1 /* Routines for discovering and unpropagating edge equivalences.
2    Copyright (C) 2005-2018 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
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation; either version 3, or (at your option)
9 any later version.
10 
11 GCC is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14 GNU General Public License for more details.
15 
16 You should have received a copy of the GNU General Public License
17 along with 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 "tree-pass.h"
27 #include "ssa.h"
28 #include "fold-const.h"
29 #include "cfganal.h"
30 #include "gimple-iterator.h"
31 #include "tree-cfg.h"
32 #include "domwalk.h"
33 #include "tree-hash-traits.h"
34 #include "tree-ssa-live.h"
35 #include "tree-ssa-coalesce.h"
36 
37 /* The basic structure describing an equivalency created by traversing
38    an edge.  Traversing the edge effectively means that we can assume
39    that we've seen an assignment LHS = RHS.  */
40 struct edge_equivalency
41 {
42   tree rhs;
43   tree lhs;
44 };
45 
46 /* This routine finds and records edge equivalences for every edge
47    in the CFG.
48 
49    When complete, each edge that creates an equivalency will have an
50    EDGE_EQUIVALENCY structure hanging off the edge's AUX field.
51    The caller is responsible for freeing the AUX fields.  */
52 
53 static void
associate_equivalences_with_edges(void)54 associate_equivalences_with_edges (void)
55 {
56   basic_block bb;
57 
58   /* Walk over each block.  If the block ends with a control statement,
59      then it might create a useful equivalence.  */
60   FOR_EACH_BB_FN (bb, cfun)
61     {
62       gimple_stmt_iterator gsi = gsi_last_bb (bb);
63       gimple *stmt;
64 
65       /* If the block does not end with a COND_EXPR or SWITCH_EXPR
66 	 then there is nothing to do.  */
67       if (gsi_end_p (gsi))
68 	continue;
69 
70       stmt = gsi_stmt (gsi);
71 
72       if (!stmt)
73 	continue;
74 
75       /* A COND_EXPR may create an equivalency in a variety of different
76 	 ways.  */
77       if (gimple_code (stmt) == GIMPLE_COND)
78 	{
79 	  edge true_edge;
80 	  edge false_edge;
81 	  struct edge_equivalency *equivalency;
82 	  enum tree_code code = gimple_cond_code (stmt);
83 
84 	  extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
85 
86 	  /* Equality tests may create one or two equivalences.  */
87 	  if (code == EQ_EXPR || code == NE_EXPR)
88 	    {
89 	      tree op0 = gimple_cond_lhs (stmt);
90 	      tree op1 = gimple_cond_rhs (stmt);
91 
92 	      /* Special case comparing booleans against a constant as we
93 		 know the value of OP0 on both arms of the branch.  i.e., we
94 		 can record an equivalence for OP0 rather than COND.  */
95 	      if (TREE_CODE (op0) == SSA_NAME
96 		  && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0)
97 		  && ssa_name_has_boolean_range (op0)
98 		  && is_gimple_min_invariant (op1)
99 		  && (integer_zerop (op1) || integer_onep (op1)))
100 		{
101 		  tree true_val = constant_boolean_node (true, TREE_TYPE (op0));
102 		  tree false_val = constant_boolean_node (false,
103 							  TREE_TYPE (op0));
104 		  if (code == EQ_EXPR)
105 		    {
106 		      equivalency = XNEW (struct edge_equivalency);
107 		      equivalency->lhs = op0;
108 		      equivalency->rhs = (integer_zerop (op1)
109 					  ? false_val
110 					  : true_val);
111 		      true_edge->aux = equivalency;
112 
113 		      equivalency = XNEW (struct edge_equivalency);
114 		      equivalency->lhs = op0;
115 		      equivalency->rhs = (integer_zerop (op1)
116 					  ? true_val
117 					  : false_val);
118 		      false_edge->aux = equivalency;
119 		    }
120 		  else
