1 /* High-level loop manipulation functions.
2    Copyright (C) 2004-2016 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 "backend.h"
24 #include "tree.h"
25 #include "gimple.h"
26 #include "cfghooks.h"
27 #include "tree-pass.h"	/* ??? for TODO_update_ssa but this isn't a pass.  */
28 #include "ssa.h"
29 #include "gimple-pretty-print.h"
30 #include "fold-const.h"
31 #include "cfganal.h"
32 #include "gimplify.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
35 #include "tree-cfg.h"
36 #include "tree-ssa-loop-ivopts.h"
37 #include "tree-ssa-loop-manip.h"
38 #include "tree-ssa-loop-niter.h"
39 #include "tree-ssa-loop.h"
40 #include "tree-into-ssa.h"
41 #include "tree-ssa.h"
42 #include "cfgloop.h"
43 #include "tree-scalar-evolution.h"
44 #include "params.h"
45 #include "tree-inline.h"
46 
47 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
48    so that we can free them all at once.  */
49 static bitmap_obstack loop_renamer_obstack;
50 
51 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
52    It is expected that neither BASE nor STEP are shared with other expressions
53    (unless the sharing rules allow this).  Use VAR as a base var_decl for it
54    (if NULL, a new temporary will be created).  The increment will occur at
55    INCR_POS (after it if AFTER is true, before it otherwise).  INCR_POS and
56    AFTER can be computed using standard_iv_increment_position.  The ssa versions
57    of the variable before and after increment will be stored in VAR_BEFORE and
58    VAR_AFTER (unless they are NULL).  */
59 
60 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)61 create_iv (tree base, tree step, tree var, struct loop *loop,
62 	   gimple_stmt_iterator *incr_pos, bool after,
63 	   tree *var_before, tree *var_after)
64 {
65   gassign *stmt;
66   gphi *phi;
67   tree initial, step1;
68   gimple_seq stmts;
69   tree vb, va;
70   enum tree_code incr_op = PLUS_EXPR;
71   edge pe = loop_preheader_edge (loop);
72 
73   if (var != NULL_TREE)
74     {
75       vb = make_ssa_name (var);
76       va = make_ssa_name (var);
77     }
78   else
79     {
80       vb = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
81       va = make_temp_ssa_name (TREE_TYPE (base), NULL, "ivtmp");
82     }
83   if (var_before)
84     *var_before = vb;
85   if (var_after)
86     *var_after = va;
87 
88   /* For easier readability of the created code, produce MINUS_EXPRs
89      when suitable.  */
90   if (TREE_CODE (step) == INTEGER_CST)
91     {
92       if (TYPE_UNSIGNED (TREE_TYPE (step)))
93 	{
94 	  step1 = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
95 	  if (tree_int_cst_lt (step1, step))
96 	    {
97 	      incr_op = MINUS_EXPR;
98 	      step = step1;
99 	    }
100 	}
101       else
102 	{
103 	  bool ovf;
104 
105 	  if (!tree_expr_nonnegative_warnv_p (step, &ovf)
106 	      && may_negate_without_overflow_p (step))
107 	    {
108 	      incr_op = MINUS_EXPR;
109 	      step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
110 	    }
111 	}
112     }
113   if (POINTER_TYPE_P (TREE_TYPE (base)))
114     {
115       if (TREE_CODE (base) == ADDR_EXPR)
116 	mark_addressable (TREE_OPERAND (base, 0));
117       step = convert_to_ptrofftype (step);
118       if (incr_op == MINUS_EXPR)
119 	step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
120       incr_op = POINTER_PLUS_EXPR;
121     }
122   /* Gimplify the step if necessary.  We put the computations in front of the
123      loop (i.e. the step should be loop invariant).  */
124   step = force_gimple_operand (step, &stmts, true, NULL_TREE);
125   if (stmts)
126     gsi_insert_seq_on_edge_immediate (pe, stmts);
127 
128   stmt = gimple_build_assign (va, incr_op, vb, step);
129   if (after)
130     gsi_insert_after (incr_pos, stmt, GSI_NEW_STMT);
131   else
132     gsi_insert_before (incr_pos, stmt, GSI_NEW_STMT);
133 
134   initial = force_gimple_operand (base, &stmts, true, var);
135   if (stmts)
136     gsi_insert_seq_on_edge_immediate (pe, stmts);
137 
138   phi = create_phi_node (vb, loop->header);
139   add_phi_arg (phi, initial, loop_preheader_edge (loop), UNKNOWN_LOCATION);
140   add_phi_arg (phi, va, loop_latch_edge (loop), UNKNOWN_LOCATION);
141 }
142 
143 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
144    both DEF_LOOP and USE_LOOP.  */
145 
146 static inline struct loop *
find_sibling_superloop(struct loop * use_loop,struct loop * def_loop)147 find_sibling_superloop (struct loop *use_loop, struct loop *def_loop)
148 {
149   unsigned ud = loop_depth (use_loop);
150   unsigned dd = loop_depth (def_loop);
151   gcc_assert (ud > 0 && dd > 0);
152   if (ud > dd)
153     use_loop = superloop_at_depth (use_loop, dd);
154   if (ud < dd)
155     def_loop = superloop_at_depth (def_loop, ud);
156   while (loop_outer (use_loop) != loop_outer (def_loop))
157     {
158       use_loop = loop_outer (use_loop);
159       def_loop = loop_outer (def_loop);
160       gcc_assert (use_loop && def_loop);
161     }
162   return use_loop;
163 }
164 
165 /* DEF_BB is a basic block containing a DEF that needs rewriting into
166    loop-closed SSA form.  USE_BLOCKS is the set of basic blocks containing
167    uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
168    USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
169    ALL_EXITS[I] is the set of all basic blocks that exit loop I.
170 
171    Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
172    or one of its loop fathers, in which DEF is live.  This set is returned
173    in the bitmap LIVE_EXITS.
174 
175    Instead of computing the complete livein set of the def, we use the loop
176    nesting tree as a form of poor man's structure analysis.  This greatly
177    speeds up the analysis, which is important because this function may be
178    called on all SSA names that need rewriting, one at a time.  */
179 
180 static void
compute_live_loop_exits(bitmap live_exits,bitmap use_blocks,bitmap * loop_exits,basic_block def_bb)181 compute_live_loop_exits (bitmap live_exits, bitmap use_blocks,
182 			 bitmap *loop_exits, basic_block def_bb)
183 {
184   unsigned i;
185   bitmap_iterator bi;
186   struct loop *def_loop = def_bb->loop_father;
187   unsigned def_loop_depth = loop_depth (def_loop);
188   bitmap def_loop_exits;
189 
190   /* Normally the work list size is bounded by the number of basic
191      blocks in the largest loop.  We don't know this number, but we
192      can be fairly sure that it will be relatively small.  */
193   auto_vec<basic_block> worklist (MAX (8, n_basic_blocks_for_fn (cfun) / 128));
194 
195   EXECUTE_IF_SET_IN_BITMAP (use_blocks, 0, i, bi)
196     {
197       basic_block use_bb = BASIC_BLOCK_FOR_FN (cfun, i);
198       struct loop *use_loop = use_bb->loop_father;
199       gcc_checking_assert (def_loop != use_loop
200 			   && ! flow_loop_nested_p (def_loop, use_loop));
201       if (! flow_loop_nested_p (use_loop, def_loop))
202 	use_bb = find_sibling_superloop (use_loop, def_loop)->header;
203       if (bitmap_set_bit (live_exits, use_bb->index))
204 	worklist.safe_push (use_bb);
205     }
206 
207   /* Iterate until the worklist is empty.  */
208   while (! worklist.is_empty ())
209     {
210       edge e;
211       edge_iterator ei;
212 
213       /* Pull a block off the worklist.  */
214       basic_block bb = worklist.pop ();
215 
216       /* Make sure we have at least enough room in the work list
217 	 for all predecessors of this block.  */
218       worklist.reserve (EDGE_COUNT (bb->preds));
219 
220       /* For each predecessor block.  */
221       FOR_EACH_EDGE (e, ei, bb->preds)
222 	{
223 	  basic_block pred = e->src;
224 	  struct loop *pred_loop = pred->loop_father;
225 	  unsigned pred_loop_depth = loop_depth (pred_loop);
226 	  bool pred_visited;
227 
228 	  /* We should have met DEF_BB along the way.  */
229 	  gcc_assert (pred != ENTRY_BLOCK_PTR_FOR_FN (cfun));
230 
231 	  if (pred_loop_depth >= def_loop_depth)
232 	    {
233 	      if (pred_loop_depth > def_loop_depth)
234 		pred_loop = superloop_at_depth (pred_loop, def_loop_depth);
235 	      /* If we've reached DEF_LOOP, our train ends here.  */
236 	      if (pred_loop == def_loop)
237 		continue;
238 	    }
239 	  else if (! flow_loop_nested_p (pred_loop, def_loop))
240 	    pred = find_sibling_superloop (pred_loop, def_loop)->header;
241 
242 	  /* Add PRED to the LIVEIN set.  PRED_VISITED is true if
243 	     we had already added PRED to LIVEIN before.  */
244 	  pred_visited = !bitmap_set_bit (live_exits, pred->index);
245 
246 	  /* If we have visited PRED before, don't add it to the worklist.
