xref: /netbsd-src/external/gpl3/gcc.old/dist/gcc/tree-ssa-loop-ivcanon.c (revision 8feb0f0b7eaff0608f8350bbfa3098827b4bb91b)
1 /* Induction variable canonicalization and loop peeling.
2    Copyright (C) 2004-2020 Free Software Foundation, Inc.
3 
4 This file is part of GCC.
5 
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
10 
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
14 for more details.
15 
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3.  If not see
18 <http://www.gnu.org/licenses/>.  */
19 
20 /* This pass detects the loops that iterate a constant number of times,
21    adds a canonical induction variable (step -1, tested against 0)
22    and replaces the exit test.  This enables the less powerful rtl
23    level analysis to use this information.
24 
25    This might spoil the code in some cases (by increasing register pressure).
26    Note that in the case the new variable is not needed, ivopts will get rid
27    of it, so it might only be a problem when there are no other linear induction
28    variables.  In that case the created optimization possibilities are likely
29    to pay up.
30 
31    We also perform
32      - complete unrolling (or peeling) when the loops is rolling few enough
33        times
34      - simple peeling (i.e. copying few initial iterations prior the loop)
35        when number of iteration estimate is known (typically by the profile
36        info).  */
37 
38 #include "config.h"
39 #include "system.h"
40 #include "coretypes.h"
41 #include "backend.h"
42 #include "tree.h"
43 #include "gimple.h"
44 #include "cfghooks.h"
45 #include "tree-pass.h"
46 #include "ssa.h"
47 #include "cgraph.h"
48 #include "gimple-pretty-print.h"
49 #include "fold-const.h"
50 #include "profile.h"
51 #include "gimple-fold.h"
52 #include "tree-eh.h"
53 #include "gimple-iterator.h"
54 #include "tree-cfg.h"
55 #include "tree-ssa-loop-manip.h"
56 #include "tree-ssa-loop-niter.h"
57 #include "tree-ssa-loop.h"
58 #include "tree-into-ssa.h"
59 #include "cfgloop.h"
60 #include "tree-chrec.h"
61 #include "tree-scalar-evolution.h"
62 #include "tree-inline.h"
63 #include "tree-cfgcleanup.h"
64 #include "builtins.h"
65 #include "tree-ssa-sccvn.h"
66 #include "dbgcnt.h"
67 
68 /* Specifies types of loops that may be unrolled.  */
69 
70 enum unroll_level
71 {
72   UL_SINGLE_ITER,	/* Only loops that exit immediately in the first
73 			   iteration.  */
74   UL_NO_GROWTH,		/* Only loops whose unrolling will not cause increase
75 			   of code size.  */
76   UL_ALL		/* All suitable loops.  */
77 };
78 
79 /* Adds a canonical induction variable to LOOP iterating NITER times.  EXIT
80    is the exit edge whose condition is replaced.  The ssa versions of the new
81    IV before and after increment will be stored in VAR_BEFORE and VAR_AFTER
82    if they are not NULL.  */
83 
84 void
85 create_canonical_iv (class loop *loop, edge exit, tree niter,
86 		     tree *var_before = NULL, tree *var_after = NULL)
87 {
88   edge in;
89   tree type, var;
90   gcond *cond;
91   gimple_stmt_iterator incr_at;
92   enum tree_code cmp;
93 
94   if (dump_file && (dump_flags & TDF_DETAILS))
95     {
96       fprintf (dump_file, "Added canonical iv to loop %d, ", loop->num);
97       print_generic_expr (dump_file, niter, TDF_SLIM);
98       fprintf (dump_file, " iterations.\n");
99     }
100 
101   cond = as_a <gcond *> (last_stmt (exit->src));
102   in = EDGE_SUCC (exit->src, 0);
103   if (in == exit)
104     in = EDGE_SUCC (exit->src, 1);
105 
106   /* Note that we do not need to worry about overflows, since
107      type of niter is always unsigned and all comparisons are
108      just for equality/nonequality -- i.e. everything works
109      with a modulo arithmetics.  */
110 
111   type = TREE_TYPE (niter);
112   niter = fold_build2 (PLUS_EXPR, type,
113 		       niter,
114 		       build_int_cst (type, 1));
115   incr_at = gsi_last_bb (in->src);
116   create_iv (niter,
117 	     build_int_cst (type, -1),
118 	     NULL_TREE, loop,
119 	     &incr_at, false, var_before, &var);
120   if (var_after)
121     *var_after = var;
122 
123   cmp = (exit->flags & EDGE_TRUE_VALUE) ? EQ_EXPR : NE_EXPR;
124   gimple_cond_set_code (cond, cmp);
125   gimple_cond_set_lhs (cond, var);
126   gimple_cond_set_rhs (cond, build_int_cst (type, 0));
127   update_stmt (cond);
128 }
129 
130 /* Describe size of loop as detected by tree_estimate_loop_size.  */
131 struct loop_size
132 {
133   /* Number of instructions in the loop.  */
134   int overall;
135 
136   /* Number of instructions that will be likely optimized out in
137      peeled iterations of loop  (i.e. computation based on induction
138      variable where induction variable starts at known constant.)  */
139   int eliminated_by_peeling;
140 
141   /* Same statistics for last iteration of loop: it is smaller because
142      instructions after exit are not executed.  */
143   int last_iteration;
144   int last_iteration_eliminated_by_peeling;
145 
146   /* If some IV computation will become constant.  */
147   bool constant_iv;
148 
149   /* Number of call stmts that are not a builtin and are pure or const
150      present on the hot path.  */
151   int num_pure_calls_on_hot_path;
152   /* Number of call stmts that are not a builtin and are not pure nor const
153      present on the hot path.  */
154   int num_non_pure_calls_on_hot_path;
155   /* Number of statements other than calls in the loop.  */
156   int non_call_stmts_on_hot_path;
157   /* Number of branches seen on the hot path.  */
158   int num_branches_on_hot_path;
159 };
160 
161 /* Return true if OP in STMT will be constant after peeling LOOP.  */
162 
163 static bool
constant_after_peeling(tree op,gimple * stmt,class loop * loop)164 constant_after_peeling (tree op, gimple *stmt, class loop *loop)
165 {
166   if (CONSTANT_CLASS_P (op))
167     return true;
168 
169   /* We can still fold accesses to constant arrays when index is known.  */
170   if (TREE_CODE (op) != SSA_NAME)
171     {
172       tree base = op;
173 
174       /* First make fast look if we see constant array inside.  */
175       while (handled_component_p (base))
176 	base = TREE_OPERAND (base, 0);
177       if ((DECL_P (base)
178 	   && ctor_for_folding (base) != error_mark_node)
179 	  || CONSTANT_CLASS_P (base))
180 	{
181 	  /* If so, see if we understand all the indices.  */
182 	  base = op;
183 	  while (handled_component_p (base))
184 	    {
185 	      if (TREE_CODE (base) == ARRAY_REF
186 		  && !constant_after_peeling (TREE_OPERAND (base, 1), stmt, loop))
187 		return false;
188 	      base = TREE_OPERAND (base, 0);
189 	    }
190 	  return true;
191 	}
192       return false;
193     }
194 
195   /* Induction variables are constants when defined in loop.  */
196   if (loop_containing_stmt (stmt) != loop)
197     return false;
198   tree ev = analyze_scalar_evolution (loop, op);
199   if (chrec_contains_undetermined (ev)
200       || chrec_contains_symbols (ev))
201     return false;
202   return true;
203 }
204 
205 /* Computes an estimated number of insns in LOOP.
206    EXIT (if non-NULL) is an exite edge that will be eliminated in all but last
207    iteration of the loop.
208    EDGE_TO_CANCEL (if non-NULL) is an non-exit edge eliminated in the last iteration
209    of loop.
