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

Subversion Repositories or1k

[/] [or1k/] [tags/] [nog_patch_52/] [or1ksim/] [cuc/] [bb.c] - Blame information for rev 897

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 
/* Print out basic blocks */
31 
void print_cuc_bb (cuc_func *f, char *s)
32 
{
33 
  int i;
34 
  printf ("------- %s -------\n", s);
35 
  for (i = 0; i < f->num_bb; i++) {
36 
    if (f->bb[i].insn) printf ("\n---- BB%-2x * %x ---- ", i, f->bb[i].cnt);
37 
    else printf ("BB%-2x: %4x-%-4x", i, f->bb[i].first, f->bb[i].last);
38 
    printf (" type %02x tmp %i ", f->bb[i].type, f->bb[i].tmp);
39 
    if (f->bb[i].next[0] >= 0) printf ("next %2x ", f->bb[i].next[0]);
40 
    else printf ("next * ");
41 
    if (f->bb[i].next[1] >= 0) printf ("%2x ", f->bb[i].next[1]);
42 
    else printf ("* ");
43 
    if (f->bb[i].prev[0] >= 0) printf ("prev %2x ", f->bb[i].prev[0]);
44 
    else printf ("prev * ");
45 
    if (f->bb[i].prev[1] >= 0) printf ("%2x\n", f->bb[i].prev[1]);
46 
    else printf ("*\n");
47 
 
48 
    if (f->bb[i].insn) print_insns (f->bb[i].insn, f->bb[i].ninsn, 0);
49 
  }
50 
  printf ("\n");
51 897 markom 
  fflush (stdout);
52 879 markom 
}
53 
 
54 
/* Copies src basic block into destination */
55 
cuc_bb *cpy_bb (cuc_bb *dest, cuc_bb *src)
56 
{
57 897 markom 
  int i, j;
58 
  assert (dest != src);
59 879 markom 
  *dest = *src;
60 
  assert (dest->insn = malloc (sizeof (cuc_insn) * src->ninsn));
61 
  for (i = 0; i < src->ninsn; i++)
62 
    dest->insn[i] = src->insn[i];
63 
  if (src->ntim) {
64 
    assert (dest->tim = malloc (sizeof (cuc_timings) * src->ntim));
65 897 markom 
    for (i = 0; i < src->ntim; i++) {
66 
      dest->tim[i] = src->tim[i];
67 
      if (src->tim[i].nshared) {
68 
        assert (dest->tim[i].shared = malloc (sizeof (int) * src->tim[i].nshared));
69 
        for (j = 0; j < src->tim[i].nshared; j++)
70 
          dest->tim[i].shared[j] = src->tim[i].shared[j];
71 
      }
72 
    }
73 879 markom 
  }
74 
}
75 
 
76 
/* Duplicates function */
77 
cuc_func *dup_func (cuc_func *f)
78 
{
79 
  cuc_func *n = (cuc_func *) malloc (sizeof (cuc_func));
80 
  int b, i;
81 
  for (b = 0; b < f->num_bb; b++) cpy_bb (&n->bb[b], &f->bb[b]);
82 
  n->num_bb = f->num_bb;
83 
  assert (n->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
84 
  for (b = 0; b < f->num_init_bb; b++) n->init_bb_reloc[b] = f->init_bb_reloc[b];
85 
  n->num_init_bb = f->num_init_bb;
86 
  for (i = 0; i < MAX_REGS; i++) n->saved_regs[i] = f->saved_regs[i];
87 
  n->start_addr = f->start_addr;
88 
  n->end_addr = f->end_addr;
89 
  n->orig_time = f->orig_time;
90 
  n->nmsched = f->nmsched;
91 
  for (i = 0; i < f->nmsched; i++) {
92 
    n->msched[i] = f->msched[i];
93 
    n->mtype[i] = f->mtype[i];
94 
  }
95 
  return n;
96 
}
97 
 
98 
/* Releases memory allocated by function */
99 
void free_func (cuc_func *f)
100 
{
101 
  int b, i;
102 
  for (b = 0; b < f->num_bb; b++) {
103 
    for (i = 0; i < f->bb[b].ninsn; i++)
104 
      dispose_list (&f->bb[b].insn[i].dep);
105 
    if (f->bb[b].insn) free (f->bb[b].insn);
106 897 markom 
    for (i = 0; i < f->bb[b].ntim; i++)
107 
      if (f->bb[b].tim[i].nshared && f->bb[b].tim[i].shared)
108 
        free (f->bb[b].tim[i].shared);
109 879 markom 
    if (f->bb[b].tim && f->bb[b].ntim) free (f->bb[b].tim);
110 
  }
111 
  free (f);
112 
}
113 
 
114 
 
115 
/* Recalculates last_used_reg */
116 
void recalc_last_used_reg (cuc_func *f, int b)
117 
{
118 
  int i;
119 
  cuc_bb *bb = &f->bb[b];
120 
 
121 
  /* rebuild last used reg array */
122 
  if (bb->insn[0].index == II_LRBB) bb->last_used_reg[LRBB_REG] = 0;
123 
  else bb->last_used_reg[LRBB_REG] = -1;
124 
 
125 
  for (i = 1; i < MAX_REGS - 1; i++) bb->last_used_reg[i] = -1;
126 
 
127 
    /* Create references */
128 
  for (i = 0; i < bb->ninsn; i++) {
129 
    int k;
130 
    /* Now check for destination operand(s) */
131 
    for (k = 0; k < MAX_OPERANDS; k++) if (bb->insn[i].opt[k] & OPT_DEST)
132 
      if ((bb->insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
133 
        && (int)bb->insn[i].op[k] >= 0) {
134 
        bb->last_used_reg[bb->insn[i].op[k]] = REF (b, i);
135 
      }
136 
  }
137 
}
138 
 
139 
/* Set the BB limits */
140 
void detect_bb (cuc_func *f)
141 
{
142 
  int i, j, end_bb = 0, eb = 0;
143 
 
144 
  /* Mark block starts/ends */
145 
  for (i = 0; i < num_insn; i++) {
146 
    if (end_bb) insn[i].type |= IT_BBSTART;
147 
    end_bb = 0;
148 
    if (insn[i].type & IT_BRANCH) {
149 
      int jt = insn[i].op[0];
150 
      insn[i].type |= IT_BBEND;
151 
      end_bb = 1;
152 
      if (jt < 0 || jt >= num_insn) {
153 
        fprintf (stderr, "Instruction #%i:Jump out of function '%s'.\n", i, insn[i].disasm);
154 
        exit (1);
155 
      }
156 
      if (jt > 0) insn[jt - 1].type |= IT_BBEND;
157 
      insn[jt].type |= IT_BBSTART;
158 
    }
159 
  }
160 
 
161 
  /* Initialize bb array */
162 
  insn[0].type |= IT_BBSTART;
163 
  insn[num_insn - 1].type |= IT_BBEND;
164 
  f->num_bb = 0;
165 
  for (i = 0; i < num_insn; i++) {
166 
    if (insn[i].type & IT_BBSTART) {
167 
      f->bb[f->num_bb].first = i;
168 
      f->bb[f->num_bb].cnt = 0;
169 
    }
170 
    /* Determine repetitions of a loop */
171 
    if (insn[i].type & IT_BBEND) {
172 
      f->bb[f->num_bb].type = 0;
173 
      f->bb[f->num_bb].last = i;
174 
      f->bb[f->num_bb].next[0] = f->bb[f->num_bb].next[1] = -1;
175 
      f->bb[f->num_bb].tmp = 0;
176 
      f->bb[f->num_bb].ntim = 0;
177 
      f->num_bb++;
178 
      assert (f->num_bb < MAX_BB);
179 
    }
180 
  }
181 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INIT");
182 879 markom 
 
183 
  /* Build forward connections between BBs */
184 
  for (i = 0; i < f->num_bb; i++)
185 
    if (insn[f->bb[i].last].type & IT_BRANCH) {
186 
      int j;
187 
      assert (insn[f->bb[i].last].index == II_BF);
188 
      /* Find block this instruction jumps to */
189 
      for (j = 0; j < f->num_bb; j++)
190 
        if (f->bb[j].first == insn[f->bb[i].last].op[0]) break;
191 
      assert (j < f->num_bb);
192 
 
