1 /* High-level loop manipulation functions.
2    Copyright (C) 2004-2013 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 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "tm.h"
24 #include "tree.h"
25 #include "tm_p.h"
26 #include "basic-block.h"
27 #include "tree-flow.h"
28 #include "dumpfile.h"
29 #include "gimple-pretty-print.h"
30 #include "cfgloop.h"
31 #include "tree-pass.h"	/* ??? for TODO_update_ssa but this isn't a pass.  */
32 #include "tree-scalar-evolution.h"
33 #include "params.h"
34 #include "tree-inline.h"
35 #include "langhooks.h"
36 
37 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
38    so that we can free them all at once.  */
39 static bitmap_obstack loop_renamer_obstack;
40 
41 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
42    It is expected that neither BASE nor STEP are shared with other expressions
43    (unless the sharing rules allow this).  Use VAR as a base var_decl for it
44    (if NULL, a new temporary will be created).  The increment will occur at
45    INCR_POS (after it if AFTER is true, before it otherwise).  INCR_POS and
46    AFTER can be computed using standard_iv_increment_position.  The ssa versions
47    of the variable before and after increment will be stored in VAR_BEFORE and
48    VAR_AFTER (unless they are NULL).  */
49 
50 void
create_iv(tree base,tree step,tree var,struct loop * loop,gimple_stmt_iterator * incr_pos,bool after,tree * var_before,tree * var_after)51 create_iv (tree base, tree step, tree var, struct loop *loop,
52 	   gimple_stmt_iterator *incr_pos, bool after,
53 	   tree *var_before, tree *var_after)
54 {
55   gimple stmt;
56   tree initial, step1;
57   gimple_seq stmts;
58   tree vb, va;
59   enum tree_code incr_op = PLUS_EXPR;
60   edge pe = loop_preheader_edge (loop);
61 
62   if (var != NULL_TREE)
63     {
64       vb = make_ssa_name (var, NULL);
65       va = make_ssa_name (var, NULL);
66     }
67   else
68     {
69       vb = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
70       va = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
71     }
72   if (var_before)
73     *var_before = vb;
74   if (var_after)
75     *var_after = va;
76 
77   /* For easier readability of the created code, produce MINUS_EXPRs
78      when suitable.  */
79   if (TREE_CODE (step) == INTEGER_CST)
80     {
81       if (TYPE_UNSIGNED (TREE_TYPE (step)))
82 	{
83 	  step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
84 	  if (tree_int_cst_lt (step1, step))
85 	    {
86 	      incr_op = MINUS_EXPR;
87 	      step = step1;
88 	    }
89 	}
90       else
91 	{
92 	  bool ovf;
93 
94 	  if (!tree_expr_nonnegative_warnv_p (step, &ovf)
95 	      && may_negate_without_overflow_p (step))
96 	    {
97 	      incr_op = MINUS_EXPR;
98 	      step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
99 	    }
100 	}
101     }
102   if (POINTER_TYPE_P (TREE_TYPE (base)))
103     {
104       if (TREE_CODE (base) == ADDR_EXPR)
105 	mark_addressable (TREE_OPERAND (base, 0));
106       step = convert_to_ptrofftype (step);
107       if (incr_op == MINUS_EXPR)
108 	step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
109       incr_op = POINTER_PLUS_EXPR;
110     }
111   /* Gimplify the step if necessary.  We put the computations in front of the
112      loop (i.e. the step should be loop invariant).  */
113   step = force_gimple_operand (step, &stmts, true, NULL_TREE);
114   if (stmts)
115     gsi_insert_seq_on_edge_immediate (pe, stmts);
116 
117   stmt = gimple_build_assign_with_ops (incr_op, va, vb, step);
118   if (after)
119     gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT);
120   else
121     gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT);
122 
123   initial = force_gimple_operand (base, &stmts, true, var);
124   if (stmts)
125     gsi_insert_seq_on_edge_immediate (pe, stmts);
126 
127   stmt = create_phi_node (vb, loop->header);
128   add_phi_arg (stmt, initial, loop_preheader_edge (loop), UNKNOWN_LOCATION);
129   add_phi_arg (stmt, va, loop_latch_edge (loop), UNKNOWN_LOCATION);
130 }
131 
132 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
133    both DEF_LOOP and USE_LOOP.  */
134 
135 static inline struct loop *
find_sibling_superloop(struct loop * use_loop,struct loop * def_loop)136 find_sibling_superloop (struct loop *use_loop, struct loop *def_loop)
137 {
138   unsigned ud = loop_depth (use_loop);
139   unsigned dd = loop_depth (def_loop);
140   gcc_assert (ud > 0 && dd > 0);
141   if (ud > dd)
142     use_loop = superloop_at_depth (use_loop, dd);
143   if (ud < dd)
144     def_loop = superloop_at_depth (def_loop, ud);
145   while (loop_outer (use_loop) != loop_outer (def_loop))
146     {
147       use_loop = loop_outer (use_loop);
148       def_loop = loop_outer (def_loop);
149       gcc_assert (use_loop && def_loop);
150     }
151   return use_loop;
152 }
153 
154 /* DEF_BB is a basic block containing a DEF that needs rewriting into
155    loop-closed SSA form.  USE_BLOCKS is the set of basic blocks containing
156    uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
157    USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
158    ALL_EXITS[I] is the set of all basic blocks that exit loop I.
159 
160    Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
161    or one of its loop fathers, in which DEF is live.  This set is returned
162    in the bitmap LIVE_EXITS.