210    Return results in SIZE, estimate benefits for complete unrolling exiting by EXIT.
211    Stop estimating after UPPER_BOUND is met.  Return true in this case.  */
212 
213 static bool
tree_estimate_loop_size(class loop * loop,edge exit,edge edge_to_cancel,struct loop_size * size,int upper_bound)214 tree_estimate_loop_size (class loop *loop, edge exit, edge edge_to_cancel,
215 			 struct loop_size *size, int upper_bound)
216 {
217   basic_block *body = get_loop_body (loop);
218   gimple_stmt_iterator gsi;
219   unsigned int i;
220   bool after_exit;
221   vec<basic_block> path = get_loop_hot_path (loop);
222 
223   size->overall = 0;
224   size->eliminated_by_peeling = 0;
225   size->last_iteration = 0;
226   size->last_iteration_eliminated_by_peeling = 0;
227   size->num_pure_calls_on_hot_path = 0;
228   size->num_non_pure_calls_on_hot_path = 0;
229   size->non_call_stmts_on_hot_path = 0;
230   size->num_branches_on_hot_path = 0;
231   size->constant_iv = 0;
232 
233   if (dump_file && (dump_flags & TDF_DETAILS))
234     fprintf (dump_file, "Estimating sizes for loop %i\n", loop->num);
235   for (i = 0; i < loop->num_nodes; i++)
236     {
237       if (edge_to_cancel && body[i] != edge_to_cancel->src
238 	  && dominated_by_p (CDI_DOMINATORS, body[i], edge_to_cancel->src))
239 	after_exit = true;
240       else
241 	after_exit = false;
242       if (dump_file && (dump_flags & TDF_DETAILS))
243 	fprintf (dump_file, " BB: %i, after_exit: %i\n", body[i]->index,
244 		 after_exit);
245 
246       for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi); gsi_next (&gsi))
247 	{
248 	  gimple *stmt = gsi_stmt (gsi);
249 	  int num = estimate_num_insns (stmt, &eni_size_weights);
250 	  bool likely_eliminated = false;
251 	  bool likely_eliminated_last = false;
252 	  bool likely_eliminated_peeled = false;
253 
254 	  if (dump_file && (dump_flags & TDF_DETAILS))
255 	    {
256 	      fprintf (dump_file, "  size: %3i ", num);
257 	      print_gimple_stmt (dump_file, gsi_stmt (gsi), 0);
258 	    }
259 
260 	  /* Look for reasons why we might optimize this stmt away. */
261 
262 	  if (!gimple_has_side_effects (stmt))
263 	    {
264 	      /* Exit conditional.  */
265 	      if (exit && body[i] == exit->src
266 		  && stmt == last_stmt (exit->src))
267 		{
268 		  if (dump_file && (dump_flags & TDF_DETAILS))
269 		    fprintf (dump_file, "   Exit condition will be eliminated "
270 			     "in peeled copies.\n");
271 		  likely_eliminated_peeled = true;
272 		}
273 	      if (edge_to_cancel && body[i] == edge_to_cancel->src
274 		  && stmt == last_stmt (edge_to_cancel->src))
275 		{
276 		  if (dump_file && (dump_flags & TDF_DETAILS))
277 		    fprintf (dump_file, "   Exit condition will be eliminated "
278 			     "in last copy.\n");
279 		  likely_eliminated_last = true;
280 		}
281 	      /* Sets of IV variables  */
282 	      if (gimple_code (stmt) == GIMPLE_ASSIGN
283 		  && constant_after_peeling (gimple_assign_lhs (stmt), stmt, loop))
284 		{
285 		  if (dump_file && (dump_flags & TDF_DETAILS))
286 		    fprintf (dump_file, "   Induction variable computation will"
287 			     " be folded away.\n");
288 		  likely_eliminated = true;
289 		}
290 	      /* Assignments of IV variables.  */
291 	      else if (gimple_code (stmt) == GIMPLE_ASSIGN
292 		       && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME
293 		       && constant_after_peeling (gimple_assign_rhs1 (stmt),
294 						  stmt, loop)
295 		       && (gimple_assign_rhs_class (stmt) != GIMPLE_BINARY_RHS
296 			   || constant_after_peeling (gimple_assign_rhs2 (stmt),
297 						      stmt, loop))
298 		       && gimple_assign_rhs_class (stmt) != GIMPLE_TERNARY_RHS)
299 		{
300 		  size->constant_iv = true;
301 		  if (dump_file && (dump_flags & TDF_DETAILS))
302 		    fprintf (dump_file,
303 			     "   Constant expression will be folded away.\n");
304 		  likely_eliminated = true;
305 		}
306 	      /* Conditionals.  */
307 	      else if ((gimple_code (stmt) == GIMPLE_COND
308 			&& constant_after_peeling (gimple_cond_lhs (stmt), stmt,
309 						   loop)
310 			&& constant_after_peeling (gimple_cond_rhs (stmt), stmt,
311 						   loop)
312 			/* We don't simplify all constant compares so make sure
313 			   they are not both constant already.  See PR70288.  */
314 			&& (! is_gimple_min_invariant (gimple_cond_lhs (stmt))
315 			    || ! is_gimple_min_invariant
316 				 (gimple_cond_rhs (stmt))))
317 		       || (gimple_code (stmt) == GIMPLE_SWITCH
318 			   && constant_after_peeling (gimple_switch_index (
319 							as_a <gswitch *>
320 							  (stmt)),
321 						      stmt, loop)
322 			   && ! is_gimple_min_invariant
323 				   (gimple_switch_index
324 				      (as_a <gswitch *> (stmt)))))
325 		{
326 		  if (dump_file && (dump_flags & TDF_DETAILS))
327 		    fprintf (dump_file, "   Constant conditional.\n");
328 		  likely_eliminated = true;
329 		}
330 	    }
331 
332 	  size->overall += num;
333 	  if (likely_eliminated || likely_eliminated_peeled)
334 	    size->eliminated_by_peeling += num;
335 	  if (!after_exit)
336 	    {
337 	      size->last_iteration += num;
338 	      if (likely_eliminated || likely_eliminated_last)
339 		size->last_iteration_eliminated_by_peeling += num;
340 	    }
341 	  if ((size->overall * 3 / 2 - size->eliminated_by_peeling
342 	      - size->last_iteration_eliminated_by_peeling) > upper_bound)
343 	    {
344               free (body);
345 	      path.release ();
346 	      return true;
347 	    }
348 	}
349     }
350   while (path.length ())
351     {
352       basic_block bb = path.pop ();
353       for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
354 	{
355 	  gimple *stmt = gsi_stmt (gsi);
356 	  if (gimple_code (stmt) == GIMPLE_CALL
357 	      && !gimple_inexpensive_call_p (as_a <gcall *>  (stmt)))
358 	    {
359 	      int flags = gimple_call_flags (stmt);
360 	      if (flags & (ECF_PURE | ECF_CONST))
361 		size->num_pure_calls_on_hot_path++;
362 	      else
363 		size->num_non_pure_calls_on_hot_path++;
364 	      size->num_branches_on_hot_path ++;
365 	    }
366 	  /* Count inexpensive calls as non-calls, because they will likely
367 	     expand inline.  */
368 	  else if (gimple_code (stmt) != GIMPLE_DEBUG)
369 	    size->non_call_stmts_on_hot_path++;
370 	  if (((gimple_code (stmt) == GIMPLE_COND
371 	        && (!constant_after_peeling (gimple_cond_lhs (stmt), stmt, loop)
372 		    || !constant_after_peeling (gimple_cond_rhs (stmt), stmt,
373 						loop)))
374 	       || (gimple_code (stmt) == GIMPLE_SWITCH
375 		   && !constant_after_peeling (gimple_switch_index (
376 						 as_a <gswitch *> (stmt)),
377 					       stmt, loop)))
378 	      && (!exit || bb != exit->src))
379 	    size->num_branches_on_hot_path++;
380 	}
381     }
382   path.release ();
383   if (dump_file && (dump_flags & TDF_DETAILS))
384     fprintf (dump_file, "size: %i-%i, last_iteration: %i-%i\n", size->overall,
385     	     size->eliminated_by_peeling, size->last_iteration,
386 	     size->last_iteration_eliminated_by_peeling);
387 
388   free (body);
389   return false;
390 }
391 
392 /* Estimate number of insns of completely unrolled loop.
393    It is (NUNROLL + 1) * size of loop body with taking into account
394    the fact that in last copy everything after exit conditional
395    is dead and that some instructions will be eliminated after
396    peeling.
397 
398    Loop body is likely going to simplify further, this is difficult
399    to guess, we just decrease the result by 1/3.  */
400 
401 static unsigned HOST_WIDE_INT
estimated_unrolled_size(struct loop_size * size,unsigned HOST_WIDE_INT nunroll)402 estimated_unrolled_size (struct loop_size *size,
403 			 unsigned HOST_WIDE_INT nunroll)
404 {
405   HOST_WIDE_INT unr_insns = ((nunroll)
406   			     * (HOST_WIDE_INT) (size->overall
407 			     			- size->eliminated_by_peeling));
408   if (!nunroll)
409     unr_insns = 0;
410   unr_insns += size->last_iteration - size->last_iteration_eliminated_by_peeling;
411 
412   unr_insns = unr_insns * 2 / 3;
413   if (unr_insns <= 0)
414     unr_insns = 1;
415 
416   return unr_insns;
417 }
418 
419 /* Loop LOOP is known to not loop.  See if there is an edge in the loop
420    body that can be remove to make the loop to always exit and at
421    the same time it does not make any code potentially executed
422    during the last iteration dead.