193 
      /* Convert the jump address to BB link */
194 
      insn[f->bb[i].last].op[0] = j; insn[f->bb[i].last].opt[0] = OPT_BB;
195 
 
196 
      /* Make a link */
197 
      f->bb[i].next[0] = j;
198 
      if (++f->bb[j].tmp > 2) eb++;
199 
      f->bb[i].next[1] = i + 1;
200 
      if (++f->bb[i + 1].tmp > 2) eb++;
201 
    } else if (f->bb[i].last == num_insn - 1) { /* Last instruction doesn't have to do anything */
202 
      f->bb[i].type |= BB_END;
203 
    } else {
204 
      f->bb[i].next[0] = i + 1;
205 
      if (++f->bb[i + 1].tmp > 2) eb++;
206 
    }
207 
 
208 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_NEXT");
209 879 markom 
 
210 
  /* Build backward connections, but first insert artificial blocks
211 
   * to handle more than 2 connections */
212 883 markom 
  cucdebug (6, "artificial %i %i\n", f->num_bb, eb);
213 879 markom 
  end_bb = f->num_bb + eb;
214 
  for (i = f->num_bb - 1; i >= 0; i--) {
215 
    j = f->bb[i].tmp;
216 
    if (f->bb[i].tmp > 2) f->bb[i].tmp = -f->bb[i].tmp;
217 
    f->bb[--end_bb] = f->bb[i];
218 
    reloc[i] = end_bb;
219 
    while (j-- > 2) {
220 
      f->bb[--end_bb].first = f->bb[i].first;
221 
      f->bb[end_bb].last = -1;
222 
      f->bb[end_bb].next[0] = -1;
223 
      f->bb[end_bb].next[1] = -1;
224 
      f->bb[end_bb].tmp = 0;
225 
      f->bb[end_bb].cnt = f->bb[i].cnt;
226 
      f->bb[end_bb].ntim = 0;
227 
    }
228 
  }
229 
  f->num_bb += eb;
230 
 
231 
  /* relocate jump instructions */
232 
  for (i = 0; i < num_insn; i++)
233 
    for (j = 0; j < MAX_OPERANDS; j++)
234 
      if (insn[i].opt[j] & OPT_BB)
235 
        insn[i].op[j] = reloc[insn[i].op[j]];
236 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT-reloc");
237 879 markom 
  for (i = 0; i < f->num_bb; i++) {
238 
    if (f->bb[i].next[0] >= 0) {
239 
      int t = reloc[f->bb[i].next[0]];
240 
      if (f->bb[t].tmp < 0) {
241 
        f->bb[t].tmp = -f->bb[t].tmp;
242 
        t -= f->bb[t].tmp - 2;
243 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
244 
      f->bb[i].next[0] = t;
245 
    }
246 
    if (f->bb[i].next[1] >= 0) {
247 
      int t = reloc[f->bb[i].next[1]];
248 
      if (f->bb[t].tmp < 0) {
249 
        f->bb[t].tmp = -f->bb[t].tmp;
250 
        t -= f->bb[t].tmp - 2;
251 
      } else if (f->bb[t].tmp > 2) t -= f->bb[t].tmp-- - 2;
252 
      f->bb[i].next[1] = t;
253 
    }
254 
    /* artificial blocks do not have relocations, hardcode them */
255 
    if (f->bb[i].last < 0) f->bb[i].next[0] = i + 1;
256 
  }
257 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_INSERT");
258 879 markom 
 
259 
  /* Uncoditional branched do not continue to next block */
260 
  for (i = 0; i < f->num_bb; i++) {
261 
    cuc_insn *ii;
262 
    if (f->bb[i].last < 0) continue;
263 
    ii = &insn[f->bb[i].last];
264 
    /* Unconditional branch? */
265 
    if (ii->type & IT_BRANCH && ii->opt[1] & OPT_CONST) {
266 
      change_insn_type (ii, II_NOP);
267 
      if (f->bb[i].next[1] == i + 1) f->bb[i].next[0] = f->bb[i].next[1];
268 
      f->bb[i].next[1] = -1;
269 
    }
270 
  }
271 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_UNCOND_JUMP");
272 879 markom 
 
273 
  /* Add backward connections */
274 
  for (i = 0; i < f->num_bb; i++)
275 
    f->bb[i].prev[0] = f->bb[i].prev[1] = -1;
276 
 
277 
  for (i = 0; i < f->num_bb; i++) {
278 
    if (f->bb[i].next[0] >= 0) {
279 
      int t = f->bb[i].next[0];
280 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
281 
      else {
282 
        assert (f->bb[t].prev[1] < 0);
283 
        f->bb[t].prev[1] = i;
284 
      }
285 
    }
286 
    if (f->bb[i].next[1] >= 0) {
287 
      int t = f->bb[i].next[1];
288 
      if (f->bb[t].prev[0] < 0) f->bb[t].prev[0] = i;
289 
      else {
290 
        assert (f->bb[t].prev[1] < 0);
291 
        f->bb[t].prev[1] = i;
292 
      }
293 
    }
294 
  }
295 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "AFTER_PREV");
296 879 markom 
}
297 
 
298 
/* Build basic blocks */
299 
void build_bb (cuc_func *f)
300 
{
301 
  int i, j, k;
302 
  for (i = 0; i < f->num_bb; i++) {
303 
    if (f->bb[i].last < 0) f->bb[i].ninsn = MAX_REGS - 1;
304 
    else f->bb[i].ninsn = f->bb[i].last - f->bb[i].first + 1 + MAX_REGS - 1;
305 
    assert (f->bb[i].ninsn >= MAX_REGS - 1);
306 
    f->bb[i].insn = (cuc_insn *) malloc (sizeof (cuc_insn) * f->bb[i].ninsn);
307 
    assert (f->bb[i].insn);
308 
    f->bb[i].nmemory = 0;
309 
    f->bb[i].unrolled = 1;
310 
 
311 
    /* Save space for conditional moves, exclude r0, place lrbb instead */
312 
    change_insn_type (&f->bb[i].insn[0], II_LRBB);
313 
    strcpy (f->bb[i].insn[0].disasm, "lrbb");
314 
    f->bb[i].insn[0].type = IT_UNUSED;
315 
    f->bb[i].insn[0].dep = NULL;
316 
    f->bb[i].insn[0].op[0] = LRBB_REG; f->bb[i].insn[0].opt[0] = OPT_REGISTER | OPT_DEST;
317 
    f->bb[i].insn[0].opt[1] = OPT_LRBB;
318 
    f->bb[i].insn[0].opt[2] = f->bb[i].insn[0].opt[3] = OPT_NONE;
319 
    for (j = 1; j < MAX_REGS - 1; j++) {
320 
      change_insn_type (&f->bb[i].insn[j], II_CMOV);
321 
      strcpy (f->bb[i].insn[j].disasm, "cmov");
322 
      f->bb[i].insn[j].type = 0;
323 
      f->bb[i].insn[j].dep = NULL;
324 
      f->bb[i].insn[j].opt[0] = f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_REGISTER;
325 
      f->bb[i].insn[j].opt[0] |= OPT_DEST;
326 
      f->bb[i].insn[j].op[0] = f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = j;
327 
      f->bb[i].insn[j].op[3] = LRBB_REG; f->bb[i].insn[j].opt[3] = OPT_REGISTER;
328 
    }
329 897 markom 
 
