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 1061

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 
    } else {
238 
      f->bb[i].next[0] = i + 1;
239 
      if (++f->bb[i + 1].tmp > 2) eb++;
240 
    }
241 
 
242 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_NEXT");
243 879 markom 
 
244 
  /* Build backward connections, but first insert artificial blocks
245 
   * to handle more than 2 connections */
246 883 markom 
  cucdebug (6, "artificial %i %i\n", f->num_bb, eb);
247 879 markom 
  end_bb = f->num_bb + eb;
248 
  for (i = f->num_bb - 1; i >= 0; i--) {
249 
    j = f->bb[i].tmp;
250 
    if (f->bb[i].tmp > 2) f->bb[i].tmp = -f->bb[i].tmp;
251 
    f->bb[--end_bb] = f->bb[i];
252 
    reloc[i] = end_bb;
253 
    while (j-- > 2) {
254 
      f->bb[--end_bb].first = f->bb[i].first;
255 
      f->bb[end_bb].last = -1;
256 
      f->bb[end_bb].next[0] = -1;
257 
      f->bb[end_bb].next[1] = -1;
258 
      f->bb[end_bb].tmp = 0;
259 
      f->bb[end_bb].cnt = f->bb[i].cnt;
260 
      f->bb[end_bb].ntim = 0;
261 
    }
262 
  }
263 
  f->num_bb += eb;
264 
 
265 
  /* relocate jump instructions */
266 
  for (i = 0; i < num_insn; i++)
267 
    for (j = 0; j < MAX_OPERANDS; j++)
268 
      if (insn[i].opt[j] & OPT_BB)
269 
        insn[i].op[j] = reloc[insn[i].op[j]];
270 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT-reloc");
271 879 markom 
  for (i = 0; i < f->num_bb; i++) {
272 
    if (f->bb[i].next[0] >= 0) {
273 
      int t = reloc[f->bb[i].next[0]];
274 
      if (f->bb[t].tmp < 0) {
275 
        f->bb[t].tmp = -f->bb[t].tmp;
276 
        t -= f->bb[t].tmp - 2;
277 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
278 
      f->bb[i].next[0] = t;
279 
    }
280 
    if (f->bb[i].next[1] >= 0) {
281 
      int t = reloc[f->bb[i].next[1]];
282 
      if (f->bb[t].tmp < 0) {
283 
        f->bb[t].tmp = -f->bb[t].tmp;
284 
        t -= f->bb[t].tmp - 2;
285 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
286 
      f->bb[i].next[1] = t;
287 
    }
288 
    /* artificial blocks do not have relocations, hardcode them */
289 
    if (f->bb[i].last < 0) f->bb[i].next[0] = i + 1;
290 
  }
291 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT");
292 879 markom 
 
293 
  /* Uncoditional branched do not continue to next block */
294 
  for (i = 0; i < f->num_bb; i++) {
295 
    cuc_insn *ii;
296 
    if (f->bb[i].last < 0) continue;
297 
    ii = &insn[f->bb[i].last];
298 
    /* Unconditional branch? */
299 
    if (ii->type & IT_BRANCH && ii->opt[1] & OPT_CONST) {
300 
      change_insn_type (ii, II_NOP);
301 1044 markom 
#if 0
302 879 markom 
      if (f->bb[i].next[1] == i + 1) f->bb[i].next[0] = f->bb[i].next[1];
303 1044 markom 
#endif
304 879 markom 
      f->bb[i].next[1] = -1;
305 
    }
306 1044 markom 
  }
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 1061 markom 
  f->bb[0].prev[0] = BBID_START;
334 932 markom 
 
335 925 markom 
  /* Add END marker */
336 1061 markom 
  assert (f->bb[f->num_bb - 1].next[0] < 0);
337 
  assert (f->bb[f->num_bb - 1].next[1] < 0);
338 
  f->bb[f->num_bb - 1].next[0] = BBID_END;
339 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_PREV");
340 879 markom 
}
341 
 
342 905 markom 
/* We do a quick check if there are some anomalies with references */
343 
void cuc_check (cuc_func *f)
344 
{
345 
  int i, j, k;
346 973 markom 
  cucdebug (1, "cuc_check\n");
347 905 markom 
  for (i = 0; i < f->num_bb; i++) {
348 930 markom 
    if (!f->bb[i].insn && f->bb[i].ninsn) goto err;
349 
    for (j = 0; j < f->bb[i].ninsn; j++) {
350 
      cuc_insn *ii = &f->bb[i].insn[j];
351 934 markom 
      if ((ii->index == II_CMOV || ii->index == II_ADD) && ii->type & IT_COND) {
352 930 markom 
        k = 0;
353 
        assert (ii->opt[k] & OPT_REGISTER);
354 
        if ((signed)ii->op[k] >= 0 && ii->op[k] != FLAG_REG && ii->op[k] != LRBB_REG) {
355 973 markom 
          cucdebug (1, "%x %x\n", ii->opt[0], ii->op[0]);
356 930 markom 
          goto err;
357 
        }
358 
      }
359 1061 markom 
      for (k = 0; k < MAX_OPERANDS; k++) {
360 930 markom 
        if (ii->opt[k] & OPT_REF) {
361 
          int t = ii->op[k];
362 
          if (REF_BB(t) >= f->num_bb || REF_I (t) >= f->bb[REF_BB(t)].ninsn
363 934 markom 
           || (ii->index == II_CMOV || ii->index == II_ADD) && (
364 930 markom 
                (f->INSN(t).type & IT_COND) != (ii->type & IT_COND) && k < 3
365 
              || !(f->INSN(t).type & IT_COND) && k == 3)) goto err;
366 905 markom 
        }
367 1061 markom 
        if (k && ii->opt[k] & OPT_DEST) {
368 
          cucdebug (1, "Destination only allowed for op0!\n");
369 
          goto err;
370 
        }
371 
      }
372 930 markom 
    }
373 905 markom 
  }
374 930 markom 
  return;
375 
err:
376 973 markom 
  cucdebug (1, "Anomaly detected at %x.%x[%i]\n", i, j, k);
377 930 markom 
  print_cuc_bb (f, "ANOMALY");
378 
  exit (1);
379 905 markom 
}
380 
 
381 879 markom 
/* Build basic blocks */
382 
void build_bb (cuc_func *f)
383 
{
384 
  int i, j, k;
385 
  for (i = 0; i < f->num_bb; i++) {
386 
    if (f->bb[i].last < 0) f->bb[i].ninsn = MAX_REGS - 1;
387 
    else f->bb[i].ninsn = f->bb[i].last - f->bb[i].first + 1 + MAX_REGS - 1;
388 
    assert (f->bb[i].ninsn >= MAX_REGS - 1);
389 
    f->bb[i].insn = (cuc_insn *) malloc (sizeof (cuc_insn) * f->bb[i].ninsn);
390 
    assert (f->bb[i].insn);
391 
    f->bb[i].nmemory = 0;
392 
    f->bb[i].unrolled = 1;
393 
 
394 
    /* Save space for conditional moves, exclude r0, place lrbb instead */
395 
    change_insn_type (&f->bb[i].insn[0], II_LRBB);
396 
    strcpy (f->bb[i].insn[0].disasm, "lrbb");
397 930 markom 
    f->bb[i].insn[0].type = IT_UNUSED | IT_COND;
398 879 markom 
    f->bb[i].insn[0].dep = NULL;
399 
    f->bb[i].insn[0].op[0] = LRBB_REG; f->bb[i].insn[0].opt[0] = OPT_REGISTER | OPT_DEST;
400 
    f->bb[i].insn[0].opt[1] = OPT_LRBB;
401 
    f->bb[i].insn[0].opt[2] = f->bb[i].insn[0].opt[3] = OPT_NONE;
402 
    for (j = 1; j < MAX_REGS - 1; j++) {
403 
      change_insn_type (&f->bb[i].insn[j], II_CMOV);
404 
      strcpy (f->bb[i].insn[j].disasm, "cmov");
405 930 markom 
      f->bb[i].insn[j].type = j == FLAG_REG || j == LRBB_REG ? IT_COND : 0;
406 879 markom 
      f->bb[i].insn[j].dep = NULL;
407 
      f->bb[i].insn[j].opt[0] = f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_REGISTER;
408 
      f->bb[i].insn[j].opt[0] |= OPT_DEST;
409 
      f->bb[i].insn[j].op[0] = f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = j;
410 
      f->bb[i].insn[j].op[3] = LRBB_REG; f->bb[i].insn[j].opt[3] = OPT_REGISTER;
411 
    }
412 897 markom 
 