163 
164    Instead of computing the complete livein set of the def, we use the loop
165    nesting tree as a form of poor man's structure analysis.  This greatly
166    speeds up the analysis, which is important because this function may be
167    called on all SSA names that need rewriting, one at a time.  */
168 
169 static void
compute_live_loop_exits(bitmap live_exits,bitmap use_blocks,bitmap * loop_exits,basic_block def_bb)170 compute_live_loop_exits (bitmap live_exits, bitmap use_blocks,
171 			 bitmap *loop_exits, basic_block def_bb)
172 {
173   unsigned i;
174   bitmap_iterator bi;
175   vec<basic_block> worklist;
176   struct loop *def_loop = def_bb->loop_father;
177   unsigned def_loop_depth = loop_depth (def_loop);
178   bitmap def_loop_exits;
179 
180   /* Normally the work list size is bounded by the number of basic
181      blocks in the largest loop.  We don't know this number, but we
182      can be fairly sure that it will be relatively small.  */
183   worklist.create (MAX (8, n_basic_blocks / 128));
184 
185   EXECUTE_IF_SET_IN_BITMAP (use_blocks, 0, i, bi)
186     {
187       basic_block use_bb = BASIC_BLOCK (i);
188       struct loop *use_loop = use_bb->loop_father;
189       gcc_checking_assert (def_loop != use_loop
190 			   && ! flow_loop_nested_p (def_loop, use_loop));
191       if (! flow_loop_nested_p (use_loop, def_loop))
192 	use_bb = find_sibling_superloop (use_loop, def_loop)->header;
193       if (bitmap_set_bit (live_exits, use_bb->index))
194 	worklist.safe_push (use_bb);
195     }
196 
197   /* Iterate until the worklist is empty.  */
198   while (! worklist.is_empty ())
199     {
200       edge e;
201       edge_iterator ei;
202 
203       /* Pull a block off the worklist.  */
204       basic_block bb = worklist.pop ();
205 
206       /* Make sure we have at least enough room in the work list
207 	 for all predecessors of this block.  */
208       worklist.reserve (EDGE_COUNT (bb->preds));
209 
210       /* For each predecessor block.  */
211       FOR_EACH_EDGE (e, ei, bb->preds)
212 	{
213 	  basic_block pred = e->src;
214 	  struct loop *pred_loop = pred->loop_father;
215 	  unsigned pred_loop_depth = loop_depth (pred_loop);
216 	  bool pred_visited;
217 
218 	  /* We should have met DEF_BB along the way.  */
219 	  gcc_assert (pred != ENTRY_BLOCK_PTR);
220 
221 	  if (pred_loop_depth >= def_loop_depth)
222 	    {
223 	      if (pred_loop_depth > def_loop_depth)
224 		pred_loop = superloop_at_depth (pred_loop, def_loop_depth);
225 	      /* If we've reached DEF_LOOP, our train ends here.  */
226 	      if (pred_loop == def_loop)
227 		continue;
228 	    }
229 	  else if (! flow_loop_nested_p (pred_loop, def_loop))
230 	    pred = find_sibling_superloop (pred_loop, def_loop)->header;
231 
232 	  /* Add PRED to the LIVEIN set.  PRED_VISITED is true if
233 	     we had already added PRED to LIVEIN before.  */
234 	  pred_visited = !bitmap_set_bit (live_exits, pred->index);
235 
236 	  /* If we have visited PRED before, don't add it to the worklist.
237 	     If BB dominates PRED, then we're probably looking at a loop.
238 	     We're only interested in looking up in the dominance tree
239 	     because DEF_BB dominates all the uses.  */
240 	  if (pred_visited || dominated_by_p (CDI_DOMINATORS, pred, bb))
241 	    continue;
242 
243 	  worklist.quick_push (pred);
244 	}
245     }
246   worklist.release ();
247 
248   def_loop_exits = BITMAP_ALLOC (&loop_renamer_obstack);
249   for (struct loop *loop = def_loop;
250        loop != current_loops->tree_root;
251        loop = loop_outer (loop))
252     bitmap_ior_into (def_loop_exits, loop_exits[loop->num]);
253   bitmap_and_into (live_exits, def_loop_exits);
254   BITMAP_FREE (def_loop_exits);
255 }
256 
257 /* Add a loop-closing PHI for VAR in basic block EXIT.  */
258 
259 static void
add_exit_phi(basic_block exit,tree var)260 add_exit_phi (basic_block exit, tree var)
261 {
262   gimple phi;
263   edge e;
264   edge_iterator ei;
265 
266 #ifdef ENABLE_CHECKING
267   /* Check that at least one of the edges entering the EXIT block exits
268      the loop, or a superloop of that loop, that VAR is defined in.  */
269   gimple def_stmt = SSA_NAME_DEF_STMT (var);
270   basic_block def_bb = gimple_bb (def_stmt);
271   FOR_EACH_EDGE (e, ei, exit->preds)
272     {
273       struct loop *aloop = find_common_loop (def_bb->loop_father,
274 					     e->src->loop_father);
275       if (!flow_bb_inside_loop_p (aloop, e->dest))
276 	break;
277     }
278 
279   gcc_checking_assert (e);
280 #endif
281 
282   phi = create_phi_node (NULL_TREE, exit);
283   create_new_def_for (var, phi, gimple_phi_result_ptr (phi));
284   FOR_EACH_EDGE (e, ei, exit->preds)
285     add_phi_arg (phi, var, e, UNKNOWN_LOCATION);
286 
287   if (dump_file && (dump_flags & TDF_DETAILS))
288     {
289       fprintf (dump_file, ";; Created LCSSA PHI: ");
290       print_gimple_stmt (dump_file, phi, 0, dump_flags);
291     }
292 }
293 
294 /* Add exit phis for VAR that is used in LIVEIN.
295    Exits of the loops are stored in LOOP_EXITS.  */
296 
297 static void
add_exit_phis_var(tree var,bitmap use_blocks,bitmap * loop_exits)298 add_exit_phis_var (tree var, bitmap use_blocks, bitmap *loop_exits)
299 {
300   unsigned index;
301   bitmap_iterator bi;
302   basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
303   bitmap live_exits = BITMAP_ALLOC (&loop_renamer_obstack);
304 
305   gcc_checking_assert (! bitmap_bit_p (use_blocks, def_bb->index));
306 
307   compute_live_loop_exits (live_exits, use_blocks, loop_exits, def_bb);
308 
309   EXECUTE_IF_SET_IN_BITMAP (live_exits, 0, index, bi)
310     {
311       add_exit_phi (BASIC_BLOCK (index), var);
312     }
313 
314   BITMAP_FREE (live_exits);
315 }
316 
317 /* Add exit phis for the names marked in NAMES_TO_RENAME.