423 
424    After complete unrolling we still may get rid of the conditional
425    on the exit in the last copy even if we have no idea what it does.
426    This is quite common case for loops of form
427 
428      int a[5];
429      for (i=0;i<b;i++)
430        a[i]=0;
431 
432    Here we prove the loop to iterate 5 times but we do not know
433    it from induction variable.
434 
435    For now we handle only simple case where there is exit condition
436    just before the latch block and the latch block contains no statements
437    with side effect that may otherwise terminate the execution of loop
438    (such as by EH or by terminating the program or longjmp).
439 
440    In the general case we may want to cancel the paths leading to statements
441    loop-niter identified as having undefined effect in the last iteration.
442    The other cases are hopefully rare and will be cleaned up later.  */
443 
444 static edge
loop_edge_to_cancel(class loop * loop)445 loop_edge_to_cancel (class loop *loop)
446 {
447   vec<edge> exits;
448   unsigned i;
449   edge edge_to_cancel;
450   gimple_stmt_iterator gsi;
451 
452   /* We want only one predecestor of the loop.  */
453   if (EDGE_COUNT (loop->latch->preds) > 1)
454     return NULL;
455 
456   exits = get_loop_exit_edges (loop);
457 
458   FOR_EACH_VEC_ELT (exits, i, edge_to_cancel)
459     {
460        /* Find the other edge than the loop exit
461           leaving the conditoinal.  */
462        if (EDGE_COUNT (edge_to_cancel->src->succs) != 2)
463          continue;
464        if (EDGE_SUCC (edge_to_cancel->src, 0) == edge_to_cancel)
465          edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 1);
466        else
467          edge_to_cancel = EDGE_SUCC (edge_to_cancel->src, 0);
468 
469       /* We only can handle conditionals.  */
470       if (!(edge_to_cancel->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE)))
471 	continue;
472 
473       /* We should never have conditionals in the loop latch. */
474       gcc_assert (edge_to_cancel->dest != loop->header);
475 
476       /* Check that it leads to loop latch.  */
477       if (edge_to_cancel->dest != loop->latch)
478         continue;
479 
480       exits.release ();
481 
482       /* Verify that the code in loop latch does nothing that may end program
483          execution without really reaching the exit.  This may include
484 	 non-pure/const function calls, EH statements, volatile ASMs etc.  */
485       for (gsi = gsi_start_bb (loop->latch); !gsi_end_p (gsi); gsi_next (&gsi))
486 	if (gimple_has_side_effects (gsi_stmt (gsi)))
487 	   return NULL;
488       return edge_to_cancel;
489     }
490   exits.release ();
491   return NULL;
492 }
493 
494 /* Remove all tests for exits that are known to be taken after LOOP was
495    peeled NPEELED times. Put gcc_unreachable before every statement
496    known to not be executed.  */
497 
498 static bool
remove_exits_and_undefined_stmts(class loop * loop,unsigned int npeeled)499 remove_exits_and_undefined_stmts (class loop *loop, unsigned int npeeled)
500 {
501   class nb_iter_bound *elt;
502   bool changed = false;
503 
504   for (elt = loop->bounds; elt; elt = elt->next)
505     {
506       /* If statement is known to be undefined after peeling, turn it
507 	 into unreachable (or trap when debugging experience is supposed
508 	 to be good).  */
509       if (!elt->is_exit
510 	  && wi::ltu_p (elt->bound, npeeled))
511 	{
512 	  gimple_stmt_iterator gsi = gsi_for_stmt (elt->stmt);
513 	  gcall *stmt = gimple_build_call
514 	      (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0);
515 	  gimple_set_location (stmt, gimple_location (elt->stmt));
516 	  gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
517 	  split_block (gimple_bb (stmt), stmt);
518 	  changed = true;
519 	  if (dump_file && (dump_flags & TDF_DETAILS))
520 	    {
521 	      fprintf (dump_file, "Forced statement unreachable: ");
522 	      print_gimple_stmt (dump_file, elt->stmt, 0);
523 	    }
524 	}
525       /* If we know the exit will be taken after peeling, update.  */
526       else if (elt->is_exit
527 	       && wi::leu_p (elt->bound, npeeled))
528 	{
529 	  basic_block bb = gimple_bb (elt->stmt);
530 	  edge exit_edge = EDGE_SUCC (bb, 0);
531 
532 	  if (dump_file && (dump_flags & TDF_DETAILS))
533 	    {
534 	      fprintf (dump_file, "Forced exit to be taken: ");
535 	      print_gimple_stmt (dump_file, elt->stmt, 0);
536 	    }
537 	  if (!loop_exit_edge_p (loop, exit_edge))
538 	    exit_edge = EDGE_SUCC (bb, 1);
539 	  exit_edge->probability = profile_probability::always ();
540 	  gcc_checking_assert (loop_exit_edge_p (loop, exit_edge));
541 	  gcond *cond_stmt = as_a <gcond *> (elt->stmt);
542 	  if (exit_edge->flags & EDGE_TRUE_VALUE)
543 	    gimple_cond_make_true (cond_stmt);
544 	  else
545 	    gimple_cond_make_false (cond_stmt);
546 	  update_stmt (cond_stmt);
547 	  changed = true;
548 	}
549     }
550   return changed;
551 }
552 
553 /* Remove all exits that are known to be never taken because of the loop bound
554    discovered.  */
555 
556 static bool
remove_redundant_iv_tests(class loop * loop)557 remove_redundant_iv_tests (class loop *loop)
558 {
559   class nb_iter_bound *elt;
560   bool changed = false;
561 
562   if (!loop->any_upper_bound)
563     return false;
564   for (elt = loop->bounds; elt; elt = elt->next)
565     {
566       /* Exit is pointless if it won't be taken before loop reaches
567 	 upper bound.  */
568       if (elt->is_exit && loop->any_upper_bound
569           && wi::ltu_p (loop->nb_iterations_upper_bound, elt->bound))
570 	{
571 	  basic_block bb = gimple_bb (elt->stmt);
572 	  edge exit_edge = EDGE_SUCC (bb, 0);
573 	  class tree_niter_desc niter;
574 
575 	  if (!loop_exit_edge_p (loop, exit_edge))
576 	    exit_edge = EDGE_SUCC (bb, 1);
577 
578 	  /* Only when we know the actual number of iterations, not
579 	     just a bound, we can remove the exit.  */
580 	  if (!number_of_iterations_exit (loop, exit_edge,
581 					  &niter, false, false)
582 	      || !integer_onep (niter.assumptions)
583 	      || !integer_zerop (niter.may_be_zero)
584 	      || !niter.niter
585 	      || TREE_CODE (niter.niter) != INTEGER_CST
586 	      || !wi::ltu_p (loop->nb_iterations_upper_bound,
587 			     wi::to_widest (niter.niter)))
588 	    continue;
589 
590 	  if (dump_file && (dump_flags & TDF_DETAILS))
591 	    {
592 	      fprintf (dump_file, "Removed pointless exit: ");
593 	      print_gimple_stmt (dump_file, elt->stmt, 0);
594 	    }
595 	  gcond *cond_stmt = as_a <gcond *> (elt->stmt);
596 	  if (exit_edge->flags & EDGE_TRUE_VALUE)
597 	    gimple_cond_make_false (cond_stmt);
598 	  else
599 	    gimple_cond_make_true (cond_stmt);
600 	  update_stmt (cond_stmt);
601 	  changed = true;
602 	}
603     }
604   return changed;
605 }
606 
607 /* Stores loops that will be unlooped and edges that will be removed
608    after we process whole loop tree. */
609 static vec<loop_p> loops_to_unloop;
610 static vec<int> loops_to_unloop_nunroll;
611 static vec<edge> edges_to_remove;
612 /* Stores loops that has been peeled.  */
613 static bitmap peeled_loops;
614 
615 /* Cancel all fully unrolled loops by putting __builtin_unreachable
616    on the latch edge.
617    We do it after all unrolling since unlooping moves basic blocks
618    across loop boundaries trashing loop closed SSA form as well
619    as SCEV info needed to be intact during unrolling.
620 
621    IRRED_INVALIDATED is used to bookkeep if information about
622    irreducible regions may become invalid as a result
623    of the transformation.
