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

Subversion Repositories or1k

[/] [or1k/] [tags/] [nog_patch_36/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 1045

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

Line No. Rev Author Line
1 879 markom 
/* bb.c -- OpenRISC Custom Unit Compiler, Basic Block handling
2 
 *    Copyright (C) 2002 Marko Mlinar, markom@opencores.org
3 
 *
4 
 *    This file is part of OpenRISC 1000 Architectural Simulator.
5 
 *
6 
 *    This program is free software; you can redistribute it and/or modify
7 
 *    it under the terms of the GNU General Public License as published by
8 
 *    the Free Software Foundation; either version 2 of the License, or
9 
 *    (at your option) any later version.
10 
 *
11 
 *    This program is distributed in the hope that it will be useful,
12 
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
13 
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14 
 *    GNU General Public License for more details.
15 
 *
16 
 *    You should have received a copy of the GNU General Public License
17 
 *    along with this program; if not, write to the Free Software
18 
 *    Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. */
19 
 
20 
#include <stdio.h>
21 
#include <stdlib.h>
22 
#include <stdarg.h>
23 
#include <assert.h>
24 897 markom 
#include "sim-config.h"
25 
#include "abstract.h"
26 879 markom 
#include "cuc.h"
27 
#include "insn.h"
28 
#include "support/profile.h"
29 
 
30 925 markom 
/* prints out bb string */
31 
void print_bb_num (int num)
32 
{
33 997 markom 
  if (num < 0) PRINTF ("*");
34 
  else if (num == BBID_END) PRINTF ("END");
35 
  else if (num == BBID_START) PRINTF ("START");
36 
  else PRINTF ("%2x", num);
37 925 markom 
}
38 
 
39 879 markom 
/* Print out basic blocks */
40 
void print_cuc_bb (cuc_func *f, char *s)
41 
{
42 
  int i;
43 997 markom 
  PRINTF ("------- %s -------\n", s);
44 879 markom 
  for (i = 0; i < f->num_bb; i++) {
45 997 markom 
    if (f->bb[i].insn) PRINTF ("\n---- BB%-2x * %x ---- ", i, f->bb[i].cnt);
46 
    else PRINTF ("BB%-2x: %4x-%-4x", i, f->bb[i].first, f->bb[i].last);
47 
    PRINTF (" type %02x tmp %i ", f->bb[i].type, f->bb[i].tmp);
48 
    PRINTF ("next "); print_bb_num (f->bb[i].next[0]);
49 
    PRINTF (" "); print_bb_num (f->bb[i].next[1]);
50 
    PRINTF (" prev "); print_bb_num (f->bb[i].prev[0]);
51 
    PRINTF (" "); print_bb_num (f->bb[i].prev[1]);
52 
    PRINTF ("\n");
53 879 markom 
 
54 
    if (f->bb[i].insn) print_insns (f->bb[i].insn, f->bb[i].ninsn, 0);
55 
  }
56 953 markom 
  if (f->nmsched) {
57 997 markom 
    PRINTF ("\nmsched: ");
58 953 markom 
    for (i = 0; i < f->nmsched; i++)
59 997 markom 
      PRINTF ("%x ", f->msched[i]);
60 
    PRINTF ("\n\n\n");
61 
  } else PRINTF ("\n");
62 897 markom 
  fflush (stdout);
63 879 markom 
}
64 
 
65 
/* Copies src basic block into destination */
66 
cuc_bb *cpy_bb (cuc_bb *dest, cuc_bb *src)
67 
{
68 897 markom 
  int i, j;
69 941 markom 
  dep_list *d;
70 897 markom 
  assert (dest != src);
71 879 markom 
  *dest = *src;
72 
  assert (dest->insn = malloc (sizeof (cuc_insn) * src->ninsn));
73 941 markom 
  for (i = 0; i < src->ninsn; i++) {
74 
    d = src->insn[i].dep;
75 879 markom 
    dest->insn[i] = src->insn[i];
76 941 markom 
    dest->insn[i].dep = NULL;
77 
    while (d) {
78 
      add_dep (&dest->insn[i].dep, d->ref);
79 
      d = d->next;
80 
    }
81 
  }
82 
 
83 
  d = src->mdep;
84 
  dest->mdep = NULL;
85 
  while (d) {
86 
    add_dep (&dest->mdep, d->ref);
87 
    d = d->next;
88 
  }
89 879 markom 
  if (src->ntim) {
90 
    assert (dest->tim = malloc (sizeof (cuc_timings) * src->ntim));
91 897 markom 
    for (i = 0; i < src->ntim; i++) {
92 
      dest->tim[i] = src->tim[i];
93 
      if (src->tim[i].nshared) {
94 
        assert (dest->tim[i].shared = malloc (sizeof (int) * src->tim[i].nshared));
95 
        for (j = 0; j < src->tim[i].nshared; j++)
96 
          dest->tim[i].shared[j] = src->tim[i].shared[j];
97 
      }
98 
    }
99 879 markom 
  }
100 
}
101 
 
102 
/* Duplicates function */
103 
cuc_func *dup_func (cuc_func *f)
104 
{
105 
  cuc_func *n = (cuc_func *) malloc (sizeof (cuc_func));
106 
  int b, i;
107 
  for (b = 0; b < f->num_bb; b++) cpy_bb (&n->bb[b], &f->bb[b]);
108 
  n->num_bb = f->num_bb;
109 
  assert (n->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
110 
  for (b = 0; b < f->num_init_bb; b++) n->init_bb_reloc[b] = f->init_bb_reloc[b];
111 
  n->num_init_bb = f->num_init_bb;
112 915 markom 
  for (i = 0; i < MAX_REGS; i++) {
113 
    n->saved_regs[i] = f->saved_regs[i];
114 
    n->lur[i] = f->lur[i];
115 
    n->used_regs[i] = f->used_regs[i];
116 
  }
117 879 markom 
  n->start_addr = f->start_addr;
118 
  n->end_addr = f->end_addr;
119 
  n->orig_time = f->orig_time;
120 
  n->nmsched = f->nmsched;
121 1041 markom 
  n->num_runs = f->num_runs;
122 879 markom 
  for (i = 0; i < f->nmsched; i++) {
123 
    n->msched[i] = f->msched[i];
124 
    n->mtype[i] = f->mtype[i];
125 
  }
126 906 markom 
  n->nfdeps = f->nfdeps;
127 
  if (f->nfdeps) {
128 
    f->fdeps = (cuc_func **) malloc (sizeof (cuc_func *) * f->nfdeps);
129 
    for (i = 0; i < f->nfdeps; i++) n->fdeps[i] = f->fdeps[i];
130 
  }
131 879 markom 
  return n;
132 
}
133 
 
134 
/* Releases memory allocated by function */
135 
void free_func (cuc_func *f)
136 
{
137 
  int b, i;
138 
  for (b = 0; b < f->num_bb; b++) {
139 
    for (i = 0; i < f->bb[b].ninsn; i++)
140 
      dispose_list (&f->bb[b].insn[i].dep);
141 
    if (f->bb[b].insn) free (f->bb[b].insn);
142 897 markom 
    for (i = 0; i < f->bb[b].ntim; i++)
143 
      if (f->bb[b].tim[i].nshared && f->bb[b].tim[i].shared)
144 
        free (f->bb[b].tim[i].shared);
145 879 markom 
    if (f->bb[b].tim && f->bb[b].ntim) free (f->bb[b].tim);
146 
  }
147 
  free (f);
148 
}
149 
 
150 
/* Recalculates last_used_reg */
151 
void recalc_last_used_reg (cuc_func *f, int b)
152 
{
153 
  int i;
154 
  cuc_bb *bb = &f->bb[b];
155 
 
156 
  /* rebuild last used reg array */
157 
  if (bb->insn[0].index == II_LRBB) bb->last_used_reg[LRBB_REG] = 0;
158 
  else bb->last_used_reg[LRBB_REG] = -1;
159 
 
160 
  for (i = 1; i < MAX_REGS - 1; i++) bb->last_used_reg[i] = -1;
161 
 
162 
    /* Create references */
163 
  for (i = 0; i < bb->ninsn; i++) {
164 
    int k;
165 
    /* Now check for destination operand(s) */
166 
    for (k = 0; k < MAX_OPERANDS; k++) if (bb->insn[i].opt[k] & OPT_DEST)
167 
      if ((bb->insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
168 
        && (int)bb->insn[i].op[k] >= 0) {
169 
        bb->last_used_reg[bb->insn[i].op[k]] = REF (b, i);
170 
      }
171 
  }
172 
}
173 
 
174 
/* Set the BB limits */
175 
void detect_bb (cuc_func *f)
176 
{
177 
  int i, j, end_bb = 0, eb = 0;
178 
 
179 
  /* Mark block starts/ends */
180 
  for (i = 0; i < num_insn; i++) {
181 
    if (end_bb) insn[i].type |= IT_BBSTART;
182 
    end_bb = 0;
183 
    if (insn[i].type & IT_BRANCH) {
184 
      int jt = insn[i].op[0];
185 
      insn[i].type |= IT_BBEND;
186 
      end_bb = 1;
187 
      if (jt < 0 || jt >= num_insn) {
188 
        fprintf (stderr, "Instruction #%i:Jump out of function '%s'.\n", i, insn[i].disasm);
189 
        exit (1);
190 
      }
191 
      if (jt > 0) insn[jt - 1].type |= IT_BBEND;
192 
      insn[jt].type |= IT_BBSTART;
193 
    }
194 
  }
195 
 
196 
  /* Initialize bb array */
197 
  insn[0].type |= IT_BBSTART;
198 
  insn[num_insn - 1].type |= IT_BBEND;
199 
  f->num_bb = 0;
200 
  for (i = 0; i < num_insn; i++) {
201 
    if (insn[i].type & IT_BBSTART) {
202 
      f->bb[f->num_bb].first = i;
203 
      f->bb[f->num_bb].cnt = 0;
204 
    }
205 
    /* Determine repetitions of a loop */
206 
    if (insn[i].type & IT_BBEND) {
207 
      f->bb[f->num_bb].type = 0;
208 
      f->bb[f->num_bb].last = i;
209 
      f->bb[f->num_bb].next[0] = f->bb[f->num_bb].next[1] = -1;
210 
      f->bb[f->num_bb].tmp = 0;
211 
      f->bb[f->num_bb].ntim = 0;
212 
      f->num_bb++;
213 
      assert (f->num_bb < MAX_BB);
214 
    }
215 
  }
216 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INIT");
217 879 markom 
 