318    Exits of the loops are stored in EXITS.  Sets of blocks where the ssa
319    names are used are stored in USE_BLOCKS.  */
320 
321 static void
add_exit_phis(bitmap names_to_rename,bitmap * use_blocks,bitmap * loop_exits)322 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap *loop_exits)
323 {
324   unsigned i;
325   bitmap_iterator bi;
326 
327   EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi)
328     {
329       add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits);
330     }
331 }
332 
333 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets.  */
334 
335 static void
get_loops_exits(bitmap * loop_exits)336 get_loops_exits (bitmap *loop_exits)
337 {
338   loop_iterator li;
339   struct loop *loop;
340   unsigned j;
341   edge e;
342 
343   FOR_EACH_LOOP (li, loop, 0)
344     {
345       vec<edge> exit_edges = get_loop_exit_edges (loop);
346       loop_exits[loop->num] = BITMAP_ALLOC (&loop_renamer_obstack);
347       FOR_EACH_VEC_ELT (exit_edges, j, e)
348         bitmap_set_bit (loop_exits[loop->num], e->dest->index);
349       exit_edges.release ();
350     }
351 }
352 
353 /* For USE in BB, if it is used outside of the loop it is defined in,
354    mark it for rewrite.  Record basic block BB where it is used
355    to USE_BLOCKS.  Record the ssa name index to NEED_PHIS bitmap.  */
356 
357 static void
find_uses_to_rename_use(basic_block bb,tree use,bitmap * use_blocks,bitmap need_phis)358 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks,
359 			 bitmap need_phis)
360 {
361   unsigned ver;
362   basic_block def_bb;
363   struct loop *def_loop;
364 
365   if (TREE_CODE (use) != SSA_NAME)
366     return;
367 
368   ver = SSA_NAME_VERSION (use);
369   def_bb = gimple_bb (SSA_NAME_DEF_STMT (use));
370   if (!def_bb)
371     return;
372   def_loop = def_bb->loop_father;
373 
374   /* If the definition is not inside a loop, it is not interesting.  */
375   if (!loop_outer (def_loop))
376     return;
377 
378   /* If the use is not outside of the loop it is defined in, it is not
379      interesting.  */
380   if (flow_bb_inside_loop_p (def_loop, bb))
381     return;
382 
383   /* If we're seeing VER for the first time, we still have to allocate
384      a bitmap for its uses.  */
385   if (bitmap_set_bit (need_phis, ver))
386     use_blocks[ver] = BITMAP_ALLOC (&loop_renamer_obstack);
387   bitmap_set_bit (use_blocks[ver], bb->index);
388 }
389 
390 /* For uses in STMT, mark names that are used outside of the loop they are
391    defined to rewrite.  Record the set of blocks in that the ssa
392    names are defined to USE_BLOCKS and the ssa names themselves to
393    NEED_PHIS.  */
394 
395 static void
find_uses_to_rename_stmt(gimple stmt,bitmap * use_blocks,bitmap need_phis)396 find_uses_to_rename_stmt (gimple stmt, bitmap *use_blocks, bitmap need_phis)
397 {
398   ssa_op_iter iter;
399   tree var;
400   basic_block bb = gimple_bb (stmt);
401 
402   if (is_gimple_debug (stmt))
403     return;
404 
405   FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
406     find_uses_to_rename_use (bb, var, use_blocks, need_phis);
407 }
408 
409 /* Marks names that are used in BB and outside of the loop they are
410    defined in for rewrite.  Records the set of blocks in that the ssa
411    names are defined to USE_BLOCKS.  Record the SSA names that will
412    need exit PHIs in NEED_PHIS.  */
413 
414 static void
find_uses_to_rename_bb(basic_block bb,bitmap * use_blocks,bitmap need_phis)415 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis)
416 {
417   gimple_stmt_iterator bsi;
418   edge e;
419   edge_iterator ei;
420 
421   FOR_EACH_EDGE (e, ei, bb->succs)
422     for (bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi); gsi_next (&bsi))
423       {
424         gimple phi = gsi_stmt (bsi);
425 	if (! virtual_operand_p (gimple_phi_result (phi)))
426 	  find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
427 				   use_blocks, need_phis);
428       }
429 
430   for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
431     find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis);
432 }
433 
434 /* Marks names that are used outside of the loop they are defined in
435    for rewrite.  Records the set of blocks in that the ssa
436    names are defined to USE_BLOCKS.  If CHANGED_BBS is not NULL,
437    scan only blocks in this set.  */
438 
439 static void
find_uses_to_rename(bitmap changed_bbs,bitmap * use_blocks,bitmap need_phis)440 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis)
441 {
442   basic_block bb;
443   unsigned index;
444   bitmap_iterator bi;
445 
446   /* ??? If CHANGED_BBS is empty we rewrite the whole function -- why?  */
447   if (changed_bbs && !bitmap_empty_p (changed_bbs))
448     {
449       EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi)
450 	{
451 	  find_uses_to_rename_bb (BASIC_BLOCK (index), use_blocks, need_phis);
452 	}
453     }
454   else
455     {
456       FOR_EACH_BB (bb)
457 	{
458 	  find_uses_to_rename_bb (bb, use_blocks, need_phis);
459 	}
460     }
461 }
462 
463 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
464    phi nodes to ensure that no variable is used outside the loop it is
465    defined in.
466 
467    This strengthening of the basic ssa form has several advantages:
468 
469    1) Updating it during unrolling/peeling/versioning is trivial, since
470       we do not need to care about the uses outside of the loop.
471       The same applies to virtual operands which are also rewritten into
472       loop closed SSA form.  Note that virtual operands are always live
473       until function exit.
474    2) The behavior of all uses of an induction variable is the same.
475       Without this, you need to distinguish the case when the variable
476       is used outside of the loop it is defined in, for example
477 
478       for (i = 0; i < 100; i++)
479 	{
480 	  for (j = 0; j < 100; j++)
481 	    {
482 	      k = i + j;
483 	      use1 (k);
484 	    }
485 	  use2 (k);
486 	}
487 
488       Looking from the outer loop with the normal SSA form, the first use of k
489       is not well-behaved, while the second one is an induction variable with
490       base 99 and step 1.
491 
492       If CHANGED_BBS is not NULL, we look for uses outside loops only in
493       the basic blocks in this set.
