
964 
      assert (f->bb[b].insn[i].index ==  II_CMOV);
965 
      assert (f->bb[b].insn[i].opt[0] == OPT_REGISTER | OPT_DEST);
966 
      assert (f->bb[b].insn[i].op[0] == i);
967 932 markom 
      if (f->bb[b].prev[0] < 0 || f->bb[b].prev[0] == BBID_START) pa = -1;
968 879 markom 
      else pa = f->bb[f->bb[b].prev[0]].last_used_reg[i];
969 932 markom 
      if (f->bb[b].prev[1] < 0 || f->bb[b].prev[1] == BBID_START) pb = -1;
970 879 markom 
      else pb = f->bb[f->bb[b].prev[1]].last_used_reg[i];
971 
 
972 
      /* We do some very simple optimizations right away to make things more readable */
973 
      if (pa < 0 && pb < 0) {
974 
        /* Was not used at all */
975 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
976 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
977 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
978 
      } else if (pa < 0) {
979 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
980 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
981 
        f->bb[b].insn[i].op[1] = pb; f->bb[b].insn[i].opt[1] = OPT_REF;
982 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
983 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
984 
      } else if (pb < 0) {
985 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
986 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
987 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
988 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
989 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
990 
      } else {
991 
        int t = REF (b, 0); /* lrbb should be first instruction */
992 
        assert (f->INSN(t).index == II_LRBB);
993 
 
994 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
995 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
996 
 
997 
        f->bb[b].insn[i].op[2] = pb; f->bb[b].insn[i].opt[2] = OPT_REF;
998 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
999 
 
1000 
        /* Update op[3] -- flag register */
1001 
        assert (f->bb[b].insn[i].opt[3] == OPT_REGISTER);
1002 
        assert (f->bb[b].insn[i].op[3] == LRBB_REG);
1003 
        assert (t >= 0);
1004 
        f->bb[b].insn[i].opt[3] = OPT_REF; /* Convert already used regs to references */
1005 
        f->bb[b].insn[i].op[3] = t;
1006 
        assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1007 
      }
1008 
    }
1009 
 
1010 
  /* assign register references */
1011 
  for (b = 0; b < f->num_bb; b++) {
1012 
    /* rebuild last used reg array */
1013 
    f->bb[b].last_used_reg[0] = -1;
1014 
    if (f->bb[b].insn[0].index == II_LRBB) f->bb[b].last_used_reg[LRBB_REG] = 0;
1015 
    else f->bb[b].last_used_reg[LRBB_REG] = -1;
1016 
 
1017 
    for (i = 1; i < MAX_REGS - 1; i++)
1018 
      f->bb[b].last_used_reg[i] = -1;
1019 
 
1020 
    /* Create references */
1021 
    for (i = 0; i < f->bb[b].ninsn; i++) {
1022 
      int k;
1023 
      /* Check for source operands first */
1024 
      for (k = 0; k < MAX_OPERANDS; k++) {
1025 
        if (!(f->bb[b].insn[i].opt[k] & OPT_DEST))
1026 
        if (f->bb[b].insn[i].opt[k] & OPT_REGISTER) {
1027 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1028 
 
1029 
          if (f->bb[b].insn[i].op[k] == 0) { /* Convert r0 to const0 */
1030 
            f->bb[b].insn[i].opt[k] = OPT_CONST;
1031 
            f->bb[b].insn[i].op[k] = 0;
1032 
          } else if (t >= 0) {
1033 
            f->bb[b].insn[i].opt[k] = OPT_REF; /* Convert already used regs to references */
1034 
            f->bb[b].insn[i].op[k] = t;
1035 
            assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1036 
            //f->INSN(t).op[0] = -1;
1037 
          }
1038 
        } else if (f->bb[b].insn[i].opt[k] & OPT_REF) {
1039 
          //f->INSN(f->bb[b].insn[i].op[k]).op[0] = -1; /* Mark referenced */
1040 
          f->INSN(f->bb[b].insn[i].op[k]).type &= ~IT_UNUSED;
1041 
        }
1042 
      }
1043 
 