218 
  /* Build forward connections between BBs */
219 
  for (i = 0; i < f->num_bb; i++)
220 
    if (insn[f->bb[i].last].type & IT_BRANCH) {
221 
      int j;
222 
      assert (insn[f->bb[i].last].index == II_BF);
223 
      /* Find block this instruction jumps to */
224 
      for (j = 0; j < f->num_bb; j++)
225 
        if (f->bb[j].first == insn[f->bb[i].last].op[0]) break;
226 
      assert (j < f->num_bb);
227 
 
228 
      /* Convert the jump address to BB link */
229 
      insn[f->bb[i].last].op[0] = j; insn[f->bb[i].last].opt[0] = OPT_BB;
230 
 
231 
      /* Make a link */
232 
      f->bb[i].next[0] = j;
233 
      if (++f->bb[j].tmp > 2) eb++;
234 
      f->bb[i].next[1] = i + 1;
235 
      if (++f->bb[i + 1].tmp > 2) eb++;
236 
    } else if (f->bb[i].last == num_insn - 1) { /* Last instruction doesn't have to do anything */
237 
      f->bb[i].type |= BB_END;
238 
    } else {
239 
      f->bb[i].next[0] = i + 1;
240 
      if (++f->bb[i + 1].tmp > 2) eb++;
241 
    }
242 
 
243 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_NEXT");
244 879 markom 
 
245 
  /* Build backward connections, but first insert artificial blocks
246 
   * to handle more than 2 connections */
247 883 markom 
  cucdebug (6, "artificial %i %i\n", f->num_bb, eb);
248 879 markom 
  end_bb = f->num_bb + eb;
249 
  for (i = f->num_bb - 1; i >= 0; i--) {
250 
    j = f->bb[i].tmp;
251 
    if (f->bb[i].tmp > 2) f->bb[i].tmp = -f->bb[i].tmp;
252 
    f->bb[--end_bb] = f->bb[i];
253 
    reloc[i] = end_bb;
254 
    while (j-- > 2) {
255 
      f->bb[--end_bb].first = f->bb[i].first;
256 
      f->bb[end_bb].last = -1;
257 925 markom 
      f->bb[end_bb].type &= ~BB_END;
258 879 markom 
      f->bb[end_bb].next[0] = -1;
259 
      f->bb[end_bb].next[1] = -1;
260 
      f->bb[end_bb].tmp = 0;
261 
      f->bb[end_bb].cnt = f->bb[i].cnt;
262 
      f->bb[end_bb].ntim = 0;
263 
    }
264 
  }
265 
  f->num_bb += eb;
266 
 
267 
  /* relocate jump instructions */
268 
  for (i = 0; i < num_insn; i++)
269 
    for (j = 0; j < MAX_OPERANDS; j++)
270 
      if (insn[i].opt[j] & OPT_BB)
271 
        insn[i].op[j] = reloc[insn[i].op[j]];
272 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT-reloc");
273 879 markom 
  for (i = 0; i < f->num_bb; i++) {
274 
    if (f->bb[i].next[0] >= 0) {
275 
      int t = reloc[f->bb[i].next[0]];
276 
      if (f->bb[t].tmp < 0) {
277 
        f->bb[t].tmp = -f->bb[t].tmp;
278 
        t -= f->bb[t].tmp - 2;
279 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
280 
      f->bb[i].next[0] = t;
281 
    }
282 
    if (f->bb[i].next[1] >= 0) {
283 
      int t = reloc[f->bb[i].next[1]];
284 
      if (f->bb[t].tmp < 0) {
285 
        f->bb[t].tmp = -f->bb[t].tmp;
286 
        t -= f->bb[t].tmp - 2;
287 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
288 
      f->bb[i].next[1] = t;
289 
    }
290 
    /* artificial blocks do not have relocations, hardcode them */
291 
    if (f->bb[i].last < 0) f->bb[i].next[0] = i + 1;
292 
  }
293 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT");
294 879 markom 
 
295 
  /* Uncoditional branched do not continue to next block */
296 
  for (i = 0; i < f->num_bb; i++) {
297 
    cuc_insn *ii;
298 
    if (f->bb[i].last < 0) continue;
299 
    ii = &insn[f->bb[i].last];
300 
    /* Unconditional branch? */
301 
    if (ii->type & IT_BRANCH && ii->opt[1] & OPT_CONST) {
302 
      change_insn_type (ii, II_NOP);
303 1044 markom 
#if 0
304 879 markom 
      if (f->bb[i].next[1] == i + 1) f->bb[i].next[0] = f->bb[i].next[1];
305 1044 markom 
#endif
306 879 markom 
      f->bb[i].next[1] = -1;
307 
    }
308 1044 markom 
  }
309 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_UNCOND_JUMP");
310 879 markom 
 
311 
  /* Add backward connections */
312 
  for (i = 0; i < f->num_bb; i++)
313 
    f->bb[i].prev[0] = f->bb[i].prev[1] = -1;
314 
 
315 
  for (i = 0; i < f->num_bb; i++) {
316 
    if (f->bb[i].next[0] >= 0) {
317 
      int t = f->bb[i].next[0];
318 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
319 
      else {
320 
        assert (f->bb[t].prev[1] < 0);
321 
        f->bb[t].prev[1] = i;
322 
      }
323 
    }
324 
    if (f->bb[i].next[1] >= 0) {
325 
      int t = f->bb[i].next[1];
326 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
327 
      else {
328 
        assert (f->bb[t].prev[1] < 0);
329 
        f->bb[t].prev[1] = i;
330 
      }
331 
    }
332 
  }
333 932 markom 
  /* Add START marker */
334 
  assert (f->bb[0].prev[0] < 0);
335 
  f->bb[0].prev[0]= BBID_START;
336 
 
337 925 markom 
  /* Add END marker */
338 
  for (i = 0; i < f->num_bb; i++)
339 
    if (f->bb[i].type & BB_END) {
340 
      assert (f->bb[i].next[0] < 0);
341 
      f->bb[i].next[0] = BBID_END;
342 
    }
343 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_PREV");
344 879 markom 
}
345 
 
346 905 markom 
/* We do a quick check if there are some anomalies with references */
347 
void cuc_check (cuc_func *f)
348 
{
349 
  int i, j, k;
350 973 markom 
  cucdebug (1, "cuc_check\n");
351 905 markom 
  for (i = 0; i < f->num_bb; i++) {
352 930 markom 
    if (!f->bb[i].insn && f->bb[i].ninsn) goto err;
353 
    for (j = 0; j < f->bb[i].ninsn; j++) {
354 
      cuc_insn *ii = &f->bb[i].insn[j];
355 934 markom 
      if ((ii->index == II_CMOV || ii->index == II_ADD) && ii->type & IT_COND) {
356 930 markom 
        k = 0;
357 
        assert (ii->opt[k] & OPT_REGISTER);
358 
        if ((signed)ii->op[k] >= 0 && ii->op[k] != FLAG_REG && ii->op[k] != LRBB_REG) {
359 973 markom 
          cucdebug (1, "%x %x\n", ii->opt[0], ii->op[0]);
360 930 markom 
          goto err;
361 
        }
362 
      }
363 905 markom 
      for (k = 0; k < MAX_OPERANDS; k++)
364 930 markom 
        if (ii->opt[k] & OPT_REF) {
365 
          int t = ii->op[k];
366 
          if (REF_BB(t) >= f->num_bb || REF_I (t) >= f->bb[REF_BB(t)].ninsn
367 934 markom 
           || (ii->index == II_CMOV || ii->index == II_ADD) && (
368 930 markom 
                (f->INSN(t).type & IT_COND) != (ii->type & IT_COND) && k < 3
369 
              || !(f->INSN(t).type & IT_COND) && k == 3)) goto err;
370 905 markom 
        }
371 930 markom 
    }
372 905 markom 
  }
373 930 markom 
  return;
374 
err:
375 973 markom 
  cucdebug (1, "Anomaly detected at %x.%x[%i]\n", i, j, k);
376 930 markom 
  print_cuc_bb (f, "ANOMALY");
377 
  exit (1);
378 905 markom 
}
379 
 
380 879 markom 
/* Build basic blocks */
381 
void build_bb (cuc_func *f)
382 
{
383 
  int i, j, k;
384 
  for (i = 0; i < f->num_bb; i++) {
385 
    if (f->bb[i].last < 0) f->bb[i].ninsn = MAX_REGS - 1;
386 
    else f->bb[i].ninsn = f->bb[i].last - f->bb[i].first + 1 + MAX_REGS - 1;
387 
    assert (f->bb[i].ninsn >= MAX_REGS - 1);
388 
    f->bb[i].insn = (cuc_insn *) malloc (sizeof (cuc_insn) * f->bb[i].ninsn);
389 
    assert (f->bb[i].insn);
390 
    f->bb[i].nmemory = 0;
391 
    f->bb[i].unrolled = 1;
392 
 