413 
    /* Relocate instructions */
414 879 markom 
    for (j = MAX_REGS - 1; j < f->bb[i].ninsn; j++) {
415 
      f->bb[i].insn[j] = insn[f->bb[i].first + j - (MAX_REGS - 1)];
416 
      for (k = 0; k < MAX_OPERANDS; k++)
417 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
418 
          int b1;
419 
          for (b1 = 0; b1 < i; b1++)
420 898 markom 
            if (f->bb[b1].first <= (signed) f->bb[i].insn[j].op[k]
421 
              && (signed)f->bb[i].insn[j].op[k] <= f->bb[b1].last) break;
422 879 markom 
          assert (b1 < f->num_bb);
423 
          f->bb[i].insn[j].op[k] = REF (b1, f->bb[i].insn[j].op[k] - f->bb[b1].first + MAX_REGS - 1);
424 
        }
425 
      if (f->bb[i].insn[j].type & IT_MEMORY) f->bb[i].nmemory++;
426 
    }
427 
  }
428 905 markom 
  cuc_check (f);
429 879 markom 
}
430 
 
431 1042 markom 
/* Does simplification on blocks A, B, C:
432 
   A->B->C, A->C to just A->B->C */
433 
static void simplify_bb (cuc_func *f, int pred, int s1, int s2, int neg)
434 
{
435 
  cuc_insn *last;
436 
  int i;
437 
  if (cuc_debug >= 3) print_cuc_bb (f, "BEFORE_SIMPLIFY");
438 
  cucdebug (3, "simplify %x->%x->%x (%i)\n", pred, s1, s2, neg);
439 
  assert (s2 != pred); /* Shouldn't occur => stupid */
440 
  f->bb[pred].next[1] = -1;
441 
  f->bb[pred].next[0] = s1;
442 
 
443 
  if (f->bb[s2].prev[0] == pred) {
444 
    f->bb[s2].prev[0] = f->bb[s2].prev[1];
445 
    f->bb[s2].prev[1] = -1;
446 
  } else if (f->bb[s2].prev[1] == pred) {
447 
    f->bb[s2].prev[1] = -1;
448 
  } else assert (0);
449 
 
450 
  last = &f->bb[pred].insn[f->bb[pred].ninsn - 1];
451 
  assert (last->type & IT_BRANCH);
452 
  for (i = 0; i < f->bb[s2].ninsn; i++) {
453 
    cuc_insn *ii= &f->bb[s2].insn[i];
454 
    if (ii->index == II_LRBB) {
455 
      change_insn_type (ii, II_CMOV);
456 
      ii->type = IT_COND;
457 
      ii->op[1] = neg ? 0 : 1; ii->opt[1] = OPT_CONST;
458 
      ii->op[2] = neg ? 1 : 0; ii->opt[2] = OPT_CONST;
459 
      ii->op[3] = last->op[1]; ii->opt[3] = last->opt[1];
460 
    }
461 
  }
462 
  change_insn_type (last, II_NOP);
463 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_SIMPLIFY");
464 
}
465 
 
466 879 markom 
/* type == 0; keep predecessor condition
467 
 * type == 1; keep successor condition
468 
 * type == 2; join loop unrolled blocks */
469 
static void join_bb (cuc_func *f, int pred, int succ, int type)
470 
{
471 905 markom 
  int i, j, k, n1, n2, ninsn, add_cond = 0;
472 879 markom 
  unsigned long cond_op, cond_opt;
473 
  cuc_insn *insn;
474 
 
475 905 markom 
  if (cuc_debug) cuc_check (f);
476 897 markom 
  cucdebug (3, "%x <= %x+%x (%i)\n", pred, pred, succ, type);
477 
  cucdebug (3, "%x %x\n", f->bb[pred].ninsn, f->bb[succ].ninsn);
478 
  if (cuc_debug >= 3) fflush (stdout);
479 879 markom 
 
480 905 markom 
  n1 = f->bb[pred].ninsn;
481 
  n2 = f->bb[succ].ninsn;
482 
  if (n1 <= 0
483 
   || !(f->bb[pred].insn[n1 - 1].type & IT_BRANCH)) type = 1;
484 879 markom 
  if (type == 0 && f->bb[succ].prev[0] == f->bb[succ].next[0]) add_cond = 1;
485 905 markom 
  if (type == 2) add_cond = 1;
486 
 
487 1045 markom 
  //assert (f->bb[pred].next[0] == f->bb[succ].next[0] || type != 2); /* not supported */
488 879 markom 
 
489 905 markom 
  ninsn = n1 + n2 + (type == 1 ? 0 : 1) + (add_cond ? MAX_REGS : 0);
490 879 markom 
 
491 
  insn = (cuc_insn *) malloc (ninsn * sizeof (cuc_insn));
492 905 markom 
  for (i = 0; i < n1; i++) insn[i] = f->bb[pred].insn[i];
493 
  /* when type == 0, we move the last (jump) instruction to the end */
494 879 markom 
  if (type == 0 || type == 2) {
495 905 markom 
    /* Move first branch instruction to the end */
496 
    assert (insn[n1 - 1].type & IT_BRANCH);
497 
    insn[ninsn - 1] = insn[n1 - 1];
498 
    cond_op = insn[n1 - 1].op[1];
499 
    cond_opt = insn[n1 - 1].opt[1];
500 
 
501 
    /* Remove old branch */
502 
    change_insn_type (&insn[n1 - 1], II_NOP);
503 879 markom 
  }
504 
  /* Copy second block */
505 905 markom 
  for (i = 0; i < n2; i++) insn[i + n1] = f->bb[succ].insn[i];
506 902 markom 
 
507 
  /* and when type == 2, we may need to add sfor instruction, to quit when either is true */
508 
  if (type == 2) {
509 905 markom 
    /* Move second branch instruction to the end */
510 
    if (insn[n1 + n2 - 1].type & IT_BRANCH) {
511 
      insn[ninsn - 1] = insn[n1 + n2 - 1];
512 
 
513 
      /* Use conditional from cmov FLAG_REG, c_p, c_s, c_p */
514 
      insn[ninsn - 1].op[1] = REF (pred, n1 + n2 + FLAG_REG); insn[ninsn - 1].opt[1] = OPT_REF;
515 
 
516 
      /* Remove old one */
517 
      change_insn_type (&insn[n1 + n2 - 1], II_NOP);
518 
    } else change_insn_type (&insn[ninsn - 1], II_NOP); /* do not use branch slot */
519 902 markom 
  }
520 
 