1044 
      /* Now check for destination operand(s) */
1045 
      for (k = 0; k < MAX_OPERANDS; k++) if (f->bb[b].insn[i].opt[k] & OPT_DEST)
1046 
        if ((f->bb[b].insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
1047 
          && (int)f->bb[b].insn[i].op[k] >= 0) {
1048 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1049 
          assert (f->bb[b].insn[i].op[k] != 0); /* r0 should never be dest */
1050 
          f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]] = REF (b, i);
1051 
        }
1052 
    }
1053 
  }
1054 
 
1055 
  /* Remove all unused lrbb */
1056 
  for (b = 0; b < f->num_bb; b++)
1057 
    for (i = 0; i < f->bb[b].ninsn; i++)
1058 
      if (f->bb[b].insn[i].type & IT_UNUSED) change_insn_type (&f->bb[b].insn[i], II_NOP);
1059 
 
1060 
  /* SSAs with final register value are marked as outputs */
1061 
  assert (f->bb[f->num_bb - 1].type & BB_END);
1062 939 markom 
  for (i = 0; i < MAX_REGS; i++) if (!caller_saved[i]) {
1063 879 markom 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1064 
    /* Mark them volatile, so optimizer does not remove them */
1065 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1066 
  }
1067 939 markom 
  {
1068 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1069 
    /* Mark them volatile, so optimizer does not remove them */
1070 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1071 
  }
1072 879 markom 
}
1073 
 
1074 897 markom 
/* split the BB, based on the group numbers in .tmp */
1075 
void expand_bb (cuc_func *f, int b)
1076 
{
1077 
  int n = f->num_bb;
1078 
  int mg = 0;
1079 
  int b1, i, j;
1080 
 
1081 
  for (i = 0; i < f->bb[b].ninsn; i++)
1082 
    if (f->bb[b].insn[i].tmp > mg) mg = f->bb[b].insn[i].tmp;
1083 
 
1084 
  /* Create copies */
1085 
  for (b1 = 1; b1 <= mg; b1++) {
1086 
    assert (f->num_bb < MAX_BB);
1087 
    cpy_bb (&f->bb[f->num_bb], &f->bb[b]);
1088 
    f->num_bb++;
1089 
  }
1090 
 
1091 
  /* Relocate */
1092 
  for (b1 = 0; b1 < f->num_bb; b1++)
1093 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1094 
      dep_list *d = f->bb[b1].insn[i].dep;
1095 
      for (j = 0; j < MAX_OPERANDS; j++)
1096 
        if (f->bb[b1].insn[i].opt[j] & OPT_REF) {
1097 
          int t = f->bb[b1].insn[i].op[j];
1098 
          if (REF_BB(t) == b && f->INSN(t).tmp != 0)
1099 
            f->bb[b1].insn[i].op[j] = REF (n + f->INSN(t).tmp - 1, REF_I(t));
1100 
        }
1101 
      while (d) {
1102 
        if (REF_BB (d->ref) == b && f->INSN(d->ref).tmp != 0)
1103 
          d->ref = REF (n + f->INSN(d->ref).tmp - 1, REF_I(d->ref));
1104 
        d = d->next;
1105 
      }
1106 
    }
1107 
 
1108 
  /* Delete unused instructions */
1109 
  for (j = 0; j <= mg; j++) {
1110 
    if (j == 0) b1 = b;
1111 
    else b1 = n + j - 1;
1112 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1113 
      if (f->bb[b1].insn[i].tmp != j)
1114 
        change_insn_type (&f->bb[b1].insn[i], II_NOP);
1115 
      f->bb[b1].insn[i].tmp = 0;
1116 
    }
1117 
    if (j < mg) {
1118 
      f->bb[b1].next[0] = n + j;
1119 
      f->bb[b1].next[1] = -1;
1120 
      f->bb[n + j].prev[0] = b1;
1121 
      f->bb[n + j].prev[1] = -1;
1122 
    } else {
1123 
      i = f->bb[b1].next[0];
1124 
      f->bb[n + j].prev[0] = j == 1 ? b : b1 - 1;
1125 
      f->bb[n + j].prev[1] = -1;
1126 925 markom 
      if (i >= 0 && i != BBID_END) {
1127 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1128 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1129 
      }
1130 
      i = f->bb[b1].next[1];
1131 925 markom 
      if (i >= 0 && i != BBID_END) {
1132 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1133 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1134 
      }
1135 
    }
1136 
  }
1137 
}
1138 
 