624    LOOP_CLOSED_SSA_INVALIDATED is used to bookkepp the case
625    when we need to go into loop closed SSA form.  */
626 
627 static void
unloop_loops(bitmap loop_closed_ssa_invalidated,bool * irred_invalidated)628 unloop_loops (bitmap loop_closed_ssa_invalidated,
629 	      bool *irred_invalidated)
630 {
631   while (loops_to_unloop.length ())
632     {
633       class loop *loop = loops_to_unloop.pop ();
634       int n_unroll = loops_to_unloop_nunroll.pop ();
635       basic_block latch = loop->latch;
636       edge latch_edge = loop_latch_edge (loop);
637       int flags = latch_edge->flags;
638       location_t locus = latch_edge->goto_locus;
639       gcall *stmt;
640       gimple_stmt_iterator gsi;
641 
642       remove_exits_and_undefined_stmts (loop, n_unroll);
643 
644       /* Unloop destroys the latch edge.  */
645       unloop (loop, irred_invalidated, loop_closed_ssa_invalidated);
646 
647       /* Create new basic block for the latch edge destination and wire
648 	 it in.  */
649       stmt = gimple_build_call (builtin_decl_implicit (BUILT_IN_UNREACHABLE), 0);
650       latch_edge = make_edge (latch, create_basic_block (NULL, NULL, latch), flags);
651       latch_edge->probability = profile_probability::never ();
652       latch_edge->flags |= flags;
653       latch_edge->goto_locus = locus;
654 
655       add_bb_to_loop (latch_edge->dest, current_loops->tree_root);
656       latch_edge->dest->count = profile_count::zero ();
657       set_immediate_dominator (CDI_DOMINATORS, latch_edge->dest, latch_edge->src);
658 
659       gsi = gsi_start_bb (latch_edge->dest);
660       gsi_insert_after (&gsi, stmt, GSI_NEW_STMT);
661     }
662   loops_to_unloop.release ();
663   loops_to_unloop_nunroll.release ();
664 
665   /* Remove edges in peeled copies.  Given remove_path removes dominated
666      regions we need to cope with removal of already removed paths.  */
667   unsigned i;
668   edge e;
669   auto_vec<int, 20> src_bbs;
670   src_bbs.reserve_exact (edges_to_remove.length ());
671   FOR_EACH_VEC_ELT (edges_to_remove, i, e)
672     src_bbs.quick_push (e->src->index);
673   FOR_EACH_VEC_ELT (edges_to_remove, i, e)
674     if (BASIC_BLOCK_FOR_FN (cfun, src_bbs[i]))
675       {
676 	bool ok = remove_path (e, irred_invalidated,
677 			       loop_closed_ssa_invalidated);
678 	gcc_assert (ok);
679       }
680   edges_to_remove.release ();
681 }
682 
683 /* Tries to unroll LOOP completely, i.e. NITER times.
684    UL determines which loops we are allowed to unroll.
685    EXIT is the exit of the loop that should be eliminated.
686    MAXITER specfy bound on number of iterations, -1 if it is
687    not known or too large for HOST_WIDE_INT.  The location
688    LOCUS corresponding to the loop is used when emitting
689    a summary of the unroll to the dump file.  */
690 
691 static bool
try_unroll_loop_completely(class loop * loop,edge exit,tree niter,bool may_be_zero,enum unroll_level ul,HOST_WIDE_INT maxiter,dump_user_location_t locus,bool allow_peel)692 try_unroll_loop_completely (class loop *loop,
693 			    edge exit, tree niter, bool may_be_zero,
694 			    enum unroll_level ul,
695 			    HOST_WIDE_INT maxiter,
696 			    dump_user_location_t locus, bool allow_peel)
697 {
698   unsigned HOST_WIDE_INT n_unroll = 0;
699   bool n_unroll_found = false;
700   edge edge_to_cancel = NULL;
701 
702   /* See if we proved number of iterations to be low constant.
703 
704      EXIT is an edge that will be removed in all but last iteration of
705      the loop.
706 
707      EDGE_TO_CACNEL is an edge that will be removed from the last iteration
708      of the unrolled sequence and is expected to make the final loop not
709      rolling.
710 
711      If the number of execution of loop is determined by standard induction
712      variable test, then EXIT and EDGE_TO_CANCEL are the two edges leaving
713      from the iv test.  */
714   if (tree_fits_uhwi_p (niter))
715     {
716       n_unroll = tree_to_uhwi (niter);
717       n_unroll_found = true;
718       edge_to_cancel = EDGE_SUCC (exit->src, 0);
719       if (edge_to_cancel == exit)
720 	edge_to_cancel = EDGE_SUCC (exit->src, 1);
721     }
722   /* We do not know the number of iterations and thus we cannot eliminate
723      the EXIT edge.  */
724   else
725     exit = NULL;
726 
727   /* See if we can improve our estimate by using recorded loop bounds.  */
728   if ((allow_peel || maxiter == 0 || ul == UL_NO_GROWTH)
729       && maxiter >= 0
730       && (!n_unroll_found || (unsigned HOST_WIDE_INT)maxiter < n_unroll))
731     {
732       n_unroll = maxiter;
733       n_unroll_found = true;
734       /* Loop terminates before the IV variable test, so we cannot
735 	 remove it in the last iteration.  */
736       edge_to_cancel = NULL;
737     }
738 
739   if (!n_unroll_found)
740     return false;
741 
742   if (!loop->unroll
743       && n_unroll > (unsigned) param_max_completely_peel_times)
744     {
745       if (dump_file && (dump_flags & TDF_DETAILS))
746 	fprintf (dump_file, "Not unrolling loop %d "
747 		 "(--param max-completely-peel-times limit reached).\n",
748 		 loop->num);
749       return false;
750     }
751 
752   if (!edge_to_cancel)
753     edge_to_cancel = loop_edge_to_cancel (loop);
754 
755   if (n_unroll)
756     {
757       if (ul == UL_SINGLE_ITER)
758 	return false;
759 
760       if (loop->unroll)
761 	{
762 	  /* If the unrolling factor is too large, bail out.  */
763 	  if (n_unroll > (unsigned)loop->unroll)
764 	    {
765 	      if (dump_file && (dump_flags & TDF_DETAILS))
766 		fprintf (dump_file,
767 			 "Not unrolling loop %d: "
768 			 "user didn't want it unrolled completely.\n",
769 			 loop->num);
770 	      return false;
771 	    }
772 	}
773       else
774 	{
775 	  struct loop_size size;
776 	  /* EXIT can be removed only if we are sure it passes first N_UNROLL
777 	     iterations.  */
778 	  bool remove_exit = (exit && niter
779 			      && TREE_CODE (niter) == INTEGER_CST
780 			      && wi::leu_p (n_unroll, wi::to_widest (niter)));
781 	  bool large
782 	    = tree_estimate_loop_size
783 		(loop, remove_exit ? exit : NULL, edge_to_cancel, &size,
784 		 param_max_completely_peeled_insns);
785 	  if (large)
786 	    {
787 	      if (dump_file && (dump_flags & TDF_DETAILS))
788 		fprintf (dump_file, "Not unrolling loop %d: it is too large.\n",
789 			 loop->num);
790 	      return false;
791 	    }
792 
793 	  unsigned HOST_WIDE_INT ninsns = size.overall;
794 	  unsigned HOST_WIDE_INT unr_insns
795 	    = estimated_unrolled_size (&size, n_unroll);
796 	  if (dump_file && (dump_flags & TDF_DETAILS))
797 	    {
798 	      fprintf (dump_file, "  Loop size: %d\n", (int) ninsns);
799 	      fprintf (dump_file, "  Estimated size after unrolling: %d\n",
800 		       (int) unr_insns);
801 	    }
802 
803 	  /* If the code is going to shrink, we don't need to be extra
804 	     cautious on guessing if the unrolling is going to be
805 	     profitable.  */
806 	  if (unr_insns
807 	      /* If there is IV variable that will become constant, we
808 		 save one instruction in the loop prologue we do not
809 		 account otherwise.  */
810 	      <= ninsns + (size.constant_iv != false))
811 	    ;
812 	  /* We unroll only inner loops, because we do not consider it
813 	     profitable otheriwse.  We still can cancel loopback edge
814 	     of not rolling loop; this is always a good idea.  */
815 	  else if (ul == UL_NO_GROWTH)
816 	    {
817 	      if (dump_file && (dump_flags & TDF_DETAILS))
818 		fprintf (dump_file, "Not unrolling loop %d: size would grow.\n",
819 			 loop->num);
820 	      return false;