521 905 markom 
#if 1
522 927 markom 
  /* LRBB at start of succ BB is not valid anymore */
523 933 markom 
  if (n1 > 0 && insn[n1].index == II_LRBB) {
524 925 markom 
    if (type == 1) {
525 
      /* We have two possibilities, how this could have happened:
526 
         1. we just moved second predecessor of succ to pred,
527 
            pred now having two predecessors => everything is ok
528 
         2. we just moved second predecessor of succ to pred,
529 
            now, having just one predecessor => LRBB is not needed anymore */
530 
      if (f->bb[pred].prev[1] < 0) { /* handle second option */
531 
        change_insn_type (&insn[n1], II_ADD);
532 
        insn[n1].op[1] = 1; insn[n1].opt[1] = OPT_CONST;
533 
        insn[n1].op[2] = 0; insn[n1].opt[2] = OPT_CONST;
534 
        insn[n1].opt[3] = OPT_NONE;
535 
      }
536 
    } else {
537 
      assert (0); /* not tested yet */
538 
      change_insn_type (&insn[n1], II_NOP);
539 
      for (i = n1; i < ninsn; i++)
540 
        if (insn[i].index == II_CMOV && insn[i].op[3] == REF (pred, n1)) {
541 
          assert (insn[i].opt[3] == OPT_REF);
542 
          insn[i].op[3] = cond_op;
543 
          insn[i].opt[3] = cond_opt;
544 
          if (f->bb[pred].next[0] != succ) {
545 
            unsigned long t; /* negate conditional -- exchange */
546 
            assert (f->bb[pred].next[1] == succ);
547 
            t = insn[i].op[1];
548 
            insn[i].op[1] = insn[i].op[2];
549 
            insn[i].op[2] = t;
550 
            t = insn[i].opt[1];
551 
            insn[i].opt[1] = insn[i].opt[2];
552 
            insn[i].opt[2] = t;
553 
          }
554 902 markom 
        }
555 925 markom 
    }
556 902 markom 
  }
557 905 markom 
#endif
558 879 markom 
 
559 
  for (i = 0; i < ninsn; i++) reloc[i] = -1;
560 
 
561 
  /* Add conditional instructions if required */
562 
  if (add_cond) {
563 
    recalc_last_used_reg (f, pred);
564 
    recalc_last_used_reg (f, succ);
565 1045 markom 
 
566 879 markom 
    /* r0 -- add nop for it */
567 905 markom 
    change_insn_type (&insn[n1 + n2], II_NOP);
568 879 markom 
    for (i = 1; i < MAX_REGS; i++) {
569 905 markom 
      cuc_insn *ii = &insn[n1 + n2 + i];
570 879 markom 
      int a = f->bb[pred].last_used_reg[i];
571 
      int b = f->bb[succ].last_used_reg[i];
572 
 
573 1045 markom 
      /* We have deleted first branch instruction, now we must setup FLAG_REG,
574 
         to point to conditional */
575 
      if (i == FLAG_REG) {
576 
        change_insn_type (ii, II_CMOV);
577 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
578 
        ii->dep = NULL;
579 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
580 
        ii->op[1] = cond_op; ii->opt[1] = cond_opt;
581 
        if (b >= 0) {
582 
          ii->op[2] = b; ii->opt[2] = OPT_REF;
583 
        } else {
584 
          ii->op[2] = cond_op; ii->opt[2] = cond_opt;
585 
        }
586 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
587 
        reloc[REF_I(a)] = REF (pred, n1 + n2 + i);
588 
      } else if (b < 0) change_insn_type (ii, II_NOP);
589 879 markom 
      else if (a < 0) {
590 
        change_insn_type (ii, II_ADD);
591 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
592 879 markom 
        ii->dep = NULL;
593 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
594 
        ii->op[1] = b; ii->opt[1] = OPT_REF;
595 
        ii->op[2] = 0; ii->opt[2] = OPT_CONST;
596 905 markom 
        ii->opt[3] = OPT_NONE;
597 879 markom 
      } else if (b >= 0) {
598 
        change_insn_type (ii, II_CMOV);
599 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
600 879 markom 
        ii->dep = NULL;
601 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
602 
        ii->op[1] = a; ii->opt[1] = OPT_REF;
603 
        ii->op[2] = b; ii->opt[2] = OPT_REF;
604 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
605 905 markom 
        reloc[REF_I(a)] = REF (pred, n1 + n2 + i);
606 879 markom 
      }
607 
      sprintf (ii->disasm, "cmov (join BB)");
608 
    }
609 
  }
610 
 
611 905 markom 
  if (cuc_debug) cuc_check (f);
612 879 markom 
  i = 0;
613 
  switch (type) {
614 
  case 0:
615 1061 markom 
    assert (f->bb[pred].next[0] >= 0);
616 879 markom 
    if (f->bb[pred].next[0] == succ) f->bb[pred].next[0] = f->bb[succ].next[0];
617 
    if (f->bb[pred].next[1] == succ) f->bb[pred].next[1] = f->bb[succ].next[0];
618 
    break;
619 
  case 1:
620 1061 markom 
    assert (f->bb[pred].next[0] >= 0 && f->bb[pred].next[0] != BBID_END);
621 879 markom 
    f->bb[pred].next[0] = f->bb[succ].next[0];
622 
    f->bb[pred].next[1] = f->bb[succ].next[1];
623 
    break;
624 905 markom 
  case 2:
625 1061 markom 
    assert (f->bb[pred].next[0] >= 0 && f->bb[pred].next[0] != BBID_END);
626 905 markom 
    f->bb[pred].next[0] = f->bb[succ].next[0];
627 
    f->bb[pred].next[1] = f->bb[succ].next[1];
628 
    break;
629 879 markom 
  }
630 924 markom 
  if (f->bb[pred].next[0] < 0) f->bb[pred].next[0] = f->bb[pred].next[1];
631 879 markom 
  if (f->bb[pred].next[0] == f->bb[pred].next[1]) f->bb[pred].next[1] = -1;
632 927 markom 
 
633 1061 markom 
  if (type == 0) assert (f->bb[succ].next[1] < 0);
634 
 
635 927 markom 
  /* We just did something stupid -- we joined two predecessors into one;
636 
     succ may need the information from which block we came.  We will repair
637 
     this by converting LRBB to CMOV */
638 930 markom 
  for (j = 0; j < 2; j++) {
639 
    int nb = f->bb[pred].next[j];
640 927 markom 
    int t;
641 
 
642 
    /* check just valid connections */
643 
    if (nb < 0 || nb == BBID_END) continue;
644 
 
645 
    /* check type */
646 
    if (f->bb[nb].prev[0] == pred && f->bb[nb].prev[1] == succ) t = 1;
647 
    else if (f->bb[nb].prev[1] == pred && f->bb[nb].prev[0] == succ) t = 0;
648 
    else continue;
649 
 
650 930 markom 
    /* check all LRBB instructions.  */
651 
    for (i = 0; i < f->bb[nb].ninsn; i++)
652 
      if (f->bb[nb].insn[i].index == II_LRBB) {
653 
        cuc_insn *lrbb =&f->bb[nb].insn[i];
654 
        change_insn_type (lrbb, II_CMOV);
655 
        lrbb->op[1] = t; lrbb->opt[1] = OPT_CONST;
656 
        lrbb->op[2] = 1 - t; lrbb->opt[2] = OPT_CONST;
657 
        lrbb->op[3] = cond_op; lrbb->opt[3] = cond_opt;
658 
        lrbb->type |= IT_COND;
659 
      }
660 927 markom 
  }
661 
 
662 879 markom 
  f->bb[succ].type = BB_DEAD;
663 997 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);
664 905 markom 
  /* remove branch instruction, if there is only one successor */
665 933 markom 
  if (f->bb[pred].next[1] < 0 && ninsn > 0 && insn[ninsn - 1].type & IT_BRANCH) {
666 925 markom 
    assert (f->bb[pred].next[0] != pred); /* end BB, loop should not be possible */
667 905 markom 
    change_insn_type (&insn[ninsn - 1], II_NOP);
668 925 markom 
  }
669 879 markom 
 