121 		    {
122 		      equivalency = XNEW (struct edge_equivalency);
123 		      equivalency->lhs = op0;
124 		      equivalency->rhs = (integer_zerop (op1)
125 					  ? true_val
126 					  : false_val);
127 		      true_edge->aux = equivalency;
128 
129 		      equivalency = XNEW (struct edge_equivalency);
130 		      equivalency->lhs = op0;
131 		      equivalency->rhs = (integer_zerop (op1)
132 					  ? false_val
133 					  : true_val);
134 		      false_edge->aux = equivalency;
135 		    }
136 		}
137 
138 	      else if (TREE_CODE (op0) == SSA_NAME
139 		       && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0)
140 		       && (is_gimple_min_invariant (op1)
141 			   || (TREE_CODE (op1) == SSA_NAME
142 			       && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op1))))
143 		{
144 		  /* For IEEE, -0.0 == 0.0, so we don't necessarily know
145 		     the sign of a variable compared against zero.  If
146 		     we're honoring signed zeros, then we cannot record
147 		     this value unless we know that the value is nonzero.  */
148 		  if (HONOR_SIGNED_ZEROS (op0)
149 		      && (TREE_CODE (op1) != REAL_CST
150 			  || real_equal (&dconst0, &TREE_REAL_CST (op1))))
151 		    continue;
152 
153 		  equivalency = XNEW (struct edge_equivalency);
154 		  equivalency->lhs = op0;
155 		  equivalency->rhs = op1;
156 		  if (code == EQ_EXPR)
157 		    true_edge->aux = equivalency;
158 		  else
159 		    false_edge->aux = equivalency;
160 
161 		}
162 	    }
163 
164 	  /* ??? TRUTH_NOT_EXPR can create an equivalence too.  */
165 	}
166 
167       /* For a SWITCH_EXPR, a case label which represents a single
168 	 value and which is the only case label which reaches the
169 	 target block creates an equivalence.  */
170       else if (gimple_code (stmt) == GIMPLE_SWITCH)
171 	{
172 	  gswitch *switch_stmt = as_a <gswitch *> (stmt);
173 	  tree cond = gimple_switch_index (switch_stmt);
174 
175 	  if (TREE_CODE (cond) == SSA_NAME
176 	      && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (cond))
177 	    {
178 	      int i, n_labels = gimple_switch_num_labels (switch_stmt);
179 	      tree *info = XCNEWVEC (tree, last_basic_block_for_fn (cfun));
180 
181 	      /* Walk over the case label vector.  Record blocks
182 		 which are reached by a single case label which represents
183 		 a single value.  */
184 	      for (i = 0; i < n_labels; i++)
185 		{
186 		  tree label = gimple_switch_label (switch_stmt, i);
187 		  basic_block bb = label_to_block (CASE_LABEL (label));
188 
189 		  if (CASE_HIGH (label)
190 		      || !CASE_LOW (label)
191 		      || info[bb->index])
192 		    info[bb->index] = error_mark_node;
193 		  else
194 		    info[bb->index] = label;
195 		}
196 
197 	      /* Now walk over the blocks to determine which ones were
198 		 marked as being reached by a useful case label.  */
199 	      for (i = 0; i < n_basic_blocks_for_fn (cfun); i++)
200 		{
201 		  tree node = info[i];
202 
203 		  if (node != NULL
204 		      && node != error_mark_node)
205 		    {
206 		      tree x = fold_convert (TREE_TYPE (cond), CASE_LOW (node));
207 		      struct edge_equivalency *equivalency;
208 
209 		      /* Record an equivalency on the edge from BB to basic
210 			 block I.  */
211 		      equivalency = XNEW (struct edge_equivalency);
212 		      equivalency->rhs = x;
213 		      equivalency->lhs = cond;
214 		      find_edge (bb, BASIC_BLOCK_FOR_FN (cfun, i))->aux =
215 			equivalency;
216 		    }
217 		}
218 	      free (info);
219 	    }
220 	}
221 
222     }
223 }
224 
225 
226 /* Translating out of SSA sometimes requires inserting copies and
227    constant initializations on edges to eliminate PHI nodes.