494 
495       UPDATE_FLAG is used in the call to update_ssa.  See
496       TODO_update_ssa* for documentation.  */
497 
498 void
rewrite_into_loop_closed_ssa(bitmap changed_bbs,unsigned update_flag)499 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag)
500 {
501   bitmap *loop_exits;
502   bitmap *use_blocks;
503   bitmap names_to_rename;
504 
505   loops_state_set (LOOP_CLOSED_SSA);
506   if (number_of_loops () <= 1)
507     return;
508 
509   /* If the pass has caused the SSA form to be out-of-date, update it
510      now.  */
511   update_ssa (update_flag);
512 
513   bitmap_obstack_initialize (&loop_renamer_obstack);
514 
515   names_to_rename = BITMAP_ALLOC (&loop_renamer_obstack);
516 
517   /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
518      that are the destination of an edge exiting loop number I.  */
519   loop_exits = XNEWVEC (bitmap, number_of_loops ());
520   get_loops_exits (loop_exits);
521 
522   /* Uses of names to rename.  We don't have to initialize this array,
523      because we know that we will only have entries for the SSA names
524      in NAMES_TO_RENAME.  */
525   use_blocks = XNEWVEC (bitmap, num_ssa_names);
526 
527   /* Find the uses outside loops.  */
528   find_uses_to_rename (changed_bbs, use_blocks, names_to_rename);
529 
530   /* Add the PHI nodes on exits of the loops for the names we need to
531      rewrite.  */
532   add_exit_phis (names_to_rename, use_blocks, loop_exits);
533 
534   bitmap_obstack_release (&loop_renamer_obstack);
535   free (use_blocks);
536   free (loop_exits);
537 
538   /* Fix up all the names found to be used outside their original
539      loops.  */
540   update_ssa (TODO_update_ssa);
541 }
542 
543 /* Check invariants of the loop closed ssa form for the USE in BB.  */
544 
545 static void
check_loop_closed_ssa_use(basic_block bb,tree use)546 check_loop_closed_ssa_use (basic_block bb, tree use)
547 {
548   gimple def;
549   basic_block def_bb;
550 
551   if (TREE_CODE (use) != SSA_NAME || virtual_operand_p (use))
552     return;
553 
554   def = SSA_NAME_DEF_STMT (use);
555   def_bb = gimple_bb (def);
556   gcc_assert (!def_bb
557 	      || flow_bb_inside_loop_p (def_bb->loop_father, bb));
558 }
559 
560 /* Checks invariants of loop closed ssa form in statement STMT in BB.  */
561 
562 static void
check_loop_closed_ssa_stmt(basic_block bb,gimple stmt)563 check_loop_closed_ssa_stmt (basic_block bb, gimple stmt)
564 {
565   ssa_op_iter iter;
566   tree var;
567 
568   if (is_gimple_debug (stmt))
569     return;
570 
571   FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
572     check_loop_closed_ssa_use (bb, var);
573 }
574 
575 /* Checks that invariants of the loop closed ssa form are preserved.
576    Call verify_ssa when VERIFY_SSA_P is true.  */
577 
578 DEBUG_FUNCTION void
verify_loop_closed_ssa(bool verify_ssa_p)579 verify_loop_closed_ssa (bool verify_ssa_p)
580 {
581   basic_block bb;
582   gimple_stmt_iterator bsi;
583   gimple phi;
584   edge e;
585   edge_iterator ei;
586 
587   if (number_of_loops () <= 1)
588     return;
589 
590   if (verify_ssa_p)
591     verify_ssa (false);
592 
593   timevar_push (TV_VERIFY_LOOP_CLOSED);
594 
595   FOR_EACH_BB (bb)
596     {
597       for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
598 	{
599 	  phi = gsi_stmt (bsi);
600 	  FOR_EACH_EDGE (e, ei, bb->preds)
601 	    check_loop_closed_ssa_use (e->src,
602 				       PHI_ARG_DEF_FROM_EDGE (phi, e));
603 	}
604 
605       for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
606 	check_loop_closed_ssa_stmt (bb, gsi_stmt (bsi));
607     }
608 
609   timevar_pop (TV_VERIFY_LOOP_CLOSED);
610 }
611 
612 /* Split loop exit edge EXIT.  The things are a bit complicated by a need to
613    preserve the loop closed ssa form.  The newly created block is returned.  */
614 
615 basic_block
split_loop_exit_edge(edge exit)616 split_loop_exit_edge (edge exit)
617 {
618   basic_block dest = exit->dest;
619   basic_block bb = split_edge (exit);
620   gimple phi, new_phi;
621   tree new_name, name;
622   use_operand_p op_p;
623   gimple_stmt_iterator psi;
624   source_location locus;
625 
626   for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi))
627     {
628       phi = gsi_stmt (psi);
629       op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb));
630       locus = gimple_phi_arg_location_from_edge (phi, single_succ_edge (bb));
631 
632       name = USE_FROM_PTR (op_p);
633 
634       /* If the argument of the PHI node is a constant, we do not need
635 	 to keep it inside loop.  */
636       if (TREE_CODE (name) != SSA_NAME)
637 	continue;
638 
639       /* Otherwise create an auxiliary phi node that will copy the value
640 	 of the SSA name out of the loop.  */
641       new_name = duplicate_ssa_name (name, NULL);
642       new_phi = create_phi_node (new_name, bb);
643       add_phi_arg (new_phi, name, exit, locus);
644       SET_USE (op_p, new_name);
645     }
646 
647   return bb;
648 }
649 
650 /* Returns the basic block in that statements should be emitted for induction
651    variables incremented at the end of the LOOP.  */
652 
653 basic_block
ip_end_pos(struct loop * loop)654 ip_end_pos (struct loop *loop)
655 {
656   return loop->latch;
657 }
658 
659 /* Returns the basic block in that statements should be emitted for induction
660    variables incremented just before exit condition of a LOOP.  */
661 
662 basic_block
ip_normal_pos(struct loop * loop)663 ip_normal_pos (struct loop *loop)
664 {
665   gimple last;
666   basic_block bb;
667   edge exit;
668 
669   if (!single_pred_p (loop->latch))
670     return NULL;
671 
672   bb = single_pred (loop->latch);
673   last = last_stmt (bb);
674   if (!last
675       || gimple_code (last) != GIMPLE_COND)
676     return NULL;
677 
678   exit = EDGE_SUCC (bb, 0);
679   if (exit->dest == loop->latch)
680     exit = EDGE_SUCC (bb, 1);
681 
682   if (flow_bb_inside_loop_p (loop, exit->dest))
683     return NULL;
684 
685   return bb;
686 }
687 
688 /* Stores the standard position for induction variable increment in LOOP
689    (just before the exit condition if it is available and latch block is empty,
690    end of the latch block otherwise) to BSI.  INSERT_AFTER is set to true if
691    the increment should be inserted after *BSI.  */
692 
693 void
standard_iv_increment_position(struct loop * loop,gimple_stmt_iterator * bsi,bool * insert_after)694 standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi,
695 				bool *insert_after)
696 {
697   basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop);
698   gimple last = last_stmt (latch);
699 
700   if (!bb
701       || (last && gimple_code (last) != GIMPLE_LABEL))
702     {
703       *bsi = gsi_last_bb (latch);
704       *insert_after = true;
705     }
706   else
707     {
708       *bsi = gsi_last_bb (bb);
709       *insert_after = false;
710     }
711 }
712 
713 /* Copies phi node arguments for duplicated blocks.  The index of the first
714    duplicated block is FIRST_NEW_BLOCK.  */
715 
716 static void
copy_phi_node_args(unsigned first_new_block)717 copy_phi_node_args (unsigned first_new_block)
718 {
719   unsigned i;
720 
721   for (i = first_new_block; i < (unsigned) last_basic_block; i++)
722     BASIC_BLOCK (i)->flags |= BB_DUPLICATED;
723 
724   for (i = first_new_block; i < (unsigned) last_basic_block; i++)
725     add_phi_args_after_copy_bb (BASIC_BLOCK (i));
726 
727   for (i = first_new_block; i < (unsigned) last_basic_block; i++)
728     BASIC_BLOCK (i)->flags &= ~BB_DUPLICATED;
729 }
730 
731 
732 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
733    updates the PHI nodes at start of the copied region.  In order to
734    achieve this, only loops whose exits all lead to the same location
735    are handled.