670 
  /* Set max count */
671 
  if (f->bb[pred].cnt < f->bb[succ].cnt) f->bb[pred].cnt = f->bb[succ].cnt;
672 
  f->bb[pred].ninsn = ninsn;
673 905 markom 
  f->bb[succ].ninsn = 0;
674 879 markom 
  free (f->bb[pred].insn); f->bb[pred].insn = NULL;
675 
  free (f->bb[succ].insn); f->bb[succ].insn = NULL;
676 
  f->bb[pred].insn = insn;
677 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
678 
    if (f->bb[i].prev[0] == succ) f->bb[i].prev[0] = pred;
679 
    if (f->bb[i].prev[1] == succ) f->bb[i].prev[1] = pred;
680 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
681 
    for (j = 0; j < f->bb[i].ninsn; j++)
682 
      for (k = 0; k < MAX_OPERANDS; k++)
683 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
684 
          /* Check if we are referencing successor BB -> relocate to second part of
685 
             the new block */
686 
          if (REF_BB (f->bb[i].insn[j].op[k]) == succ) {
687 905 markom 
            int t = f->bb[i].insn[j].op[k];
688 
            int ndest = REF (pred, REF_I (t) + n1);
689 997 markom 
            //PRINTF ("%x: %x %x\n", REF(i, j), t, ndest);
690 879 markom 
 
691 905 markom 
            /* We've found a reference to succ. block, being removed, relocate */
692 
            f->bb[i].insn[j].op[k] = ndest;
693 879 markom 
          } else if (REF_BB(f->bb[i].insn[j].op[k]) == pred) {
694 
            if (i != pred && reloc[REF_I(f->bb[i].insn[j].op[k])] >= 0) {
695 
              f->bb[i].insn[j].op[k] = reloc[REF_I(f->bb[i].insn[j].op[k])];
696 
            }
697 
          }
698 
        }
699 
  }
700 
 
701 905 markom 
  if (cuc_debug) cuc_check (f);
702 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "join");
703 879 markom 
}
704 
 
705 
/* Optimize basic blocks */
706 931 markom 
int optimize_bb (cuc_func *f)
707 879 markom 
{
708 931 markom 
  int modified = 0;
709 879 markom 
  int i, j;
710 
remove_lrbb:
711 
  /* we can remove lrbb instructions from blocks with just one predecessor */
712 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
713 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* exactly one predecessor */
714 
      for (j = 0; j < f->bb[i].ninsn; j++)
715 
        if (f->bb[i].insn[j].index == II_LRBB) {
716 
          cuc_insn *t;
717 925 markom 
          cucdebug (4, "-lrbb %x.%x\n", i, j);
718 879 markom 
 
719 
          /* Change to add LRBB, 0, 0 */
720 
          change_insn_type (&f->bb[i].insn[j], II_ADD);
721 
          f->bb[i].insn[j].type &= ~IT_VOLATILE;
722 
          f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_CONST;
723 
          f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = 0; /* always use left block */
724 
          f->bb[i].insn[j].opt[3] = OPT_NONE;
725 931 markom 
          modified = 1;
726 934 markom 
          if (f->bb[i].prev[0] != BBID_START && f->bb[f->bb[i].prev[0]].ninsn > 0) {
727 932 markom 
            t = &f->bb[f->bb[i].prev[0]].insn[f->bb[f->bb[i].prev[0]].ninsn - 1];
728 879 markom 
 
729 932 markom 
            /* If the predecessor still has a conditional jump instruction, we must be careful.
730 
               If next[0] == next[1] join them. Now we will link lrbb and correct the situation */
731 
            if (t->type & IT_BRANCH) { /* We must set a reference to branch result */
732 
              f->bb[i].insn[j].opt[1] = t->opt[1];
733 
              f->bb[i].insn[j].op[1] = t->op[1];
734 
              /* sometimes branch is not needed anymore */
735 
              if (f->bb[f->bb[i].prev[0]].next[1] < 0) change_insn_type (t, II_NOP);
736 
            }
737 879 markom 
          }
738 
        }
739 
    }
740 
  }
741 
 
742 927 markom 
  /* Ordering of joining types is cruical -- we should concat all directly connected BBs
743 
     together first, so when we do a type != 1 joining, we can remove LRBB, directly by
744 
     looking at number of its predeccessors */
745 879 markom 
 
746 
  /* Type 1 joining
747 
     1. link between pred & succ
748 
     2. no other pred's successors
749 925 markom 
     3. no other succ's predecessors, except if pred has max one */
750 932 markom 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
751 
    int p = f->bb[i].prev[0];
752 
    if (p < 0 || p == BBID_START) continue;
753 925 markom 
    /* one successor and max sum of 3 predecessors */
754 932 markom 
    if (f->bb[p].next[0] >= 0 && f->bb[p].next[1] < 0
755 
     && (f->bb[p].prev[1] < 0 || f->bb[i].prev[1] < 0)) {
756 925 markom 
      /* First we will move all predecessors from succ to pred, and then we will do
757 
         real type 1 joining */
758 932 markom 
      if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START) {
759 
        int p1 = f->bb[i].prev[1];
760 925 markom 
        /* joining is surely not worth another extra memory access */
761 932 markom 
        if (f->bb[p].nmemory) continue;
762 
        if (f->bb[p].prev[0] >= 0) {
763 
           assert (f->bb[p].prev[1] < 0);
764 
           f->bb[p].prev[1] = p1;
765 
        } else f->bb[p].prev[0] = p1;
766 
        if (f->bb[p1].next[0] == i) f->bb[p1].next[0] = p;
767 
        else if (f->bb[p1].next[1] == i) f->bb[p1].next[1] = p;
768 925 markom 
        else assert (0);
769 
        f->bb[i].prev[1] = -1;
770 
      }
771 932 markom 
      assert (p >= 0 && f->bb[i].prev[1] < 0); /* one predecessor */
772 
      join_bb (f, p, i, 1);
773 931 markom 
      modified = 1;
774 879 markom 
      goto remove_lrbb;
775 
    }
776 932 markom 
  }
777 927 markom 
 
778 
  /* Type 0 joining
779 
     1. link between pred & succ
780 
     2. no memory accesses in succ
781 
     3. optional pred's second successors
782 
     4. max. one succ's successors */
783 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
784 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START
785 
     && f->bb[i].prev[1] < 0 /* one predecessor */
786 927 markom 
     && f->bb[i].next[1] < 0 /* max. one successor */
787 
     && f->bb[i].nmemory == 0) {                  /* and no memory acceses */
788 
      join_bb (f, f->bb[i].prev[0], i, 0);
789 931 markom 
      modified = 1;
790 927 markom 
      goto remove_lrbb;
791 
    }
792 879 markom 
 
793 
  /* Type 2 joining
794 
     1. link between pred & succ
795 
     2. succ has exactly one predeccessor
796 
     3. pred & succ share common successor
797 
     4. optional succ's second successor */
798 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
799 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* one predecessor */
800 
      int p = f->bb[i].prev[0];
801 932 markom 
      if (p == BBID_START) continue;
802 905 markom 
#if 0 /* not yet supported */
803 
      if (f->bb[p].next[0] == i
804 
       && (f->bb[i].next[1] == f->bb[p].next[1]
805 
        || f->bb[i].next[1] == f->bb[p].next[0])) {
806 
        join_bb (f, p, i, 2);
807 897 markom 
        goto remove_lrbb;
808 
      }
809 905 markom 
#endif
810 
      if (f->bb[p].next[1] == i
811 1045 markom 
       && (f->bb[p].next[0] == f->bb[i].next[1]
812 
        || f->bb[p].next[0] == f->bb[i].next[0])) {
813 905 markom 
        join_bb (f, p, i, 2);
814 931 markom 
        modified = 1;
815 905 markom 
        goto remove_lrbb;
816 
      }
817 879 markom 
    }
818 1042 markom 
 
819 
  /* BB simplify:
820 
     1. a block has exactly 2 successors A and B
821 
     2. A has exactly one successor -- B
822 
     3. A has no memory accesses
823 
     to:
824 
     flow always goes though A, LRBB is replaced by current block conditional
825 
    */
826 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
827 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END
828 
      && f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END) {
829 
      int a = f->bb[i].next[0];
830 
      int b = f->bb[i].next[1];
831 
      int neg = 0;
832 
      /* Exchange? */
833 
      if (f->bb[b].next[0] == a && f->bb[b].next[1] < 0) {
834 
        int t = a;
835 
        a = b;
836 
        b = t;
837 
        neg = 1;
838 
      }
839 
      /* Do the simplification if possible */
840 
      if (f->bb[a].next[0] == b && f->bb[a].next[1] < 0
841 
       && f->bb[a].nmemory == 0) {
842 
        simplify_bb (f, i, a, b, neg);
843 
        modified = 1;
844 
        goto remove_lrbb;
845 
      }
846 
    }
847 
 