247 	     If BB dominates PRED, then we're probably looking at a loop.
248 	     We're only interested in looking up in the dominance tree
249 	     because DEF_BB dominates all the uses.  */
250 	  if (pred_visited || dominated_by_p (CDI_DOMINATORS, pred, bb))
251 	    continue;
252 
253 	  worklist.quick_push (pred);
254 	}
255     }
256 
257   def_loop_exits = BITMAP_ALLOC (&loop_renamer_obstack);
258   for (struct loop *loop = def_loop;
259        loop != current_loops->tree_root;
260        loop = loop_outer (loop))
261     bitmap_ior_into (def_loop_exits, loop_exits[loop->num]);
262   bitmap_and_into (live_exits, def_loop_exits);
263   BITMAP_FREE (def_loop_exits);
264 }
265 
266 /* Add a loop-closing PHI for VAR in basic block EXIT.  */
267 
268 static void
add_exit_phi(basic_block exit,tree var)269 add_exit_phi (basic_block exit, tree var)
270 {
271   gphi *phi;
272   edge e;
273   edge_iterator ei;
274 
275   /* Check that at least one of the edges entering the EXIT block exits
276      the loop, or a superloop of that loop, that VAR is defined in.  */
277   if (flag_checking)
278     {
279       gimple *def_stmt = SSA_NAME_DEF_STMT (var);
280       basic_block def_bb = gimple_bb (def_stmt);
281       FOR_EACH_EDGE (e, ei, exit->preds)
282 	{
283 	  struct loop *aloop = find_common_loop (def_bb->loop_father,
284 						 e->src->loop_father);
285 	  if (!flow_bb_inside_loop_p (aloop, e->dest))
286 	    break;
287 	}
288       gcc_assert (e);
289     }
290 
291   phi = create_phi_node (NULL_TREE, exit);
292   create_new_def_for (var, phi, gimple_phi_result_ptr (phi));
293   FOR_EACH_EDGE (e, ei, exit->preds)
294     add_phi_arg (phi, var, e, UNKNOWN_LOCATION);
295 
296   if (dump_file && (dump_flags & TDF_DETAILS))
297     {
298       fprintf (dump_file, ";; Created LCSSA PHI: ");
299       print_gimple_stmt (dump_file, phi, 0, dump_flags);
300     }
301 }
302 
303 /* Add exit phis for VAR that is used in LIVEIN.
304    Exits of the loops are stored in LOOP_EXITS.  */
305 
306 static void
add_exit_phis_var(tree var,bitmap use_blocks,bitmap * loop_exits)307 add_exit_phis_var (tree var, bitmap use_blocks, bitmap *loop_exits)
308 {
309   unsigned index;
310   bitmap_iterator bi;
311   basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var));
312   bitmap live_exits = BITMAP_ALLOC (&loop_renamer_obstack);
313 
314   gcc_checking_assert (! bitmap_bit_p (use_blocks, def_bb->index));
315 
316   compute_live_loop_exits (live_exits, use_blocks, loop_exits, def_bb);
317 
318   EXECUTE_IF_SET_IN_BITMAP (live_exits, 0, index, bi)
319     {
320       add_exit_phi (BASIC_BLOCK_FOR_FN (cfun, index), var);
321     }
322 
323   BITMAP_FREE (live_exits);
324 }
325 
326 /* Add exit phis for the names marked in NAMES_TO_RENAME.
327    Exits of the loops are stored in EXITS.  Sets of blocks where the ssa
328    names are used are stored in USE_BLOCKS.  */
329 
330 static void
add_exit_phis(bitmap names_to_rename,bitmap * use_blocks,bitmap * loop_exits)331 add_exit_phis (bitmap names_to_rename, bitmap *use_blocks, bitmap *loop_exits)
332 {
333   unsigned i;
334   bitmap_iterator bi;
335 
336   EXECUTE_IF_SET_IN_BITMAP (names_to_rename, 0, i, bi)
337     {
338       add_exit_phis_var (ssa_name (i), use_blocks[i], loop_exits);
339     }
340 }
341 
342 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets.  */
343 
344 static void
get_loops_exits(bitmap * loop_exits)345 get_loops_exits (bitmap *loop_exits)
346 {
347   struct loop *loop;
348   unsigned j;
349   edge e;
350 
351   FOR_EACH_LOOP (loop, 0)
352     {
353       vec<edge> exit_edges = get_loop_exit_edges (loop);
354       loop_exits[loop->num] = BITMAP_ALLOC (&loop_renamer_obstack);
355       FOR_EACH_VEC_ELT (exit_edges, j, e)
356         bitmap_set_bit (loop_exits[loop->num], e->dest->index);
357       exit_edges.release ();
358     }
359 }
360 
361 /* For USE in BB, if it is used outside of the loop it is defined in,
362    mark it for rewrite.  Record basic block BB where it is used
363    to USE_BLOCKS.  Record the ssa name index to NEED_PHIS bitmap.
