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

Subversion Repositories or1k

[/] [or1k/] [tags/] [rel-0-3-0-rc3/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 930

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

powered by: WebSVN 2.1.0

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