1139 879 markom 
/* Scans sequence of BBs and set bb[].cnt */
1140 
void generate_bb_seq (cuc_func *f, char *mp_filename, char *bb_filename)
1141 
{
1142 
  FILE *fi, *fo;
1143 
  struct mprofentry_struct *buf;
1144 
  const int bufsize = 256;
1145 
  unsigned long *bb_start;
1146 
  unsigned long *bb_end;
1147 
  int b, i, r;
1148 
  int curbb, prevbb = -1;
1149 
  unsigned long addr = -1;
1150 
  unsigned long prevaddr = -1;
1151 897 markom 
  int mssum = 0;
1152 
  int mlsum = 0;
1153 
  int mscnt = 0;
1154 
  int mlcnt = 0;
1155 940 markom 
  int reopened = 0;
1156 879 markom 
 
1157 940 markom 
  /* Use already opened stream? */
1158 
  if (runtime.sim.fmprof) {
1159 
    fi = runtime.sim.fmprof;
1160 
    reopened = 1;
1161 
    rewind (fi);
1162 
  } else assert (fi = fopen (mp_filename, "rb"));
1163 879 markom 
  assert (fo = fopen (bb_filename, "wb+"));
1164 
 
1165 
  assert (bb_start = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1166 
  assert (bb_end = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1167 
  for (b = 0; b < f->num_bb; b++) {
1168 
    bb_start[b] = f->start_addr + f->bb[b].first * 4;
1169 
    bb_end[b] = f->start_addr + f->bb[b].last * 4;
1170 997 markom 
    //PRINTF ("%i %x %x\n", b, bb_start[b], bb_end[b]);
1171 879 markom 
    f->bb[0].cnt = 0;
1172 
  }
1173 
 
1174 
  buf = (struct mprofentry_struct *) malloc (sizeof (struct mprofentry_struct) * bufsize);
1175 
  assert (buf);
1176 
 
1177 997 markom 
  //PRINTF ("BBSEQ:\n");
1178 879 markom 
  do {
1179 
    r = fread (buf, sizeof (struct mprofentry_struct), bufsize, fi);
1180 997 markom 
    //PRINTF ("r%i : ", r);
1181 879 markom 
    for (i = 0; i < r; i++) {
1182 
      if (buf[i].type & MPROF_FETCH) {
1183 997 markom 
        //PRINTF ("%x, ", buf[i].addr);
1184 879 markom 
        if (buf[i].addr >= f->start_addr && buf[i].addr <= f->end_addr) {
1185 
          assert (buf[i].type & MPROF_32);
1186 
          prevaddr = addr;
1187 
          addr = buf[i].addr;
1188 
          for (b = 0; b < f->num_bb; b++)
1189 
            if (bb_start[b] <= addr && addr <= bb_end[b]) break;
1190 
          assert (b < f->num_bb);
1191 
          curbb = b;
1192 
          if (prevaddr + 4 != addr) prevbb = -1;
1193 
        } else curbb = -1;
1194 
 
1195 
#warning TODO: do not count interrupts
1196 
        if (curbb != prevbb && curbb >= 0) {
1197 
          fwrite (&curbb, sizeof (unsigned long), 1, fo);
1198 997 markom 
          //PRINTF (" [%i] ", curbb);
1199 879 markom 
          f->bb[curbb].cnt++;
1200 
          prevbb = curbb;
1201 
        }
1202 897 markom 
      } else {
1203 
        if (verify_memoryarea(buf[i].addr))
1204 
          if (buf[i].type & MPROF_WRITE) mscnt++, mssum += cur_area->delayw;
1205 1041 markom 
          else mlcnt++, mlsum += cur_area->delayr;
1206 879 markom 
      }
1207 
    }
1208 997 markom 
    //PRINTF ("\n");
1209 879 markom 
  } while (r == bufsize);
1210 997 markom 
  //PRINTF ("\n");
1211 879 markom 
 