393 
    /* Save space for conditional moves, exclude r0, place lrbb instead */
394 
    change_insn_type (&f->bb[i].insn[0], II_LRBB);
395 
    strcpy (f->bb[i].insn[0].disasm, "lrbb");
396 930 markom 
    f->bb[i].insn[0].type = IT_UNUSED | IT_COND;
397 879 markom 
    f->bb[i].insn[0].dep = NULL;
398 
    f->bb[i].insn[0].op[0] = LRBB_REG; f->bb[i].insn[0].opt[0] = OPT_REGISTER | OPT_DEST;
399 
    f->bb[i].insn[0].opt[1] = OPT_LRBB;
400 
    f->bb[i].insn[0].opt[2] = f->bb[i].insn[0].opt[3] = OPT_NONE;
401 
    for (j = 1; j < MAX_REGS - 1; j++) {
402 
      change_insn_type (&f->bb[i].insn[j], II_CMOV);
403 
      strcpy (f->bb[i].insn[j].disasm, "cmov");
404 930 markom 
      f->bb[i].insn[j].type = j == FLAG_REG || j == LRBB_REG ? IT_COND : 0;
405 879 markom 
      f->bb[i].insn[j].dep = NULL;
406 
      f->bb[i].insn[j].opt[0] = f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_REGISTER;
407 
      f->bb[i].insn[j].opt[0] |= OPT_DEST;
408 
      f->bb[i].insn[j].op[0] = f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = j;
409 
      f->bb[i].insn[j].op[3] = LRBB_REG; f->bb[i].insn[j].opt[3] = OPT_REGISTER;
410 
    }
411 897 markom 
 
412 
    /* Relocate instructions */
413 879 markom 
    for (j = MAX_REGS - 1; j < f->bb[i].ninsn; j++) {
414 
      f->bb[i].insn[j] = insn[f->bb[i].first + j - (MAX_REGS - 1)];
415 
      for (k = 0; k < MAX_OPERANDS; k++)
416 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
417 
          int b1;
418 
          for (b1 = 0; b1 < i; b1++)
419 898 markom 
            if (f->bb[b1].first <= (signed) f->bb[i].insn[j].op[k]
420 
              && (signed)f->bb[i].insn[j].op[k] <= f->bb[b1].last) break;
421 879 markom 
          assert (b1 < f->num_bb);
422 
          f->bb[i].insn[j].op[k] = REF (b1, f->bb[i].insn[j].op[k] - f->bb[b1].first + MAX_REGS - 1);
423 
        }
424 
      if (f->bb[i].insn[j].type & IT_MEMORY) f->bb[i].nmemory++;
425 
    }
426 
  }
427 905 markom 
  cuc_check (f);
428 879 markom 
}
429 
 
430 1042 markom 
/* Does simplification on blocks A, B, C:
431 
   A->B->C, A->C to just A->B->C */
432 
static void simplify_bb (cuc_func *f, int pred, int s1, int s2, int neg)
433 
{
434 
  cuc_insn *last;
435 
  int i;
436 
  if (cuc_debug >= 3) print_cuc_bb (f, "BEFORE_SIMPLIFY");
437 
  cucdebug (3, "simplify %x->%x->%x (%i)\n", pred, s1, s2, neg);
438 
  assert (s2 != pred); /* Shouldn't occur => stupid */
439 
  f->bb[pred].next[1] = -1;
440 
  f->bb[pred].next[0] = s1;
441 
 
442 
  if (f->bb[s2].prev[0] == pred) {
443 
    f->bb[s2].prev[0] = f->bb[s2].prev[1];
444 
    f->bb[s2].prev[1] = -1;
445 
  } else if (f->bb[s2].prev[1] == pred) {
446 
    f->bb[s2].prev[1] = -1;
447 
  } else assert (0);
448 
 
449 
  last = &f->bb[pred].insn[f->bb[pred].ninsn - 1];
450 
  assert (last->type & IT_BRANCH);
451 
  for (i = 0; i < f->bb[s2].ninsn; i++) {
452 
    cuc_insn *ii= &f->bb[s2].insn[i];
453 
    if (ii->index == II_LRBB) {
454 
      change_insn_type (ii, II_CMOV);
455 
      ii->type = IT_COND;
456 
      ii->op[1] = neg ? 0 : 1; ii->opt[1] = OPT_CONST;
457 
      ii->op[2] = neg ? 1 : 0; ii->opt[2] = OPT_CONST;
458 
      ii->op[3] = last->op[1]; ii->opt[3] = last->opt[1];
459 
    }
460 
  }
461 
  change_insn_type (last, II_NOP);
462 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_SIMPLIFY");
463 
}
464 
 
465 879 markom 
/* type == 0; keep predecessor condition
466 
 * type == 1; keep successor condition
467 
 * type == 2; join loop unrolled blocks */
468 
static void join_bb (cuc_func *f, int pred, int succ, int type)
469 
{
470 905 markom 
  int i, j, k, n1, n2, ninsn, add_cond = 0;
471 879 markom 
  unsigned long cond_op, cond_opt;
472 
  cuc_insn *insn;
473 
 
474 905 markom 
  if (cuc_debug) cuc_check (f);
475 897 markom 
  cucdebug (3, "%x <= %x+%x (%i)\n", pred, pred, succ, type);
476 
  cucdebug (3, "%x %x\n", f->bb[pred].ninsn, f->bb[succ].ninsn);
477 
  if (cuc_debug >= 3) fflush (stdout);
478 879 markom 
 
479 905 markom 
  n1 = f->bb[pred].ninsn;
480 
  n2 = f->bb[succ].ninsn;
481 
  if (n1 <= 0
482 
   || !(f->bb[pred].insn[n1 - 1].type & IT_BRANCH)) type = 1;
483 879 markom 
  if (type == 0 && f->bb[succ].prev[0] == f->bb[succ].next[0]) add_cond = 1;
484 905 markom 
  if (type == 2) add_cond = 1;
485 
 
486 1045 markom 
  //assert (f->bb[pred].next[0] == f->bb[succ].next[0] || type != 2); /* not supported */
487 879 markom 
 
488 905 markom 
  ninsn = n1 + n2 + (type == 1 ? 0 : 1) + (add_cond ? MAX_REGS : 0);
489 879 markom 
 
490 
  insn = (cuc_insn *) malloc (ninsn * sizeof (cuc_insn));
491 905 markom 
  for (i = 0; i < n1; i++) insn[i] = f->bb[pred].insn[i];
492 
  /* when type == 0, we move the last (jump) instruction to the end */
493 879 markom 
  if (type == 0 || type == 2) {
494 905 markom 
    /* Move first branch instruction to the end */
495 
    assert (insn[n1 - 1].type & IT_BRANCH);
496 
    insn[ninsn - 1] = insn[n1 - 1];
497 
    cond_op = insn[n1 - 1].op[1];
498 
    cond_opt = insn[n1 - 1].opt[1];
499 
 
500 
    /* Remove old branch */
501 
    change_insn_type (&insn[n1 - 1], II_NOP);
502 879 markom 
  }
503 
  /* Copy second block */
504 905 markom 
  for (i = 0; i < n2; i++) insn[i + n1] = f->bb[succ].insn[i];
505 902 markom 
 
506 
  /* and when type == 2, we may need to add sfor instruction, to quit when either is true */
507 
  if (type == 2) {
508 905 markom 
    /* Move second branch instruction to the end */
509 
    if (insn[n1 + n2 - 1].type & IT_BRANCH) {
510 
      insn[ninsn - 1] = insn[n1 + n2 - 1];
511 
 
512 
      /* Use conditional from cmov FLAG_REG, c_p, c_s, c_p */
513 
      insn[ninsn - 1].op[1] = REF (pred, n1 + n2 + FLAG_REG); insn[ninsn - 1].opt[1] = OPT_REF;
514 
 
515 
      /* Remove old one */
516 
      change_insn_type (&insn[n1 + n2 - 1], II_NOP);
517 
    } else change_insn_type (&insn[ninsn - 1], II_NOP); /* do not use branch slot */
518 902 markom 
  }
519 
 
520 905 markom 
#if 1
521 927 markom 
  /* LRBB at start of succ BB is not valid anymore */
522 933 markom 
  if (n1 > 0 && insn[n1].index == II_LRBB) {
523 925 markom 
    if (type == 1) {
524 
      /* We have two possibilities, how this could have happened:
525 
         1. we just moved second predecessor of succ to pred,
526 
            pred now having two predecessors => everything is ok
527 
         2. we just moved second predecessor of succ to pred,
528 
            now, having just one predecessor => LRBB is not needed anymore */
529 
      if (f->bb[pred].prev[1] < 0) { /* handle second option */
530 
        change_insn_type (&insn[n1], II_ADD);
531 
        insn[n1].op[1] = 1; insn[n1].opt[1] = OPT_CONST;
532 
        insn[n1].op[2] = 0; insn[n1].opt[2] = OPT_CONST;
533 
        insn[n1].opt[3] = OPT_NONE;
534 
      }
535 
    } else {
536 
      assert (0); /* not tested yet */
537 
      change_insn_type (&insn[n1], II_NOP);
538 
      for (i = n1; i < ninsn; i++)
539 
        if (insn[i].index == II_CMOV && insn[i].op[3] == REF (pred, n1)) {
540 
          assert (insn[i].opt[3] == OPT_REF);
541 
          insn[i].op[3] = cond_op;
542 
          insn[i].opt[3] = cond_opt;
543 
          if (f->bb[pred].next[0] != succ) {
544 
            unsigned long t; /* negate conditional -- exchange */
545 
            assert (f->bb[pred].next[1] == succ);
546 
            t = insn[i].op[1];
547 
            insn[i].op[1] = insn[i].op[2];
548 
            insn[i].op[2] = t;
549 
            t = insn[i].opt[1];
550 
            insn[i].opt[1] = insn[i].opt[2];
551 
            insn[i].opt[2] = t;
552 
          }
553 902 markom 
        }
554 925 markom 
    }
555 902 markom 
  }
556 905 markom 
#endif
557 879 markom 
 
558 
  for (i = 0; i < ninsn; i++) reloc[i] = -1;
559 
 
560 
  /* Add conditional instructions if required */
561 
  if (add_cond) {
562 
    recalc_last_used_reg (f, pred);
563 
    recalc_last_used_reg (f, succ);
564 1045 markom 
 
565 879 markom 
    /* r0 -- add nop for it */
566 905 markom 
    change_insn_type (&insn[n1 + n2], II_NOP);
567 879 markom 
    for (i = 1; i < MAX_REGS; i++) {
568 905 markom 
      cuc_insn *ii = &insn[n1 + n2 + i];
569 879 markom 
      int a = f->bb[pred].last_used_reg[i];
570 
      int b = f->bb[succ].last_used_reg[i];
571 
 