848 931 markom 
  return modified;
849 879 markom 
}
850 
 
851 
/* Removes BBs marked as dead */
852 931 markom 
int remove_dead_bb (cuc_func *f)
853 879 markom 
{
854 
  int i, j, k, d = 0;
855 
 
856 
  for (i = 0; i < f->num_bb; i++) if (f->bb[i].type & BB_DEAD) {
857 
    if (f->bb[i].insn) free (f->bb[i].insn);
858 
    f->bb[i].insn = NULL;
859 
    reloc[i] = -1;
860 
  } else {
861 
    reloc[i] = d;
862 
    f->bb[d++] = f->bb[i];
863 
  }
864 931 markom 
  if (f->num_bb == d) return 0;
865 879 markom 
  f->num_bb = d;
866 
 
867 
  /* relocate initial blocks */
868 
  for (i = 0; i < f->num_init_bb; i++)
869 
    f->init_bb_reloc[i] = reloc[f->init_bb_reloc[i]];
870 
 
871 
  /* repair references */
872 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
873 973 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]);
874 925 markom 
          fflush (stdout);
875 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START)
876 
      assert ((f->bb[i].prev[0] = reloc[f->bb[i].prev[0]]) >= 0);
877 
    if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START)
878 
      assert ((f->bb[i].prev[1] = reloc[f->bb[i].prev[1]]) >= 0);
879 925 markom 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END)
880 
      assert ((f->bb[i].next[0] = reloc[f->bb[i].next[0]]) >= 0);
881 
    if (f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END)
882 
      assert ((f->bb[i].next[1] = reloc[f->bb[i].next[1]]) >= 0);
883 879 markom 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
884 
    if (f->bb[i].next[0] == f->bb[i].next[1]) f->bb[i].next[1] = -1;
885 
 
886 
    for (j = 0; j < f->bb[i].ninsn; j++)
887 
      for (k = 0; k < MAX_OPERANDS; k++)
888 925 markom 
        if (f->bb[i].insn[j].opt[k] & OPT_BB && (signed)f->bb[i].insn[j].op[k] >= 0) {
889 
          if (f->bb[i].insn[j].op[k] != BBID_END)
890 
            assert ((f->bb[i].insn[j].op[k] = reloc[f->bb[i].insn[j].op[k]]) >= 0);
891 
        } else if (f->bb[i].insn[j].opt[k] & OPT_REF) {
892 879 markom 
          int t = f->bb[i].insn[j].op[k];
893 
          assert (reloc[REF_BB(t)] >= 0);
894 
          f->bb[i].insn[j].op[k] = REF (reloc[REF_BB(t)], REF_I (t));
895 
        }
896 
  }
897 931 markom 
  return 1;
898 879 markom 
}
899 
 
900 
/* Recursive calculation of dependencies */
901 
static int reg_dep_rec (cuc_func *f, int cur)
902 
{
903 
  int i, j;
904 
  cuc_insn *insn = f->bb[cur].insn;
905 
 
906 997 markom 
  //PRINTF ("\n %i", cur);
907 879 markom 
  /* Spread only, do not loop */
908 
  if (f->bb[cur].tmp) return;
909 
  f->bb[cur].tmp = 1;
910 997 markom 
  //PRINTF ("!   ");
911 879 markom 
 
912 
  for (i = 0; i < f->bb[cur].ninsn; i++) {
913 
    /* Check for destination operand(s) */
914 
    for (j = 0; j < MAX_OPERANDS; j++) if (insn[i].opt[j] & OPT_DEST)
915 
      if ((insn[i].opt[j] & ~OPT_DEST) == OPT_REGISTER && (signed)insn[i].op[j] >= 0) {
916 997 markom 
        //PRINTF ("%i:%i,%x ", insn[i].op[j], i, REF (cur, i));
917 879 markom 
        assert (insn[i].op[j] > 0 && insn[i].op[j] < MAX_REGS); /* r0 should never be dest */
918 
        f->bb[cur].last_used_reg[insn[i].op[j]] = REF (cur, i);
919 
      }
920 
  }
921 
 
922 925 markom 
  if (f->bb[cur].next[0] >= 0 && f->bb[cur].next[0] != BBID_END)
923 
    reg_dep_rec (f, f->bb[cur].next[0]);
924 
  if (f->bb[cur].next[1] >= 0 && f->bb[cur].next[1] != BBID_END)
925 
    reg_dep_rec (f, f->bb[cur].next[1]);
926 879 markom 
}
927 
 
928 
/* Detect register dependencies */
929 
void reg_dep (cuc_func *f)
930 
{
931 
  int i, b, c;
932 
 
933 
  /* Set dead blocks */
934 
  for (b = 0; b < f->num_bb; b++) {
935 
    f->bb[b].tmp = 0;
936 
    for (i = 0; i < MAX_REGS; i++) f->bb[b].last_used_reg[i] = -1;
937 
  }
938 
 
939 
  /* Start with first block and set dependecies of all reachable blocks */
940 
  /* At the same time set last_used_regs */
941 
  reg_dep_rec (f, 0);
942 
 
943 
  for (i = 0; i < f->num_bb; i++)
944 
    if (f->bb[i].tmp) f->bb[i].tmp = 0;
945 
    else f->bb[i].type |= BB_DEAD;
946 
 
947 
  /* Detect loops; mark BBs where loops must be broken */
948 
  for (c = 0; c < f->num_bb; c++) {
949 
    int min = 3, minb;
950 
 
951 
    /* search though all non-visited for minimum number of unvisited predecessors */
952 
    for (b = 0; b < f->num_bb; b++) if (!f->bb[b].tmp) {
953 
      int tmp = 0;
954 932 markom 
      if (f->bb[b].prev[0] >= 0 && f->bb[b].prev[0] != BBID_START
955 
       && !f->bb[f->bb[b].prev[0]].tmp) tmp++;
956 
      if (f->bb[b].prev[1] >= 0 && f->bb[b].prev[1] != BBID_START
957 
       && !f->bb[f->bb[b].prev[1]].tmp) tmp++;
958 879 markom 
      if (tmp < min) {
959 
        minb = b;
960 
        min = tmp;
961 
        if (tmp == 0) break; /* We already have the best one */
962 
      }
963 
    }
964 
    b = minb;
965 
    f->bb[b].tmp = 1; /* Mark visited */
966 883 markom 
    cucdebug (3, "minb %i min %i\n", minb, min);
967 879 markom 
    if (min) { /* We just broke the loop */
968 
      f->bb[b].type |= BB_INLOOP;
969 
    }
970 
  }
971 
 
972 
  /* Set real predecessors in cmov instructions to previous blocks */
973 
  for (b = 0; b < f->num_bb; b++)
974 
    for (i = 1; i < MAX_REGS - 1; i++) {
975 
      int pa, pb;
976 
      assert (f->bb[b].insn[i].index ==  II_CMOV);
977 
      assert (f->bb[b].insn[i].opt[0] == OPT_REGISTER | OPT_DEST);
978 
      assert (f->bb[b].insn[i].op[0] == i);
979 932 markom 
      if (f->bb[b].prev[0] < 0 || f->bb[b].prev[0] == BBID_START) pa = -1;
980 879 markom 
      else pa = f->bb[f->bb[b].prev[0]].last_used_reg[i];
981 932 markom 
      if (f->bb[b].prev[1] < 0 || f->bb[b].prev[1] == BBID_START) pb = -1;
982 879 markom 
      else pb = f->bb[f->bb[b].prev[1]].last_used_reg[i];
983 
 