1212 897 markom 
  runtime.cuc.mdelay[0] = (1. * mlsum) / mlcnt;
1213 1041 markom 
  runtime.cuc.mdelay[1] = (1. * mssum) / mscnt;
1214 897 markom 
  runtime.cuc.mdelay[2] = runtime.cuc.mdelay[3] = 1;
1215 883 markom 
  f->num_runs = f->bb[0].cnt;
1216 940 markom 
  if (!reopened) fclose (fi);
1217 879 markom 
  fclose (fo);
1218 
  free (buf);
1219 
  free (bb_end);
1220 
  free (bb_start);
1221 
 
1222 
  /* Initialize basic block relocations */
1223 
  f->num_init_bb = f->num_bb;
1224 997 markom 
  //PRINTF ("num_init_bb = %i\n", f->num_init_bb);
1225 879 markom 
  assert (f->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
1226 
  for (b = 0; b < f->num_init_bb; b++) f->init_bb_reloc[b] = b;
1227 
}
1228 
 
1229 
/* Scans sequence of BBs and set counts for pre/unrolled loop for BB b */
1230 
void count_bb_seq (cuc_func *f, int b, char *bb_filename, int *counts, int preroll, int unroll)
1231 
{
1232 
  FILE *fi;
1233 
  const int bufsize = 256;
1234 
  int i, r;
1235 
  int *buf;
1236 
  int cnt = 0;
1237 
  int times = preroll - 1 + unroll;
1238 
 
1239 
  assert (fi = fopen (bb_filename, "rb"));
1240 
  for (i = 0; i < times; i++) counts[i] = 0;
1241 
  assert (buf = (int *) malloc (sizeof (int) * bufsize));
1242 
 
1243 
  do {
1244 
    r = fread (buf, sizeof (int), bufsize, fi);
1245 
    for (i = 0; i < r; i++) {
1246 
      /* count consecutive acesses */
1247 
      if (f->init_bb_reloc[buf[i]] == b) {
1248 
        counts[cnt]++;
1249 
        if (++cnt >= times) cnt = preroll - 1;
1250 
      } else cnt = 0;
1251 
    }
1252 
  } while (r == bufsize);
1253 
 
1254 
  log ("Counts %i,%i :", preroll, unroll);
1255 
  for (i = 0; i < times; i++) log ("%x ", counts[i]);
1256 
  log ("\n");
1257 
 
1258 
  fclose (fi);
1259 
  free (buf);
1260 
}
1261 
 
1262 
/* relocate all accesses inside of BB b to back/fwd */
1263 
static void relocate_bb (cuc_bb *bb, int b, int back, int fwd)
1264 
{
1265 
  int i, j;
1266 
  for (i = 0; i < bb->ninsn; i++)
1267 
    for (j = 0; j < MAX_OPERANDS; j++)
1268 
      if (bb->insn[i].opt[j] & OPT_REF
1269 
       && REF_BB (bb->insn[i].op[j]) == b) {
1270 
        int t = REF_I (bb->insn[i].op[j]);
1271 
        if (t < i) bb->insn[i].op[j] = REF (back, t);
1272 
        else bb->insn[i].op[j] = REF (fwd, t);
1273 
      }
1274 
}
1275 
 
1276 
/* Unroll loop b unroll times and return new function. Original
1277 
   function is unmodified. */
1278 
static cuc_func *unroll_loop (cuc_func *f, int b, int unroll)
1279 
{
1280 
  int b1, t, i, j, prevb, prevart_b;
1281 
  cuc_func *n = dup_func (f);
1282 
  cuc_bb *ob = &f->bb[b];
1283 
  cuc_insn *ii;
1284 
 
1285 
  assert (unroll > 1);
1286 997 markom 
  //PRINTF ("unroll BB%i x %i (num_bb %i)\n", b, unroll, n->num_bb);
1287 879 markom 
  unroll--;
1288 
  assert (n->num_bb + unroll * 2 < MAX_BB);
1289 
 