364    Note that for USEs in phis, BB should be the src of the edge corresponding to
365    the use, rather than the bb containing the phi.  */
366 
367 static void
find_uses_to_rename_use(basic_block bb,tree use,bitmap * use_blocks,bitmap need_phis)368 find_uses_to_rename_use (basic_block bb, tree use, bitmap *use_blocks,
369 			 bitmap need_phis)
370 {
371   unsigned ver;
372   basic_block def_bb;
373   struct loop *def_loop;
374 
375   if (TREE_CODE (use) != SSA_NAME)
376     return;
377 
378   ver = SSA_NAME_VERSION (use);
379   def_bb = gimple_bb (SSA_NAME_DEF_STMT (use));
380   if (!def_bb)
381     return;
382   def_loop = def_bb->loop_father;
383 
384   /* If the definition is not inside a loop, it is not interesting.  */
385   if (!loop_outer (def_loop))
386     return;
387 
388   /* If the use is not outside of the loop it is defined in, it is not
389      interesting.  */
390   if (flow_bb_inside_loop_p (def_loop, bb))
391     return;
392 
393   /* If we're seeing VER for the first time, we still have to allocate
394      a bitmap for its uses.  */
395   if (bitmap_set_bit (need_phis, ver))
396     use_blocks[ver] = BITMAP_ALLOC (&loop_renamer_obstack);
397   bitmap_set_bit (use_blocks[ver], bb->index);
398 }
399 
400 /* For uses matching USE_FLAGS in STMT, mark names that are used outside of the
401    loop they are defined to rewrite.  Record the set of blocks in which the ssa
402    names are used to USE_BLOCKS, and the ssa names themselves to NEED_PHIS.  */
403 
404 static void
find_uses_to_rename_stmt(gimple * stmt,bitmap * use_blocks,bitmap need_phis,int use_flags)405 find_uses_to_rename_stmt (gimple *stmt, bitmap *use_blocks, bitmap need_phis,
406 			  int use_flags)
407 {
408   ssa_op_iter iter;
409   tree var;
410   basic_block bb = gimple_bb (stmt);
411 
412   if (is_gimple_debug (stmt))
413     return;
414 
415   /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows SSA_OP_VIRTUAL_USES
416      only.  */
417   if (use_flags == SSA_OP_VIRTUAL_USES)
418     {
419       tree vuse = gimple_vuse (stmt);
420       if (vuse != NULL_TREE)
421 	find_uses_to_rename_use (bb, gimple_vuse (stmt), use_blocks, need_phis);
422     }
423   else
424     FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, use_flags)
425       find_uses_to_rename_use (bb, var, use_blocks, need_phis);
426 }
427 
428 /* Marks names matching USE_FLAGS that are used in BB and outside of the loop
429    they are defined in for rewrite.  Records the set of blocks in which the ssa
430    names are used to USE_BLOCKS.  Record the SSA names that will
431    need exit PHIs in NEED_PHIS.  */
432 
433 static void
find_uses_to_rename_bb(basic_block bb,bitmap * use_blocks,bitmap need_phis,int use_flags)434 find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis,
435 			int use_flags)
436 {
437   edge e;
438   edge_iterator ei;
439   bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0;
440   bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0;
441 
442   FOR_EACH_EDGE (e, ei, bb->succs)
443     for (gphi_iterator bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi);
444 	 gsi_next (&bsi))
445       {
446         gphi *phi = bsi.phi ();
447 	bool virtual_p = virtual_operand_p (gimple_phi_result (phi));
448 	if ((virtual_p && do_virtuals)
449 	    || (!virtual_p && do_nonvirtuals))
450 	  find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
451 				   use_blocks, need_phis);
452       }
453 
454   for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
455        gsi_next (&bsi))
456     find_uses_to_rename_stmt (gsi_stmt (bsi), use_blocks, need_phis,
457 			      use_flags);
458 }
459 
460 /* Marks names matching USE_FLAGS that are used outside of the loop they are
461    defined in for rewrite.  Records the set of blocks in which the ssa names are
462    used to USE_BLOCKS.  Record the SSA names that will need exit PHIs in
463    NEED_PHIS.  If CHANGED_BBS is not NULL, scan only blocks in this set.  */
464 
465 static void
find_uses_to_rename(bitmap changed_bbs,bitmap * use_blocks,bitmap need_phis,int use_flags)466 find_uses_to_rename (bitmap changed_bbs, bitmap *use_blocks, bitmap need_phis,
467 		     int use_flags)
468 {
469   basic_block bb;
470   unsigned index;
471   bitmap_iterator bi;
472 
473   if (changed_bbs)
474     EXECUTE_IF_SET_IN_BITMAP (changed_bbs, 0, index, bi)
475       find_uses_to_rename_bb (BASIC_BLOCK_FOR_FN (cfun, index), use_blocks,
476 			      need_phis, use_flags);
477   else
478     FOR_EACH_BB_FN (bb, cfun)
479       find_uses_to_rename_bb (bb, use_blocks, need_phis, use_flags);
480 }
481 
482 /* Mark uses of DEF that are used outside of the loop they are defined in for
483    rewrite.  Record the set of blocks in which the ssa names are used to
484    USE_BLOCKS.  Record the SSA names that will need exit PHIs in NEED_PHIS.  */
485 
486 static void
find_uses_to_rename_def(tree def,bitmap * use_blocks,bitmap need_phis)487 find_uses_to_rename_def (tree def, bitmap *use_blocks, bitmap need_phis)
488 {
489   gimple *use_stmt;
490   imm_use_iterator imm_iter;
491 
492   FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
493     {
494       basic_block use_bb = gimple_bb (use_stmt);
495 
496       use_operand_p use_p;
497       FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
498 	{
499 	  if (gimple_code (use_stmt) == GIMPLE_PHI)
500 	    {
501 	      edge e = gimple_phi_arg_edge (as_a <gphi *> (use_stmt),
502 					    PHI_ARG_INDEX_FROM_USE (use_p));
503 	      use_bb = e->src;
504 	    }
505 	  find_uses_to_rename_use (use_bb, USE_FROM_PTR (use_p), use_blocks,
506 				   need_phis);
507 	}
508     }
509 }
510 
511 /* Marks names matching USE_FLAGS that are defined in LOOP and used outside of
512    it for rewrite.  Records the set of blocks in which the ssa names are used to
513    USE_BLOCKS.  Record the SSA names that will need exit PHIs in NEED_PHIS.  */
514 
515 static void
find_uses_to_rename_in_loop(struct loop * loop,bitmap * use_blocks,bitmap need_phis,int use_flags)516 find_uses_to_rename_in_loop (struct loop *loop, bitmap *use_blocks,
517 			     bitmap need_phis, int use_flags)
518 {
519   bool do_virtuals = (use_flags & SSA_OP_VIRTUAL_USES) != 0;
520   bool do_nonvirtuals = (use_flags & SSA_OP_USE) != 0;
521   int def_flags = ((do_virtuals ? SSA_OP_VIRTUAL_DEFS : 0)
522 		   | (do_nonvirtuals ? SSA_OP_DEF : 0));
523 
524 
525   basic_block *bbs = get_loop_body (loop);
526 
527   for (unsigned int i = 0; i < loop->num_nodes; i++)
528     {
529       basic_block bb = bbs[i];
530 
531       for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
532 	   gsi_next (&bsi))
533 	{
534 	  gphi *phi = bsi.phi ();
535 	  tree res = gimple_phi_result (phi);
536 	  bool virtual_p = virtual_operand_p (res);
537 	  if ((virtual_p && do_virtuals)
538 	      || (!virtual_p && do_nonvirtuals))
539 	    find_uses_to_rename_def (res, use_blocks, need_phis);
540       }
541 
542       for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
543 	   gsi_next (&bsi))
544 	{
545 	  gimple *stmt = gsi_stmt (bsi);
546 	  /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows
547 	     SSA_OP_VIRTUAL_DEFS only.  */
548 	  if (def_flags == SSA_OP_VIRTUAL_DEFS)
549 	    {
550 	      tree vdef = gimple_vdef (stmt);
551 	      if (vdef != NULL)
552 		find_uses_to_rename_def (vdef, use_blocks, need_phis);
553 	    }
554 	  else
555 	    {
556 	      tree var;
557 	      ssa_op_iter iter;
558 	      FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, def_flags)
559 		find_uses_to_rename_def (var, use_blocks, need_phis);
560 	    }
561 	}
562     }
563 
564   XDELETEVEC (bbs);
565 }
566 
567 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
568    phi nodes to ensure that no variable is used outside the loop it is
569    defined in.
570 
571    This strengthening of the basic ssa form has several advantages:
572 
573    1) Updating it during unrolling/peeling/versioning is trivial, since
574       we do not need to care about the uses outside of the loop.