984 
      /* We do some very simple optimizations right away to make things more readable */
985 
      if (pa < 0 && pb < 0) {
986 
        /* Was not used at all */
987 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
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 if (pa < 0) {
991 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
992 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
993 
        f->bb[b].insn[i].op[1] = pb; f->bb[b].insn[i].opt[1] = OPT_REF;
994 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
995 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
996 
      } else if (pb < 0) {
997 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
998 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
999 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
1000 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
1001 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
1002 
      } else {
1003 
        int t = REF (b, 0); /* lrbb should be first instruction */
1004 
        assert (f->INSN(t).index == II_LRBB);
1005 
 
1006 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
1007 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
1008 
 
1009 
        f->bb[b].insn[i].op[2] = pb; f->bb[b].insn[i].opt[2] = OPT_REF;
1010 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
1011 
 
1012 
        /* Update op[3] -- flag register */
1013 
        assert (f->bb[b].insn[i].opt[3] == OPT_REGISTER);
1014 
        assert (f->bb[b].insn[i].op[3] == LRBB_REG);
1015 
        assert (t >= 0);
1016 
        f->bb[b].insn[i].opt[3] = OPT_REF; /* Convert already used regs to references */
1017 
        f->bb[b].insn[i].op[3] = t;
1018 
        assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1019 
      }
1020 
    }
1021 
 
1022 
  /* assign register references */
1023 
  for (b = 0; b < f->num_bb; b++) {
1024 
    /* rebuild last used reg array */
1025 
    f->bb[b].last_used_reg[0] = -1;
1026 
    if (f->bb[b].insn[0].index == II_LRBB) f->bb[b].last_used_reg[LRBB_REG] = 0;
1027 
    else f->bb[b].last_used_reg[LRBB_REG] = -1;
1028 
 
1029 
    for (i = 1; i < MAX_REGS - 1; i++)
1030 
      f->bb[b].last_used_reg[i] = -1;
1031 
 
1032 
    /* Create references */
1033 
    for (i = 0; i < f->bb[b].ninsn; i++) {
1034 
      int k;
1035 
      /* Check for source operands first */
1036 
      for (k = 0; k < MAX_OPERANDS; k++) {
1037 
        if (!(f->bb[b].insn[i].opt[k] & OPT_DEST))
1038 
        if (f->bb[b].insn[i].opt[k] & OPT_REGISTER) {
1039 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1040 
 
1041 
          if (f->bb[b].insn[i].op[k] == 0) { /* Convert r0 to const0 */
1042 
            f->bb[b].insn[i].opt[k] = OPT_CONST;
1043 
            f->bb[b].insn[i].op[k] = 0;
1044 
          } else if (t >= 0) {
1045 
            f->bb[b].insn[i].opt[k] = OPT_REF; /* Convert already used regs to references */
1046 
            f->bb[b].insn[i].op[k] = t;
1047 
            assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1048 
            //f->INSN(t).op[0] = -1;
1049 
          }
1050 
        } else if (f->bb[b].insn[i].opt[k] & OPT_REF) {
1051 
          //f->INSN(f->bb[b].insn[i].op[k]).op[0] = -1; /* Mark referenced */
1052 
          f->INSN(f->bb[b].insn[i].op[k]).type &= ~IT_UNUSED;
1053 
        }
1054 
      }
1055 
 
1056 
      /* Now check for destination operand(s) */
1057 
      for (k = 0; k < MAX_OPERANDS; k++) if (f->bb[b].insn[i].opt[k] & OPT_DEST)
1058 
        if ((f->bb[b].insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
1059 
          && (int)f->bb[b].insn[i].op[k] >= 0) {
1060 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1061 
          assert (f->bb[b].insn[i].op[k] != 0); /* r0 should never be dest */
1062 
          f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]] = REF (b, i);
1063 
        }
1064 
    }
1065 
  }
1066 
 
1067 
  /* Remove all unused lrbb */
1068 
  for (b = 0; b < f->num_bb; b++)
1069 
    for (i = 0; i < f->bb[b].ninsn; i++)
1070 
      if (f->bb[b].insn[i].type & IT_UNUSED) change_insn_type (&f->bb[b].insn[i], II_NOP);
1071 
 
1072 
  /* SSAs with final register value are marked as outputs */
1073 1061 markom 
  assert (f->bb[f->num_bb - 1].next[0] == BBID_END);
1074 939 markom 
  for (i = 0; i < MAX_REGS; i++) if (!caller_saved[i]) {
1075 879 markom 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1076 
    /* Mark them volatile, so optimizer does not remove them */
1077 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1078 
  }
1079 939 markom 
  {
1080 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1081 
    /* Mark them volatile, so optimizer does not remove them */
1082 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1083 
  }
1084 879 markom 
}
1085 
 
1086 897 markom 
/* split the BB, based on the group numbers in .tmp */
1087 
void expand_bb (cuc_func *f, int b)
1088 
{
1089 
  int n = f->num_bb;
1090 
  int mg = 0;
1091 
  int b1, i, j;
1092 
 
1093 
  for (i = 0; i < f->bb[b].ninsn; i++)
1094 
    if (f->bb[b].insn[i].tmp > mg) mg = f->bb[b].insn[i].tmp;
1095 
 
1096 
  /* Create copies */
1097 
  for (b1 = 1; b1 <= mg; b1++) {
1098 
    assert (f->num_bb < MAX_BB);
1099 
    cpy_bb (&f->bb[f->num_bb], &f->bb[b]);
1100 
    f->num_bb++;
1101 
  }
1102 
 
1103 
  /* Relocate */
1104 
  for (b1 = 0; b1 < f->num_bb; b1++)
1105 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1106 
      dep_list *d = f->bb[b1].insn[i].dep;
1107 
      for (j = 0; j < MAX_OPERANDS; j++)
1108 
        if (f->bb[b1].insn[i].opt[j] & OPT_REF) {
1109 
          int t = f->bb[b1].insn[i].op[j];
1110 
          if (REF_BB(t) == b && f->INSN(t).tmp != 0)
1111 
            f->bb[b1].insn[i].op[j] = REF (n + f->INSN(t).tmp - 1, REF_I(t));
1112 
        }
1113 
      while (d) {
1114 
        if (REF_BB (d->ref) == b && f->INSN(d->ref).tmp != 0)
1115 
          d->ref = REF (n + f->INSN(d->ref).tmp - 1, REF_I(d->ref));
1116 
        d = d->next;
1117 
      }
1118 
    }
1119 
 
1120 
  /* Delete unused instructions */
1121 
  for (j = 0; j <= mg; j++) {
1122 
    if (j == 0) b1 = b;
1123 
    else b1 = n + j - 1;
1124 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1125 
      if (f->bb[b1].insn[i].tmp != j)
1126 
        change_insn_type (&f->bb[b1].insn[i], II_NOP);
1127 
      f->bb[b1].insn[i].tmp = 0;
1128 
    }
1129 
    if (j < mg) {
1130 
      f->bb[b1].next[0] = n + j;
1131 
      f->bb[b1].next[1] = -1;
1132 
      f->bb[n + j].prev[0] = b1;
1133 
      f->bb[n + j].prev[1] = -1;
1134 
    } else {
1135 
      i = f->bb[b1].next[0];
1136 
      f->bb[n + j].prev[0] = j == 1 ? b : b1 - 1;
1137 
      f->bb[n + j].prev[1] = -1;
1138 925 markom 
      if (i >= 0 && i != BBID_END) {
1139 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1140 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1141 
      }
1142 
      i = f->bb[b1].next[1];
1143 925 markom 
      if (i >= 0 && i != BBID_END) {
1144 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1145 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1146 
      }
1147 
    }
1148 
  }
1149 
}
1150 
 