1290 
  prevb = b;
1291 
  prevart_b = b;
1292 
  /* Duplicate the BB */
1293 
  for (t = 0; t < unroll; t++) {
1294 
    cuc_bb *pb = &n->bb[prevart_b];
1295 
    /* Add new block and set links */
1296 
    b1 = n->num_bb++;
1297 
    cpy_bb (&n->bb[b1], ob);
1298 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1299 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1300 
 
1301 
    /* Set predecessor's successor */
1302 
    if (n->bb[prevb].next[0] == b) {
1303 
      n->bb[prevb].next[0] = b1;
1304 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1305 
      else pb->next[1] = b1 + 1;
1306 
      n->bb[b1].next[1] = b1 + 1;
1307 
    } else if (n->bb[prevb].next[1] == b) {
1308 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1309 
      else pb->next[1] = b1 + 1;
1310 
      n->bb[b1].next[0] = b1 + 1;
1311 
      n->bb[prevb].next[1] = b1;
1312 
    } else assert (0);
1313 
 
1314 
    /* Set predecessor */
1315 
    n->bb[b1].prev[0] = prevb;
1316 
    n->bb[b1].prev[1] = -1;
1317 
 
1318 
    /* Relocate backward references to current instance and forward references
1319 
       to previous one */
1320 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1321 
 
1322 
    /* add artificial block, just to join accesses */
1323 
    b1 = n->num_bb++;
1324 
    cpy_bb (&n->bb[b1], ob);
1325 
    n->bb[b1].cnt = 0;
1326 
 
1327 
    for (i = 0; i < ob->ninsn - 1; i++) {
1328 
      ii = &n->bb[b1].insn[i];
1329 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1330 
        change_insn_type (ii, II_CMOV);
1331 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1332 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1333 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1334 
 
1335 
        /* Take left one, if we should have finished the first iteration*/
1336 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1337 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1338 
        } else {
1339 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1340 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1341 
        }
1342 
        ii->dep = NULL;
1343 930 markom 
        ii->type = ob->insn[i].type & IT_COND;
1344 879 markom 
      } else {
1345 
        change_insn_type (ii, II_NOP);
1346 
      }
1347 
    }
1348 
 
1349 928 markom 
    /* Add conditional or instruction at the end, prioritizing flags */
1350 879 markom 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1351 928 markom 
    change_insn_type (ii, II_CMOV);
1352 879 markom 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1353 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1354 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1355 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1356 
    } else {
1357 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1358 
      ii->opt[1] = OPT_REF;
1359 
    }
1360 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1361 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1362 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1363 
    } else {
1364 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1365 
      ii->opt[2] = OPT_REF;
1366 
    }
1367 928 markom 
    /* {z = x || y;} is same as {z = x ? x : y;} */
1368 
    ii->op[3] = ii->op[1]; ii->opt[3] = ii->opt[1];
1369 933 markom 
    ii->type = IT_COND;
1370 879 markom 
 
1371 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1372 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1373 
    n->bb[b1].prev[0] = prevart_b;
1374 
    n->bb[b1].prev[1] = b1 - 1;
1375 925 markom 
    n->bb[b1].next[0] = -1;
1376 879 markom 
    n->bb[b1].next[1] = -1;
1377 
 
1378 
    prevb = b1 - 1;
1379 
    prevart_b = b1;
1380 
  }
1381 
  if (ob->type & BB_END) {
1382 
    n->bb[prevart_b].type |= BB_END;
1383 
    n->bb[b].type &= ~BB_END;
1384 
  }
1385 
 
1386 925 markom 
  n->bb[prevart_b].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1387 879 markom 
  //print_cuc_bb (n, "unroll1");
1388 
  /* repair BB after loop, to point back to latest artificial BB */
1389 
  b1 = n->bb[prevart_b].next[0];
1390 925 markom 
  if (b1 >= 0 && b1 != BBID_END) {
1391 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1392 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1393 879 markom 
    else assert (0);
1394 
  }
1395 
 
1396 
  /* Relink back to start of the loop */
1397 
  /* Set predecessor's successor */
1398 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = b;
1399 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = b;
1400 
  else assert (0);
1401 
 