821 	    }
822 	  /* Outer loops tend to be less interesting candidates for
823 	     complete unrolling unless we can do a lot of propagation
824 	     into the inner loop body.  For now we disable outer loop
825 	     unrolling when the code would grow.  */
826 	  else if (loop->inner)
827 	    {
828 	      if (dump_file && (dump_flags & TDF_DETAILS))
829 		fprintf (dump_file, "Not unrolling loop %d: "
830 			 "it is not innermost and code would grow.\n",
831 			 loop->num);
832 	      return false;
833 	    }
834 	  /* If there is call on a hot path through the loop, then
835 	     there is most probably not much to optimize.  */
836 	  else if (size.num_non_pure_calls_on_hot_path)
837 	    {
838 	      if (dump_file && (dump_flags & TDF_DETAILS))
839 		fprintf (dump_file, "Not unrolling loop %d: "
840 			 "contains call and code would grow.\n",
841 			 loop->num);
842 	      return false;
843 	    }
844 	  /* If there is pure/const call in the function, then we can
845 	     still optimize the unrolled loop body if it contains some
846 	     other interesting code than the calls and code storing or
847 	     cumulating the return value.  */
848 	  else if (size.num_pure_calls_on_hot_path
849 		   /* One IV increment, one test, one ivtmp store and
850 		      one useful stmt.  That is about minimal loop
851 		      doing pure call.  */
852 		   && (size.non_call_stmts_on_hot_path
853 		       <= 3 + size.num_pure_calls_on_hot_path))
854 	    {
855 	      if (dump_file && (dump_flags & TDF_DETAILS))
856 		fprintf (dump_file, "Not unrolling loop %d: "
857 			 "contains just pure calls and code would grow.\n",
858 			 loop->num);
859 	      return false;
860 	    }
861 	  /* Complete unrolling is major win when control flow is
862 	     removed and one big basic block is created.  If the loop
863 	     contains control flow the optimization may still be a win
864 	     because of eliminating the loop overhead but it also may
865 	     blow the branch predictor tables.  Limit number of
866 	     branches on the hot path through the peeled sequence.  */
867 	  else if (size.num_branches_on_hot_path * (int)n_unroll
868 		   > param_max_peel_branches)
869 	    {
870 	      if (dump_file && (dump_flags & TDF_DETAILS))
871 		fprintf (dump_file, "Not unrolling loop %d: "
872 			 "number of branches on hot path in the unrolled "
873 			 "sequence reaches --param max-peel-branches limit.\n",
874 			 loop->num);
875 	      return false;
876 	    }
877 	  else if (unr_insns
878 		   > (unsigned) param_max_completely_peeled_insns)
879 	    {
880 	      if (dump_file && (dump_flags & TDF_DETAILS))
881 		fprintf (dump_file, "Not unrolling loop %d: "
882 			 "number of insns in the unrolled sequence reaches "
883 			 "--param max-completely-peeled-insns limit.\n",
884 			 loop->num);
885 	      return false;
886 	    }
887 	}
888 
889       if (!dbg_cnt (gimple_unroll))
890 	return false;
891 
892       initialize_original_copy_tables ();
893       auto_sbitmap wont_exit (n_unroll + 1);
894       if (exit && niter
895 	  && TREE_CODE (niter) == INTEGER_CST
896 	  && wi::leu_p (n_unroll, wi::to_widest (niter)))
897 	{
898 	  bitmap_ones (wont_exit);
899 	  if (wi::eq_p (wi::to_widest (niter), n_unroll)
900 	      || edge_to_cancel)
901 	    bitmap_clear_bit (wont_exit, 0);
902 	}
903       else
904 	{
905 	  exit = NULL;
906 	  bitmap_clear (wont_exit);
907 	}
908       if (may_be_zero)
909 	bitmap_clear_bit (wont_exit, 1);
910 
911       if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop),
912 						 n_unroll, wont_exit,
913 						 exit, &edges_to_remove,
914 						 DLTHE_FLAG_UPDATE_FREQ
915 						 | DLTHE_FLAG_COMPLETTE_PEEL))
916 	{
917           free_original_copy_tables ();
918 	  if (dump_file && (dump_flags & TDF_DETAILS))
919 	    fprintf (dump_file, "Failed to duplicate the loop\n");
920 	  return false;
921 	}
922 
923       free_original_copy_tables ();
924     }
925 
926   /* Remove the conditional from the last copy of the loop.  */
927   if (edge_to_cancel)
928     {
929       gcond *cond = as_a <gcond *> (last_stmt (edge_to_cancel->src));
930       force_edge_cold (edge_to_cancel, true);
931       if (edge_to_cancel->flags & EDGE_TRUE_VALUE)
932 	gimple_cond_make_false (cond);
933       else
934 	gimple_cond_make_true (cond);
935       update_stmt (cond);
936       /* Do not remove the path, as doing so may remove outer loop and
937 	 confuse bookkeeping code in tree_unroll_loops_completely.  */
938     }
939 
940   /* Store the loop for later unlooping and exit removal.  */
941   loops_to_unloop.safe_push (loop);
942   loops_to_unloop_nunroll.safe_push (n_unroll);
943 
944   if (dump_enabled_p ())
945     {
946       if (!n_unroll)
947         dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus,
948                          "loop turned into non-loop; it never loops\n");
949       else
950         {
951           dump_printf_loc (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, locus,
952                            "loop with %d iterations completely unrolled",
953 			   (int) n_unroll);
954           if (loop->header->count.initialized_p ())
955             dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS,
956                          " (header execution count %d)",
957                          (int)loop->header->count.to_gcov_type ());
958           dump_printf (MSG_OPTIMIZED_LOCATIONS | TDF_DETAILS, "\n");
959         }
960     }
961 
962   if (dump_file && (dump_flags & TDF_DETAILS))
963     {
964       if (exit)
965         fprintf (dump_file, "Exit condition of peeled iterations was "
966 		 "eliminated.\n");
967       if (edge_to_cancel)
968         fprintf (dump_file, "Last iteration exit edge was proved true.\n");
969       else
970         fprintf (dump_file, "Latch of last iteration was marked by "
971 		 "__builtin_unreachable ().\n");
972     }
973 
974   return true;
975 }
976 
977 /* Return number of instructions after peeling.  */
978 static unsigned HOST_WIDE_INT
estimated_peeled_sequence_size(struct loop_size * size,unsigned HOST_WIDE_INT npeel)979 estimated_peeled_sequence_size (struct loop_size *size,
980 			        unsigned HOST_WIDE_INT npeel)
981 {
982   return MAX (npeel * (HOST_WIDE_INT) (size->overall
983 			     	       - size->eliminated_by_peeling), 1);
984 }
985 
986 /* If the loop is expected to iterate N times and is
987    small enough, duplicate the loop body N+1 times before
988    the loop itself.  This way the hot path will never
989    enter the loop.
990    Parameters are the same as for try_unroll_loops_completely */
991 
992 static bool
try_peel_loop(class loop * loop,edge exit,tree niter,bool may_be_zero,HOST_WIDE_INT maxiter)993 try_peel_loop (class loop *loop,
994 	       edge exit, tree niter, bool may_be_zero,
995 	       HOST_WIDE_INT maxiter)
996 {
997   HOST_WIDE_INT npeel;
998   struct loop_size size;
999   int peeled_size;
1000 
1001   if (!flag_peel_loops
1002       || param_max_peel_times <= 0
1003       || !peeled_loops)
1004     return false;
1005 
1006   if (bitmap_bit_p (peeled_loops, loop->num))
1007     {
1008       if (dump_file)
1009         fprintf (dump_file, "Not peeling: loop is already peeled\n");
1010       return false;
1011     }
1012 
1013   /* We don't peel loops that will be unrolled as this can duplicate a
1014      loop more times than the user requested.  */
1015   if (loop->unroll)
1016     {
1017       if (dump_file)
1018         fprintf (dump_file, "Not peeling: user didn't want it peeled.\n");
1019       return false;
1020     }
1021 
1022   /* Peel only innermost loops.