575       The same applies to virtual operands which are also rewritten into
576       loop closed SSA form.  Note that virtual operands are always live
577       until function exit.
578    2) The behavior of all uses of an induction variable is the same.
579       Without this, you need to distinguish the case when the variable
580       is used outside of the loop it is defined in, for example
581 
582       for (i = 0; i < 100; i++)
583 	{
584 	  for (j = 0; j < 100; j++)
585 	    {
586 	      k = i + j;
587 	      use1 (k);
588 	    }
589 	  use2 (k);
590 	}
591 
592       Looking from the outer loop with the normal SSA form, the first use of k
593       is not well-behaved, while the second one is an induction variable with
594       base 99 and step 1.
595 
596       If LOOP is non-null, only rewrite uses that have defs in LOOP.  Otherwise,
597       if CHANGED_BBS is not NULL, we look for uses outside loops only in the
598       basic blocks in this set.
599 
600       USE_FLAGS allows us to specify whether we want virtual, non-virtual or
601       both variables rewritten.
602 
603       UPDATE_FLAG is used in the call to update_ssa.  See
604       TODO_update_ssa* for documentation.  */
605 
606 void
rewrite_into_loop_closed_ssa_1(bitmap changed_bbs,unsigned update_flag,int use_flags,struct loop * loop)607 rewrite_into_loop_closed_ssa_1 (bitmap changed_bbs, unsigned update_flag,
608 				int use_flags, struct loop *loop)
609 {
610   bitmap *use_blocks;
611   bitmap names_to_rename;
612 
613   loops_state_set (LOOP_CLOSED_SSA);
614   if (number_of_loops (cfun) <= 1)
615     return;
616 
617   /* If the pass has caused the SSA form to be out-of-date, update it
618      now.  */
619   if (update_flag != 0)
620     update_ssa (update_flag);
621   else if (flag_checking)
622     verify_ssa (true, true);
623 
624   bitmap_obstack_initialize (&loop_renamer_obstack);
625 
626   names_to_rename = BITMAP_ALLOC (&loop_renamer_obstack);
627 
628   /* Uses of names to rename.  We don't have to initialize this array,
629      because we know that we will only have entries for the SSA names
630      in NAMES_TO_RENAME.  */
631   use_blocks = XNEWVEC (bitmap, num_ssa_names);
632 
633   if (loop != NULL)
634     {
635       gcc_assert (changed_bbs == NULL);
636       find_uses_to_rename_in_loop (loop, use_blocks, names_to_rename,
637 				   use_flags);
638     }
639   else
640     {
641       gcc_assert (loop == NULL);
642       find_uses_to_rename (changed_bbs, use_blocks, names_to_rename, use_flags);
643     }
644 
645   if (!bitmap_empty_p (names_to_rename))
646     {
647       /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
648 	 that are the destination of an edge exiting loop number I.  */
649       bitmap *loop_exits = XNEWVEC (bitmap, number_of_loops (cfun));
650       get_loops_exits (loop_exits);
651 
652       /* Add the PHI nodes on exits of the loops for the names we need to
653 	 rewrite.  */
654       add_exit_phis (names_to_rename, use_blocks, loop_exits);
655 
656       free (loop_exits);
657 
658       /* Fix up all the names found to be used outside their original
659 	 loops.  */
660       update_ssa (TODO_update_ssa);
661     }
662 
663   bitmap_obstack_release (&loop_renamer_obstack);
664   free (use_blocks);
665 }
666 
667 /* Rewrites the non-virtual defs and uses into a loop closed ssa form.  If
668    CHANGED_BBS is not NULL, we look for uses outside loops only in the basic
669    blocks in this set.  UPDATE_FLAG is used in the call to update_ssa.  See
670    TODO_update_ssa* for documentation.  */
671 
672 void
rewrite_into_loop_closed_ssa(bitmap changed_bbs,unsigned update_flag)673 rewrite_into_loop_closed_ssa (bitmap changed_bbs, unsigned update_flag)
674 {
675   rewrite_into_loop_closed_ssa_1 (changed_bbs, update_flag, SSA_OP_USE, NULL);
676 }
677 
678 /* Rewrites virtual defs and uses with def in LOOP into loop closed ssa
679    form.  */
680 
681 void
rewrite_virtuals_into_loop_closed_ssa(struct loop * loop)682 rewrite_virtuals_into_loop_closed_ssa (struct loop *loop)
683 {
684   rewrite_into_loop_closed_ssa_1 (NULL, 0, SSA_OP_VIRTUAL_USES, loop);
685 }
686 
687 /* Check invariants of the loop closed ssa form for the USE in BB.  */
688 
689 static void
check_loop_closed_ssa_use(basic_block bb,tree use)690 check_loop_closed_ssa_use (basic_block bb, tree use)
691 {
692   gimple *def;
693   basic_block def_bb;
694 
695   if (TREE_CODE (use) != SSA_NAME || virtual_operand_p (use))
696     return;
697 
698   def = SSA_NAME_DEF_STMT (use);
699   def_bb = gimple_bb (def);
700   gcc_assert (!def_bb
701 	      || flow_bb_inside_loop_p (def_bb->loop_father, bb));
702 }
703 
704 /* Checks invariants of loop closed ssa form in statement STMT in BB.  */
705 
706 static void
check_loop_closed_ssa_stmt(basic_block bb,gimple * stmt)707 check_loop_closed_ssa_stmt (basic_block bb, gimple *stmt)
708 {
709   ssa_op_iter iter;
710   tree var;
711 
712   if (is_gimple_debug (stmt))
713     return;
714 
715   FOR_EACH_SSA_TREE_OPERAND (var, stmt, iter, SSA_OP_USE)
716     check_loop_closed_ssa_use (bb, var);
717 }
718 
719 /* Checks that invariants of the loop closed ssa form are preserved.