1402 
  /* Set predecessor */
1403 
  if (n->bb[b].prev[0] == b) n->bb[b].prev[0] = prevb;
1404 
  else if (n->bb[b].prev[1] == b) n->bb[b].prev[1] = prevb;
1405 
  else assert (0);
1406 
 
1407 
  //print_cuc_bb (n, "unroll2");
1408 
 
1409 
  /* Relocate backward references to current instance and forward references
1410 
     to previous one */
1411 
  relocate_bb (&n->bb[b], b, b, prevb);
1412 
 
1413 
  /* Relocate all other blocks to point to latest prevart_b */
1414 
  for (i = 0; i < f->num_bb; i++)
1415 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1416 
 
1417 
  return n;
1418 
}
1419 
 
1420 
/* Preroll loop b preroll times and return new function. Original
1421 
   function is unmodified. */
1422 
static cuc_func *preroll_loop (cuc_func *f, int b, int preroll)
1423 
{
1424 
  int b1, t, i, j, prevb, prevart_b;
1425 
  cuc_func *n = dup_func (f);
1426 
  cuc_bb *ob = &f->bb[b];
1427 
  cuc_insn *ii;
1428 
 
1429 
  assert (preroll > 1);
1430 997 markom 
  //PRINTF ("preroll BB%i x %i (num_bb %i)\n", b, preroll, n->num_bb);
1431 879 markom 
  preroll--;
1432 
  assert (n->num_bb + preroll * 2 < MAX_BB);
1433 
 
1434 
  prevb = b;
1435 
  prevart_b = b;
1436 
  /* Duplicate the BB */
1437 
  for (t = 0; t < preroll; t++) {
1438 
    cuc_bb *pb = &n->bb[prevart_b];
1439 
    /* Add new block and set links */
1440 
    b1 = n->num_bb++;
1441 
    cpy_bb (&n->bb[b1], ob);
1442 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1443 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1444 
 
1445 
    /* Set predecessor's successor */
1446 
    if (n->bb[prevb].next[0] == b) {
1447 
      n->bb[prevb].next[0] = b1;
1448 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1449 
      else pb->next[1] = b1 + 1;
1450 
      n->bb[b1].next[1] = b1 + 1;
1451 
    } else if (n->bb[prevb].next[1] == b) {
1452 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1453 
      else pb->next[1] = b1 + 1;
1454 
      n->bb[b1].next[0] = b1 + 1;
1455 
      n->bb[prevb].next[1] = b1;
1456 
    } else assert (0);
1457 
 
1458 
    /* Set predecessor */
1459 
    n->bb[b1].prev[0] = prevb;
1460 
    n->bb[b1].prev[1] = -1;
1461 
 
1462 
    /* Relocate backward references to current instance and forward references
1463 
       to previous one */
1464 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1465 
 
1466 
    /* add artificial block, just to join accesses */
1467 
    b1 = n->num_bb++;
1468 
    cpy_bb (&n->bb[b1], ob);
1469 
    n->bb[b1].cnt = 0;
1470 
 
1471 
    for (i = 0; i < ob->ninsn - 1; i++) {
1472 
      ii = &n->bb[b1].insn[i];
1473 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1474 
        change_insn_type (ii, II_CMOV);
1475 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1476 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1477 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1478 
 
1479 
        /* Take left one, if we should have finished the first iteration*/
1480 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1481 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1482 
        } else {
1483 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1484 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1485 
        }
1486 
        ii->dep = NULL;
1487 930 markom 
        ii->type = ob->insn[i].type & IT_COND;
1488 879 markom 
      } else {
1489 
        change_insn_type (ii, II_NOP);
1490 
      }
1491 
    }
1492 
 
1493 928 markom 
    /* Add conditional or instruction at the end, prioritizing flags */
1494 879 markom 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1495 928 markom 
    change_insn_type (ii, II_CMOV);
1496 879 markom 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1497 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1498 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1499 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1500 
    } else {
1501 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1502 
      ii->opt[1] = OPT_REF;
1503 
    }
1504 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1505 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1506 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1507 
    } else {
1508 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1509 
      ii->opt[2] = OPT_REF;
1510 
    }
1511 928 markom 
    /* {z = x || y;} is same as {z = x ? x : y;} */
1512 
    ii->op[3] = ii->op[1]; ii->opt[3] = ii->opt[1];
1513 933 markom 
    ii->type = IT_COND;
1514 879 markom 
 
