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

Subversion Repositories or1k

[/] [or1k/] [branches/] [stable_0_2_x/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 997

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

powered by: WebSVN 2.1.0

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