720    Call verify_ssa when VERIFY_SSA_P is true.  */
721 
722 DEBUG_FUNCTION void
verify_loop_closed_ssa(bool verify_ssa_p)723 verify_loop_closed_ssa (bool verify_ssa_p)
724 {
725   basic_block bb;
726   edge e;
727   edge_iterator ei;
728 
729   if (number_of_loops (cfun) <= 1)
730     return;
731 
732   if (verify_ssa_p)
733     verify_ssa (false, true);
734 
735   timevar_push (TV_VERIFY_LOOP_CLOSED);
736 
737   FOR_EACH_BB_FN (bb, cfun)
738     {
739       for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
740 	   gsi_next (&bsi))
741 	{
742 	  gphi *phi = bsi.phi ();
743 	  FOR_EACH_EDGE (e, ei, bb->preds)
744 	    check_loop_closed_ssa_use (e->src,
745 				       PHI_ARG_DEF_FROM_EDGE (phi, e));
746 	}
747 
748       for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);
749 	   gsi_next (&bsi))
750 	check_loop_closed_ssa_stmt (bb, gsi_stmt (bsi));
751     }
752 
753   timevar_pop (TV_VERIFY_LOOP_CLOSED);
754 }
755 
756 /* Split loop exit edge EXIT.  The things are a bit complicated by a need to
757    preserve the loop closed ssa form.  The newly created block is returned.  */
758 
759 basic_block
split_loop_exit_edge(edge exit)760 split_loop_exit_edge (edge exit)
761 {
762   basic_block dest = exit->dest;
763   basic_block bb = split_edge (exit);
764   gphi *phi, *new_phi;
765   tree new_name, name;
766   use_operand_p op_p;
767   gphi_iterator psi;
768   source_location locus;
769 
770   for (psi = gsi_start_phis (dest); !gsi_end_p (psi); gsi_next (&psi))
771     {
772       phi = psi.phi ();
773       op_p = PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (bb));
774       locus = gimple_phi_arg_location_from_edge (phi, single_succ_edge (bb));
775 
776       name = USE_FROM_PTR (op_p);
777 
778       /* If the argument of the PHI node is a constant, we do not need
779 	 to keep it inside loop.  */
780       if (TREE_CODE (name) != SSA_NAME)
781 	continue;
782 
783       /* Otherwise create an auxiliary phi node that will copy the value
784 	 of the SSA name out of the loop.  */
785       new_name = duplicate_ssa_name (name, NULL);
786       new_phi = create_phi_node (new_name, bb);
787       add_phi_arg (new_phi, name, exit, locus);
788       SET_USE (op_p, new_name);
789     }
790 
791   return bb;
792 }
793 
794 /* Returns the basic block in that statements should be emitted for induction
795    variables incremented at the end of the LOOP.  */
796 
797 basic_block
ip_end_pos(struct loop * loop)798 ip_end_pos (struct loop *loop)
799 {
800   return loop->latch;
801 }
802 
803 /* Returns the basic block in that statements should be emitted for induction
804    variables incremented just before exit condition of a LOOP.  */
805 
806 basic_block
ip_normal_pos(struct loop * loop)807 ip_normal_pos (struct loop *loop)
808 {
809   gimple *last;
810   basic_block bb;
811   edge exit;
812 
813   if (!single_pred_p (loop->latch))
814     return NULL;
815 
816   bb = single_pred (loop->latch);
817   last = last_stmt (bb);
818   if (!last
819       || gimple_code (last) != GIMPLE_COND)
820     return NULL;
821 
822   exit = EDGE_SUCC (bb, 0);
823   if (exit->dest == loop->latch)
824     exit = EDGE_SUCC (bb, 1);
825 
826   if (flow_bb_inside_loop_p (loop, exit->dest))
827     return NULL;
828 
829   return bb;
830 }
831 
832 /* Stores the standard position for induction variable increment in LOOP
833    (just before the exit condition if it is available and latch block is empty,
834    end of the latch block otherwise) to BSI.  INSERT_AFTER is set to true if
835    the increment should be inserted after *BSI.  */
836 
837 void
standard_iv_increment_position(struct loop * loop,gimple_stmt_iterator * bsi,bool * insert_after)838 standard_iv_increment_position (struct loop *loop, gimple_stmt_iterator *bsi,
839 				bool *insert_after)
840 {
841   basic_block bb = ip_normal_pos (loop), latch = ip_end_pos (loop);
842   gimple *last = last_stmt (latch);
843 
844   if (!bb
845       || (last && gimple_code (last) != GIMPLE_LABEL))
846     {
847       *bsi = gsi_last_bb (latch);
848       *insert_after = true;
849     }
850   else
851     {
852       *bsi = gsi_last_bb (bb);
853       *insert_after = false;
854     }
855 }
856 
857 /* Copies phi node arguments for duplicated blocks.  The index of the first
858    duplicated block is FIRST_NEW_BLOCK.  */
859 
860 static void
copy_phi_node_args(unsigned first_new_block)861 copy_phi_node_args (unsigned first_new_block)
862 {
863   unsigned i;
864 
865   for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
866     BASIC_BLOCK_FOR_FN (cfun, i)->flags |= BB_DUPLICATED;
867 
868   for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
869     add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun, i));
870 
871   for (i = first_new_block; i < (unsigned) last_basic_block_for_fn (cfun); i++)
872     BASIC_BLOCK_FOR_FN (cfun, i)->flags &= ~BB_DUPLICATED;
873 }
874 
875 
876 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
877    updates the PHI nodes at start of the copied region.  In order to
878    achieve this, only loops whose exits all lead to the same location
879    are handled.
880 
881    Notice that we do not completely update the SSA web after
882    duplication.  The caller is responsible for calling update_ssa
883    after the loop has been duplicated.  */
884 
885 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)886 gimple_duplicate_loop_to_header_edge (struct loop *loop, edge e,
887 				    unsigned int ndupl, sbitmap wont_exit,
888 				    edge orig, vec<edge> *to_remove,
889 				    int flags)
890 {
891   unsigned first_new_block;
892 
893   if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES))
894     return false;
895   if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
896     return false;
897 
898   first_new_block = last_basic_block_for_fn (cfun);
899   if (!duplicate_loop_to_header_edge (loop, e, ndupl, wont_exit,
900 				      orig, to_remove, flags))
901     return false;
902 
903   /* Readd the removed phi args for e.  */
904   flush_pending_stmts (e);
905 
906   /* Copy the phi node arguments.  */
907   copy_phi_node_args (first_new_block);
908 
909   scev_reset ();
910 
911   return true;
912 }
913 
914 /* Returns true if we can unroll LOOP FACTOR times.  Number
915    of iterations of the loop is returned in NITER.  */
916 
917 bool
can_unroll_loop_p(struct loop * loop,unsigned factor,struct tree_niter_desc * niter)918 can_unroll_loop_p (struct loop *loop, unsigned factor,
919 		   struct tree_niter_desc *niter)
920 {
921   edge exit;
922 
923   /* Check whether unrolling is possible.  We only want to unroll loops
924      for that we are able to determine number of iterations.  We also
925      want to split the extra iterations of the loop from its end,
926      therefore we require that the loop has precisely one
927      exit.  */
928 
929   exit = single_dom_exit (loop);
930   if (!exit)
931     return false;
932 
933   if (!number_of_iterations_exit (loop, exit, niter, false)
934       || niter->cmp == ERROR_MARK
935       /* Scalar evolutions analysis might have copy propagated
936 	 the abnormal ssa names into these expressions, hence
937 	 emitting the computations based on them during loop
938 	 unrolling might create overlapping life ranges for
939 	 them, and failures in out-of-ssa.  */
940       || contains_abnormal_ssa_name_p (niter->may_be_zero)
941       || contains_abnormal_ssa_name_p (niter->control.base)
942       || contains_abnormal_ssa_name_p (niter->control.step)
943       || contains_abnormal_ssa_name_p (niter->bound))
944     return false;
945 
946   /* And of course, we must be able to duplicate the loop.  */
947   if (!can_duplicate_loop_p (loop))
948     return false;
949 
950   /* The final loop should be small enough.  */
951   if (tree_num_loop_insns (loop, &eni_size_weights) * factor
952       > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS))
953     return false;
954 
955   return true;
956 }
957 
958 /* Determines the conditions that control execution of LOOP unrolled FACTOR
959    times.  DESC is number of iterations of LOOP.  ENTER_COND is set to
960    condition that must be true if the main loop can be entered.
961    EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
962    how the exit from the unrolled loop should be controlled.  */
963 
964 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)965 determine_exit_conditions (struct loop *loop, struct tree_niter_desc *desc,
966 			   unsigned factor, tree *enter_cond,
967 			   tree *exit_base, tree *exit_step,
968 			   enum tree_code *exit_cmp, tree *exit_bound)
969 {
970   gimple_seq stmts;
971   tree base = desc->control.base;
972   tree step = desc->control.step;
973   tree bound = desc->bound;
974   tree type = TREE_TYPE (step);
975   tree bigstep, delta;
976   tree min = lower_bound_in_type (type, type);
977   tree max = upper_bound_in_type (type, type);
978   enum tree_code cmp = desc->cmp;
979   tree cond = boolean_true_node, assum;
980 
981   /* For pointers, do the arithmetics in the type of step.  */
982   base = fold_convert (type, base);
983   bound = fold_convert (type, bound);
984 
985   *enter_cond = boolean_false_node;
986   *exit_base = NULL_TREE;
987   *exit_step = NULL_TREE;
988   *exit_cmp = ERROR_MARK;
989   *exit_bound = NULL_TREE;
990   gcc_assert (cmp != ERROR_MARK);
991 
992   /* We only need to be correct when we answer question
993      "Do at least FACTOR more iterations remain?" in the unrolled loop.