572 1045 markom 
      /* We have deleted first branch instruction, now we must setup FLAG_REG,
573 
         to point to conditional */
574 
      if (i == FLAG_REG) {
575 
        change_insn_type (ii, II_CMOV);
576 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
577 
        ii->dep = NULL;
578 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
579 
        ii->op[1] = cond_op; ii->opt[1] = cond_opt;
580 
        if (b >= 0) {
581 
          ii->op[2] = b; ii->opt[2] = OPT_REF;
582 
        } else {
583 
          ii->op[2] = cond_op; ii->opt[2] = cond_opt;
584 
        }
585 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
586 
        reloc[REF_I(a)] = REF (pred, n1 + n2 + i);
587 
      } else if (b < 0) change_insn_type (ii, II_NOP);
588 879 markom 
      else if (a < 0) {
589 
        change_insn_type (ii, II_ADD);
590 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
591 879 markom 
        ii->dep = NULL;
592 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
593 
        ii->op[1] = b; ii->opt[1] = OPT_REF;
594 
        ii->op[2] = 0; ii->opt[2] = OPT_CONST;
595 905 markom 
        ii->opt[3] = OPT_NONE;
596 879 markom 
      } else if (b >= 0) {
597 
        change_insn_type (ii, II_CMOV);
598 930 markom 
        ii->type = i == FLAG_REG || i == LRBB_REG ? IT_COND : 0;
599 879 markom 
        ii->dep = NULL;
600 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
601 
        ii->op[1] = a; ii->opt[1] = OPT_REF;
602 
        ii->op[2] = b; ii->opt[2] = OPT_REF;
603 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
604 905 markom 
        reloc[REF_I(a)] = REF (pred, n1 + n2 + i);
605 879 markom 
      }
606 
      sprintf (ii->disasm, "cmov (join BB)");
607 
    }
608 
  }
609 
 
610 905 markom 
  if (cuc_debug) cuc_check (f);
611 925 markom 
  if (f->bb[succ].type & BB_END) {
612 
    f->bb[pred].type |= BB_END;
613 933 markom 
    if (ninsn > 0 && insn[ninsn - 1].type & IT_BRANCH && insn[ninsn - 1].op[0] == succ) {
614 925 markom 
      assert (insn[ninsn - 1].opt[0] & OPT_BB);
615 
      insn[ninsn - 1].op[0] = BBID_END;
616 
    }
617 
  }
618 879 markom 
  i = 0;
619 925 markom 
  assert (f->bb[pred].next[0] >= 0 && f->bb[pred].next[0] != BBID_END);
620 879 markom 
  switch (type) {
621 
  case 0:
622 
    if (f->bb[pred].next[0] == succ) f->bb[pred].next[0] = f->bb[succ].next[0];
623 
    if (f->bb[pred].next[1] == succ) f->bb[pred].next[1] = f->bb[succ].next[0];
624 
    assert (f->bb[succ].next[1] < 0);
625 
    break;
626 
  case 1:
627 
    f->bb[pred].next[0] = f->bb[succ].next[0];
628 
    f->bb[pred].next[1] = f->bb[succ].next[1];
629 
    break;
630 905 markom 
  case 2:
631 
    f->bb[pred].next[0] = f->bb[succ].next[0];
632 
    f->bb[pred].next[1] = f->bb[succ].next[1];
633 
    break;
634 879 markom 
  }
635 924 markom 
  if (f->bb[pred].next[0] < 0) f->bb[pred].next[0] = f->bb[pred].next[1];
636 879 markom 
  if (f->bb[pred].next[0] == f->bb[pred].next[1]) f->bb[pred].next[1] = -1;
637 927 markom 
 
638 
  /* We just did something stupid -- we joined two predecessors into one;
639 
     succ may need the information from which block we came.  We will repair
640 
     this by converting LRBB to CMOV */
641 930 markom 
  for (j = 0; j < 2; j++) {
642 
    int nb = f->bb[pred].next[j];
643 927 markom 
    int t;
644 
 
645 
    /* check just valid connections */
646 
    if (nb < 0 || nb == BBID_END) continue;
647 
 
648 
    /* check type */
649 
    if (f->bb[nb].prev[0] == pred && f->bb[nb].prev[1] == succ) t = 1;
650 
    else if (f->bb[nb].prev[1] == pred && f->bb[nb].prev[0] == succ) t = 0;
651 
    else continue;
652 
 
653 930 markom 
    /* check all LRBB instructions.  */
654 
    for (i = 0; i < f->bb[nb].ninsn; i++)
655 
      if (f->bb[nb].insn[i].index == II_LRBB) {
656 
        cuc_insn *lrbb =&f->bb[nb].insn[i];
657 
        change_insn_type (lrbb, II_CMOV);
658 
        lrbb->op[1] = t; lrbb->opt[1] = OPT_CONST;
659 
        lrbb->op[2] = 1 - t; lrbb->opt[2] = OPT_CONST;
660 
        lrbb->op[3] = cond_op; lrbb->opt[3] = cond_opt;
661 
        lrbb->type |= IT_COND;
662 
      }
663 927 markom 
  }
664 
 
665 879 markom 
  f->bb[succ].type = BB_DEAD;
666 997 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);
667 905 markom 
  /* remove branch instruction, if there is only one successor */
668 933 markom 
  if (f->bb[pred].next[1] < 0 && ninsn > 0 && insn[ninsn - 1].type & IT_BRANCH) {
669 925 markom 
    assert (f->bb[pred].next[0] != pred); /* end BB, loop should not be possible */
670 905 markom 
    change_insn_type (&insn[ninsn - 1], II_NOP);
671 925 markom 
  }
672 879 markom 
 
673 
  /* Set max count */
674 
  if (f->bb[pred].cnt < f->bb[succ].cnt) f->bb[pred].cnt = f->bb[succ].cnt;
675 
  f->bb[pred].ninsn = ninsn;
676 905 markom 
  f->bb[succ].ninsn = 0;
677 879 markom 
  free (f->bb[pred].insn); f->bb[pred].insn = NULL;
678 
  free (f->bb[succ].insn); f->bb[succ].insn = NULL;
679 
  f->bb[pred].insn = insn;
680 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
681 
    if (f->bb[i].prev[0] == succ) f->bb[i].prev[0] = pred;
682 
    if (f->bb[i].prev[1] == succ) f->bb[i].prev[1] = pred;
683 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
684 
    for (j = 0; j < f->bb[i].ninsn; j++)
685 
      for (k = 0; k < MAX_OPERANDS; k++)
686 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
687 
          /* Check if we are referencing successor BB -> relocate to second part of
688 
             the new block */
689 
          if (REF_BB (f->bb[i].insn[j].op[k]) == succ) {
690 905 markom 
            int t = f->bb[i].insn[j].op[k];
691 
            int ndest = REF (pred, REF_I (t) + n1);
692 997 markom 
            //PRINTF ("%x: %x %x\n", REF(i, j), t, ndest);
693 879 markom 
 
694 905 markom 
            /* We've found a reference to succ. block, being removed, relocate */
695 
            f->bb[i].insn[j].op[k] = ndest;
696 879 markom 
          } else if (REF_BB(f->bb[i].insn[j].op[k]) == pred) {
697 
            if (i != pred && reloc[REF_I(f->bb[i].insn[j].op[k])] >= 0) {
698 
              f->bb[i].insn[j].op[k] = reloc[REF_I(f->bb[i].insn[j].op[k])];
699 
            }
700 
          }
701 
        }
702 
  }
703 
 
704 905 markom 
  if (cuc_debug) cuc_check (f);
705 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "join");
706 879 markom 
}
707 
 
708 
/* Optimize basic blocks */
709 931 markom 
int optimize_bb (cuc_func *f)
710 879 markom 
{
711 931 markom 
  int modified = 0;
712 879 markom 
  int i, j;
713 
remove_lrbb:
714 
  /* we can remove lrbb instructions from blocks with just one predecessor */
715 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
716 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* exactly one predecessor */
717 
      for (j = 0; j < f->bb[i].ninsn; j++)
718 
        if (f->bb[i].insn[j].index == II_LRBB) {
719 
          cuc_insn *t;
720 925 markom 
          cucdebug (4, "-lrbb %x.%x\n", i, j);
721 879 markom 
 
722 
          /* Change to add LRBB, 0, 0 */
723 
          change_insn_type (&f->bb[i].insn[j], II_ADD);
724 
          f->bb[i].insn[j].type &= ~IT_VOLATILE;
725 
          f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_CONST;
726 
          f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = 0; /* always use left block */
727 
          f->bb[i].insn[j].opt[3] = OPT_NONE;
728 931 markom 
          modified = 1;
729 934 markom 
          if (f->bb[i].prev[0] != BBID_START && f->bb[f->bb[i].prev[0]].ninsn > 0) {
730 932 markom 
            t = &f->bb[f->bb[i].prev[0]].insn[f->bb[f->bb[i].prev[0]].ninsn - 1];
731 879 markom 
 
732 932 markom 
            /* If the predecessor still has a conditional jump instruction, we must be careful.
733 
               If next[0] == next[1] join them. Now we will link lrbb and correct the situation */
734 
            if (t->type & IT_BRANCH) { /* We must set a reference to branch result */
735 
              f->bb[i].insn[j].opt[1] = t->opt[1];
736 
              f->bb[i].insn[j].op[1] = t->op[1];
737 
              /* sometimes branch is not needed anymore */
738 
              if (f->bb[f->bb[i].prev[0]].next[1] < 0) change_insn_type (t, II_NOP);
739 
            }
740 879 markom 
          }
741 
        }
742 
    }
743 
  }
744 
 
745 927 markom 
  /* Ordering of joining types is cruical -- we should concat all directly connected BBs
746 
     together first, so when we do a type != 1 joining, we can remove LRBB, directly by
747 
     looking at number of its predeccessors */
748 879 markom 
 