736 
737    Notice that we do not completely update the SSA web after
738    duplication.  The caller is responsible for calling update_ssa
739    after the loop has been duplicated.  */
740 
741 bool
gimple_duplicate_loop_to_header_edge(struct loop * loop,edge e,unsigned int ndupl,sbitmap wont_exit,edge orig,vec<edge> * to_remove,int flags)742 gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e,
743 				    unsigned int ndupl, sbitmap wont_exit,
744 				    edge orig, vec<edge> *to_remove,
745 				    int flags)
746 {
747   unsigned first_new_block;
748 
749   if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
750     return false;
751   if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
752     return false;
753 
754 #ifdef ENABLE_CHECKING
755   /* ???  This forces needless update_ssa calls after processing each
756      loop instead of just once after processing all loops.  We should
757      instead verify that loop-closed SSA form is up-to-date for LOOP
758      only (and possibly SSA form).  For now just skip verifying if
759      there are to-be renamed variables.  */
760   if (!need_ssa_update_p (cfun)
761       && loops_state_satisfies_p (LOOP_CLOSED_SSA))
762     verify_loop_closed_ssa (true);
763 #endif
764 
765   first_new_block = last_basic_block;
766   if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit,
767 				      orig, to_remove, flags))
768     return false;
769 
770   /* Readd the removed phi args for e.  */
771   flush_pending_stmts (e);
772 
773   /* Copy the phi node arguments.  */
774   copy_phi_node_args (first_new_block);
775 
776   scev_reset ();
777 
778   return true;
779 }
780 
781 /* Returns true if we can unroll LOOP FACTOR times.  Number
782    of iterations of the loop is returned in NITER.  */
783 
784 bool
can_unroll_loop_p(struct loop * loop,unsigned factor,struct tree_niter_desc * niter)785 can_unroll_loop_p (struct loop *loop, unsigned factor,
786 		   struct tree_niter_desc *niter)
787 {
788   edge exit;
789 
790   /* Check whether unrolling is possible.  We only want to unroll loops
791      for that we are able to determine number of iterations.  We also
792      want to split the extra iterations of the loop from its end,
793      therefore we require that the loop has precisely one
794      exit.  */
795 
796   exit = single_dom_exit (loop);
797   if (!exit)
798     return false;
799 
800   if (!number_of_iterations_exit (loop, exit, niter, false)
801       || niter->cmp == ERROR_MARK
802       /* Scalar evolutions analysis might have copy propagated
803 	 the abnormal ssa names into these expressions, hence
804 	 emitting the computations based on them during loop
805 	 unrolling might create overlapping life ranges for
806 	 them, and failures in out-of-ssa.  */
807       || contains_abnormal_ssa_name_p (niter->may_be_zero)
808       || contains_abnormal_ssa_name_p (niter->control.base)
809       || contains_abnormal_ssa_name_p (niter->control.step)
810       || contains_abnormal_ssa_name_p (niter->bound))
811     return false;
812 
813   /* And of course, we must be able to duplicate the loop.  */
814   if (!can_duplicate_loop_p (loop))
815     return false;
816 
817   /* The final loop should be small enough.  */
818   if (tree_num_loop_insns (loop, &eni_size_weights) * factor
819       > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS))
820     return false;
821 
822   return true;
823 }
824 
825 /* Determines the conditions that control execution of LOOP unrolled FACTOR
826    times.  DESC is number of iterations of LOOP.  ENTER_COND is set to
827    condition that must be true if the main loop can be entered.
828    EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
829    how the exit from the unrolled loop should be controlled.  */
830 
831 static void
determine_exit_conditions(struct loop * loop,struct tree_niter_desc * desc,unsigned factor,tree * enter_cond,tree * exit_base,tree * exit_step,enum tree_code * exit_cmp,tree * exit_bound)832 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc,
833 			   unsigned factor, tree *enter_cond,
834 			   tree *exit_base, tree *exit_step,
835 			   enum tree_code *exit_cmp, tree *exit_bound)
836 {
837   gimple_seq stmts;
838   tree base = desc->control.base;
839   tree step = desc->control.step;
840   tree bound = desc->bound;
841   tree type = TREE_TYPE (step);
842   tree bigstep, delta;
843   tree min = lower_bound_in_type (type, type);
844   tree max = upper_bound_in_type (type, type);
845   enum tree_code cmp = desc->cmp;
846   tree cond = boolean_true_node, assum;
847 
848   /* For pointers, do the arithmetics in the type of step.  */
849   base = fold_convert (type, base);
850   bound = fold_convert (type, bound);
851 
852   *enter_cond = boolean_false_node;
853   *exit_base = NULL_TREE;
854   *exit_step = NULL_TREE;
855   *exit_cmp = ERROR_MARK;
856   *exit_bound = NULL_TREE;
857   gcc_assert (cmp != ERROR_MARK);
858 
859   /* We only need to be correct when we answer question
860      "Do at least FACTOR more iterations remain?" in the unrolled loop.
861      Thus, transforming BASE + STEP * i <> BOUND to
862      BASE + STEP * i < BOUND is ok.  */
863   if (cmp == NE_EXPR)
864     {
865       if (tree_int_cst_sign_bit (step))
866 	cmp = GT_EXPR;
867       else
868 	cmp = LT_EXPR;
869     }
870   else if (cmp == LT_EXPR)
871     {
872       gcc_assert (!tree_int_cst_sign_bit (step));
873     }
874   else if (cmp == GT_EXPR)
875     {
876       gcc_assert (tree_int_cst_sign_bit (step));
877     }
878   else
879     gcc_unreachable ();
880 
881   /* The main body of the loop may be entered iff:
882 
883      1) desc->may_be_zero is false.
884      2) it is possible to check that there are at least FACTOR iterations
885 	of the loop, i.e., BOUND - step * FACTOR does not overflow.