228 
229    In some cases those copies and constant initializations are
230    redundant because the target already has the value on the
231    RHS of the assignment.
232 
233    We previously tried to catch these cases after translating
234    out of SSA form.  However, that code often missed cases.  Worse
235    yet, the cases it missed were also often missed by the RTL
236    optimizers.  Thus the resulting code had redundant instructions.
237 
238    This pass attempts to detect these situations before translating
239    out of SSA form.
240 
241    The key concept that this pass is built upon is that these
242    redundant copies and constant initializations often occur
243    due to constant/copy propagating equivalences resulting from
244    COND_EXPRs and SWITCH_EXPRs.
245 
246    We want to do those propagations as they can sometimes allow
247    the SSA optimizers to do a better job.  However, in the cases
248    where such propagations do not result in further optimization,
249    we would like to "undo" the propagation to avoid the redundant
250    copies and constant initializations.
251 
252    This pass works by first associating equivalences with edges in
253    the CFG.  For example, the edge leading from a SWITCH_EXPR to
254    its associated CASE_LABEL will have an equivalency between
255    SWITCH_COND and the value in the case label.
256 
257    Once we have found the edge equivalences, we proceed to walk
258    the CFG in dominator order.  As we traverse edges we record
259    equivalences associated with those edges we traverse.
260 
261    When we encounter a PHI node, we walk its arguments to see if we
262    have an equivalence for the PHI argument.  If so, then we replace
263    the argument.
264 
265    Equivalences are looked up based on their value (think of it as
266    the RHS of an assignment).   A value may be an SSA_NAME or an
267    invariant.  We may have several SSA_NAMEs with the same value,
268    so with each value we have a list of SSA_NAMEs that have the
269    same value.  */
270 
271 /* Traits for the hash_map to record the value to SSA name equivalences
272    mapping.  */
273 struct ssa_equip_hash_traits : default_hash_traits <tree>
274 {
hashssa_equip_hash_traits275   static inline hashval_t hash (value_type value)
276     { return iterative_hash_expr (value, 0); }
equalssa_equip_hash_traits277   static inline bool equal (value_type existing, value_type candidate)
278     { return operand_equal_p (existing, candidate, 0); }
279 };
280 
281 typedef hash_map<tree, auto_vec<tree>,
282 		 simple_hashmap_traits <ssa_equip_hash_traits,
283 					auto_vec <tree> > > val_ssa_equiv_t;
284 
285 /* Global hash table implementing a mapping from invariant values
286    to a list of SSA_NAMEs which have the same value.  We might be
287    able to reuse tree-vn for this code.  */
288 val_ssa_equiv_t *val_ssa_equiv;
289 
290 static void uncprop_into_successor_phis (basic_block);
291 
292 /* Remove the most recently recorded equivalency for VALUE.  */
293 
294 static void
remove_equivalence(tree value)295 remove_equivalence (tree value)
296 {
297     val_ssa_equiv->get (value)->pop ();
298 }
299 
300 /* Record EQUIVALENCE = VALUE into our hash table.  */
301 
302 static void
record_equiv(tree value,tree equivalence)303 record_equiv (tree value, tree equivalence)
304 {
305   val_ssa_equiv->get_or_insert (value).safe_push (equivalence);
306 }
307 
308 class uncprop_dom_walker : public dom_walker
309 {
310 public:
uncprop_dom_walker(cdi_direction direction)311   uncprop_dom_walker (cdi_direction direction) : dom_walker (direction) {}
312 
313   virtual edge before_dom_children (basic_block);