749 
  /* Type 1 joining
750 
     1. link between pred & succ
751 
     2. no other pred's successors
752 925 markom 
     3. no other succ's predecessors, except if pred has max one */
753 932 markom 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
754 
    int p = f->bb[i].prev[0];
755 
    if (p < 0 || p == BBID_START) continue;
756 925 markom 
    /* one successor and max sum of 3 predecessors */
757 932 markom 
    if (f->bb[p].next[0] >= 0 && f->bb[p].next[1] < 0
758 
     && (f->bb[p].prev[1] < 0 || f->bb[i].prev[1] < 0)) {
759 925 markom 
      /* First we will move all predecessors from succ to pred, and then we will do
760 
         real type 1 joining */
761 932 markom 
      if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START) {
762 
        int p1 = f->bb[i].prev[1];
763 925 markom 
        /* joining is surely not worth another extra memory access */
764 932 markom 
        if (f->bb[p].nmemory) continue;
765 
        if (f->bb[p].prev[0] >= 0) {
766 
           assert (f->bb[p].prev[1] < 0);
767 
           f->bb[p].prev[1] = p1;
768 
        } else f->bb[p].prev[0] = p1;
769 
        if (f->bb[p1].next[0] == i) f->bb[p1].next[0] = p;
770 
        else if (f->bb[p1].next[1] == i) f->bb[p1].next[1] = p;
771 925 markom 
        else assert (0);
772 
        f->bb[i].prev[1] = -1;
773 
      }
774 932 markom 
      assert (p >= 0 && f->bb[i].prev[1] < 0); /* one predecessor */
775 
      join_bb (f, p, i, 1);
776 931 markom 
      modified = 1;
777 879 markom 
      goto remove_lrbb;
778 
    }
779 932 markom 
  }
780 927 markom 
 
781 
  /* Type 0 joining
782 
     1. link between pred & succ
783 
     2. no memory accesses in succ
784 
     3. optional pred's second successors
785 
     4. max. one succ's successors */
786 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
787 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START
788 
     && f->bb[i].prev[1] < 0 /* one predecessor */
789 927 markom 
     && f->bb[i].next[1] < 0 /* max. one successor */
790 
     && f->bb[i].nmemory == 0) {                  /* and no memory acceses */
791 
      join_bb (f, f->bb[i].prev[0], i, 0);
792 931 markom 
      modified = 1;
793 927 markom 
      goto remove_lrbb;
794 
    }
795 879 markom 
 
796 
  /* Type 2 joining
797 
     1. link between pred & succ
798 
     2. succ has exactly one predeccessor
799 
     3. pred & succ share common successor
800 
     4. optional succ's second successor */
801 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
802 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* one predecessor */
803 
      int p = f->bb[i].prev[0];
804 932 markom 
      if (p == BBID_START) continue;
805 905 markom 
#if 0 /* not yet supported */
806 
      if (f->bb[p].next[0] == i
807 
       && (f->bb[i].next[1] == f->bb[p].next[1]
808 
        || f->bb[i].next[1] == f->bb[p].next[0])) {
809 
        join_bb (f, p, i, 2);
810 897 markom 
        goto remove_lrbb;
811 
      }
812 905 markom 
#endif
813 
      if (f->bb[p].next[1] == i
814 1045 markom 
       && (f->bb[p].next[0] == f->bb[i].next[1]
815 
        || f->bb[p].next[0] == f->bb[i].next[0])) {
816 905 markom 
        join_bb (f, p, i, 2);
817 931 markom 
        modified = 1;
818 905 markom 
        goto remove_lrbb;
819 
      }
820 879 markom 
    }
821 1042 markom 
 
822 
  /* BB simplify:
823 
     1. a block has exactly 2 successors A and B
824 
     2. A has exactly one successor -- B
825 
     3. A has no memory accesses
826 
     to:
827 
     flow always goes though A, LRBB is replaced by current block conditional
828 
    */
829 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
830 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END
831 
      && f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END) {
832 
      int a = f->bb[i].next[0];
833 
      int b = f->bb[i].next[1];
834 
      int neg = 0;
835 
      /* Exchange? */
836 
      if (f->bb[b].next[0] == a && f->bb[b].next[1] < 0) {
837 
        int t = a;
838 
        a = b;
839 
        b = t;
840 
        neg = 1;
841 
      }
842 
      /* Do the simplification if possible */
843 
      if (f->bb[a].next[0] == b && f->bb[a].next[1] < 0
844 
       && f->bb[a].nmemory == 0) {
845 
        simplify_bb (f, i, a, b, neg);
846 
        modified = 1;
847 
        goto remove_lrbb;
848 
      }
849 
    }
850 
 
851 931 markom 
  return modified;
852 879 markom 
}
853 
 
854 
/* Removes BBs marked as dead */
855 931 markom 
int remove_dead_bb (cuc_func *f)
856 879 markom 
{
857 
  int i, j, k, d = 0;
858 
 
859 
  for (i = 0; i < f->num_bb; i++) if (f->bb[i].type & BB_DEAD) {
860 
    if (f->bb[i].insn) free (f->bb[i].insn);
861 
    f->bb[i].insn = NULL;
862 
    reloc[i] = -1;
863 
  } else {
864 
    reloc[i] = d;
865 
    f->bb[d++] = f->bb[i];
866 
  }
867 931 markom 
  if (f->num_bb == d) return 0;
868 879 markom 
  f->num_bb = d;
869 
 
870 
  /* relocate initial blocks */
871 
  for (i = 0; i < f->num_init_bb; i++)
872 
    f->init_bb_reloc[i] = reloc[f->init_bb_reloc[i]];
873 
 
874 
  /* repair references */
875 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
876 973 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]);
877 925 markom 
          fflush (stdout);
878 932 markom 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[0] != BBID_START)
879 
      assert ((f->bb[i].prev[0] = reloc[f->bb[i].prev[0]]) >= 0);
880 
    if (f->bb[i].prev[1] >= 0 && f->bb[i].prev[1] != BBID_START)
881 
      assert ((f->bb[i].prev[1] = reloc[f->bb[i].prev[1]]) >= 0);
882 925 markom 
    if (f->bb[i].next[0] >= 0 && f->bb[i].next[0] != BBID_END)
883 
      assert ((f->bb[i].next[0] = reloc[f->bb[i].next[0]]) >= 0);
884 
    if (f->bb[i].next[1] >= 0 && f->bb[i].next[1] != BBID_END)
885 
      assert ((f->bb[i].next[1] = reloc[f->bb[i].next[1]]) >= 0);
886 879 markom 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
887 
    if (f->bb[i].next[0] == f->bb[i].next[1]) f->bb[i].next[1] = -1;
888 
 
889 
    for (j = 0; j < f->bb[i].ninsn; j++)
890 
      for (k = 0; k < MAX_OPERANDS; k++)
891 925 markom 
        if (f->bb[i].insn[j].opt[k] & OPT_BB && (signed)f->bb[i].insn[j].op[k] >= 0) {
892 
          if (f->bb[i].insn[j].op[k] != BBID_END)
893 
            assert ((f->bb[i].insn[j].op[k] = reloc[f->bb[i].insn[j].op[k]]) >= 0);
894 
        } else if (f->bb[i].insn[j].opt[k] & OPT_REF) {
895 879 markom 
          int t = f->bb[i].insn[j].op[k];
896 
          assert (reloc[REF_BB(t)] >= 0);
897 
          f->bb[i].insn[j].op[k] = REF (reloc[REF_BB(t)], REF_I (t));
898 
        }
899 
  }
900 931 markom 
  return 1;
901 879 markom 
}
902 
 
903 
/* Recursive calculation of dependencies */
904 
static int reg_dep_rec (cuc_func *f, int cur)
905 
{
906 
  int i, j;
907 
  cuc_insn *insn = f->bb[cur].insn;
908 
 
909 997 markom 
  //PRINTF ("\n %i", cur);
910 879 markom 
  /* Spread only, do not loop */
911 
  if (f->bb[cur].tmp) return;
912 
  f->bb[cur].tmp = 1;
913 997 markom 
  //PRINTF ("!   ");
914 879 markom 
 
915 
  for (i = 0; i < f->bb[cur].ninsn; i++) {
916 
    /* Check for destination operand(s) */
917 
    for (j = 0; j < MAX_OPERANDS; j++) if (insn[i].opt[j] & OPT_DEST)
918 
      if ((insn[i].opt[j] & ~OPT_DEST) == OPT_REGISTER && (signed)insn[i].op[j] >= 0) {
919 997 markom 
        //PRINTF ("%i:%i,%x ", insn[i].op[j], i, REF (cur, i));
920 879 markom 
        assert (insn[i].op[j] > 0 && insn[i].op[j] < MAX_REGS); /* r0 should never be dest */
921 
        f->bb[cur].last_used_reg[insn[i].op[j]] = REF (cur, i);
922 
      }
923 
  }
924 
 
925 925 markom 
  if (f->bb[cur].next[0] >= 0 && f->bb[cur].next[0] != BBID_END)
926 
    reg_dep_rec (f, f->bb[cur].next[0]);
927 
  if (f->bb[cur].next[1] >= 0 && f->bb[cur].next[1] != BBID_END)
928 
    reg_dep_rec (f, f->bb[cur].next[1]);
929 879 markom 
}
930 
 
931 
/* Detect register dependencies */
932 
void reg_dep (cuc_func *f)
933 
{
934 
  int i, b, c;
935 
 
936 
  /* Set dead blocks */
937 
  for (b = 0; b < f->num_bb; b++) {
938 
    f->bb[b].tmp = 0;
939 
    for (i = 0; i < MAX_REGS; i++) f->bb[b].last_used_reg[i] = -1;
940 
  }
941 
 
942 
  /* Start with first block and set dependecies of all reachable blocks */
943 
  /* At the same time set last_used_regs */
944 
  reg_dep_rec (f, 0);
945 
 