886      3) # of iterations is at least FACTOR  */
887 
888   if (!integer_zerop (desc->may_be_zero))
889     cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
890 			invert_truthvalue (desc->may_be_zero),
891 			cond);
892 
893   bigstep = fold_build2 (MULT_EXPR, type, step,
894 			 build_int_cst_type (type, factor));
895   delta = fold_build2 (MINUS_EXPR, type, bigstep, step);
896   if (cmp == LT_EXPR)
897     assum = fold_build2 (GE_EXPR, boolean_type_node,
898 			 bound,
899 			 fold_build2 (PLUS_EXPR, type, min, delta));
900   else
901     assum = fold_build2 (LE_EXPR, boolean_type_node,
902 			 bound,
903 			 fold_build2 (PLUS_EXPR, type, max, delta));
904   cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
905 
906   bound = fold_build2 (MINUS_EXPR, type, bound, delta);
907   assum = fold_build2 (cmp, boolean_type_node, base, bound);
908   cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
909 
910   cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE);
911   if (stmts)
912     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
913   /* cond now may be a gimple comparison, which would be OK, but also any
914      other gimple rhs (say a && b).  In this case we need to force it to
915      operand.  */
916   if (!is_gimple_condexpr (cond))
917     {
918       cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
919       if (stmts)
920 	gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
921     }
922   *enter_cond = cond;
923 
924   base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE);
925   if (stmts)
926     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
927   bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE);
928   if (stmts)
929     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
930 
931   *exit_base = base;
932   *exit_step = bigstep;
933   *exit_cmp = cmp;
934   *exit_bound = bound;
935 }
936 
937 /* Scales the frequencies of all basic blocks in LOOP that are strictly
938    dominated by BB by NUM/DEN.  */
939 
940 static void
scale_dominated_blocks_in_loop(struct loop * loop,basic_block bb,int num,int den)941 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb,
942 				int num, int den)
943 {
944   basic_block son;
945 
946   if (den == 0)
947     return;
948 
949   for (son = first_dom_son (CDI_DOMINATORS, bb);
950        son;
951        son = next_dom_son (CDI_DOMINATORS, son))
952     {
953       if (!flow_bb_inside_loop_p (loop, son))
954 	continue;
955       scale_bbs_frequencies_int (&son, 1, num, den);
956       scale_dominated_blocks_in_loop (loop, son, num, den);
957     }
958 }
959 
960 /* Unroll LOOP FACTOR times.  DESC describes number of iterations of LOOP.
961    EXIT is the exit of the loop to that DESC corresponds.
962 
963    If N is number of iterations of the loop and MAY_BE_ZERO is the condition
964    under that loop exits in the first iteration even if N != 0,
965 
966    while (1)
967      {
968        x = phi (init, next);
969 
970        pre;
971        if (st)
972          break;
973        post;
974      }
975 
976    becomes (with possibly the exit conditions formulated a bit differently,
977    avoiding the need to create a new iv):
978 
979    if (MAY_BE_ZERO || N < FACTOR)
980      goto rest;
981 
982    do
983      {
984        x = phi (init, next);
985 
986        pre;
987        post;
988        pre;
989        post;
990        ...
991        pre;
992        post;
993        N -= FACTOR;
994 
995      } while (N >= FACTOR);
996 
997    rest:
998      init' = phi (init, x);
999 
1000    while (1)
1001      {
1002        x = phi (init', next);
1003 
1004        pre;
1005        if (st)
1006          break;
1007        post;
1008      }
1009 
1010    Before the loop is unrolled, TRANSFORM is called for it (only for the
1011    unrolled loop, but not for its versioned copy).  DATA is passed to
1012    TRANSFORM.  */
1013 
1014 /* Probability in % that the unrolled loop is entered.  Just a guess.  */
1015 #define PROB_UNROLLED_LOOP_ENTERED 90
1016 
1017 void
tree_transform_and_unroll_loop(struct loop * loop,unsigned factor,edge exit,struct tree_niter_desc * desc,transform_callback transform,void * data)1018 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor,
1019 				edge exit, struct tree_niter_desc *desc,
1020 				transform_callback transform,
1021 				void *data)
1022 {
1023   gimple exit_if;
1024   tree ctr_before, ctr_after;
1025   tree enter_main_cond, exit_base, exit_step, exit_bound;
1026   enum tree_code exit_cmp;
1027   gimple phi_old_loop, phi_new_loop, phi_rest;
1028   gimple_stmt_iterator psi_old_loop, psi_new_loop;
1029   tree init, next, new_init;
1030   struct loop *new_loop;
1031   basic_block rest, exit_bb;
1032   edge old_entry, new_entry, old_latch, precond_edge, new_exit;
1033   edge new_nonexit, e;
1034   gimple_stmt_iterator bsi;
1035   use_operand_p op;
1036   bool ok;
1037   unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h;
1038   unsigned new_est_niter, i, prob;
1039   unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
1040   sbitmap wont_exit;
1041   vec<edge> to_remove = vNULL;
1042 
1043   est_niter = expected_loop_iterations (loop);
1044   determine_exit_conditions (loop, desc, factor,
1045 			     &enter_main_cond, &exit_base, &exit_step,
1046 			     &exit_cmp, &exit_bound);
1047 
1048   /* Let us assume that the unrolled loop is quite likely to be entered.  */
1049   if (integer_nonzerop (enter_main_cond))
1050     prob_entry = REG_BR_PROB_BASE;
1051   else
1052     prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100;
1053 
1054   /* The values for scales should keep profile consistent, and somewhat close
1055      to correct.