1151 879 markom 
/* Scans sequence of BBs and set bb[].cnt */
1152 
void generate_bb_seq (cuc_func *f, char *mp_filename, char *bb_filename)
1153 
{
1154 
  FILE *fi, *fo;
1155 
  struct mprofentry_struct *buf;
1156 
  const int bufsize = 256;
1157 
  unsigned long *bb_start;
1158 
  unsigned long *bb_end;
1159 
  int b, i, r;
1160 
  int curbb, prevbb = -1;
1161 
  unsigned long addr = -1;
1162 
  unsigned long prevaddr = -1;
1163 897 markom 
  int mssum = 0;
1164 
  int mlsum = 0;
1165 
  int mscnt = 0;
1166 
  int mlcnt = 0;
1167 940 markom 
  int reopened = 0;
1168 879 markom 
 
1169 940 markom 
  /* Use already opened stream? */
1170 
  if (runtime.sim.fmprof) {
1171 
    fi = runtime.sim.fmprof;
1172 
    reopened = 1;
1173 
    rewind (fi);
1174 
  } else assert (fi = fopen (mp_filename, "rb"));
1175 879 markom 
  assert (fo = fopen (bb_filename, "wb+"));
1176 
 
1177 
  assert (bb_start = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1178 
  assert (bb_end = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1179 
  for (b = 0; b < f->num_bb; b++) {
1180 
    bb_start[b] = f->start_addr + f->bb[b].first * 4;
1181 
    bb_end[b] = f->start_addr + f->bb[b].last * 4;
1182 997 markom 
    //PRINTF ("%i %x %x\n", b, bb_start[b], bb_end[b]);
1183 879 markom 
    f->bb[0].cnt = 0;
1184 
  }
1185 
 
1186 
  buf = (struct mprofentry_struct *) malloc (sizeof (struct mprofentry_struct) * bufsize);
1187 
  assert (buf);
1188 
 
1189 997 markom 
  //PRINTF ("BBSEQ:\n");
1190 879 markom 
  do {
1191 
    r = fread (buf, sizeof (struct mprofentry_struct), bufsize, fi);
1192 997 markom 
    //PRINTF ("r%i : ", r);
1193 879 markom 
    for (i = 0; i < r; i++) {
1194 
      if (buf[i].type & MPROF_FETCH) {
1195 997 markom 
        //PRINTF ("%x, ", buf[i].addr);
1196 879 markom 
        if (buf[i].addr >= f->start_addr && buf[i].addr <= f->end_addr) {
1197 
          assert (buf[i].type & MPROF_32);
1198 
          prevaddr = addr;
1199 
          addr = buf[i].addr;
1200 
          for (b = 0; b < f->num_bb; b++)
1201 
            if (bb_start[b] <= addr && addr <= bb_end[b]) break;
1202 
          assert (b < f->num_bb);
1203 
          curbb = b;
1204 
          if (prevaddr + 4 != addr) prevbb = -1;
1205 
        } else curbb = -1;
1206 
 
1207 
#warning TODO: do not count interrupts
1208 
        if (curbb != prevbb && curbb >= 0) {
1209 
          fwrite (&curbb, sizeof (unsigned long), 1, fo);
1210 997 markom 
          //PRINTF (" [%i] ", curbb);
1211 879 markom 
          f->bb[curbb].cnt++;
1212 
          prevbb = curbb;
1213 
        }
1214 897 markom 
      } else {
1215 
        if (verify_memoryarea(buf[i].addr))
1216 
          if (buf[i].type & MPROF_WRITE) mscnt++, mssum += cur_area->delayw;
1217 1041 markom 
          else mlcnt++, mlsum += cur_area->delayr;
1218 879 markom 
      }
1219 
    }
1220 997 markom 
    //PRINTF ("\n");
1221 879 markom 
  } while (r == bufsize);
1222 997 markom 
  //PRINTF ("\n");
1223 879 markom 
 
1224 897 markom 
  runtime.cuc.mdelay[0] = (1. * mlsum) / mlcnt;
1225 1041 markom 
  runtime.cuc.mdelay[1] = (1. * mssum) / mscnt;
1226 897 markom 
  runtime.cuc.mdelay[2] = runtime.cuc.mdelay[3] = 1;
1227 883 markom 
  f->num_runs = f->bb[0].cnt;
1228 940 markom 
  if (!reopened) fclose (fi);
1229 879 markom 
  fclose (fo);
1230 
  free (buf);
1231 
  free (bb_end);
1232 
  free (bb_start);
1233 
 
1234 
  /* Initialize basic block relocations */
1235 
  f->num_init_bb = f->num_bb;
1236 997 markom 
  //PRINTF ("num_init_bb = %i\n", f->num_init_bb);
1237 879 markom 
  assert (f->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
1238 
  for (b = 0; b < f->num_init_bb; b++) f->init_bb_reloc[b] = b;
1239 
}
1240 
 
1241 
/* Scans sequence of BBs and set counts for pre/unrolled loop for BB b */
1242 
void count_bb_seq (cuc_func *f, int b, char *bb_filename, int *counts, int preroll, int unroll)
1243 
{
1244 
  FILE *fi;
1245 
  const int bufsize = 256;
1246 
  int i, r;
1247 
  int *buf;
1248 
  int cnt = 0;
1249 
  int times = preroll - 1 + unroll;
1250 
 
1251 
  assert (fi = fopen (bb_filename, "rb"));
1252 
  for (i = 0; i < times; i++) counts[i] = 0;
1253 
  assert (buf = (int *) malloc (sizeof (int) * bufsize));
1254 
 
1255 
  do {
1256 
    r = fread (buf, sizeof (int), bufsize, fi);
1257 
    for (i = 0; i < r; i++) {
1258 
      /* count consecutive acesses */
1259 
      if (f->init_bb_reloc[buf[i]] == b) {
1260 
        counts[cnt]++;
1261 
        if (++cnt >= times) cnt = preroll - 1;
1262 
      } else cnt = 0;
1263 
    }
1264 
  } while (r == bufsize);
1265 
 
1266 
  log ("Counts %i,%i :", preroll, unroll);
1267 
  for (i = 0; i < times; i++) log ("%x ", counts[i]);
1268 
  log ("\n");
1269 
 
1270 
  fclose (fi);
1271 
  free (buf);
1272 
}
1273 
 
1274 
/* relocate all accesses inside of BB b to back/fwd */
1275 
static void relocate_bb (cuc_bb *bb, int b, int back, int fwd)
1276 
{
1277 
  int i, j;
1278 
  for (i = 0; i < bb->ninsn; i++)
1279 
    for (j = 0; j < MAX_OPERANDS; j++)
1280 
      if (bb->insn[i].opt[j] & OPT_REF
1281 
       && REF_BB (bb->insn[i].op[j]) == b) {
1282 
        int t = REF_I (bb->insn[i].op[j]);
1283 
        if (t < i) bb->insn[i].op[j] = REF (back, t);
1284 
        else bb->insn[i].op[j] = REF (fwd, t);
1285 
      }
1286 
}
1287 
 
1288 1061 markom 
/* Preroll if type == 1 or unroll if type == 0 loop in BB b `ntimes' times and return
1289 
   new function. Original function is unmodified. */
1290 
static cuc_func *roll_loop (cuc_func *f, int b, int ntimes, int type)
1291 879 markom 
{
1292 
  int b1, t, i, j, prevb, prevart_b;
1293 
  cuc_func *n = dup_func (f);
1294 
  cuc_bb *ob = &f->bb[b];
1295 
  cuc_insn *ii;
1296 
 
1297 1061 markom 
  assert (ntimes > 1);
1298 
  cucdebug (3, "roll type = %i, BB%i x %i (num_bb %i)\n", type, b, ntimes, n->num_bb);
1299 
  ntimes--;
1300 
  assert (n->num_bb + ntimes * 2 < MAX_BB);
1301 879 markom 
 
1302 
  prevb = b;
1303 
  prevart_b = b;
1304 1061 markom 
 
1305 
  /* point to first artificial block */
1306 
  if (n->bb[b].next[0] != b) {
1307 
    n->bb[b].next[0] = n->num_bb + 1;
1308 
  } else if (n->bb[b].next[1] != b) {
1309 
    n->bb[b].next[1] = n->num_bb + 1;
1310 
  }
1311 
 