946 
  for (i = 0; i < f->num_bb; i++)
947 
    if (f->bb[i].tmp) f->bb[i].tmp = 0;
948 
    else f->bb[i].type |= BB_DEAD;
949 
 
950 
  /* Detect loops; mark BBs where loops must be broken */
951 
  for (c = 0; c < f->num_bb; c++) {
952 
    int min = 3, minb;
953 
 
954 
    /* search though all non-visited for minimum number of unvisited predecessors */
955 
    for (b = 0; b < f->num_bb; b++) if (!f->bb[b].tmp) {
956 
      int tmp = 0;
957 932 markom 
      if (f->bb[b].prev[0] >= 0 && f->bb[b].prev[0] != BBID_START
958 
       && !f->bb[f->bb[b].prev[0]].tmp) tmp++;
959 
      if (f->bb[b].prev[1] >= 0 && f->bb[b].prev[1] != BBID_START
960 
       && !f->bb[f->bb[b].prev[1]].tmp) tmp++;
961 879 markom 
      if (tmp < min) {
962 
        minb = b;
963 
        min = tmp;
964 
        if (tmp == 0) break; /* We already have the best one */
965 
      }
966 
    }
967 
    b = minb;
968 
    f->bb[b].tmp = 1; /* Mark visited */
969 883 markom 
    cucdebug (3, "minb %i min %i\n", minb, min);
970 879 markom 
    if (min) { /* We just broke the loop */
971 
      f->bb[b].type |= BB_INLOOP;
972 
    }
973 
  }
974 
 
975 
  /* Set real predecessors in cmov instructions to previous blocks */
976 
  for (b = 0; b < f->num_bb; b++)
977 
    for (i = 1; i < MAX_REGS - 1; i++) {
978 
      int pa, pb;
979 
      assert (f->bb[b].insn[i].index ==  II_CMOV);
980 
      assert (f->bb[b].insn[i].opt[0] == OPT_REGISTER | OPT_DEST);
981 
      assert (f->bb[b].insn[i].op[0] == i);
982 932 markom 
      if (f->bb[b].prev[0] < 0 || f->bb[b].prev[0] == BBID_START) pa = -1;
983 879 markom 
      else pa = f->bb[f->bb[b].prev[0]].last_used_reg[i];
984 932 markom 
      if (f->bb[b].prev[1] < 0 || f->bb[b].prev[1] == BBID_START) pb = -1;
985 879 markom 
      else pb = f->bb[f->bb[b].prev[1]].last_used_reg[i];
986 
 
987 
      /* We do some very simple optimizations right away to make things more readable */
988 
      if (pa < 0 && pb < 0) {
989 
        /* Was not used at all */
990 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
991 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
992 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
993 
      } else if (pa < 0) {
994 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
995 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
996 
        f->bb[b].insn[i].op[1] = pb; f->bb[b].insn[i].opt[1] = OPT_REF;
997 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
998 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
999 
      } else if (pb < 0) {
1000 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
1001 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
1002 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
1003 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
1004 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
1005 
      } else {
1006 
        int t = REF (b, 0); /* lrbb should be first instruction */
1007 
        assert (f->INSN(t).index == II_LRBB);
1008 
 
1009 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
1010 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
1011 
 
1012 
        f->bb[b].insn[i].op[2] = pb; f->bb[b].insn[i].opt[2] = OPT_REF;
1013 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
1014 
 
1015 
        /* Update op[3] -- flag register */
1016 
        assert (f->bb[b].insn[i].opt[3] == OPT_REGISTER);
1017 
        assert (f->bb[b].insn[i].op[3] == LRBB_REG);
1018 
        assert (t >= 0);
1019 
        f->bb[b].insn[i].opt[3] = OPT_REF; /* Convert already used regs to references */
1020 
        f->bb[b].insn[i].op[3] = t;
1021 
        assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1022 
      }
1023 
    }
1024 
 
1025 
  /* assign register references */
1026 
  for (b = 0; b < f->num_bb; b++) {
1027 
    /* rebuild last used reg array */
1028 
    f->bb[b].last_used_reg[0] = -1;
1029 
    if (f->bb[b].insn[0].index == II_LRBB) f->bb[b].last_used_reg[LRBB_REG] = 0;
1030 
    else f->bb[b].last_used_reg[LRBB_REG] = -1;
1031 
 
1032 
    for (i = 1; i < MAX_REGS - 1; i++)
1033 
      f->bb[b].last_used_reg[i] = -1;
1034 
 
1035 
    /* Create references */
1036 
    for (i = 0; i < f->bb[b].ninsn; i++) {
1037 
      int k;
1038 
      /* Check for source operands first */
1039 
      for (k = 0; k < MAX_OPERANDS; k++) {
1040 
        if (!(f->bb[b].insn[i].opt[k] & OPT_DEST))
1041 
        if (f->bb[b].insn[i].opt[k] & OPT_REGISTER) {
1042 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1043 
 
1044 
          if (f->bb[b].insn[i].op[k] == 0) { /* Convert r0 to const0 */
1045 
            f->bb[b].insn[i].opt[k] = OPT_CONST;
1046 
            f->bb[b].insn[i].op[k] = 0;
1047 
          } else if (t >= 0) {
1048 
            f->bb[b].insn[i].opt[k] = OPT_REF; /* Convert already used regs to references */
1049 
            f->bb[b].insn[i].op[k] = t;
1050 
            assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
1051 
            //f->INSN(t).op[0] = -1;
1052 
          }
1053 
        } else if (f->bb[b].insn[i].opt[k] & OPT_REF) {
1054 
          //f->INSN(f->bb[b].insn[i].op[k]).op[0] = -1; /* Mark referenced */
1055 
          f->INSN(f->bb[b].insn[i].op[k]).type &= ~IT_UNUSED;
1056 
        }
1057 
      }
1058 
 
1059 
      /* Now check for destination operand(s) */
1060 
      for (k = 0; k < MAX_OPERANDS; k++) if (f->bb[b].insn[i].opt[k] & OPT_DEST)
1061 
        if ((f->bb[b].insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
1062 
          && (int)f->bb[b].insn[i].op[k] >= 0) {
1063 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
1064 
          assert (f->bb[b].insn[i].op[k] != 0); /* r0 should never be dest */
1065 
          f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]] = REF (b, i);
1066 
        }
1067 
    }
1068 
  }
1069 
 
1070 
  /* Remove all unused lrbb */
1071 
  for (b = 0; b < f->num_bb; b++)
1072 
    for (i = 0; i < f->bb[b].ninsn; i++)
1073 
      if (f->bb[b].insn[i].type & IT_UNUSED) change_insn_type (&f->bb[b].insn[i], II_NOP);
1074 
 
1075 
  /* SSAs with final register value are marked as outputs */
1076 
  assert (f->bb[f->num_bb - 1].type & BB_END);
1077 939 markom 
  for (i = 0; i < MAX_REGS; i++) if (!caller_saved[i]) {
1078 879 markom 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1079 
    /* Mark them volatile, so optimizer does not remove them */
1080 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1081 
  }
1082 939 markom 
  {
1083 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
1084 
    /* Mark them volatile, so optimizer does not remove them */
1085 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
1086 
  }
1087 879 markom 
}
1088 
 
1089 897 markom 
/* split the BB, based on the group numbers in .tmp */
1090 
void expand_bb (cuc_func *f, int b)
1091 
{
1092 
  int n = f->num_bb;
1093 
  int mg = 0;
1094 
  int b1, i, j;
1095 
 
1096 
  for (i = 0; i < f->bb[b].ninsn; i++)
1097 
    if (f->bb[b].insn[i].tmp > mg) mg = f->bb[b].insn[i].tmp;
1098 
 
1099 
  /* Create copies */
1100 
  for (b1 = 1; b1 <= mg; b1++) {
1101 
    assert (f->num_bb < MAX_BB);
1102 
    cpy_bb (&f->bb[f->num_bb], &f->bb[b]);
1103 
    f->num_bb++;
1104 
  }
1105 
 
1106 
  /* Relocate */
1107 
  for (b1 = 0; b1 < f->num_bb; b1++)
1108 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1109 
      dep_list *d = f->bb[b1].insn[i].dep;
1110 
      for (j = 0; j < MAX_OPERANDS; j++)
1111 
        if (f->bb[b1].insn[i].opt[j] & OPT_REF) {
1112 
          int t = f->bb[b1].insn[i].op[j];
1113 
          if (REF_BB(t) == b && f->INSN(t).tmp != 0)
1114 
            f->bb[b1].insn[i].op[j] = REF (n + f->INSN(t).tmp - 1, REF_I(t));
1115 
        }
1116 
      while (d) {
1117 
        if (REF_BB (d->ref) == b && f->INSN(d->ref).tmp != 0)
1118 
          d->ref = REF (n + f->INSN(d->ref).tmp - 1, REF_I(d->ref));
1119 
        d = d->next;
1120 
      }
1121 
    }
1122 
 
1123 
  /* Delete unused instructions */
1124 
  for (j = 0; j <= mg; j++) {
1125 
    if (j == 0) b1 = b;
1126 
    else b1 = n + j - 1;
1127 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
1128 
      if (f->bb[b1].insn[i].tmp != j)
1129 
        change_insn_type (&f->bb[b1].insn[i], II_NOP);
1130 
      f->bb[b1].insn[i].tmp = 0;
1131 
    }
1132 
    if (j < mg) {
1133 
      f->bb[b1].next[0] = n + j;
1134 
      f->bb[b1].next[1] = -1;
1135 
      f->bb[n + j].prev[0] = b1;
1136 
      f->bb[n + j].prev[1] = -1;
1137 
    } else {
1138 
      i = f->bb[b1].next[0];
1139 
      f->bb[n + j].prev[0] = j == 1 ? b : b1 - 1;
1140 
      f->bb[n + j].prev[1] = -1;
1141 925 markom 
      if (i >= 0 && i != BBID_END) {
1142 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1143 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1144 
      }
1145 
      i = f->bb[b1].next[1];
1146 925 markom 
      if (i >= 0 && i != BBID_END) {
1147 897 markom 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
1148 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
1149 
      }
1150 
    }
1151 
  }
1152 
}
1153 
 