1056 
1057      TODO: The current value of SCALE_REST makes it appear that the loop that
1058      is created by splitting the remaining iterations of the unrolled loop is
1059      executed the same number of times as the original loop, and with the same
1060      frequencies, which is obviously wrong.  This does not appear to cause
1061      problems, so we do not bother with fixing it for now.  To make the profile
1062      correct, we would need to change the probability of the exit edge of the
1063      loop, and recompute the distribution of frequencies in its body because
1064      of this change (scale the frequencies of blocks before and after the exit
1065      by appropriate factors).  */
1066   scale_unrolled = prob_entry;
1067   scale_rest = REG_BR_PROB_BASE;
1068 
1069   new_loop = loop_version (loop, enter_main_cond, NULL,
1070 			   prob_entry, scale_unrolled, scale_rest, true);
1071   gcc_assert (new_loop != NULL);
1072   update_ssa (TODO_update_ssa);
1073 
1074   /* Determine the probability of the exit edge of the unrolled loop.  */
1075   new_est_niter = est_niter / factor;
1076 
1077   /* Without profile feedback, loops for that we do not know a better estimate
1078      are assumed to roll 10 times.  When we unroll such loop, it appears to
1079      roll too little, and it may even seem to be cold.  To avoid this, we
1080      ensure that the created loop appears to roll at least 5 times (but at
1081      most as many times as before unrolling).  */
1082   if (new_est_niter < 5)
1083     {
1084       if (est_niter < 5)
1085 	new_est_niter = est_niter;
1086       else
1087 	new_est_niter = 5;
1088     }
1089 
1090   /* Prepare the cfg and update the phi nodes.  Move the loop exit to the
1091      loop latch (and make its condition dummy, for the moment).  */
1092   rest = loop_preheader_edge (new_loop)->src;
1093   precond_edge = single_pred_edge (rest);
1094   split_edge (loop_latch_edge (loop));
1095   exit_bb = single_pred (loop->latch);
1096 
1097   /* Since the exit edge will be removed, the frequency of all the blocks
1098      in the loop that are dominated by it must be scaled by
1099      1 / (1 - exit->probability).  */
1100   scale_dominated_blocks_in_loop (loop, exit->src,
1101 				  REG_BR_PROB_BASE,
1102 				  REG_BR_PROB_BASE - exit->probability);
1103 
1104   bsi = gsi_last_bb (exit_bb);
1105   exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node,
1106 			       integer_zero_node,
1107 			       NULL_TREE, NULL_TREE);
1108 
1109   gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT);
1110   new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr);
1111   rescan_loop_exit (new_exit, true, false);
1112 
1113   /* Set the probability of new exit to the same of the old one.  Fix
1114      the frequency of the latch block, by scaling it back by
1115      1 - exit->probability.  */
1116   new_exit->count = exit->count;
1117   new_exit->probability = exit->probability;
1118   new_nonexit = single_pred_edge (loop->latch);
1119   new_nonexit->probability = REG_BR_PROB_BASE - exit->probability;
1120   new_nonexit->flags = EDGE_TRUE_VALUE;
1121   new_nonexit->count -= exit->count;
1122   if (new_nonexit->count < 0)
1123     new_nonexit->count = 0;
1124   scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1125 			     REG_BR_PROB_BASE);
1126 
1127   old_entry = loop_preheader_edge (loop);
1128   new_entry = loop_preheader_edge (new_loop);
1129   old_latch = loop_latch_edge (loop);
1130   for (psi_old_loop = gsi_start_phis (loop->header),
1131        psi_new_loop = gsi_start_phis (new_loop->header);
1132        !gsi_end_p (psi_old_loop);
1133        gsi_next (&psi_old_loop), gsi_next (&psi_new_loop))
1134     {
1135       phi_old_loop = gsi_stmt (psi_old_loop);
1136       phi_new_loop = gsi_stmt (psi_new_loop);
1137 
1138       init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry);
1139       op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry);
1140       gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op)));
1141       next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch);
1142 
1143       /* Prefer using original variable as a base for the new ssa name.
1144 	 This is necessary for virtual ops, and useful in order to avoid
1145 	 losing debug info for real ops.  */
1146       if (TREE_CODE (next) == SSA_NAME
1147 	  && useless_type_conversion_p (TREE_TYPE (next),
1148 					TREE_TYPE (init)))
1149 	new_init = copy_ssa_name (next, NULL);
1150       else if (TREE_CODE (init) == SSA_NAME
1151 	       && useless_type_conversion_p (TREE_TYPE (init),
1152 					     TREE_TYPE (next)))
1153 	new_init = copy_ssa_name (init, NULL);
1154       else if (useless_type_conversion_p (TREE_TYPE (next), TREE_TYPE (init)))
1155 	new_init = make_temp_ssa_name (TREE_TYPE (next), NULL, "unrinittmp");
1156       else
1157 	new_init = make_temp_ssa_name (TREE_TYPE (init), NULL, "unrinittmp");
1158 
1159       phi_rest = create_phi_node (new_init, rest);
1160 
1161       add_phi_arg (phi_rest, init, precond_edge, UNKNOWN_LOCATION);
1162       add_phi_arg (phi_rest, next, new_exit, UNKNOWN_LOCATION);
1163       SET_USE (op, new_init);
1164     }
1165 
1166   remove_path (exit);
1167 
1168   /* Transform the loop.  */
1169   if (transform)
1170     (*transform) (loop, data);
1171 
1172   /* Unroll the loop and remove the exits in all iterations except for the
1173      last one.  */
1174   wont_exit = sbitmap_alloc (factor);
1175   bitmap_ones (wont_exit);
1176   bitmap_clear_bit (wont_exit, factor - 1);
1177 
1178   ok = gimple_duplicate_loop_to_header_edge
1179 	  (loop, loop_latch_edge (loop), factor - 1,
1180 	   wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ);
1181   free (wont_exit);
1182   gcc_assert (ok);
1183 
1184   FOR_EACH_VEC_ELT (to_remove, i, e)
1185     {
1186       ok = remove_path (e);
1187       gcc_assert (ok);
1188     }
1189   to_remove.release ();
1190   update_ssa (TODO_update_ssa);
1191 
1192   /* Ensure that the frequencies in the loop match the new estimated
1193      number of iterations, and change the probability of the new
1194      exit edge.  */
1195   freq_h = loop->header->frequency;
1196   freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop));
1197   if (freq_h != 0)
1198     scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h);
1199 
1200   exit_bb = single_pred (loop->latch);
1201   new_exit = find_edge (exit_bb, rest);
1202   new_exit->count = loop_preheader_edge (loop)->count;
1203   new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);
1204 
1205   rest->count += new_exit->count;
1206   rest->frequency += EDGE_FREQUENCY (new_exit);
1207 
1208   new_nonexit = single_pred_edge (loop->latch);
1209   prob = new_nonexit->probability;