330 
    /* Relocate instructions */
331 879 markom 
    for (j = MAX_REGS - 1; j < f->bb[i].ninsn; j++) {
332 
      f->bb[i].insn[j] = insn[f->bb[i].first + j - (MAX_REGS - 1)];
333 
      for (k = 0; k < MAX_OPERANDS; k++)
334 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
335 
          int b1;
336 
          for (b1 = 0; b1 < i; b1++)
337 
            if (f->bb[b1].first <= f->bb[i].insn[j].op[k]
338 
              && f->bb[i].insn[j].op[k] <= f->bb[b1].last) break;
339 
          assert (b1 < f->num_bb);
340 
          f->bb[i].insn[j].op[k] = REF (b1, f->bb[i].insn[j].op[k] - f->bb[b1].first + MAX_REGS - 1);
341 
        }
342 
      if (f->bb[i].insn[j].type & IT_MEMORY) f->bb[i].nmemory++;
343 
    }
344 
  }
345 897 markom 
 
346 
  /* We do a quick check if there are some anomalies */
347 
  for (i = 0; i < f->num_bb; i++)
348 
    for (j = 0; j < f->bb[i].ninsn; j++)
349 
      for (k = 0; k < MAX_OPERANDS; k++)
350 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
351 
          int t = f->bb[i].insn[j].op[k];
352 
          assert (REF_I (t) < f->bb[REF_BB(t)].ninsn);
353 
        }
354 879 markom 
}
355 
 
356 
/* type == 0; keep predecessor condition
357 
 * type == 1; keep successor condition
358 
 * type == 2; join loop unrolled blocks */
359 
static void join_bb (cuc_func *f, int pred, int succ, int type)
360 
{
361 
  int i, j, k, add, ninsn, add_cond = 0;
362 
  unsigned long cond_op, cond_opt;
363 
  cuc_insn *insn;
364 
 
365 897 markom 
  cucdebug (3, "%x <= %x+%x (%i)\n", pred, pred, succ, type);
366 
  cucdebug (3, "%x %x\n", f->bb[pred].ninsn, f->bb[succ].ninsn);
367 
  if (cuc_debug >= 3) fflush (stdout);
368 879 markom 
 
369 
  add = f->bb[pred].ninsn;
370 
  if (f->bb[pred].ninsn <= 0
371 
   || !(f->bb[pred].insn[f->bb[pred].ninsn - 1].type & IT_BRANCH)) type = 1;
372 
  if (type == 0 && f->bb[succ].prev[0] == f->bb[succ].next[0]) add_cond = 1;
373 
 
374 
  ninsn = f->bb[pred].ninsn + f->bb[succ].ninsn + (type == 0 || type == 2 ? 1 : 0)
375 
          + (add_cond ? MAX_REGS : 0);
376 
 
377 
  insn = (cuc_insn *) malloc (ninsn * sizeof (cuc_insn));
378 
  for (i = 0; i < add; i++) insn[i] = f->bb[pred].insn[i];
379 
  /* when type == 0, we copy the last (jump) instruction to the end */
380 
  if (type == 0 || type == 2) {
381 
    insn[ninsn - 1] = insn[add - 1];
382 
    cond_op = insn[add - 1].op[1];
383 
    cond_opt = insn[add - 1].opt[1];
384 
    change_insn_type (&insn[add - 1], II_NOP);
385 
    /* and when type == 2, we must add sfor instruction, to quit when either is true */
386 
    if (type == 2) {
387 
      /* TODO */
388 
      assert (0);
389 
    }
390 
  }
391 
  /* Copy second block */
392 
  for (i = 0; i < f->bb[succ].ninsn; i++) insn[i + f->bb[pred].ninsn] = f->bb[succ].insn[i];
393 
 
394 
  for (i = 0; i < ninsn; i++) reloc[i] = -1;
395 
 
396 
  /* Add conditional instructions if required */
397 
  if (add_cond) {
398 
    recalc_last_used_reg (f, pred);
399 
    recalc_last_used_reg (f, succ);
400 
 
401 
    /* r0 -- add nop for it */
402 
    change_insn_type (&insn[add + f->bb[succ].ninsn], II_NOP);
403 
    for (i = 1; i < MAX_REGS; i++) {
404 
      cuc_insn *ii = &insn[add + f->bb[succ].ninsn + i];
405 
      int a = f->bb[pred].last_used_reg[i];
406 
      int b = f->bb[succ].last_used_reg[i];
407 
 
408 
      if (b < 0) change_insn_type (ii, II_NOP);
409 
      else if (a < 0) {
410 
        change_insn_type (ii, II_ADD);
411 
        ii->type = 0;
412 
        ii->dep = NULL;
413 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
414 
        ii->op[1] = b; ii->opt[1] = OPT_REF;
415 
        ii->op[2] = 0; ii->opt[2] = OPT_CONST;
416 
        ii->op[3] = OPT_NONE;
417 
      } else if (b >= 0) {
418 
        change_insn_type (ii, II_CMOV);
419 
        ii->type = 0;
420 
        ii->dep = NULL;
421 
        ii->op[0] = i; ii->opt[0] = OPT_REGISTER | OPT_DEST;
422 
        ii->op[1] = a; ii->opt[1] = OPT_REF;
423 
        ii->op[2] = b; ii->opt[2] = OPT_REF;
424 
        ii->op[3] = cond_op; ii->opt[3] = cond_opt;
425 
        reloc[REF_I(a)] = REF (pred, add + f->bb[succ].ninsn + i);
426 
      }
427 
      sprintf (ii->disasm, "cmov (join BB)");
428 
    }
429 
  }
430 
 
431 
  f->bb[pred].type |= f->bb[succ].type;
432 
  i = 0;
433 
  assert (f->bb[pred].next[0] >= 0);
434 
  switch (type) {
435 
  case 0:
436 
    if (f->bb[pred].next[0] == succ) f->bb[pred].next[0] = f->bb[succ].next[0];
437 
    if (f->bb[pred].next[1] == succ) f->bb[pred].next[1] = f->bb[succ].next[0];
438 
    assert (f->bb[succ].next[1] < 0);
439 
    break;
440 
  case 1:
441 
    f->bb[pred].next[0] = f->bb[succ].next[0];
442 
    f->bb[pred].next[1] = f->bb[succ].next[1];
443 
    break;
444 
  }
445 
  if (f->bb[pred].next[0] == f->bb[pred].next[1]) f->bb[pred].next[1] = -1;
446 
  f->bb[succ].type = BB_DEAD;
447 
 
448 
  /* Set max count */
449 
  if (f->bb[pred].cnt < f->bb[succ].cnt) f->bb[pred].cnt = f->bb[succ].cnt;
450 
  f->bb[pred].ninsn = ninsn;
451 
  free (f->bb[pred].insn); f->bb[pred].insn = NULL;
452 
  free (f->bb[succ].insn); f->bb[succ].insn = NULL;
453 
  f->bb[pred].insn = insn;
454 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
455 
    if (f->bb[i].prev[0] == succ) f->bb[i].prev[0] = pred;
456 
    if (f->bb[i].prev[1] == succ) f->bb[i].prev[1] = pred;
457 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
458 
    for (j = 0; j < f->bb[i].ninsn; j++)
459 
      for (k = 0; k < MAX_OPERANDS; k++)
460 
        if (f->bb[i].insn[j].opt[k] & OPT_REF) {
461 
          /* Check if we are referencing successor BB -> relocate to second part of
462 
             the new block */
463 
          if (REF_BB (f->bb[i].insn[j].op[k]) == succ) {
464 
            if (type == 0 && f->bb[i].insn[j].op[j] == REF (succ, ninsn - 1))
465 
              f->bb[i].insn[j].op[j] = REF (pred, f->bb[pred].ninsn);
466 
            else {
467 
              int t = f->bb[i].insn[j].op[k];
468 
              int ndest = REF (pred, REF_I (t) + add);
469 
 