1023      While the code is perfectly capable of peeling non-innermost loops,
1024      the heuristics would probably need some improvements. */
1025   if (loop->inner)
1026     {
1027       if (dump_file)
1028 	fprintf (dump_file, "Not peeling: outer loop\n");
1029       return false;
1030     }
1031 
1032   if (!optimize_loop_for_speed_p (loop))
1033     {
1034       if (dump_file)
1035 	fprintf (dump_file, "Not peeling: cold loop\n");
1036       return false;
1037     }
1038 
1039   /* Check if there is an estimate on the number of iterations.  */
1040   npeel = estimated_loop_iterations_int (loop);
1041   if (npeel < 0)
1042     npeel = likely_max_loop_iterations_int (loop);
1043   if (npeel < 0)
1044     {
1045       if (dump_file)
1046         fprintf (dump_file, "Not peeling: number of iterations is not "
1047 	         "estimated\n");
1048       return false;
1049     }
1050   if (maxiter >= 0 && maxiter <= npeel)
1051     {
1052       if (dump_file)
1053 	fprintf (dump_file, "Not peeling: upper bound is known so can "
1054 		 "unroll completely\n");
1055       return false;
1056     }
1057 
1058   /* We want to peel estimated number of iterations + 1 (so we never
1059      enter the loop on quick path).  Check against PARAM_MAX_PEEL_TIMES
1060      and be sure to avoid overflows.  */
1061   if (npeel > param_max_peel_times - 1)
1062     {
1063       if (dump_file)
1064 	fprintf (dump_file, "Not peeling: rolls too much "
1065 		 "(%i + 1 > --param max-peel-times)\n", (int) npeel);
1066       return false;
1067     }
1068   npeel++;
1069 
1070   /* Check peeled loops size.  */
1071   tree_estimate_loop_size (loop, exit, NULL, &size,
1072 			   param_max_peeled_insns);
1073   if ((peeled_size = estimated_peeled_sequence_size (&size, (int) npeel))
1074       > param_max_peeled_insns)
1075     {
1076       if (dump_file)
1077 	fprintf (dump_file, "Not peeling: peeled sequence size is too large "
1078 		 "(%i insns > --param max-peel-insns)", peeled_size);
1079       return false;
1080     }
1081 
1082   if (!dbg_cnt (gimple_unroll))
1083     return false;
1084 
1085   /* Duplicate possibly eliminating the exits.  */
1086   initialize_original_copy_tables ();
1087   auto_sbitmap wont_exit (npeel + 1);
1088   if (exit && niter
1089       && TREE_CODE (niter) == INTEGER_CST
1090       && wi::leu_p (npeel, wi::to_widest (niter)))
1091     {
1092       bitmap_ones (wont_exit);
1093       bitmap_clear_bit (wont_exit, 0);
1094     }
1095   else
1096     {
1097       exit = NULL;
1098       bitmap_clear (wont_exit);
1099     }
1100   if (may_be_zero)
1101     bitmap_clear_bit (wont_exit, 1);
1102   if (!gimple_duplicate_loop_to_header_edge (loop, loop_preheader_edge (loop),
1103 					     npeel, wont_exit,
1104 					     exit, &edges_to_remove,
1105 					     DLTHE_FLAG_UPDATE_FREQ))
1106     {
1107       free_original_copy_tables ();
1108       return false;
1109     }
1110   free_original_copy_tables ();
1111   if (dump_file && (dump_flags & TDF_DETAILS))
1112     {
1113       fprintf (dump_file, "Peeled loop %d, %i times.\n",
1114 	       loop->num, (int) npeel);
1115     }
1116   if (loop->any_estimate)
1117     {
1118       if (wi::ltu_p (npeel, loop->nb_iterations_estimate))
1119         loop->nb_iterations_estimate -= npeel;
1120       else
1121 	loop->nb_iterations_estimate = 0;
1122     }
1123   if (loop->any_upper_bound)
1124     {
1125       if (wi::ltu_p (npeel, loop->nb_iterations_upper_bound))
1126         loop->nb_iterations_upper_bound -= npeel;
1127       else
1128         loop->nb_iterations_upper_bound = 0;
1129     }
1130   if (loop->any_likely_upper_bound)
1131     {
1132       if (wi::ltu_p (npeel, loop->nb_iterations_likely_upper_bound))
1133 	loop->nb_iterations_likely_upper_bound -= npeel;
1134       else
1135 	{
1136 	  loop->any_estimate = true;
1137 	  loop->nb_iterations_estimate = 0;
1138 	  loop->nb_iterations_likely_upper_bound = 0;
1139 	}
1140     }
1141   profile_count entry_count = profile_count::zero ();
1142 
1143   edge e;
1144   edge_iterator ei;
1145   FOR_EACH_EDGE (e, ei, loop->header->preds)
1146     if (e->src != loop->latch)
1147       {
1148 	if (e->src->count.initialized_p ())
1149 	  entry_count += e->src->count;
1150 	gcc_assert (!flow_bb_inside_loop_p (loop, e->src));
1151       }
1152   profile_probability p;
1153   p = entry_count.probability_in (loop->header->count);
1154   scale_loop_profile (loop, p, 0);
1155   bitmap_set_bit (peeled_loops, loop->num);
1156   return true;
1157 }
1158 /* Adds a canonical induction variable to LOOP if suitable.
1159    CREATE_IV is true if we may create a new iv.  UL determines
1160    which loops we are allowed to completely unroll.  If TRY_EVAL is true, we try
1161    to determine the number of iterations of a loop by direct evaluation.
1162    Returns true if cfg is changed.   */
1163 
1164 static bool
canonicalize_loop_induction_variables(class loop * loop,bool create_iv,enum unroll_level ul,bool try_eval,bool allow_peel)1165 canonicalize_loop_induction_variables (class loop *loop,
1166 				       bool create_iv, enum unroll_level ul,
1167 				       bool try_eval, bool allow_peel)
1168 {
1169   edge exit = NULL;
1170   tree niter;
1171   HOST_WIDE_INT maxiter;
1172   bool modified = false;
1173   dump_user_location_t locus;
1174   class tree_niter_desc niter_desc;
1175   bool may_be_zero = false;
1176 
1177   /* For unrolling allow conditional constant or zero iterations, thus
1178      perform loop-header copying on-the-fly.  */
1179   exit = single_exit (loop);
1180   niter = chrec_dont_know;
1181   if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false))
1182     {
1183       niter = niter_desc.niter;
1184       may_be_zero
1185 	= niter_desc.may_be_zero && !integer_zerop (niter_desc.may_be_zero);
1186     }
1187   if (TREE_CODE (niter) == INTEGER_CST)
1188     locus = last_stmt (exit->src);
1189   else
1190     {
1191       /* For non-constant niter fold may_be_zero into niter again.  */
1192       if (may_be_zero)
1193 	{
1194 	  if (COMPARISON_CLASS_P (niter_desc.may_be_zero))
1195 	    niter = fold_build3 (COND_EXPR, TREE_TYPE (niter),
1196 				 niter_desc.may_be_zero,
1197 				 build_int_cst (TREE_TYPE (niter), 0), niter);
1198 	  else
1199 	    niter = chrec_dont_know;
1200 	  may_be_zero = false;
1201 	}
1202 
1203       /* If the loop has more than one exit, try checking all of them
1204 	 for # of iterations determinable through scev.  */
1205       if (!exit)
1206 	niter = find_loop_niter (loop, &exit);
1207 
1208       /* Finally if everything else fails, try brute force evaluation.  */
1209       if (try_eval
1210 	  && (chrec_contains_undetermined (niter)
1211 	      || TREE_CODE (niter) != INTEGER_CST))
1212 	niter = find_loop_niter_by_eval (loop, &exit);
1213 
1214       if (exit)
1215         locus = last_stmt (exit->src);
1216 
1217       if (TREE_CODE (niter) != INTEGER_CST)
1218 	exit = NULL;
1219     }
1220 
1221   /* We work exceptionally hard here to estimate the bound
1222      by find_loop_niter_by_eval.  Be sure to keep it for future.  */
1223   if (niter && TREE_CODE (niter) == INTEGER_CST)
1224     {
1225       vec<edge> exits = get_loop_exit_edges  (loop);
1226       record_niter_bound (loop, wi::to_widest (niter),
1227 			  exit == single_likely_exit (loop, exits), true);
1228       exits.release ();
1229     }
1230 
1231   /* Force re-computation of loop bounds so we can remove redundant exits.  */
1232   maxiter = max_loop_iterations_int (loop);
1233 
1234   if (dump_file && (dump_flags & TDF_DETAILS)
1235       && TREE_CODE (niter) == INTEGER_CST)
1236     {
1237       fprintf (dump_file, "Loop %d iterates ", loop->num);
1238       print_generic_expr (dump_file, niter, TDF_SLIM);
1239       fprintf (dump_file, " times.\n");
1240     }
1241   if (dump_file && (dump_flags & TDF_DETAILS)
1242       && maxiter >= 0)
1243     {
1244       fprintf (dump_file, "Loop %d iterates at most %i times.\n", loop->num,
1245 	       (int)maxiter);
1246     }
1247   if (dump_file && (dump_flags & TDF_DETAILS)
1248       && likely_max_loop_iterations_int (loop) >= 0)
1249     {
1250       fprintf (dump_file, "Loop %d likely iterates at most %i times.\n",
1251 	       loop->num, (int)likely_max_loop_iterations_int (loop));
1252     }
1253 
1254   /* Remove exits that are known to be never taken based on loop bound.