1154 879 markom 
/* Scans sequence of BBs and set bb[].cnt */
1155 
void generate_bb_seq (cuc_func *f, char *mp_filename, char *bb_filename)
1156 
{
1157 
  FILE *fi, *fo;
1158 
  struct mprofentry_struct *buf;
1159 
  const int bufsize = 256;
1160 
  unsigned long *bb_start;
1161 
  unsigned long *bb_end;
1162 
  int b, i, r;
1163 
  int curbb, prevbb = -1;
1164 
  unsigned long addr = -1;
1165 
  unsigned long prevaddr = -1;
1166 897 markom 
  int mssum = 0;
1167 
  int mlsum = 0;
1168 
  int mscnt = 0;
1169 
  int mlcnt = 0;
1170 940 markom 
  int reopened = 0;
1171 879 markom 
 
1172 940 markom 
  /* Use already opened stream? */
1173 
  if (runtime.sim.fmprof) {
1174 
    fi = runtime.sim.fmprof;
1175 
    reopened = 1;
1176 
    rewind (fi);
1177 
  } else assert (fi = fopen (mp_filename, "rb"));
1178 879 markom 
  assert (fo = fopen (bb_filename, "wb+"));
1179 
 
1180 
  assert (bb_start = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1181 
  assert (bb_end = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
1182 
  for (b = 0; b < f->num_bb; b++) {
1183 
    bb_start[b] = f->start_addr + f->bb[b].first * 4;
1184 
    bb_end[b] = f->start_addr + f->bb[b].last * 4;
1185 997 markom 
    //PRINTF ("%i %x %x\n", b, bb_start[b], bb_end[b]);
1186 879 markom 
    f->bb[0].cnt = 0;
1187 
  }
1188 
 
1189 
  buf = (struct mprofentry_struct *) malloc (sizeof (struct mprofentry_struct) * bufsize);
1190 
  assert (buf);
1191 
 
1192 997 markom 
  //PRINTF ("BBSEQ:\n");
1193 879 markom 
  do {
1194 
    r = fread (buf, sizeof (struct mprofentry_struct), bufsize, fi);
1195 997 markom 
    //PRINTF ("r%i : ", r);
1196 879 markom 
    for (i = 0; i < r; i++) {
1197 
      if (buf[i].type & MPROF_FETCH) {
1198 997 markom 
        //PRINTF ("%x, ", buf[i].addr);
1199 879 markom 
        if (buf[i].addr >= f->start_addr && buf[i].addr <= f->end_addr) {
1200 
          assert (buf[i].type & MPROF_32);
1201 
          prevaddr = addr;
1202 
          addr = buf[i].addr;
1203 
          for (b = 0; b < f->num_bb; b++)
1204 
            if (bb_start[b] <= addr && addr <= bb_end[b]) break;
1205 
          assert (b < f->num_bb);
1206 
          curbb = b;
1207 
          if (prevaddr + 4 != addr) prevbb = -1;
1208 
        } else curbb = -1;
1209 
 
1210 
#warning TODO: do not count interrupts
1211 
        if (curbb != prevbb && curbb >= 0) {
1212 
          fwrite (&curbb, sizeof (unsigned long), 1, fo);
1213 997 markom 
          //PRINTF (" [%i] ", curbb);
1214 879 markom 
          f->bb[curbb].cnt++;
1215 
          prevbb = curbb;
1216 
        }
1217 897 markom 
      } else {
1218 
        if (verify_memoryarea(buf[i].addr))
1219 
          if (buf[i].type & MPROF_WRITE) mscnt++, mssum += cur_area->delayw;
1220 1041 markom 
          else mlcnt++, mlsum += cur_area->delayr;
1221 879 markom 
      }
1222 
    }
1223 997 markom 
    //PRINTF ("\n");
1224 879 markom 
  } while (r == bufsize);
1225 997 markom 
  //PRINTF ("\n");
1226 879 markom 
 
1227 897 markom 
  runtime.cuc.mdelay[0] = (1. * mlsum) / mlcnt;
1228 1041 markom 
  runtime.cuc.mdelay[1] = (1. * mssum) / mscnt;
1229 897 markom 
  runtime.cuc.mdelay[2] = runtime.cuc.mdelay[3] = 1;
1230 883 markom 
  f->num_runs = f->bb[0].cnt;
1231 940 markom 
  if (!reopened) fclose (fi);
1232 879 markom 
  fclose (fo);
1233 
  free (buf);
1234 
  free (bb_end);
1235 
  free (bb_start);
1236 
 
1237 
  /* Initialize basic block relocations */
1238 
  f->num_init_bb = f->num_bb;
1239 997 markom 
  //PRINTF ("num_init_bb = %i\n", f->num_init_bb);
1240 879 markom 
  assert (f->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
1241 
  for (b = 0; b < f->num_init_bb; b++) f->init_bb_reloc[b] = b;
1242 
}
1243 
 
1244 
/* Scans sequence of BBs and set counts for pre/unrolled loop for BB b */
1245 
void count_bb_seq (cuc_func *f, int b, char *bb_filename, int *counts, int preroll, int unroll)
1246 
{
1247 
  FILE *fi;
1248 
  const int bufsize = 256;
1249 
  int i, r;
1250 
  int *buf;
1251 
  int cnt = 0;
1252 
  int times = preroll - 1 + unroll;
1253 
 
1254 
  assert (fi = fopen (bb_filename, "rb"));
1255 
  for (i = 0; i < times; i++) counts[i] = 0;
1256 
  assert (buf = (int *) malloc (sizeof (int) * bufsize));
1257 
 
1258 
  do {
1259 
    r = fread (buf, sizeof (int), bufsize, fi);
1260 
    for (i = 0; i < r; i++) {
1261 
      /* count consecutive acesses */
1262 
      if (f->init_bb_reloc[buf[i]] == b) {
1263 
        counts[cnt]++;
1264 
        if (++cnt >= times) cnt = preroll - 1;
1265 
      } else cnt = 0;
1266 
    }
1267 
  } while (r == bufsize);
1268 
 
1269 
  log ("Counts %i,%i :", preroll, unroll);
1270 
  for (i = 0; i < times; i++) log ("%x ", counts[i]);
1271 
  log ("\n");
1272 
 
1273 
  fclose (fi);
1274 
  free (buf);
1275 
}
1276 
 
1277 
/* relocate all accesses inside of BB b to back/fwd */
1278 
static void relocate_bb (cuc_bb *bb, int b, int back, int fwd)
1279 
{
1280 
  int i, j;
1281 
  for (i = 0; i < bb->ninsn; i++)
1282 
    for (j = 0; j < MAX_OPERANDS; j++)
1283 
      if (bb->insn[i].opt[j] & OPT_REF
1284 
       && REF_BB (bb->insn[i].op[j]) == b) {
1285 
        int t = REF_I (bb->insn[i].op[j]);
1286 
        if (t < i) bb->insn[i].op[j] = REF (back, t);
1287 
        else bb->insn[i].op[j] = REF (fwd, t);
1288 
      }
1289 
}
1290 
 
1291 
/* Unroll loop b unroll times and return new function. Original
1292 
   function is unmodified. */
1293 
static cuc_func *unroll_loop (cuc_func *f, int b, int unroll)
1294 
{
1295 
  int b1, t, i, j, prevb, prevart_b;
1296 
  cuc_func *n = dup_func (f);
1297 
  cuc_bb *ob = &f->bb[b];
1298 
  cuc_insn *ii;
1299 
 
1300 
  assert (unroll > 1);
1301 997 markom 
  //PRINTF ("unroll BB%i x %i (num_bb %i)\n", b, unroll, n->num_bb);
1302 879 markom 
  unroll--;
1303 
  assert (n->num_bb + unroll * 2 < MAX_BB);
1304 
 
1305 
  prevb = b;
1306 
  prevart_b = b;
1307 
  /* Duplicate the BB */
1308 
  for (t = 0; t < unroll; t++) {
1309 
    cuc_bb *pb = &n->bb[prevart_b];
1310 
    /* Add new block and set links */
1311 
    b1 = n->num_bb++;
1312 
    cpy_bb (&n->bb[b1], ob);
1313 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1314 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1315 
 
1316 
    /* Set predecessor's successor */
1317 
    if (n->bb[prevb].next[0] == b) {
1318 
      n->bb[prevb].next[0] = b1;
1319 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1320 
      else pb->next[1] = b1 + 1;
1321 
      n->bb[b1].next[1] = b1 + 1;
1322 
    } else if (n->bb[prevb].next[1] == b) {
1323 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1324 
      else pb->next[1] = b1 + 1;
1325 
      n->bb[b1].next[0] = b1 + 1;
1326 
      n->bb[prevb].next[1] = b1;
1327 
    } else assert (0);
1328 
 
1329 
    /* Set predecessor */
1330 
    n->bb[b1].prev[0] = prevb;
1331 
    n->bb[b1].prev[1] = -1;
1332 
 
1333 
    /* Relocate backward references to current instance and forward references
1334 
       to previous one */
1335 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1336 
 
1337 
    /* add artificial block, just to join accesses */
1338 
    b1 = n->num_bb++;
1339 
    cpy_bb (&n->bb[b1], ob);
1340 
    n->bb[b1].cnt = 0;
1341 
 
1342 
    for (i = 0; i < ob->ninsn - 1; i++) {
1343 
      ii = &n->bb[b1].insn[i];
1344 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1345 
        change_insn_type (ii, II_CMOV);
1346 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1347 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1348 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1349 
 
1350 
        /* Take left one, if we should have finished the first iteration*/
1351 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1352 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1353 
        } else {
1354 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1355 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1356 
        }
1357 
        ii->dep = NULL;
1358 930 markom 
        ii->type = ob->insn[i].type & IT_COND;
1359 879 markom 
      } else {
1360 
        change_insn_type (ii, II_NOP);
1361 
      }
1362 
    }
1363 
 