470 
              /* We've found a reference to succ. block, being removed, relocate */
471 
              if (add_cond && i != succ && !(i == pred && j >= ninsn - MAX_REGS)) {
472 
                //printf ("%x!\n", t);
473 
                //printf ("%x!\n", f->INSN(ndest).op[0]);
474 
                /* interblock dependency should have physical register attached */
475 
                assert (f->INSN(ndest).opt[0] == OPT_DEST | OPT_REGISTER);
476 
                assert (f->INSN(ndest).op[0] >= 0);
477 
                f->bb[i].insn[j].op[k] = REF (pred, add + f->bb[succ].ninsn + f->INSN(ndest).op[0]);
478 
              } else f->bb[i].insn[j].op[k] = ndest;
479 
            }
480 
          } else if (REF_BB(f->bb[i].insn[j].op[k]) == pred) {
481 
            if (i != pred && reloc[REF_I(f->bb[i].insn[j].op[k])] >= 0) {
482 
              f->bb[i].insn[j].op[k] = reloc[REF_I(f->bb[i].insn[j].op[k])];
483 
            }
484 
          }
485 
        }
486 
  }
487 
 
488 883 markom 
  if (cuc_debug >= 3) print_cuc_bb (f, "join");
489 879 markom 
}
490 
 
491 
/* Optimize basic blocks */
492 
void optimize_bb (cuc_func *f)
493 
{
494 
  int i, j;
495 
remove_lrbb:
496 
  /* we can remove lrbb instructions from blocks with just one predecessor */
497 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
498 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* exactly one predecessor */
499 
      for (j = 0; j < f->bb[i].ninsn; j++)
500 
        if (f->bb[i].insn[j].index == II_LRBB) {
501 
          cuc_insn *t;
502 883 markom 
          cucdebug (4, "-lrbb %i.%i\n", i, j);
503 879 markom 
 
504 
          /* Change to add LRBB, 0, 0 */
505 
          change_insn_type (&f->bb[i].insn[j], II_ADD);
506 
          f->bb[i].insn[j].type &= ~IT_VOLATILE;
507 
          t = &f->bb[f->bb[i].prev[0]].insn[f->bb[f->bb[i].prev[0]].ninsn - 1];
508 
          f->bb[i].insn[j].opt[1] = f->bb[i].insn[j].opt[2] = OPT_CONST;
509 
          f->bb[i].insn[j].op[1] = f->bb[i].insn[j].op[2] = 0; /* always use left block */
510 
          f->bb[i].insn[j].opt[3] = OPT_NONE;
511 
 
512 
          /* If the predecessor still has a conditional jump instruction, we must be careful.
513 
             This could only have occured when we found out that next[0] == next[1] and have
514 
             joined them. Now we will link lrbb and correct the situation */
515 
          if (t->type & IT_BRANCH) { /* We must set a reference to branch result */
516 
            f->bb[i].insn[j].opt[1] = t->opt[1];
517 
            f->bb[i].insn[j].op[1] = t->op[1];
518 
            change_insn_type (t, II_NOP);
519 
          }
520 
        }
521 
    }
522 
  }
523 
 
524 
  /* Type 0 joining
525 
     1. link between pred & succ
526 
     2. no memory accesses in succ
527 
     3. optional pred's second successors
528 
     4. max. one succ's successors */
529 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
530 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0 /* one predecessor */
531 
     && f->bb[i].next[1] < 0 /* max. one successor */
532 
     && f->bb[i].nmemory == 0) {                  /* and no memory acceses */
533 
      join_bb (f, f->bb[i].prev[0], i, 0);
534 
      goto remove_lrbb;
535 
    }
536 
 
537 
  /* Type 1 joining
538 
     1. link between pred & succ
539 
     2. no other pred's successors
540 
     3. no other succ's predecessors */
541 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
542 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0 /* one predecessor */
543 
     && f->bb[f->bb[i].prev[0]].next[0] >= 0 && f->bb[f->bb[i].prev[0]].next[1] < 0) { /* one successor */
544 
      join_bb (f, f->bb[i].prev[0], i, 1);
545 
      goto remove_lrbb;
546 
    }
547 
 
548 
#if 1
549 
  /* Type 2 joining
550 
     1. link between pred & succ
551 
     2. succ has exactly one predeccessor
552 
     3. pred & succ share common successor
553 
     4. optional succ's second successor */
554 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD))
555 
    if (f->bb[i].prev[0] >= 0 && f->bb[i].prev[1] < 0) { /* one predecessor */
556 
      int p = f->bb[i].prev[0];
557 897 markom 
      if (f->bb[p].next[0] == i && f->bb[p].next[1] == f->bb[p].next[1]) {
558 
        join_bb (f, f->bb[i].prev[0], i, 2);
559 
        goto remove_lrbb;
560 
      }
561 879 markom 
    }
562 
#endif
563 
}
564 
 
565 
/* Removes BBs marked as dead */
566 
void remove_dead_bb (cuc_func *f)
567 
{
568 
  int i, j, k, d = 0;
569 
 
570 
  for (i = 0; i < f->num_bb; i++) if (f->bb[i].type & BB_DEAD) {
571 
    if (f->bb[i].insn) free (f->bb[i].insn);
572 
    f->bb[i].insn = NULL;
573 
    reloc[i] = -1;
574 
  } else {
575 
    reloc[i] = d;
576 
    f->bb[d++] = f->bb[i];
577 
  }
578 
  f->num_bb = d;
579 
 
580 
  /* relocate initial blocks */
581 
  for (i = 0; i < f->num_init_bb; i++)
582 
    f->init_bb_reloc[i] = reloc[f->init_bb_reloc[i]];
583 
 
584 
  /* repair references */
585 
  for (i = 0; i < f->num_bb; i++) if (!(f->bb[i].type & BB_DEAD)) {
586 
    if (f->bb[i].prev[0] >= 0) assert ((f->bb[i].prev[0] = reloc[f->bb[i].prev[0]]) >= 0);
587 
    if (f->bb[i].prev[1] >= 0) assert ((f->bb[i].prev[1] = reloc[f->bb[i].prev[1]]) >= 0);
588 
    if (f->bb[i].next[0] >= 0) assert ((f->bb[i].next[0] = reloc[f->bb[i].next[0]]) >= 0);
589 
    if (f->bb[i].next[1] >= 0) assert ((f->bb[i].next[1] = reloc[f->bb[i].next[1]]) >= 0);
590 
    if (f->bb[i].prev[0] == f->bb[i].prev[1]) f->bb[i].prev[1] = -1;
591 
    if (f->bb[i].next[0] == f->bb[i].next[1]) f->bb[i].next[1] = -1;
592 
 
593 
    for (j = 0; j < f->bb[i].ninsn; j++)
594 
      for (k = 0; k < MAX_OPERANDS; k++)
595 
        if (f->bb[i].insn[j].opt[k] & OPT_BB && (signed)f->bb[i].insn[j].op[k] >= 0)
596 
          assert ((f->bb[i].insn[j].op[k] = reloc[f->bb[i].insn[j].op[k]]) >= 0);
597 
        else if (f->bb[i].insn[j].opt[k] & OPT_REF) {
598 
          int t = f->bb[i].insn[j].op[k];
599 
          assert (reloc[REF_BB(t)] >= 0);
600 
          f->bb[i].insn[j].op[k] = REF (reloc[REF_BB(t)], REF_I (t));
601 
        }
602 
  }
603 
}
604 
 
605 
/* Recursive calculation of dependencies */
606 
static int reg_dep_rec (cuc_func *f, int cur)
607 
{
608 
  int i, j;
609 
  cuc_insn *insn = f->bb[cur].insn;
610 
 
611 
  //printf ("\n %i", cur);
612 
  /* Spread only, do not loop */
613 
  if (f->bb[cur].tmp) return;
614 
  f->bb[cur].tmp = 1;
615 
  //printf ("!   ");
616 
 
617 
  for (i = 0; i < f->bb[cur].ninsn; i++) {
618 
    /* Check for destination operand(s) */
619 
    for (j = 0; j < MAX_OPERANDS; j++) if (insn[i].opt[j] & OPT_DEST)
620 
      if ((insn[i].opt[j] & ~OPT_DEST) == OPT_REGISTER && (signed)insn[i].op[j] >= 0) {
621 
        //printf ("%i:%i,%x ", insn[i].op[j], i, REF (cur, i));
622 
        assert (insn[i].op[j] > 0 && insn[i].op[j] < MAX_REGS); /* r0 should never be dest */
623 
        f->bb[cur].last_used_reg[insn[i].op[j]] = REF (cur, i);
624 
      }
625 
  }
626 
 