994      Thus, transforming BASE + STEP * i <> BOUND to
995      BASE + STEP * i < BOUND is ok.  */
996   if (cmp == NE_EXPR)
997     {
998       if (tree_int_cst_sign_bit (step))
999 	cmp = GT_EXPR;
1000       else
1001 	cmp = LT_EXPR;
1002     }
1003   else if (cmp == LT_EXPR)
1004     {
1005       gcc_assert (!tree_int_cst_sign_bit (step));
1006     }
1007   else if (cmp == GT_EXPR)
1008     {
1009       gcc_assert (tree_int_cst_sign_bit (step));
1010     }
1011   else
1012     gcc_unreachable ();
1013 
1014   /* The main body of the loop may be entered iff:
1015 
1016      1) desc->may_be_zero is false.
1017      2) it is possible to check that there are at least FACTOR iterations
1018 	of the loop, i.e., BOUND - step * FACTOR does not overflow.
1019      3) # of iterations is at least FACTOR  */
1020 
1021   if (!integer_zerop (desc->may_be_zero))
1022     cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
1023 			invert_truthvalue (desc->may_be_zero),
1024 			cond);
1025 
1026   bigstep = fold_build2 (MULT_EXPR, type, step,
1027 			 build_int_cst_type (type, factor));
1028   delta = fold_build2 (MINUS_EXPR, type, bigstep, step);
1029   if (cmp == LT_EXPR)
1030     assum = fold_build2 (GE_EXPR, boolean_type_node,
1031 			 bound,
1032 			 fold_build2 (PLUS_EXPR, type, min, delta));
1033   else
1034     assum = fold_build2 (LE_EXPR, boolean_type_node,
1035 			 bound,
1036 			 fold_build2 (PLUS_EXPR, type, max, delta));
1037   cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
1038 
1039   bound = fold_build2 (MINUS_EXPR, type, bound, delta);
1040   assum = fold_build2 (cmp, boolean_type_node, base, bound);
1041   cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, assum, cond);
1042 
1043   cond = force_gimple_operand (unshare_expr (cond), &stmts, false, NULL_TREE);
1044   if (stmts)
1045     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1046   /* cond now may be a gimple comparison, which would be OK, but also any
1047      other gimple rhs (say a && b).  In this case we need to force it to
1048      operand.  */
1049   if (!is_gimple_condexpr (cond))
1050     {
1051       cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
1052       if (stmts)
1053 	gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1054     }
1055   *enter_cond = cond;
1056 
1057   base = force_gimple_operand (unshare_expr (base), &stmts, true, NULL_TREE);
1058   if (stmts)
1059     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1060   bound = force_gimple_operand (unshare_expr (bound), &stmts, true, NULL_TREE);
1061   if (stmts)
1062     gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1063 
1064   *exit_base = base;
1065   *exit_step = bigstep;
1066   *exit_cmp = cmp;
1067   *exit_bound = bound;
1068 }
1069 
1070 /* Scales the frequencies of all basic blocks in LOOP that are strictly
1071    dominated by BB by NUM/DEN.  */
1072 
1073 static void
scale_dominated_blocks_in_loop(struct loop * loop,basic_block bb,int num,int den)1074 scale_dominated_blocks_in_loop (struct loop *loop, basic_block bb,
1075 				int num, int den)
1076 {
1077   basic_block son;
1078 
1079   if (den == 0)
1080     return;
1081 
1082   for (son = first_dom_son (CDI_DOMINATORS, bb);
1083        son;
1084        son = next_dom_son (CDI_DOMINATORS, son))
1085     {
1086       if (!flow_bb_inside_loop_p (loop, son))
1087 	continue;
1088       scale_bbs_frequencies_int (&son, 1, num, den);
1089       scale_dominated_blocks_in_loop (loop, son, num, den);
1090     }
1091 }
1092 
1093 /* Unroll LOOP FACTOR times.  DESC describes number of iterations of LOOP.
1094    EXIT is the exit of the loop to that DESC corresponds.
1095 
1096    If N is number of iterations of the loop and MAY_BE_ZERO is the condition
1097    under that loop exits in the first iteration even if N != 0,
1098 
1099    while (1)
1100      {
1101        x = phi (init, next);
1102 
1103        pre;
1104        if (st)
1105          break;
1106        post;
1107      }
1108 
1109    becomes (with possibly the exit conditions formulated a bit differently,
1110    avoiding the need to create a new iv):
1111 
1112    if (MAY_BE_ZERO || N < FACTOR)
1113      goto rest;
1114 
1115    do
1116      {
1117        x = phi (init, next);
1118 
1119        pre;
1120        post;
1121        pre;
1122        post;
1123        ...
1124        pre;
1125        post;
1126        N -= FACTOR;
1127 
1128      } while (N >= FACTOR);
1129 
1130    rest:
1131      init' = phi (init, x);
1132 
1133    while (1)
1134      {
1135        x = phi (init', next);
1136 
1137        pre;
1138        if (st)
1139          break;
1140        post;
1141      }
1142 
1143    Before the loop is unrolled, TRANSFORM is called for it (only for the
1144    unrolled loop, but not for its versioned copy).  DATA is passed to
1145    TRANSFORM.  */
1146 
1147 /* Probability in % that the unrolled loop is entered.  Just a guess.  */
1148 #define PROB_UNROLLED_LOOP_ENTERED 90
1149 
1150 void
tree_transform_and_unroll_loop(struct loop * loop,unsigned factor,edge exit,struct tree_niter_desc * desc,transform_callback transform,void * data)1151 tree_transform_and_unroll_loop (struct loop *loop, unsigned factor,
1152 				edge exit, struct tree_niter_desc *desc,
1153 				transform_callback transform,
1154 				void *data)
1155 {
1156   gcond *exit_if;
1157   tree ctr_before, ctr_after;
1158   tree enter_main_cond, exit_base, exit_step, exit_bound;
1159   enum tree_code exit_cmp;
1160   gphi *phi_old_loop, *phi_new_loop, *phi_rest;
1161   gphi_iterator psi_old_loop, psi_new_loop;
1162   tree init, next, new_init;
1163   struct loop *new_loop;
1164   basic_block rest, exit_bb;
1165   edge old_entry, new_entry, old_latch, precond_edge, new_exit;
1166   edge new_nonexit, e;
1167   gimple_stmt_iterator bsi;
1168   use_operand_p op;
1169   bool ok;
1170   unsigned est_niter, prob_entry, scale_unrolled, scale_rest, freq_e, freq_h;
1171   unsigned new_est_niter, i, prob;
1172   unsigned irr = loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP;
1173   sbitmap wont_exit;
1174   auto_vec<edge> to_remove;
1175 
1176   est_niter = expected_loop_iterations (loop);
1177   determine_exit_conditions (loop, desc, factor,
1178 			     &enter_main_cond, &exit_base, &exit_step,
1179 			     &exit_cmp, &exit_bound);
1180 
1181   /* Let us assume that the unrolled loop is quite likely to be entered.  */
1182   if (integer_nonzerop (enter_main_cond))
1183     prob_entry = REG_BR_PROB_BASE;
1184   else
1185     prob_entry = PROB_UNROLLED_LOOP_ENTERED * REG_BR_PROB_BASE / 100;
1186 
1187   /* The values for scales should keep profile consistent, and somewhat close
1188      to correct.