1255      Needs to be called after compilation of max_loop_iterations_int that
1256      populates the loop bounds.  */
1257   modified |= remove_redundant_iv_tests (loop);
1258 
1259   if (try_unroll_loop_completely (loop, exit, niter, may_be_zero, ul,
1260 				  maxiter, locus, allow_peel))
1261     return true;
1262 
1263   if (create_iv
1264       && niter && !chrec_contains_undetermined (niter)
1265       && exit && just_once_each_iteration_p (loop, exit->src))
1266     {
1267       tree iv_niter = niter;
1268       if (may_be_zero)
1269 	{
1270 	  if (COMPARISON_CLASS_P (niter_desc.may_be_zero))
1271 	    iv_niter = fold_build3 (COND_EXPR, TREE_TYPE (iv_niter),
1272 				    niter_desc.may_be_zero,
1273 				    build_int_cst (TREE_TYPE (iv_niter), 0),
1274 				    iv_niter);
1275 	  else
1276 	    iv_niter = NULL_TREE;
1277 	}
1278       if (iv_niter)
1279 	create_canonical_iv (loop, exit, iv_niter);
1280     }
1281 
1282   if (ul == UL_ALL)
1283     modified |= try_peel_loop (loop, exit, niter, may_be_zero, maxiter);
1284 
1285   return modified;
1286 }
1287 
1288 /* The main entry point of the pass.  Adds canonical induction variables
1289    to the suitable loops.  */
1290 
1291 unsigned int
canonicalize_induction_variables(void)1292 canonicalize_induction_variables (void)
1293 {
1294   class loop *loop;
1295   bool changed = false;
1296   bool irred_invalidated = false;
1297   bitmap loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL);
1298 
1299   estimate_numbers_of_iterations (cfun);
1300 
1301   FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
1302     {
1303       changed |= canonicalize_loop_induction_variables (loop,
1304 							true, UL_SINGLE_ITER,
1305 							true, false);
1306     }
1307   gcc_assert (!need_ssa_update_p (cfun));
1308 
1309   unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated);
1310   if (irred_invalidated
1311       && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
1312     mark_irreducible_loops ();
1313 
1314   /* Clean up the information about numbers of iterations, since brute force
1315      evaluation could reveal new information.  */
1316   free_numbers_of_iterations_estimates (cfun);
1317   scev_reset ();
1318 
1319   if (!bitmap_empty_p (loop_closed_ssa_invalidated))
1320     {
1321       gcc_checking_assert (loops_state_satisfies_p (LOOP_CLOSED_SSA));
1322       rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1323     }
1324   BITMAP_FREE (loop_closed_ssa_invalidated);
1325 
1326   if (changed)
1327     return TODO_cleanup_cfg;
1328   return 0;
1329 }
1330 
1331 /* Process loops from innermost to outer, stopping at the innermost
1332    loop we unrolled.  */
1333 
1334 static bool
tree_unroll_loops_completely_1(bool may_increase_size,bool unroll_outer,bitmap father_bbs,class loop * loop)1335 tree_unroll_loops_completely_1 (bool may_increase_size, bool unroll_outer,
1336 				bitmap father_bbs, class loop *loop)
1337 {
1338   class loop *loop_father;
1339   bool changed = false;
1340   class loop *inner;
1341   enum unroll_level ul;
1342   unsigned num = number_of_loops (cfun);
1343 
1344   /* Process inner loops first.  Don't walk loops added by the recursive
1345      calls because SSA form is not up-to-date.  They can be handled in the
1346      next iteration.  */
1347   bitmap child_father_bbs = NULL;
1348   for (inner = loop->inner; inner != NULL; inner = inner->next)
1349     if ((unsigned) inner->num < num)
1350       {
1351 	if (!child_father_bbs)
1352 	  child_father_bbs = BITMAP_ALLOC (NULL);
1353 	if (tree_unroll_loops_completely_1 (may_increase_size, unroll_outer,
1354 					    child_father_bbs, inner))
1355 	  {
1356 	    bitmap_ior_into (father_bbs, child_father_bbs);
1357 	    bitmap_clear (child_father_bbs);
1358 	    changed = true;
1359 	  }
1360       }
1361   if (child_father_bbs)
1362     BITMAP_FREE (child_father_bbs);
1363 
1364   /* If we changed an inner loop we cannot process outer loops in this
1365      iteration because SSA form is not up-to-date.  Continue with
1366      siblings of outer loops instead.  */
1367   if (changed)
1368     {
1369       /* If we are recorded as father clear all other fathers that
1370          are necessarily covered already to avoid redundant work.  */
1371       if (bitmap_bit_p (father_bbs, loop->header->index))
1372 	{
1373 	  bitmap_clear (father_bbs);
1374 	  bitmap_set_bit (father_bbs, loop->header->index);
1375 	}
1376       return true;
1377     }
1378 
1379   /* Don't unroll #pragma omp simd loops until the vectorizer
1380      attempts to vectorize those.  */
1381   if (loop->force_vectorize)
1382     return false;
1383 
1384   /* Try to unroll this loop.  */
1385   loop_father = loop_outer (loop);
1386   if (!loop_father)
1387     return false;
1388 
1389   if (loop->unroll > 1)
1390     ul = UL_ALL;
1391   else if (may_increase_size && optimize_loop_nest_for_speed_p (loop)
1392       /* Unroll outermost loops only if asked to do so or they do
1393 	 not cause code growth.  */
1394       && (unroll_outer || loop_outer (loop_father)))
1395     ul = UL_ALL;
1396   else
1397     ul = UL_NO_GROWTH;
1398 
1399   if (canonicalize_loop_induction_variables
1400         (loop, false, ul, !flag_tree_loop_ivcanon, unroll_outer))
1401     {
1402       /* If we'll continue unrolling, we need to propagate constants
1403 	 within the new basic blocks to fold away induction variable
1404 	 computations; otherwise, the size might blow up before the
1405 	 iteration is complete and the IR eventually cleaned up.  */
1406       if (loop_outer (loop_father))
1407 	{
1408 	  /* Once we process our father we will have processed
1409 	     the fathers of our children as well, so avoid doing
1410 	     redundant work and clear fathers we've gathered sofar.  */
1411 	  bitmap_clear (father_bbs);
1412 	  bitmap_set_bit (father_bbs, loop_father->header->index);
1413 	}
1414 
1415       return true;
1416     }
1417 
1418   return false;
1419 }
1420 
1421 /* Unroll LOOPS completely if they iterate just few times.  Unless
1422    MAY_INCREASE_SIZE is true, perform the unrolling only if the
1423    size of the code does not increase.  */
1424 
1425 static unsigned int
tree_unroll_loops_completely(bool may_increase_size,bool unroll_outer)1426 tree_unroll_loops_completely (bool may_increase_size, bool unroll_outer)
1427 {
1428   bitmap father_bbs = BITMAP_ALLOC (NULL);
1429   bool changed;
1430   int iteration = 0;
1431   bool irred_invalidated = false;
1432 
1433   estimate_numbers_of_iterations (cfun);
1434 
1435   do
1436     {
1437       changed = false;
1438       bitmap loop_closed_ssa_invalidated = NULL;
1439 
1440       if (loops_state_satisfies_p (LOOP_CLOSED_SSA))
1441 	loop_closed_ssa_invalidated = BITMAP_ALLOC (NULL);
1442 
1443       free_numbers_of_iterations_estimates (cfun);
1444       estimate_numbers_of_iterations (cfun);
1445 
1446       changed = tree_unroll_loops_completely_1 (may_increase_size,
1447 						unroll_outer, father_bbs,
1448 						current_loops->tree_root);
1449       if (changed)
1450 	{
1451 	  unsigned i;
1452 
1453           unloop_loops (loop_closed_ssa_invalidated, &irred_invalidated);
1454 
1455 	  /* We cannot use TODO_update_ssa_no_phi because VOPS gets confused.  */
1456 	  if (loop_closed_ssa_invalidated
1457 	      && !bitmap_empty_p (loop_closed_ssa_invalidated))
1458             rewrite_into_loop_closed_ssa (loop_closed_ssa_invalidated,
1459 					  TODO_update_ssa);
1460 	  else
1461 	    update_ssa (TODO_update_ssa);
1462 
1463 	  /* father_bbs is a bitmap of loop father header BB indices.