627 
  if (f->bb[cur].next[0] >= 0) reg_dep_rec (f, f->bb[cur].next[0]);
628 
  if (f->bb[cur].next[1] >= 0) reg_dep_rec (f, f->bb[cur].next[1]);
629 
}
630 
 
631 
/* Detect register dependencies */
632 
void reg_dep (cuc_func *f)
633 
{
634 
  int i, b, c;
635 
 
636 
  /* Set dead blocks */
637 
  for (b = 0; b < f->num_bb; b++) {
638 
    f->bb[b].tmp = 0;
639 
    for (i = 0; i < MAX_REGS; i++) f->bb[b].last_used_reg[i] = -1;
640 
  }
641 
 
642 
  /* Start with first block and set dependecies of all reachable blocks */
643 
  /* At the same time set last_used_regs */
644 
  reg_dep_rec (f, 0);
645 
 
646 
  for (i = 0; i < f->num_bb; i++)
647 
    if (f->bb[i].tmp) f->bb[i].tmp = 0;
648 
    else f->bb[i].type |= BB_DEAD;
649 
 
650 
  /* Detect loops; mark BBs where loops must be broken */
651 
  for (c = 0; c < f->num_bb; c++) {
652 
    int min = 3, minb;
653 
 
654 
    /* search though all non-visited for minimum number of unvisited predecessors */
655 
    for (b = 0; b < f->num_bb; b++) if (!f->bb[b].tmp) {
656 
      int tmp = 0;
657 
      if (f->bb[b].prev[0] >= 0 && !f->bb[f->bb[b].prev[0]].tmp) tmp++;
658 
      if (f->bb[b].prev[1] >= 0 && !f->bb[f->bb[b].prev[1]].tmp) tmp++;
659 
      if (tmp < min) {
660 
        minb = b;
661 
        min = tmp;
662 
        if (tmp == 0) break; /* We already have the best one */
663 
      }
664 
    }
665 
    b = minb;
666 
    f->bb[b].tmp = 1; /* Mark visited */
667 883 markom 
    cucdebug (3, "minb %i min %i\n", minb, min);
668 879 markom 
    if (min) { /* We just broke the loop */
669 
      f->bb[b].type |= BB_INLOOP;
670 
    }
671 
  }
672 
 
673 
  /* Set real predecessors in cmov instructions to previous blocks */
674 
  for (b = 0; b < f->num_bb; b++)
675 
    for (i = 1; i < MAX_REGS - 1; i++) {
676 
      int pa, pb;
677 
      assert (f->bb[b].insn[i].index ==  II_CMOV);
678 
      assert (f->bb[b].insn[i].opt[0] == OPT_REGISTER | OPT_DEST);
679 
      assert (f->bb[b].insn[i].op[0] == i);
680 
      if (f->bb[b].prev[0] < 0) pa = -1;
681 
      else pa = f->bb[f->bb[b].prev[0]].last_used_reg[i];
682 
      if (f->bb[b].prev[1] < 0) pb = -1;
683 
      else pb = f->bb[f->bb[b].prev[1]].last_used_reg[i];
684 
 
685 
      /* We do some very simple optimizations right away to make things more readable */
686 
      if (pa < 0 && pb < 0) {
687 
        /* Was not used at all */
688 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
689 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
690 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
691 
      } else if (pa < 0) {
692 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
693 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
694 
        f->bb[b].insn[i].op[1] = pb; f->bb[b].insn[i].opt[1] = OPT_REF;
695 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
696 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
697 
      } else if (pb < 0) {
698 
        change_insn_type (&f->bb[b].insn[i], II_ADD);
699 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
700 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
701 
        f->bb[b].insn[i].op[2] = 0; f->bb[b].insn[i].opt[2] = OPT_CONST;
702 
        f->bb[b].insn[i].opt[3] = OPT_NONE;
703 
      } else {
704 
        int t = REF (b, 0); /* lrbb should be first instruction */
705 
        assert (f->INSN(t).index == II_LRBB);
706 
 
707 
        f->bb[b].insn[i].op[1] = pa; f->bb[b].insn[i].opt[1] = OPT_REF;
708 
        assert (f->INSN(pa).opt[0] == (OPT_REGISTER | OPT_DEST));
709 
 
710 
        f->bb[b].insn[i].op[2] = pb; f->bb[b].insn[i].opt[2] = OPT_REF;
711 
        assert (f->INSN(pb).opt[0] == (OPT_REGISTER | OPT_DEST));
712 
 
713 
        /* Update op[3] -- flag register */
714 
        assert (f->bb[b].insn[i].opt[3] == OPT_REGISTER);
715 
        assert (f->bb[b].insn[i].op[3] == LRBB_REG);
716 
        assert (t >= 0);
717 
        f->bb[b].insn[i].opt[3] = OPT_REF; /* Convert already used regs to references */
718 
        f->bb[b].insn[i].op[3] = t;
719 
        assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
720 
      }
721 
    }
722 
 
723 
  /* assign register references */
724 
  for (b = 0; b < f->num_bb; b++) {
725 
    /* rebuild last used reg array */
726 
    f->bb[b].last_used_reg[0] = -1;
727 
    if (f->bb[b].insn[0].index == II_LRBB) f->bb[b].last_used_reg[LRBB_REG] = 0;
728 
    else f->bb[b].last_used_reg[LRBB_REG] = -1;
729 
 
730 
    for (i = 1; i < MAX_REGS - 1; i++)
731 
      f->bb[b].last_used_reg[i] = -1;
732 
 
733 
    /* Create references */
734 
    for (i = 0; i < f->bb[b].ninsn; i++) {
735 
      int k;
736 
      /* Check for source operands first */
737 
      for (k = 0; k < MAX_OPERANDS; k++) {
738 
        if (!(f->bb[b].insn[i].opt[k] & OPT_DEST))
739 
        if (f->bb[b].insn[i].opt[k] & OPT_REGISTER) {
740 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
741 
 
742 
          if (f->bb[b].insn[i].op[k] == 0) { /* Convert r0 to const0 */
743 
            f->bb[b].insn[i].opt[k] = OPT_CONST;
744 
            f->bb[b].insn[i].op[k] = 0;
745 
          } else if (t >= 0) {
746 
            f->bb[b].insn[i].opt[k] = OPT_REF; /* Convert already used regs to references */
747 
            f->bb[b].insn[i].op[k] = t;
748 
            assert (f->INSN(t).opt[0] == (OPT_REGISTER | OPT_DEST));
749 
            //f->INSN(t).op[0] = -1;
750 
          }
751 
        } else if (f->bb[b].insn[i].opt[k] & OPT_REF) {
752 
          //f->INSN(f->bb[b].insn[i].op[k]).op[0] = -1; /* Mark referenced */
753 
          f->INSN(f->bb[b].insn[i].op[k]).type &= ~IT_UNUSED;
754 
        }
755 
      }
756 
 
757 
      /* Now check for destination operand(s) */
758 
      for (k = 0; k < MAX_OPERANDS; k++) if (f->bb[b].insn[i].opt[k] & OPT_DEST)
759 
        if ((f->bb[b].insn[i].opt[k] & ~OPT_DEST) == OPT_REGISTER
760 
          && (int)f->bb[b].insn[i].op[k] >= 0) {
761 
          int t = f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]];
762 
          assert (f->bb[b].insn[i].op[k] != 0); /* r0 should never be dest */
763 
          f->bb[b].last_used_reg[f->bb[b].insn[i].op[k]] = REF (b, i);
764 
        }
765 
    }
766 
  }
767 
 
768 
  /* Remove all unused lrbb */
769 
  for (b = 0; b < f->num_bb; b++)
770 
    for (i = 0; i < f->bb[b].ninsn; i++)
771 
      if (f->bb[b].insn[i].type & IT_UNUSED) change_insn_type (&f->bb[b].insn[i], II_NOP);
772 
 