1189 
1190      TODO: The current value of SCALE_REST makes it appear that the loop that
1191      is created by splitting the remaining iterations of the unrolled loop is
1192      executed the same number of times as the original loop, and with the same
1193      frequencies, which is obviously wrong.  This does not appear to cause
1194      problems, so we do not bother with fixing it for now.  To make the profile
1195      correct, we would need to change the probability of the exit edge of the
1196      loop, and recompute the distribution of frequencies in its body because
1197      of this change (scale the frequencies of blocks before and after the exit
1198      by appropriate factors).  */
1199   scale_unrolled = prob_entry;
1200   scale_rest = REG_BR_PROB_BASE;
1201 
1202   new_loop = loop_version (loop, enter_main_cond, NULL,
1203 			   prob_entry, scale_unrolled, scale_rest, true);
1204   gcc_assert (new_loop != NULL);
1205   update_ssa (TODO_update_ssa);
1206 
1207   /* Determine the probability of the exit edge of the unrolled loop.  */
1208   new_est_niter = est_niter / factor;
1209 
1210   /* Without profile feedback, loops for that we do not know a better estimate
1211      are assumed to roll 10 times.  When we unroll such loop, it appears to
1212      roll too little, and it may even seem to be cold.  To avoid this, we
1213      ensure that the created loop appears to roll at least 5 times (but at
1214      most as many times as before unrolling).  */
1215   if (new_est_niter < 5)
1216     {
1217       if (est_niter < 5)
1218 	new_est_niter = est_niter;
1219       else
1220 	new_est_niter = 5;
1221     }
1222 
1223   /* Prepare the cfg and update the phi nodes.  Move the loop exit to the
1224      loop latch (and make its condition dummy, for the moment).  */
1225   rest = loop_preheader_edge (new_loop)->src;
1226   precond_edge = single_pred_edge (rest);
1227   split_edge (loop_latch_edge (loop));
1228   exit_bb = single_pred (loop->latch);
1229 
1230   /* Since the exit edge will be removed, the frequency of all the blocks
1231      in the loop that are dominated by it must be scaled by
1232      1 / (1 - exit->probability).  */
1233   scale_dominated_blocks_in_loop (loop, exit->src,
1234 				  REG_BR_PROB_BASE,
1235 				  REG_BR_PROB_BASE - exit->probability);
1236 
1237   bsi = gsi_last_bb (exit_bb);
1238   exit_if = gimple_build_cond (EQ_EXPR, integer_zero_node,
1239 			       integer_zero_node,
1240 			       NULL_TREE, NULL_TREE);
1241 
1242   gsi_insert_after (&bsi, exit_if, GSI_NEW_STMT);
1243   new_exit = make_edge (exit_bb, rest, EDGE_FALSE_VALUE | irr);
1244   rescan_loop_exit (new_exit, true, false);
1245 
1246   /* Set the probability of new exit to the same of the old one.  Fix
1247      the frequency of the latch block, by scaling it back by
1248      1 - exit->probability.  */
1249   new_exit->count = exit->count;
1250   new_exit->probability = exit->probability;
1251   new_nonexit = single_pred_edge (loop->latch);
1252   new_nonexit->probability = REG_BR_PROB_BASE - exit->probability;
1253   new_nonexit->flags = EDGE_TRUE_VALUE;
1254   new_nonexit->count -= exit->count;
1255   if (new_nonexit->count < 0)
1256     new_nonexit->count = 0;
1257   scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1258 			     REG_BR_PROB_BASE);
1259 
1260   old_entry = loop_preheader_edge (loop);
1261   new_entry = loop_preheader_edge (new_loop);
1262   old_latch = loop_latch_edge (loop);
1263   for (psi_old_loop = gsi_start_phis (loop->header),
1264        psi_new_loop = gsi_start_phis (new_loop->header);
1265        !gsi_end_p (psi_old_loop);
1266        gsi_next (&psi_old_loop), gsi_next (&psi_new_loop))
1267     {
1268       phi_old_loop = psi_old_loop.phi ();
1269       phi_new_loop = psi_new_loop.phi ();
1270 
1271       init = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_entry);
1272       op = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop, new_entry);
1273       gcc_assert (operand_equal_for_phi_arg_p (init, USE_FROM_PTR (op)));
1274       next = PHI_ARG_DEF_FROM_EDGE (phi_old_loop, old_latch);
1275 
1276       /* Prefer using original variable as a base for the new ssa name.
1277 	 This is necessary for virtual ops, and useful in order to avoid
1278 	 losing debug info for real ops.  */
1279       if (TREE_CODE (next) == SSA_NAME
1280 	  && useless_type_conversion_p (TREE_TYPE (next),
1281 					TREE_TYPE (init)))
1282 	new_init = copy_ssa_name (next);
1283       else if (TREE_CODE (init) == SSA_NAME
1284 	       && useless_type_conversion_p (TREE_TYPE (init),
1285 					     TREE_TYPE (next)))
1286 	new_init = copy_ssa_name (init);
1287       else if (useless_type_conversion_p (TREE_TYPE (next), TREE_TYPE (init)))
1288 	new_init = make_temp_ssa_name (TREE_TYPE (next), NULL, "unrinittmp");
1289       else
1290 	new_init = make_temp_ssa_name (TREE_TYPE (init), NULL, "unrinittmp");
1291 
1292       phi_rest = create_phi_node (new_init, rest);
1293 
1294       add_phi_arg (phi_rest, init, precond_edge, UNKNOWN_LOCATION);
1295       add_phi_arg (phi_rest, next, new_exit, UNKNOWN_LOCATION);
1296       SET_USE (op, new_init);
1297     }
1298 
1299   remove_path (exit);
1300 
1301   /* Transform the loop.  */
1302   if (transform)
1303     (*transform) (loop, data);
1304 
1305   /* Unroll the loop and remove the exits in all iterations except for the
1306      last one.  */
1307   wont_exit = sbitmap_alloc (factor);
1308   bitmap_ones (wont_exit);
1309   bitmap_clear_bit (wont_exit, factor - 1);
1310 
1311   ok = gimple_duplicate_loop_to_header_edge
1312 	  (loop, loop_latch_edge (loop), factor - 1,
1313 	   wont_exit, new_exit, &to_remove, DLTHE_FLAG_UPDATE_FREQ);
1314   free (wont_exit);
1315   gcc_assert (ok);
1316 
1317   FOR_EACH_VEC_ELT (to_remove, i, e)
1318     {
1319       ok = remove_path (e);
1320       gcc_assert (ok);
1321     }
1322   update_ssa (TODO_update_ssa);
1323 
1324   /* Ensure that the frequencies in the loop match the new estimated
1325      number of iterations, and change the probability of the new
1326      exit edge.  */
1327   freq_h = loop->header->frequency;
1328   freq_e = EDGE_FREQUENCY (loop_preheader_edge (loop));
1329   if (freq_h != 0)
1330     scale_loop_frequencies (loop, freq_e * (new_est_niter + 1), freq_h);
1331 
1332   exit_bb = single_pred (loop->latch);
1333   new_exit = find_edge (exit_bb, rest);
1334   new_exit->count = loop_preheader_edge (loop)->count;
1335   new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);
1336 
1337   rest->count += new_exit->count;
1338   rest->frequency += EDGE_FREQUENCY (new_exit);
1339 
1340   new_nonexit = single_pred_edge (loop->latch);
1341   prob = new_nonexit->probability;
1342   new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