1464 	     Translate that to what non-root loops these BBs belong to now.  */
1465 	  bitmap_iterator bi;
1466 	  bitmap fathers = BITMAP_ALLOC (NULL);
1467 	  EXECUTE_IF_SET_IN_BITMAP (father_bbs, 0, i, bi)
1468 	    {
1469 	      basic_block unrolled_loop_bb = BASIC_BLOCK_FOR_FN (cfun, i);
1470 	      if (! unrolled_loop_bb)
1471 		continue;
1472 	      if (loop_outer (unrolled_loop_bb->loop_father))
1473 		bitmap_set_bit (fathers,
1474 				unrolled_loop_bb->loop_father->num);
1475 	    }
1476 	  bitmap_clear (father_bbs);
1477 	  /* Propagate the constants within the new basic blocks.  */
1478 	  EXECUTE_IF_SET_IN_BITMAP (fathers, 0, i, bi)
1479 	    {
1480 	      loop_p father = get_loop (cfun, i);
1481 	      bitmap exit_bbs = BITMAP_ALLOC (NULL);
1482 	      loop_exit *exit = father->exits->next;
1483 	      while (exit->e)
1484 		{
1485 		  bitmap_set_bit (exit_bbs, exit->e->dest->index);
1486 		  exit = exit->next;
1487 		}
1488 	      do_rpo_vn (cfun, loop_preheader_edge (father), exit_bbs);
1489 	    }
1490 	  BITMAP_FREE (fathers);
1491 
1492 	  /* This will take care of removing completely unrolled loops
1493 	     from the loop structures so we can continue unrolling now
1494 	     innermost loops.  */
1495 	  if (cleanup_tree_cfg ())
1496 	    update_ssa (TODO_update_ssa_only_virtuals);
1497 
1498 	  /* Clean up the information about numbers of iterations, since
1499 	     complete unrolling might have invalidated it.  */
1500 	  scev_reset ();
1501 	  if (flag_checking && loops_state_satisfies_p (LOOP_CLOSED_SSA))
1502 	    verify_loop_closed_ssa (true);
1503 	}
1504       if (loop_closed_ssa_invalidated)
1505         BITMAP_FREE (loop_closed_ssa_invalidated);
1506     }
1507   while (changed
1508 	 && ++iteration <= param_max_unroll_iterations);
1509 
1510   BITMAP_FREE (father_bbs);
1511 
1512   if (irred_invalidated
1513       && loops_state_satisfies_p (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS))
1514     mark_irreducible_loops ();
1515 
1516   return 0;
1517 }
1518 
1519 /* Canonical induction variable creation pass.  */
1520 
1521 namespace {
1522 
1523 const pass_data pass_data_iv_canon =
1524 {
1525   GIMPLE_PASS, /* type */
1526   "ivcanon", /* name */
1527   OPTGROUP_LOOP, /* optinfo_flags */
1528   TV_TREE_LOOP_IVCANON, /* tv_id */
1529   ( PROP_cfg | PROP_ssa ), /* properties_required */
1530   0, /* properties_provided */
1531   0, /* properties_destroyed */
1532   0, /* todo_flags_start */
1533   0, /* todo_flags_finish */
1534 };
1535 
1536 class pass_iv_canon : public gimple_opt_pass
1537 {
1538 public:
pass_iv_canon(gcc::context * ctxt)1539   pass_iv_canon (gcc::context *ctxt)
1540     : gimple_opt_pass (pass_data_iv_canon, ctxt)
1541   {}
1542 
1543   /* opt_pass methods: */
gate(function *)1544   virtual bool gate (function *) { return flag_tree_loop_ivcanon != 0; }
1545   virtual unsigned int execute (function *fun);
1546 
1547 }; // class pass_iv_canon
1548 
1549 unsigned int
execute(function * fun)1550 pass_iv_canon::execute (function *fun)
1551 {
1552   if (number_of_loops (fun) <= 1)
1553     return 0;
1554 
1555   return canonicalize_induction_variables ();
1556 }
1557 
1558 } // anon namespace
1559 
1560 gimple_opt_pass *
make_pass_iv_canon(gcc::context * ctxt)1561 make_pass_iv_canon (gcc::context *ctxt)
1562 {
1563   return new pass_iv_canon (ctxt);
1564 }
1565 
1566 /* Complete unrolling of loops.  */
1567 
1568 namespace {
1569 
1570 const pass_data pass_data_complete_unroll =
1571 {
1572   GIMPLE_PASS, /* type */
1573   "cunroll", /* name */
1574   OPTGROUP_LOOP, /* optinfo_flags */
1575   TV_COMPLETE_UNROLL, /* tv_id */
1576   ( PROP_cfg | PROP_ssa ), /* properties_required */
1577   0, /* properties_provided */
1578   0, /* properties_destroyed */
1579   0, /* todo_flags_start */
1580   0, /* todo_flags_finish */
1581 };
1582 
1583 class pass_complete_unroll : public gimple_opt_pass
1584 {
1585 public:
pass_complete_unroll(gcc::context * ctxt)1586   pass_complete_unroll (gcc::context *ctxt)
1587     : gimple_opt_pass (pass_data_complete_unroll, ctxt)
1588   {}
1589 
1590   /* opt_pass methods: */
1591   virtual unsigned int execute (function *);
1592 
1593 }; // class pass_complete_unroll
1594 
1595 unsigned int
execute(function * fun)1596 pass_complete_unroll::execute (function *fun)
1597 {
1598   if (number_of_loops (fun) <= 1)
1599     return 0;
1600 
1601   /* If we ever decide to run loop peeling more than once, we will need to
1602      track loops already peeled in loop structures themselves to avoid
1603      re-peeling the same loop multiple times.  */
1604   if (flag_peel_loops)
1605     peeled_loops = BITMAP_ALLOC (NULL);
1606   unsigned int val = tree_unroll_loops_completely (flag_cunroll_grow_size,
1607 						   true);
1608   if (peeled_loops)
1609     {
1610       BITMAP_FREE (peeled_loops);
1611       peeled_loops = NULL;
1612     }
1613   return val;
1614 }
1615 
1616 } // anon namespace
1617 
1618 gimple_opt_pass *
make_pass_complete_unroll(gcc::context * ctxt)1619 make_pass_complete_unroll (gcc::context *ctxt)
1620 {
1621   return new pass_complete_unroll (ctxt);
1622 }
1623 
1624 /* Complete unrolling of inner loops.  */
1625 
1626 namespace {
1627 
1628 const pass_data pass_data_complete_unrolli =
1629 {
1630   GIMPLE_PASS, /* type */
1631   "cunrolli", /* name */
1632   OPTGROUP_LOOP, /* optinfo_flags */
1633   TV_COMPLETE_UNROLL, /* tv_id */
1634   ( PROP_cfg | PROP_ssa ), /* properties_required */
1635   0, /* properties_provided */
1636   0, /* properties_destroyed */
1637   0, /* todo_flags_start */
1638   0, /* todo_flags_finish */
1639 };
1640 
1641 class pass_complete_unrolli : public gimple_opt_pass
1642 {
1643 public:
pass_complete_unrolli(gcc::context * ctxt)1644   pass_complete_unrolli (gcc::context *ctxt)
1645     : gimple_opt_pass (pass_data_complete_unrolli, ctxt)
1646   {}
1647 
1648   /* opt_pass methods: */
gate(function *)1649   virtual bool gate (function *) { return optimize >= 2; }
1650   virtual unsigned int execute (function *);
1651 
1652 }; // class pass_complete_unrolli
1653 
1654 unsigned int
execute(function * fun)1655 pass_complete_unrolli::execute (function *fun)
1656 {
1657   unsigned ret = 0;
1658 
1659   loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
1660   if (number_of_loops (fun) > 1)
1661     {
1662       scev_initialize ();
1663       ret = tree_unroll_loops_completely (optimize >= 3, false);
1664       scev_finalize ();
1665     }
1666   loop_optimizer_finalize ();
1667 
1668   return ret;
1669 }
1670 
1671 } // anon namespace
1672 
1673 gimple_opt_pass *
make_pass_complete_unrolli(gcc::context * ctxt)1674 make_pass_complete_unrolli (gcc::context *ctxt)
1675 {
1676   return new pass_complete_unrolli (ctxt);
1677 }
1678 
1679 
1680