773 
  /* SSAs with final register value are marked as outputs */
774 
  assert (f->bb[f->num_bb - 1].type & BB_END);
775 
  for (i = 0; i < MAX_REGS; i++) if (!call_saved[i]) {
776 
    int t = f->bb[f->num_bb - 1].last_used_reg[i];
777 
    /* Mark them volatile, so optimizer does not remove them */
778 
    if (t >= 0) f->bb[REF_BB(t)].insn[REF_I(t)].type |= IT_OUTPUT;
779 
  }
780 
}
781 
 
782 897 markom 
/* split the BB, based on the group numbers in .tmp */
783 
void expand_bb (cuc_func *f, int b)
784 
{
785 
  int n = f->num_bb;
786 
  int mg = 0;
787 
  int b1, i, j;
788 
 
789 
  for (i = 0; i < f->bb[b].ninsn; i++)
790 
    if (f->bb[b].insn[i].tmp > mg) mg = f->bb[b].insn[i].tmp;
791 
 
792 
  /* Create copies */
793 
  for (b1 = 1; b1 <= mg; b1++) {
794 
    assert (f->num_bb < MAX_BB);
795 
    cpy_bb (&f->bb[f->num_bb], &f->bb[b]);
796 
    f->num_bb++;
797 
  }
798 
 
799 
  /* Relocate */
800 
  for (b1 = 0; b1 < f->num_bb; b1++)
801 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
802 
      dep_list *d = f->bb[b1].insn[i].dep;
803 
      for (j = 0; j < MAX_OPERANDS; j++)
804 
        if (f->bb[b1].insn[i].opt[j] & OPT_REF) {
805 
          int t = f->bb[b1].insn[i].op[j];
806 
          if (REF_BB(t) == b && f->INSN(t).tmp != 0)
807 
            f->bb[b1].insn[i].op[j] = REF (n + f->INSN(t).tmp - 1, REF_I(t));
808 
        }
809 
      while (d) {
810 
        if (REF_BB (d->ref) == b && f->INSN(d->ref).tmp != 0)
811 
          d->ref = REF (n + f->INSN(d->ref).tmp - 1, REF_I(d->ref));
812 
        d = d->next;
813 
      }
814 
    }
815 
 
816 
  /* Delete unused instructions */
817 
  for (j = 0; j <= mg; j++) {
818 
    if (j == 0) b1 = b;
819 
    else b1 = n + j - 1;
820 
    for (i = 0; i < f->bb[b1].ninsn; i++) {
821 
      if (f->bb[b1].insn[i].tmp != j)
822 
        change_insn_type (&f->bb[b1].insn[i], II_NOP);
823 
      f->bb[b1].insn[i].tmp = 0;
824 
    }
825 
    if (j < mg) {
826 
      f->bb[b1].next[0] = n + j;
827 
      f->bb[b1].next[1] = -1;
828 
      f->bb[n + j].prev[0] = b1;
829 
      f->bb[n + j].prev[1] = -1;
830 
    } else {
831 
      i = f->bb[b1].next[0];
832 
      f->bb[n + j].prev[0] = j == 1 ? b : b1 - 1;
833 
      f->bb[n + j].prev[1] = -1;
834 
      if (i >= 0) {
835 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
836 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
837 
      }
838 
      i = f->bb[b1].next[1];
839 
      if (i >= 0) {
840 
        if (f->bb[i].prev[0] == b) f->bb[i].prev[0] = b1;
841 
        if (f->bb[i].prev[1] == b) f->bb[i].prev[1] = b1;
842 
      }
843 
    }
844 
  }
845 
}
846 
 
847 879 markom 
/* Scans sequence of BBs and set bb[].cnt */
848 
void generate_bb_seq (cuc_func *f, char *mp_filename, char *bb_filename)
849 
{
850 
  FILE *fi, *fo;
851 
  struct mprofentry_struct *buf;
852 
  const int bufsize = 256;
853 
  unsigned long *bb_start;
854 
  unsigned long *bb_end;
855 
  int b, i, r;
856 
  int curbb, prevbb = -1;
857 
  unsigned long addr = -1;
858 
  unsigned long prevaddr = -1;
859 897 markom 
  int mssum = 0;
860 
  int mlsum = 0;
861 
  int mscnt = 0;
862 
  int mlcnt = 0;
863 879 markom 
 
864 
  assert (fi = fopen (mp_filename, "rb"));
865 
  assert (fo = fopen (bb_filename, "wb+"));
866 
 
867 
  assert (bb_start = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
868 
  assert (bb_end = (unsigned long *) malloc (sizeof (unsigned long) * f->num_bb));
869 
  for (b = 0; b < f->num_bb; b++) {
870 
    bb_start[b] = f->start_addr + f->bb[b].first * 4;
871 
    bb_end[b] = f->start_addr + f->bb[b].last * 4;
872 
    //printf ("%i %x %x\n", b, bb_start[b], bb_end[b]);
873 
    f->bb[0].cnt = 0;
874 
  }
875 
 
876 
  buf = (struct mprofentry_struct *) malloc (sizeof (struct mprofentry_struct) * bufsize);
877 
  assert (buf);
878 
 
879 
  //printf ("BBSEQ:\n");
880 
  do {
881 
    r = fread (buf, sizeof (struct mprofentry_struct), bufsize, fi);
882 
    //printf ("r%i : ", r);
883 
    for (i = 0; i < r; i++) {
884 
      if (buf[i].type & MPROF_FETCH) {
885 
        //printf ("%x, ", buf[i].addr);
886 
        if (buf[i].addr >= f->start_addr && buf[i].addr <= f->end_addr) {
887 
          assert (buf[i].type & MPROF_32);
888 
          prevaddr = addr;
889 
          addr = buf[i].addr;
890 
          for (b = 0; b < f->num_bb; b++)
891 
            if (bb_start[b] <= addr && addr <= bb_end[b]) break;
892 
          assert (b < f->num_bb);
893 
          curbb = b;
894 
          if (prevaddr + 4 != addr) prevbb = -1;
895 
        } else curbb = -1;
896 
 
897 
#warning TODO: do not count interrupts
898 
        if (curbb != prevbb && curbb >= 0) {
899 
          fwrite (&curbb, sizeof (unsigned long), 1, fo);
900 
          //printf (" [%i] ", curbb);
901 
          f->bb[curbb].cnt++;
902 
          prevbb = curbb;
903 
        }
904 897 markom 
      } else {
905 
        if (verify_memoryarea(buf[i].addr))
906 
          if (buf[i].type & MPROF_WRITE) mscnt++, mssum += cur_area->delayw;
907 
          else mlcnt++, mlsum += cur_area->delayw;
908 879 markom 
      }
909 
    }
910 
    //printf ("\n");
911 
  } while (r == bufsize);
912 
  //printf ("\n");
913 
 
914 897 markom 
  runtime.cuc.mdelay[0] = (1. * mlsum) / mlcnt;
915 
  runtime.cuc.mdelay[1] = (1. * mlsum) / mlcnt;
916 
  runtime.cuc.mdelay[2] = runtime.cuc.mdelay[3] = 1;
917 883 markom 
  f->num_runs = f->bb[0].cnt;
918 879 markom 
  fclose (fi);
919 
  fclose (fo);
920 
  free (buf);
921 
  free (bb_end);
922 
  free (bb_start);
923 
 
924 
  /* Initialize basic block relocations */
925 
  f->num_init_bb = f->num_bb;
926 
  //printf ("num_init_bb = %i\n", f->num_init_bb);
927 
  assert (f->init_bb_reloc = (int *)malloc (sizeof (int) * f->num_init_bb));
928 
  for (b = 0; b < f->num_init_bb; b++) f->init_bb_reloc[b] = b;
929 
}
930 
 
931 
/* Scans sequence of BBs and set counts for pre/unrolled loop for BB b */
932 
void count_bb_seq (cuc_func *f, int b, char *bb_filename, int *counts, int preroll, int unroll)
933 
{
934 
  FILE *fi;
935 
  const int bufsize = 256;
936 
  int i, r;
937 
  int *buf;
938 
  int cnt = 0;
939 
  int times = preroll - 1 + unroll;
940 
 