1210   new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
1211   new_nonexit->count = exit_bb->count - new_exit->count;
1212   if (new_nonexit->count < 0)
1213     new_nonexit->count = 0;
1214   if (prob > 0)
1215     scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1216 			       prob);
1217 
1218   /* Finally create the new counter for number of iterations and add the new
1219      exit instruction.  */
1220   bsi = gsi_last_nondebug_bb (exit_bb);
1221   exit_if = gsi_stmt (bsi);
1222   create_iv (exit_base, exit_step, NULL_TREE, loop,
1223 	     &bsi, false, &ctr_before, &ctr_after);
1224   gimple_cond_set_code (exit_if, exit_cmp);
1225   gimple_cond_set_lhs (exit_if, ctr_after);
1226   gimple_cond_set_rhs (exit_if, exit_bound);
1227   update_stmt (exit_if);
1228 
1229 #ifdef ENABLE_CHECKING
1230   verify_flow_info ();
1231   verify_loop_structure ();
1232   verify_loop_closed_ssa (true);
1233 #endif
1234 }
1235 
1236 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1237    want to transform the loop before unrolling.  The meaning
1238    of the arguments is the same as for tree_transform_and_unroll_loop.  */
1239 
1240 void
tree_unroll_loop(struct loop * loop,unsigned factor,edge exit,struct tree_niter_desc * desc)1241 tree_unroll_loop (struct loop *loop, unsigned factor,
1242 		  edge exit, struct tree_niter_desc *desc)
1243 {
1244   tree_transform_and_unroll_loop (loop, factor, exit, desc,
1245 				  NULL, NULL);
1246 }
1247 
1248 /* Rewrite the phi node at position PSI in function of the main
1249    induction variable MAIN_IV and insert the generated code at GSI.  */
1250 
1251 static void
rewrite_phi_with_iv(loop_p loop,gimple_stmt_iterator * psi,gimple_stmt_iterator * gsi,tree main_iv)1252 rewrite_phi_with_iv (loop_p loop,
1253 		     gimple_stmt_iterator *psi,
1254 		     gimple_stmt_iterator *gsi,
1255 		     tree main_iv)
1256 {
1257   affine_iv iv;
1258   gimple stmt, phi = gsi_stmt (*psi);
1259   tree atype, mtype, val, res = PHI_RESULT (phi);
1260 
1261   if (virtual_operand_p (res) || res == main_iv)
1262     {
1263       gsi_next (psi);
1264       return;
1265     }
1266 
1267   if (!simple_iv (loop, loop, res, &iv, true))
1268     {
1269       gsi_next (psi);
1270       return;
1271     }
1272 
1273   remove_phi_node (psi, false);
1274 
1275   atype = TREE_TYPE (res);
1276   mtype = POINTER_TYPE_P (atype) ? sizetype : atype;
1277   val = fold_build2 (MULT_EXPR, mtype, unshare_expr (iv.step),
1278 		     fold_convert (mtype, main_iv));
1279   val = fold_build2 (POINTER_TYPE_P (atype)
1280 		     ? POINTER_PLUS_EXPR : PLUS_EXPR,
1281 		     atype, unshare_expr (iv.base), val);
1282   val = force_gimple_operand_gsi (gsi, val, false, NULL_TREE, true,
1283 				  GSI_SAME_STMT);
1284   stmt = gimple_build_assign (res, val);
1285   gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1286   SSA_NAME_DEF_STMT (res) = stmt;
1287 }
1288 
1289 /* Rewrite all the phi nodes of LOOP in function of the main induction
1290    variable MAIN_IV.  */
1291 
1292 static void
rewrite_all_phi_nodes_with_iv(loop_p loop,tree main_iv)1293 rewrite_all_phi_nodes_with_iv (loop_p loop, tree main_iv)
1294 {
1295   unsigned i;
1296   basic_block *bbs = get_loop_body_in_dom_order (loop);
1297   gimple_stmt_iterator psi;
1298 
1299   for (i = 0; i < loop->num_nodes; i++)
1300     {
1301       basic_block bb = bbs[i];
1302       gimple_stmt_iterator gsi = gsi_after_labels (bb);
1303 
1304       if (bb->loop_father != loop)
1305 	continue;
1306 
1307       for (psi = gsi_start_phis (bb); !gsi_end_p (psi); )
1308 	rewrite_phi_with_iv (loop, &psi, &gsi, main_iv);
1309     }
1310 
1311   free (bbs);
1312 }
1313 
1314 /* Bases all the induction variables in LOOP on a single induction
1315    variable (unsigned with base 0 and step 1), whose final value is
1316    compared with *NIT.  When the IV type precision has to be larger
1317    than *NIT type precision, *NIT is converted to the larger type, the
1318    conversion code is inserted before the loop, and *NIT is updated to
1319    the new definition.  When BUMP_IN_LATCH is true, the induction
1320    variable is incremented in the loop latch, otherwise it is
1321    incremented in the loop header.  Return the induction variable that
1322    was created.  */
1323 
1324 tree
canonicalize_loop_ivs(struct loop * loop,tree * nit,bool bump_in_latch)1325 canonicalize_loop_ivs (struct loop *loop, tree *nit, bool bump_in_latch)
1326 {
1327   unsigned precision = TYPE_PRECISION (TREE_TYPE (*nit));
1328   unsigned original_precision = precision;
1329   tree type, var_before;
1330   gimple_stmt_iterator gsi, psi;
1331   gimple stmt;
1332   edge exit = single_dom_exit (loop);
1333   gimple_seq stmts;
1334   enum machine_mode mode;
1335   bool unsigned_p = false;
1336 
1337   for (psi = gsi_start_phis (loop->header);
1338        !gsi_end_p (psi); gsi_next (&psi))
1339     {
1340       gimple phi = gsi_stmt (psi);
1341       tree res = PHI_RESULT (phi);
1342       bool uns;
1343 
1344       type = TREE_TYPE (res);
1345       if (virtual_operand_p (res)
1346 	  || (!INTEGRAL_TYPE_P (type)
1347 	      && !POINTER_TYPE_P (type))
1348 	  || TYPE_PRECISION (type) < precision)
1349 	continue;
1350 
1351       uns = POINTER_TYPE_P (type) | TYPE_UNSIGNED (type);
1352 
1353       if (TYPE_PRECISION (type) > precision)
1354 	unsigned_p = uns;
1355       else
1356 	unsigned_p |= uns;
1357 
1358       precision = TYPE_PRECISION (type);
1359     }
1360 
1361   mode = smallest_mode_for_size (precision, MODE_INT);
1362   precision = GET_MODE_PRECISION (mode);
1363   type = build_nonstandard_integer_type (precision, unsigned_p);
1364 
1365   if (original_precision != precision)
1366     {
1367       *nit = fold_convert (type, *nit);
1368       *nit = force_gimple_operand (*nit, &stmts, true, NULL_TREE);
1369       if (stmts)
1370 	gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1371     }
1372 
1373   if (bump_in_latch)
1374     gsi = gsi_last_bb (loop->latch);
1375   else
1376     gsi = gsi_last_nondebug_bb (loop->header);
1377   create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
1378 	     loop, &gsi, bump_in_latch, &var_before, NULL);
1379 
1380   rewrite_all_phi_nodes_with_iv (loop, var_before);
1381 
1382   stmt = last_stmt (exit->src);
1383   /* Make the loop exit if the control condition is not satisfied.  */
1384   if (exit->flags & EDGE_TRUE_VALUE)
1385     {
1386       edge te, fe;
1387 
1388       extract_true_false_edges_from_block (exit->src, &te, &fe);
1389       te->flags = EDGE_FALSE_VALUE;
1390       fe->flags = EDGE_TRUE_VALUE;
1391     }
1392   gimple_cond_set_code (stmt, LT_EXPR);
1393   gimple_cond_set_lhs (stmt, var_before);
1394   gimple_cond_set_rhs (stmt, *nit);
1395   update_stmt (stmt);
1396 
1397   return var_before;
1398 }
1399