1312 879 markom 
  /* Duplicate the BB */
1313 1061 markom 
  for (t = 0; t < ntimes; t++) {
1314 879 markom 
    cuc_bb *pb = &n->bb[prevart_b];
1315 
    /* Add new block and set links */
1316 
    b1 = n->num_bb++;
1317 
    cpy_bb (&n->bb[b1], ob);
1318 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1319 1061 markom 
    n->bb[b1].type &= ~BB_INLOOP;
1320 
    print_cuc_bb (n, "prerollA");
1321 879 markom 
 
1322 1061 markom 
    printf ("prevb %i b1 %i prevart %i\n", prevb, b1, prevart_b);
1323 879 markom 
    /* Set predecessor's successor */
1324 
    if (n->bb[prevb].next[0] == b) {
1325 
      n->bb[prevb].next[0] = b1;
1326 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1327 
      else pb->next[1] = b1 + 1;
1328 
      n->bb[b1].next[1] = b1 + 1;
1329 
    } else if (n->bb[prevb].next[1] == b) {
1330 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1331 
      else pb->next[1] = b1 + 1;
1332 
      n->bb[b1].next[0] = b1 + 1;
1333 
      n->bb[prevb].next[1] = b1;
1334 
    } else assert (0);
1335 
 
1336 
    /* Set predecessor */
1337 
    n->bb[b1].prev[0] = prevb;
1338 
    n->bb[b1].prev[1] = -1;
1339 
 
1340 
    /* Relocate backward references to current instance and forward references
1341 
       to previous one */
1342 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1343 
 
1344 
    /* add artificial block, just to join accesses */
1345 
    b1 = n->num_bb++;
1346 
    cpy_bb (&n->bb[b1], ob);
1347 
    n->bb[b1].cnt = 0;
1348 
 
1349 
    for (i = 0; i < ob->ninsn - 1; i++) {
1350 
      ii = &n->bb[b1].insn[i];
1351 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1352 
        change_insn_type (ii, II_CMOV);
1353 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1354 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1355 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1356 
 
1357 
        /* Take left one, if we should have finished the first iteration*/
1358 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1359 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1360 
        } else {
1361 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1362 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1363 
        }
1364 
        ii->dep = NULL;
1365 930 markom 
        ii->type = ob->insn[i].type & IT_COND;
1366 879 markom 
      } else {
1367 
        change_insn_type (ii, II_NOP);
1368 
      }
1369 
    }
1370 
 
1371 928 markom 
    /* Add conditional or instruction at the end, prioritizing flags */
1372 879 markom 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1373 928 markom 
    change_insn_type (ii, II_CMOV);
1374 879 markom 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1375 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1376 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1377 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1378 
    } else {
1379 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1380 
      ii->opt[1] = OPT_REF;
1381 
    }
1382 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1383 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1384 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1385 
    } else {
1386 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1387 
      ii->opt[2] = OPT_REF;
1388 
    }
1389 928 markom 
    /* {z = x || y;} is same as {z = x ? x : y;} */
1390 
    ii->op[3] = ii->op[1]; ii->opt[3] = ii->opt[1];
1391 933 markom 
    ii->type = IT_COND;
1392 879 markom 
 
1393 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1394 1061 markom 
    n->bb[b1].type &= ~BB_INLOOP;
1395 879 markom 
    n->bb[b1].prev[0] = prevart_b;
1396 
    n->bb[b1].prev[1] = b1 - 1;
1397 925 markom 
    n->bb[b1].next[0] = -1;
1398 879 markom 
    n->bb[b1].next[1] = -1;
1399 1061 markom 
 
1400 879 markom 
    prevb = b1 - 1;
1401 
    prevart_b = b1;
1402 1061 markom 
    print_cuc_bb (n, "prerollB");
1403 879 markom 
  }
1404 1061 markom 
 
1405 
  print_cuc_bb (n, "preroll0");
1406 
  n->bb[prevart_b].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1407 879 markom 
 
1408 1061 markom 
  print_cuc_bb (n, "preroll1");
1409 879 markom 
  /* repair BB after loop, to point back to latest artificial BB */
1410 
  b1 = n->bb[prevart_b].next[0];
1411 925 markom 
  if (b1 >= 0 && b1 != BBID_END) {
1412 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1413 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1414 879 markom 
    else assert (0);
1415 
  }
1416 
 
1417 1061 markom 
  if (type) {
1418 
    /* Relink to itself */
1419 
    /* Set predecessor's successor */
1420 
    if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = prevb;
1421 
    else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = prevb;
1422 
    else assert (0);
1423 
    n->bb[prevb].prev[1] = prevb;
1424 879 markom 
 
1425 1061 markom 
    /* Set predecessor */
1426 
    if (n->bb[b].prev[0] == b) {
1427 
      n->bb[b].prev[0] = n->bb[b].prev[1];
1428 
      n->bb[b].prev[1] = -1;
1429 
    } else if (n->bb[b].prev[1] == b) n->bb[b].prev[1] = -1;
1430 
    else assert (0);
1431 
  } else {
1432 
    /* Relink back to start of the loop */
1433 879 markom 
    /* Set predecessor's successor */
1434 1061 markom 
    if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = b;
1435 
    else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = b;
1436 
    else assert (0);
1437 879 markom 
 
1438 
    /* Set predecessor */
1439 1061 markom 
    if (n->bb[b].prev[0] == b) n->bb[b].prev[0] = prevb;
1440 
    else if (n->bb[b].prev[1] == b) n->bb[b].prev[1] = prevb;
1441 879 markom 
    else assert (0);
1442 
  }
1443 
 
1444 1061 markom 
  print_cuc_bb (n, "preroll2");
1445 879 markom 
 
1446 
  /* Relocate backward references to current instance and forward references
1447 
     to previous one */
1448 
  relocate_bb (&n->bb[b], b, b, prevb);
1449 
 
1450 
  /* Relocate all other blocks to point to latest prevart_b */
1451 
  for (i = 0; i < f->num_bb; i++)
1452 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1453 
 
1454 
  return n;
1455 
}
1456 
 
1457 
/* Unroll loop b unroll times and return new function. Original
1458 
   function is unmodified. */
1459 
cuc_func *preunroll_loop (cuc_func *f, int b, int preroll, int unroll, char *bb_filename)
1460 
{
1461 
  int b1, i;
1462 
  cuc_func *n, *t;
1463 
  int *counts;
1464 
  int *bb_reloc;
1465 
 
1466 
  if (preroll > 1) {
1467 1061 markom 
    t = roll_loop (f, b, preroll, 1);
1468 879 markom 
    b1 = t->num_bb - 2;
1469 
    if (unroll > 1) {
1470 
      //print_cuc_bb (t, "preunroll1");
1471 1061 markom 
      n = roll_loop (t, b1, unroll, 0);
1472 879 markom 
      free_func (t);
1473 
    } else n = t;
1474 
  } else {
1475 
    b1 = b;
1476 1061 markom 
    if (unroll > 1) n = roll_loop (f, b1, unroll, 0);
1477 897 markom 
    else return dup_func (f);
1478 879 markom 
  }
1479 
 
1480 897 markom 
  /* Assign new counts to functions */
1481 879 markom 
  assert (counts = (int *)malloc (sizeof (int) * (preroll - 1 + unroll)));
1482 
  count_bb_seq (n, b, bb_filename, counts, preroll, unroll);
1483 
  for (i = 0; i < preroll - 1 + unroll; i++) {
1484 
    if (i == 0) b1 = b;
1485 
    else b1 = f->num_bb + (i - 1) * 2;
1486 
    n->bb[b1].cnt = counts[i];
1487 
  }
1488 
 
1489 
  //print_cuc_bb (n, "preunroll");
1490 
  free (counts);
1491 
  return n;
1492 
}
1493 933 markom 
 

powered by: WebSVN 2.1.0

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