1343   new_nonexit->count = exit_bb->count - new_exit->count;
1344   if (new_nonexit->count < 0)
1345     new_nonexit->count = 0;
1346   if (prob > 0)
1347     scale_bbs_frequencies_int (&loop->latch, 1, new_nonexit->probability,
1348 			       prob);
1349 
1350   /* Finally create the new counter for number of iterations and add the new
1351      exit instruction.  */
1352   bsi = gsi_last_nondebug_bb (exit_bb);
1353   exit_if = as_a <gcond *> (gsi_stmt (bsi));
1354   create_iv (exit_base, exit_step, NULL_TREE, loop,
1355 	     &bsi, false, &ctr_before, &ctr_after);
1356   gimple_cond_set_code (exit_if, exit_cmp);
1357   gimple_cond_set_lhs (exit_if, ctr_after);
1358   gimple_cond_set_rhs (exit_if, exit_bound);
1359   update_stmt (exit_if);
1360 
1361   checking_verify_flow_info ();
1362   checking_verify_loop_structure ();
1363   checking_verify_loop_closed_ssa (true);
1364 }
1365 
1366 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1367    want to transform the loop before unrolling.  The meaning
1368    of the arguments is the same as for tree_transform_and_unroll_loop.  */
1369 
1370 void
tree_unroll_loop(struct loop * loop,unsigned factor,edge exit,struct tree_niter_desc * desc)1371 tree_unroll_loop (struct loop *loop, unsigned factor,
1372 		  edge exit, struct tree_niter_desc *desc)
1373 {
1374   tree_transform_and_unroll_loop (loop, factor, exit, desc,
1375 				  NULL, NULL);
1376 }
1377 
1378 /* Rewrite the phi node at position PSI in function of the main
1379    induction variable MAIN_IV and insert the generated code at GSI.  */
1380 
1381 static void
rewrite_phi_with_iv(loop_p loop,gphi_iterator * psi,gimple_stmt_iterator * gsi,tree main_iv)1382 rewrite_phi_with_iv (loop_p loop,
1383 		     gphi_iterator *psi,
1384 		     gimple_stmt_iterator *gsi,
1385 		     tree main_iv)
1386 {
1387   affine_iv iv;
1388   gassign *stmt;
1389   gphi *phi = psi->phi ();
1390   tree atype, mtype, val, res = PHI_RESULT (phi);
1391 
1392   if (virtual_operand_p (res) || res == main_iv)
1393     {
1394       gsi_next (psi);
1395       return;
1396     }
1397 
1398   if (!simple_iv (loop, loop, res, &iv, true))
1399     {
1400       gsi_next (psi);
1401       return;
1402     }
1403 
1404   remove_phi_node (psi, false);
1405 
1406   atype = TREE_TYPE (res);
1407   mtype = POINTER_TYPE_P (atype) ? sizetype : atype;
1408   val = fold_build2 (MULT_EXPR, mtype, unshare_expr (iv.step),
1409 		     fold_convert (mtype, main_iv));
1410   val = fold_build2 (POINTER_TYPE_P (atype)
1411 		     ? POINTER_PLUS_EXPR : PLUS_EXPR,
1412 		     atype, unshare_expr (iv.base), val);
1413   val = force_gimple_operand_gsi (gsi, val, false, NULL_TREE, true,
1414 				  GSI_SAME_STMT);
1415   stmt = gimple_build_assign (res, val);
1416   gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1417 }
1418 
1419 /* Rewrite all the phi nodes of LOOP in function of the main induction
1420    variable MAIN_IV.  */
1421 
1422 static void
rewrite_all_phi_nodes_with_iv(loop_p loop,tree main_iv)1423 rewrite_all_phi_nodes_with_iv (loop_p loop, tree main_iv)
1424 {
1425   unsigned i;
1426   basic_block *bbs = get_loop_body_in_dom_order (loop);
1427   gphi_iterator psi;
1428 
1429   for (i = 0; i < loop->num_nodes; i++)
1430     {
1431       basic_block bb = bbs[i];
1432       gimple_stmt_iterator gsi = gsi_after_labels (bb);
1433 
1434       if (bb->loop_father != loop)
1435 	continue;
1436 
1437       for (psi = gsi_start_phis (bb); !gsi_end_p (psi); )
1438 	rewrite_phi_with_iv (loop, &psi, &gsi, main_iv);
1439     }
1440 
1441   free (bbs);
1442 }
1443 
1444 /* Bases all the induction variables in LOOP on a single induction variable
1445    (with base 0 and step 1), whose final value is compared with *NIT.  When the
1446    IV type precision has to be larger than *NIT type precision, *NIT is
1447    converted to the larger type, the conversion code is inserted before the
1448    loop, and *NIT is updated to the new definition.  When BUMP_IN_LATCH is true,
1449    the induction variable is incremented in the loop latch, otherwise it is
1450    incremented in the loop header.  Return the induction variable that was
1451    created.  */
1452 
1453 tree
canonicalize_loop_ivs(struct loop * loop,tree * nit,bool bump_in_latch)1454 canonicalize_loop_ivs (struct loop *loop, tree *nit, bool bump_in_latch)
1455 {
1456   unsigned precision = TYPE_PRECISION (TREE_TYPE (*nit));
1457   unsigned original_precision = precision;
1458   tree type, var_before;
1459   gimple_stmt_iterator gsi;
1460   gphi_iterator psi;
1461   gcond *stmt;
1462   edge exit = single_dom_exit (loop);
1463   gimple_seq stmts;
1464   machine_mode mode;
1465   bool unsigned_p = false;
1466 
1467   for (psi = gsi_start_phis (loop->header);
1468        !gsi_end_p (psi); gsi_next (&psi))
1469     {
1470       gphi *phi = psi.phi ();
1471       tree res = PHI_RESULT (phi);
1472       bool uns;
1473 
1474       type = TREE_TYPE (res);
1475       if (virtual_operand_p (res)
1476 	  || (!INTEGRAL_TYPE_P (type)
1477 	      && !POINTER_TYPE_P (type))
1478 	  || TYPE_PRECISION (type) < precision)
1479 	continue;
1480 
1481       uns = POINTER_TYPE_P (type) | TYPE_UNSIGNED (type);
1482 
1483       if (TYPE_PRECISION (type) > precision)
1484 	unsigned_p = uns;
1485       else
1486 	unsigned_p |= uns;
1487 
1488       precision = TYPE_PRECISION (type);
1489     }
1490 
1491   mode = smallest_mode_for_size (precision, MODE_INT);
1492   precision = GET_MODE_PRECISION (mode);
1493   type = build_nonstandard_integer_type (precision, unsigned_p);
1494 
1495   if (original_precision != precision)
1496     {
1497       *nit = fold_convert (type, *nit);
1498       *nit = force_gimple_operand (*nit, &stmts, true, NULL_TREE);
1499       if (stmts)
1500 	gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts);
1501     }
1502 
1503   if (bump_in_latch)
1504     gsi = gsi_last_bb (loop->latch);
1505   else
1506     gsi = gsi_last_nondebug_bb (loop->header);
1507   create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
1508 	     loop, &gsi, bump_in_latch, &var_before, NULL);
1509 
1510   rewrite_all_phi_nodes_with_iv (loop, var_before);
1511 
1512   stmt = as_a <gcond *> (last_stmt (exit->src));
1513   /* Make the loop exit if the control condition is not satisfied.  */
1514   if (exit->flags & EDGE_TRUE_VALUE)
1515     {
1516       edge te, fe;
1517 
1518       extract_true_false_edges_from_block (exit->src, &te, &fe);
1519       te->flags = EDGE_FALSE_VALUE;
1520       fe->flags = EDGE_TRUE_VALUE;
1521     }
1522   gimple_cond_set_code (stmt, LT_EXPR);
1523   gimple_cond_set_lhs (stmt, var_before);
1524   gimple_cond_set_rhs (stmt, *nit);
1525   update_stmt (stmt);
1526 
1527   return var_before;
1528 }
1529