1515 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1516 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1517 
    n->bb[b1].prev[0] = prevart_b;
1518 
    n->bb[b1].prev[1] = b1 - 1;
1519 925 markom 
    n->bb[b1].next[0] = -1;
1520 879 markom 
    n->bb[b1].next[1] = -1;
1521 
 
1522 
    prevb = b1 - 1;
1523 
    prevart_b = b1;
1524 
  }
1525 
  if (ob->type & BB_END) {
1526 
    n->bb[prevart_b].type |= BB_END;
1527 
    n->bb[b].type &= ~BB_END;
1528 
  }
1529 925 markom 
  n->bb[prevart_b].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1530 879 markom 
 
1531 
  //print_cuc_bb (n, "preroll1");
1532 
  /* repair BB after loop, to point back to latest artificial BB */
1533 
  b1 = n->bb[prevart_b].next[0];
1534 925 markom 
  if (b1 >= 0 && b1 != BBID_END) {
1535 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1536 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1537 879 markom 
    else assert (0);
1538 
  }
1539 
 
1540 
  /* Relink to itself */
1541 
  /* Set predecessor's successor */
1542 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = prevb;
1543 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = prevb;
1544 
  else assert (0);
1545 
  n->bb[prevb].prev[1] = prevb;
1546 
 
1547 
  if (n->bb[b].prev[0] == b) {
1548 
    n->bb[b].prev[0] = n->bb[b].prev[1];
1549 
    n->bb[b].prev[1] = -1;
1550 
  } else if (n->bb[b].prev[1] == b) {
1551 
    n->bb[b].prev[1] = -1;
1552 
  }
1553 
 
1554 
  //print_cuc_bb (n, "preroll2");
1555 
 
1556 
  /* Relocate backward references to current instance and forward references
1557 
     to previous one */
1558 
  relocate_bb (&n->bb[b], b, b, prevb);
1559 
 
1560 
  /* Relocate all other blocks to point to latest prevart_b */
1561 
  for (i = 0; i < f->num_bb; i++)
1562 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1563 
 
1564 
  return n;
1565 
}
1566 
 
1567 
/* Unroll loop b unroll times and return new function. Original
1568 
   function is unmodified. */
1569 
cuc_func *preunroll_loop (cuc_func *f, int b, int preroll, int unroll, char *bb_filename)
1570 
{
1571 
  int b1, i;
1572 
  cuc_func *n, *t;
1573 
  int *counts;
1574 
  int *bb_reloc;
1575 
 
1576 
  if (preroll > 1) {
1577 
    t = preroll_loop (f, b, preroll);
1578 
    b1 = t->num_bb - 2;
1579 
    if (unroll > 1) {
1580 
      //print_cuc_bb (t, "preunroll1");
1581 
      n = unroll_loop (t, b1, unroll);
1582 
      free_func (t);
1583 
    } else n = t;
1584 
  } else {
1585 
    b1 = b;
1586 897 markom 
    if (unroll > 1) n = unroll_loop (f, b1, unroll);
1587 
    else return dup_func (f);
1588 879 markom 
  }
1589 
 
1590 897 markom 
  /* Assign new counts to functions */
1591 879 markom 
  assert (counts = (int *)malloc (sizeof (int) * (preroll - 1 + unroll)));
1592 
  count_bb_seq (n, b, bb_filename, counts, preroll, unroll);
1593 
  for (i = 0; i < preroll - 1 + unroll; i++) {
1594 
    if (i == 0) b1 = b;
1595 
    else b1 = f->num_bb + (i - 1) * 2;
1596 
    n->bb[b1].cnt = counts[i];
1597 
  }
1598 
 
1599 
  //print_cuc_bb (n, "preunroll");
1600 
  free (counts);
1601 
  return n;
1602 
}
1603 933 markom 
 

powered by: WebSVN 2.1.0

© copyright 1999-2024 OpenCores.org, equivalent to Oliscience, all rights reserved. OpenCores®, registered trademark.