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

Subversion Repositories or1k

[/] [or1k/] [tags/] [rel-0-3-0-rc2/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 1042

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

powered by: WebSVN 2.1.0

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