941 
  assert (fi = fopen (bb_filename, "rb"));
942 
  for (i = 0; i < times; i++) counts[i] = 0;
943 
  assert (buf = (int *) malloc (sizeof (int) * bufsize));
944 
 
945 
  do {
946 
    r = fread (buf, sizeof (int), bufsize, fi);
947 
    for (i = 0; i < r; i++) {
948 
      /* count consecutive acesses */
949 
      if (f->init_bb_reloc[buf[i]] == b) {
950 
        counts[cnt]++;
951 
        if (++cnt >= times) cnt = preroll - 1;
952 
      } else cnt = 0;
953 
    }
954 
  } while (r == bufsize);
955 
 
956 
  log ("Counts %i,%i :", preroll, unroll);
957 
  for (i = 0; i < times; i++) log ("%x ", counts[i]);
958 
  log ("\n");
959 
 
960 
  fclose (fi);
961 
  free (buf);
962 
}
963 
 
964 
/* relocate all accesses inside of BB b to back/fwd */
965 
static void relocate_bb (cuc_bb *bb, int b, int back, int fwd)
966 
{
967 
  int i, j;
968 
  for (i = 0; i < bb->ninsn; i++)
969 
    for (j = 0; j < MAX_OPERANDS; j++)
970 
      if (bb->insn[i].opt[j] & OPT_REF
971 
       && REF_BB (bb->insn[i].op[j]) == b) {
972 
        int t = REF_I (bb->insn[i].op[j]);
973 
        if (t < i) bb->insn[i].op[j] = REF (back, t);
974 
        else bb->insn[i].op[j] = REF (fwd, t);
975 
      }
976 
}
977 
 
978 
/* Unroll loop b unroll times and return new function. Original
979 
   function is unmodified. */
980 
static cuc_func *unroll_loop (cuc_func *f, int b, int unroll)
981 
{
982 
  int b1, t, i, j, prevb, prevart_b;
983 
  cuc_func *n = dup_func (f);
984 
  cuc_bb *ob = &f->bb[b];
985 
  cuc_insn *ii;
986 
 
987 
  assert (unroll > 1);
988 
  //printf ("unroll BB%i x %i (num_bb %i)\n", b, unroll, n->num_bb);
989 
  unroll--;
990 
  assert (n->num_bb + unroll * 2 < MAX_BB);
991 
 
992 
  prevb = b;
993 
  prevart_b = b;
994 
  /* Duplicate the BB */
995 
  for (t = 0; t < unroll; t++) {
996 
    cuc_bb *pb = &n->bb[prevart_b];
997 
    /* Add new block and set links */
998 
    b1 = n->num_bb++;
999 
    cpy_bb (&n->bb[b1], ob);
1000 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1001 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1002 
 
1003 
    /* Set predecessor's successor */
1004 
    if (n->bb[prevb].next[0] == b) {
1005 
      n->bb[prevb].next[0] = b1;
1006 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1007 
      else pb->next[1] = b1 + 1;
1008 
      n->bb[b1].next[1] = b1 + 1;
1009 
    } else if (n->bb[prevb].next[1] == b) {
1010 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1011 
      else pb->next[1] = b1 + 1;
1012 
      n->bb[b1].next[0] = b1 + 1;
1013 
      n->bb[prevb].next[1] = b1;
1014 
    } else assert (0);
1015 
 
1016 
    /* Set predecessor */
1017 
    n->bb[b1].prev[0] = prevb;
1018 
    n->bb[b1].prev[1] = -1;
1019 
 
1020 
    /* Relocate backward references to current instance and forward references
1021 
       to previous one */
1022 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1023 
 
1024 
    /* add artificial block, just to join accesses */
1025 
    b1 = n->num_bb++;
1026 
    cpy_bb (&n->bb[b1], ob);
1027 
    n->bb[b1].cnt = 0;
1028 
 
1029 
    for (i = 0; i < ob->ninsn - 1; i++) {
1030 
      ii = &n->bb[b1].insn[i];
1031 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1032 
        change_insn_type (ii, II_CMOV);
1033 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1034 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1035 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1036 
 
1037 
        /* Take left one, if we should have finished the first iteration*/
1038 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1039 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1040 
        } else {
1041 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1042 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1043 
        }
1044 
        ii->dep = NULL;
1045 
        ii->type = 0;
1046 
      } else {
1047 
        change_insn_type (ii, II_NOP);
1048 
      }
1049 
    }
1050 
 
1051 
    /* Add sfor instruction at the end, prioritizing flags */
1052 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1053 
    change_insn_type (ii, II_SFOR);
1054 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1055 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1056 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1057 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1058 
    } else {
1059 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1060 
      ii->opt[1] = OPT_REF;
1061 
    }
1062 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1063 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1064 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1065 
    } else {
1066 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1067 
      ii->opt[2] = OPT_REF;
1068 
    }
1069 
    ii->opt[3] = OPT_NONE;
1070 
    ii->type = IT_COND;
1071 
 
1072 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1073 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1074 
    n->bb[b1].prev[0] = prevart_b;
1075 
    n->bb[b1].prev[1] = b1 - 1;
1076 
    n->bb[b1].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1077 
    n->bb[b1].next[1] = -1;
1078 
 
1079 
    prevb = b1 - 1;
1080 
    prevart_b = b1;
1081 
  }
1082 
  if (ob->type & BB_END) {
1083 
    n->bb[prevart_b].type |= BB_END;
1084 
    n->bb[b].type &= ~BB_END;
1085 
  }
1086 
 
1087 
  //print_cuc_bb (n, "unroll1");
1088 
  /* repair BB after loop, to point back to latest artificial BB */
1089 
  b1 = n->bb[prevart_b].next[0];
1090 
  if (b1 >= 0) {
1091 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1092 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1093 879 markom 
    else assert (0);
1094 
  }
1095 
 
1096 
  /* Relink back to start of the loop */
1097 
  /* Set predecessor's successor */
1098 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = b;
1099 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = b;
1100 
  else assert (0);
1101 
 
1102 
  /* Set predecessor */
1103 
  if (n->bb[b].prev[0] == b) n->bb[b].prev[0] = prevb;
1104 
  else if (n->bb[b].prev[1] == b) n->bb[b].prev[1] = prevb;
1105 
  else assert (0);
1106 
 
1107 
  //print_cuc_bb (n, "unroll2");
1108 
 
1109 
  /* Relocate backward references to current instance and forward references
1110 
     to previous one */
1111 
  relocate_bb (&n->bb[b], b, b, prevb);
1112 
 
1113 
  /* Relocate all other blocks to point to latest prevart_b */
1114 
  for (i = 0; i < f->num_bb; i++)
1115 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1116 
 
1117 
  return n;
1118 
}
1119 
 
1120 
/* Preroll loop b preroll times and return new function. Original
1121 
   function is unmodified. */
1122 
static cuc_func *preroll_loop (cuc_func *f, int b, int preroll)
1123 
{
1124 
  int b1, t, i, j, prevb, prevart_b;
1125 
  cuc_func *n = dup_func (f);
1126 
  cuc_bb *ob = &f->bb[b];
1127 
  cuc_insn *ii;
1128 
 
1129 
  assert (preroll > 1);
1130 
  //printf ("preroll BB%i x %i (num_bb %i)\n", b, preroll, n->num_bb);
1131 
  preroll--;
1132 
  assert (n->num_bb + preroll * 2 < MAX_BB);
1133 
 
1134 
  prevb = b;
1135 
  prevart_b = b;
1136 
  /* Duplicate the BB */
1137 
  for (t = 0; t < preroll; t++) {
1138 
    cuc_bb *pb = &n->bb[prevart_b];
1139 
    /* Add new block and set links */
1140 
    b1 = n->num_bb++;
1141 
    cpy_bb (&n->bb[b1], ob);
1142 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1143 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1144 
 