1364 928 markom 
    /* Add conditional or instruction at the end, prioritizing flags */
1365 879 markom 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1366 928 markom 
    change_insn_type (ii, II_CMOV);
1367 879 markom 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1368 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1369 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1370 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1371 
    } else {
1372 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1373 
      ii->opt[1] = OPT_REF;
1374 
    }
1375 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1376 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1377 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1378 
    } else {
1379 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1380 
      ii->opt[2] = OPT_REF;
1381 
    }
1382 928 markom 
    /* {z = x || y;} is same as {z = x ? x : y;} */
1383 
    ii->op[3] = ii->op[1]; ii->opt[3] = ii->opt[1];
1384 933 markom 
    ii->type = IT_COND;
1385 879 markom 
 
1386 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1387 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1388 
    n->bb[b1].prev[0] = prevart_b;
1389 
    n->bb[b1].prev[1] = b1 - 1;
1390 925 markom 
    n->bb[b1].next[0] = -1;
1391 879 markom 
    n->bb[b1].next[1] = -1;
1392 
 
1393 
    prevb = b1 - 1;
1394 
    prevart_b = b1;
1395 
  }
1396 
  if (ob->type & BB_END) {
1397 
    n->bb[prevart_b].type |= BB_END;
1398 
    n->bb[b].type &= ~BB_END;
1399 
  }
1400 
 
1401 925 markom 
  n->bb[prevart_b].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1402 879 markom 
  //print_cuc_bb (n, "unroll1");
1403 
  /* repair BB after loop, to point back to latest artificial BB */
1404 
  b1 = n->bb[prevart_b].next[0];
1405 925 markom 
  if (b1 >= 0 && b1 != BBID_END) {
1406 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1407 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1408 879 markom 
    else assert (0);
1409 
  }
1410 
 
1411 
  /* Relink back to start of the loop */
1412 
  /* Set predecessor's successor */
1413 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = b;
1414 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = b;
1415 
  else assert (0);
1416 
 
1417 
  /* Set predecessor */
1418 
  if (n->bb[b].prev[0] == b) n->bb[b].prev[0] = prevb;
1419 
  else if (n->bb[b].prev[1] == b) n->bb[b].prev[1] = prevb;
1420 
  else assert (0);
1421 
 
1422 
  //print_cuc_bb (n, "unroll2");
1423 
 
1424 
  /* Relocate backward references to current instance and forward references
1425 
     to previous one */
1426 
  relocate_bb (&n->bb[b], b, b, prevb);
1427 
 
1428 
  /* Relocate all other blocks to point to latest prevart_b */
1429 
  for (i = 0; i < f->num_bb; i++)
1430 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1431 
 
1432 
  return n;
1433 
}
1434 
 
1435 
/* Preroll loop b preroll times and return new function. Original
1436 
   function is unmodified. */
1437 
static cuc_func *preroll_loop (cuc_func *f, int b, int preroll)
1438 
{
1439 
  int b1, t, i, j, prevb, prevart_b;
1440 
  cuc_func *n = dup_func (f);
1441 
  cuc_bb *ob = &f->bb[b];
1442 
  cuc_insn *ii;
1443 
 
1444 
  assert (preroll > 1);
1445 997 markom 
  //PRINTF ("preroll BB%i x %i (num_bb %i)\n", b, preroll, n->num_bb);
1446 879 markom 
  preroll--;
1447 
  assert (n->num_bb + preroll * 2 < MAX_BB);
1448 
 
1449 
  prevb = b;
1450 
  prevart_b = b;
1451 
  /* Duplicate the BB */
1452 
  for (t = 0; t < preroll; t++) {
1453 
    cuc_bb *pb = &n->bb[prevart_b];
1454 
    /* Add new block and set links */
1455 
    b1 = n->num_bb++;
1456 
    cpy_bb (&n->bb[b1], ob);
1457 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1458 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1459 
 
1460 
    /* Set predecessor's successor */
1461 
    if (n->bb[prevb].next[0] == b) {
1462 
      n->bb[prevb].next[0] = b1;
1463 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1464 
      else pb->next[1] = b1 + 1;
1465 
      n->bb[b1].next[1] = b1 + 1;
1466 
    } else if (n->bb[prevb].next[1] == b) {
1467 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1468 
      else pb->next[1] = b1 + 1;
1469 
      n->bb[b1].next[0] = b1 + 1;
1470 
      n->bb[prevb].next[1] = b1;
1471 
    } else assert (0);
1472 
 
1473 
    /* Set predecessor */
1474 
    n->bb[b1].prev[0] = prevb;
1475 
    n->bb[b1].prev[1] = -1;
1476 
 
1477 
    /* Relocate backward references to current instance and forward references
1478 
       to previous one */
1479 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1480 
 
1481 
    /* add artificial block, just to join accesses */
1482 
    b1 = n->num_bb++;
1483 
    cpy_bb (&n->bb[b1], ob);
1484 
    n->bb[b1].cnt = 0;
1485 
 
1486 
    for (i = 0; i < ob->ninsn - 1; i++) {
1487 
      ii = &n->bb[b1].insn[i];
1488 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1489 
        change_insn_type (ii, II_CMOV);
1490 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1491 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1492 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1493 
 
1494 
        /* Take left one, if we should have finished the first iteration*/
1495 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1496 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1497 
        } else {
1498 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1499 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1500 
        }
1501 
        ii->dep = NULL;
1502 930 markom 
        ii->type = ob->insn[i].type & IT_COND;
1503 879 markom 
      } else {
1504 
        change_insn_type (ii, II_NOP);
1505 
      }
1506 
    }
1507 
 
1508 928 markom 
    /* Add conditional or instruction at the end, prioritizing flags */
1509 879 markom 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1510 928 markom 
    change_insn_type (ii, II_CMOV);
1511 879 markom 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1512 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1513 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1514 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1515 
    } else {
1516 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1517 
      ii->opt[1] = OPT_REF;
1518 
    }
1519 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1520 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1521 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1522 
    } else {
1523 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1524 
      ii->opt[2] = OPT_REF;
1525 
    }
1526 928 markom 
    /* {z = x || y;} is same as {z = x ? x : y;} */
1527 
    ii->op[3] = ii->op[1]; ii->opt[3] = ii->opt[1];
1528 933 markom 
    ii->type = IT_COND;
1529 879 markom 
 
1530 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1531 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1532 
    n->bb[b1].prev[0] = prevart_b;
1533 
    n->bb[b1].prev[1] = b1 - 1;
1534 925 markom 
    n->bb[b1].next[0] = -1;
1535 879 markom 
    n->bb[b1].next[1] = -1;
1536 
 
1537 
    prevb = b1 - 1;
1538 
    prevart_b = b1;
1539 
  }
1540 
  if (ob->type & BB_END) {
1541 
    n->bb[prevart_b].type |= BB_END;
1542 
    n->bb[b].type &= ~BB_END;
1543 
  }
1544 925 markom 
  n->bb[prevart_b].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1545 879 markom 
 
1546 
  //print_cuc_bb (n, "preroll1");
1547 
  /* repair BB after loop, to point back to latest artificial BB */
1548 
  b1 = n->bb[prevart_b].next[0];
1549 925 markom 
  if (b1 >= 0 && b1 != BBID_END) {
1550 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1551 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1552 879 markom 
    else assert (0);
1553 
  }
1554 
 
1555 
  /* Relink to itself */
1556 
  /* Set predecessor's successor */
1557 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = prevb;
1558 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = prevb;
1559 
  else assert (0);
1560 
  n->bb[prevb].prev[1] = prevb;
1561 
 
1562 
  if (n->bb[b].prev[0] == b) {
1563 
    n->bb[b].prev[0] = n->bb[b].prev[1];
1564 
    n->bb[b].prev[1] = -1;
1565 
  } else if (n->bb[b].prev[1] == b) {
1566 
    n->bb[b].prev[1] = -1;
1567 
  }
1568 
 
1569 
  //print_cuc_bb (n, "preroll2");
1570 
 
1571 
  /* Relocate backward references to current instance and forward references
1572 
     to previous one */
1573 
  relocate_bb (&n->bb[b], b, b, prevb);
1574 
 
1575 
  /* Relocate all other blocks to point to latest prevart_b */
1576 
  for (i = 0; i < f->num_bb; i++)
1577 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1578 
 
1579 
  return n;
1580 
}
1581 
 
1582 
/* Unroll loop b unroll times and return new function. Original
1583 
   function is unmodified. */
1584 
cuc_func *preunroll_loop (cuc_func *f, int b, int preroll, int unroll, char *bb_filename)
1585 
{
1586 
  int b1, i;
1587 
  cuc_func *n, *t;
1588 
  int *counts;
1589 
  int *bb_reloc;
1590 
 
1591 
  if (preroll > 1) {
1592 
    t = preroll_loop (f, b, preroll);
1593 
    b1 = t->num_bb - 2;
1594 
    if (unroll > 1) {
1595 
      //print_cuc_bb (t, "preunroll1");
1596 
      n = unroll_loop (t, b1, unroll);
1597 
      free_func (t);
1598 
    } else n = t;
1599 
  } else {
1600 
    b1 = b;
1601 897 markom 
    if (unroll > 1) n = unroll_loop (f, b1, unroll);
1602 
    else return dup_func (f);
1603 879 markom 
  }
1604 
 
1605 897 markom 
  /* Assign new counts to functions */
1606 879 markom 
  assert (counts = (int *)malloc (sizeof (int) * (preroll - 1 + unroll)));
1607 
  count_bb_seq (n, b, bb_filename, counts, preroll, unroll);
1608 
  for (i = 0; i < preroll - 1 + unroll; i++) {
1609 
    if (i == 0) b1 = b;
1610 
    else b1 = f->num_bb + (i - 1) * 2;
1611 
    n->bb[b1].cnt = counts[i];
1612 
  }
1613 
 
1614 
  //print_cuc_bb (n, "preunroll");
1615 
  free (counts);
1616 
  return n;
1617 
}
1618 933 markom 
 

powered by: WebSVN 2.1.0

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