1145 
    /* Set predecessor's successor */
1146 
    if (n->bb[prevb].next[0] == b) {
1147 
      n->bb[prevb].next[0] = b1;
1148 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1149 
      else pb->next[1] = b1 + 1;
1150 
      n->bb[b1].next[1] = b1 + 1;
1151 
    } else if (n->bb[prevb].next[1] == b) {
1152 
      if (pb->next[0] < 0) pb->next[0] = b1 + 1;
1153 
      else pb->next[1] = b1 + 1;
1154 
      n->bb[b1].next[0] = b1 + 1;
1155 
      n->bb[prevb].next[1] = b1;
1156 
    } else assert (0);
1157 
 
1158 
    /* Set predecessor */
1159 
    n->bb[b1].prev[0] = prevb;
1160 
    n->bb[b1].prev[1] = -1;
1161 
 
1162 
    /* Relocate backward references to current instance and forward references
1163 
       to previous one */
1164 
    relocate_bb (&n->bb[b1], b, b1, prevb);
1165 
 
1166 
    /* add artificial block, just to join accesses */
1167 
    b1 = n->num_bb++;
1168 
    cpy_bb (&n->bb[b1], ob);
1169 
    n->bb[b1].cnt = 0;
1170 
 
1171 
    for (i = 0; i < ob->ninsn - 1; i++) {
1172 
      ii = &n->bb[b1].insn[i];
1173 
      if (ob->insn[i].opt[0] & OPT_DEST) {
1174 
        change_insn_type (ii, II_CMOV);
1175 
        ii->op[0] = -1; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1176 
        ii->op[1] = REF (prevart_b, i); ii->opt[1] = OPT_REF;
1177 
        ii->op[2] = REF (b1 - 1, i); ii->opt[2] = OPT_REF;
1178 
 
1179 
        /* Take left one, if we should have finished the first iteration*/
1180 
        if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1181 
          ii->op[3] = pb->insn[pb->ninsn - 1].op[1]; ii->opt[3] = pb->insn[pb->ninsn - 1].opt[1];
1182 
        } else {
1183 
          assert (pb->insn[pb->ninsn - 1].type & IT_COND);
1184 
          ii->op[3] = REF (prevart_b, pb->ninsn - 1); ii->opt[3] = OPT_REF;
1185 
        }
1186 
        ii->dep = NULL;
1187 
        ii->type = 0;
1188 
      } else {
1189 
        change_insn_type (ii, II_NOP);
1190 
      }
1191 
    }
1192 
 
1193 
    /* Add sfor instruction at the end, prioritizing flags */
1194 
    ii = &n->bb[b1].insn[ob->ninsn - 1];
1195 
    change_insn_type (ii, II_SFOR);
1196 
    ii->op[0] = FLAG_REG; ii->opt[0] = OPT_REGISTER | OPT_DEST;
1197 
    if (pb->insn[pb->ninsn - 1].type & IT_BRANCH) {
1198 
      ii->op[1] = pb->insn[pb->ninsn - 1].op[1];
1199 
      ii->opt[1] = pb->insn[pb->ninsn - 1].opt[1];
1200 
    } else {
1201 
      ii->op[1] = REF (prevart_b, pb->ninsn - 1);
1202 
      ii->opt[1] = OPT_REF;
1203 
    }
1204 
    if (n->bb[b1 - 1].insn[pb->ninsn - 1].type & IT_BRANCH) {
1205 
      ii->op[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].op[1];
1206 
      ii->opt[2] = n->bb[b1 - 1].insn[pb->ninsn - 1].opt[1];
1207 
    } else {
1208 
      ii->op[2] = REF (b1 - 1, pb->ninsn - 1);
1209 
      ii->opt[2] = OPT_REF;
1210 
    }
1211 
    ii->opt[3] = OPT_NONE;
1212 
    ii->type = IT_COND;
1213 
 
1214 
    /* Only one should be in loop, so we remove any INLOOP flags from duplicates */
1215 
    n->bb[b1].type &= ~(BB_END | BB_INLOOP);
1216 
    n->bb[b1].prev[0] = prevart_b;
1217 
    n->bb[b1].prev[1] = b1 - 1;
1218 
    n->bb[b1].next[0] = ob->next[0] == b ? ob->next[1] : ob->next[0];
1219 
    n->bb[b1].next[1] = -1;
1220 
 
1221 
    prevb = b1 - 1;
1222 
    prevart_b = b1;
1223 
  }
1224 
  if (ob->type & BB_END) {
1225 
    n->bb[prevart_b].type |= BB_END;
1226 
    n->bb[b].type &= ~BB_END;
1227 
  }
1228 
 
1229 
  //print_cuc_bb (n, "preroll1");
1230 
  /* repair BB after loop, to point back to latest artificial BB */
1231 
  b1 = n->bb[prevart_b].next[0];
1232 
  if (b1 >= 0) {
1233 897 markom 
    if (n->bb[b1].prev[0] == b) n->bb[b1].prev[0] = prevart_b;
1234 
    else if (n->bb[b1].prev[1] == b) n->bb[b1].prev[1] = prevart_b;
1235 879 markom 
    else assert (0);
1236 
  }
1237 
 
1238 
  /* Relink to itself */
1239 
  /* Set predecessor's successor */
1240 
  if (n->bb[prevb].next[0] == b) n->bb[prevb].next[0] = prevb;
1241 
  else if (n->bb[prevb].next[1] == b) n->bb[prevb].next[1] = prevb;
1242 
  else assert (0);
1243 
  n->bb[prevb].prev[1] = prevb;
1244 
 
1245 
  if (n->bb[b].prev[0] == b) {
1246 
    n->bb[b].prev[0] = n->bb[b].prev[1];
1247 
    n->bb[b].prev[1] = -1;
1248 
  } else if (n->bb[b].prev[1] == b) {
1249 
    n->bb[b].prev[1] = -1;
1250 
  }
1251 
 
1252 
  //print_cuc_bb (n, "preroll2");
1253 
 
1254 
  /* Relocate backward references to current instance and forward references
1255 
     to previous one */
1256 
  relocate_bb (&n->bb[b], b, b, prevb);
1257 
 
1258 
  /* Relocate all other blocks to point to latest prevart_b */
1259 
  for (i = 0; i < f->num_bb; i++)
1260 
    if (i != b) relocate_bb (&n->bb[i], b, prevart_b, prevart_b);
1261 
 
1262 
  return n;
1263 
}
1264 
 
1265 
/* Unroll loop b unroll times and return new function. Original
1266 
   function is unmodified. */
1267 
cuc_func *preunroll_loop (cuc_func *f, int b, int preroll, int unroll, char *bb_filename)
1268 
{
1269 
  int b1, i;
1270 
  cuc_func *n, *t;
1271 
  int *counts;
1272 
  int *bb_reloc;
1273 
 
1274 
  if (preroll > 1) {
1275 
    t = preroll_loop (f, b, preroll);
1276 
    b1 = t->num_bb - 2;
1277 
    if (unroll > 1) {
1278 
      //print_cuc_bb (t, "preunroll1");
1279 
      n = unroll_loop (t, b1, unroll);
1280 
      free_func (t);
1281 
    } else n = t;
1282 
  } else {
1283 
    b1 = b;
1284 897 markom 
    if (unroll > 1) n = unroll_loop (f, b1, unroll);
1285 
    else return dup_func (f);
1286 879 markom 
  }
1287 
 
1288 897 markom 
  /* Assign new counts to functions */
1289 879 markom 
  assert (counts = (int *)malloc (sizeof (int) * (preroll - 1 + unroll)));
1290 
  count_bb_seq (n, b, bb_filename, counts, preroll, unroll);
1291 
  for (i = 0; i < preroll - 1 + unroll; i++) {
1292 
    if (i == 0) b1 = b;
1293 
    else b1 = f->num_bb + (i - 1) * 2;
1294 
    n->bb[b1].cnt = counts[i];
1295 
  }
1296 
 
1297 
  //print_cuc_bb (n, "preunroll");
1298 
  free (counts);
1299 
  return n;
1300 
}

powered by: WebSVN 2.1.0

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