314   virtual void after_dom_children (basic_block);
315 
316 private:
317 
318   /* As we enter each block we record the value for any edge equivalency
319      leading to this block.  If no such edge equivalency exists, then we
320      record NULL.  These equivalences are live until we leave the dominator
321      subtree rooted at the block where we record the equivalency.  */
322   auto_vec<tree, 2> m_equiv_stack;
323 };
324 
325 /* We have finished processing the dominator children of BB, perform
326    any finalization actions in preparation for leaving this node in
327    the dominator tree.  */
328 
329 void
after_dom_children(basic_block bb ATTRIBUTE_UNUSED)330 uncprop_dom_walker::after_dom_children (basic_block bb ATTRIBUTE_UNUSED)
331 {
332   /* Pop the topmost value off the equiv stack.  */
333   tree value = m_equiv_stack.pop ();
334 
335   /* If that value was non-null, then pop the topmost equivalency off
336      its equivalency stack.  */
337   if (value != NULL)
338     remove_equivalence (value);
339 }
340 
341 /* Unpropagate values from PHI nodes in successor blocks of BB.  */
342 
343 static void
uncprop_into_successor_phis(basic_block bb)344 uncprop_into_successor_phis (basic_block bb)
345 {
346   edge e;
347   edge_iterator ei;
348 
349   /* For each successor edge, first temporarily record any equivalence
350      on that edge.  Then unpropagate values in any PHI nodes at the
351      destination of the edge.  Then remove the temporary equivalence.  */
352   FOR_EACH_EDGE (e, ei, bb->succs)
353     {
354       gimple_seq phis = phi_nodes (e->dest);
355       gimple_stmt_iterator gsi;
356 
357       /* If there are no PHI nodes in this destination, then there is
358 	 no sense in recording any equivalences.  */
359       if (gimple_seq_empty_p (phis))
360 	continue;
361 
362       /* Record any equivalency associated with E.  */
363       if (e->aux)
364 	{
365 	  struct edge_equivalency *equiv = (struct edge_equivalency *) e->aux;
366 	  record_equiv (equiv->rhs, equiv->lhs);
367 	}
368 
369       /* Walk over the PHI nodes, unpropagating values.  */
370       for (gsi = gsi_start (phis) ; !gsi_end_p (gsi); gsi_next (&gsi))
371 	{
372 	  gimple *phi = gsi_stmt (gsi);
373 	  tree arg = PHI_ARG_DEF (phi, e->dest_idx);
374 	  tree res = PHI_RESULT (phi);
375 
376 	  /* If the argument is not an invariant and can be potentially
377 	     coalesced with the result, then there's no point in
378 	     un-propagating the argument.  */
379 	  if (!is_gimple_min_invariant (arg)
380 	      && gimple_can_coalesce_p (arg, res))
381 	    continue;
382 
383 	  /* Lookup this argument's value in the hash table.  */
384 	  vec<tree> *equivalences = val_ssa_equiv->get (arg);
385 	  if (equivalences)
386 	    {
387 	      /* Walk every equivalence with the same value.  If we find
388 		 one that can potentially coalesce with the PHI rsult,
389 		 then replace the value in the argument with its equivalent
390 		 SSA_NAME.  Use the most recent equivalence as hopefully
391 		 that results in shortest lifetimes.  */
392 	      for (int j = equivalences->length () - 1; j >= 0; j--)
393 		{
394 		  tree equiv = (*equivalences)[j];
395 
396 		  if (gimple_can_coalesce_p (equiv, res))
397 		    {
398 		      SET_PHI_ARG_DEF (phi, e->dest_idx, equiv);
399 		      break;
400 		    }
401 		}
402 	    }
403 	}
404 
405       /* If we had an equivalence associated with this edge, remove it.  */
406       if (e->aux)
407 	{
408 	  struct edge_equivalency *equiv = (struct edge_equivalency *) e->aux;
409 	  remove_equivalence (equiv->rhs);
410 	}
411     }
412 }
413 
414 edge
before_dom_children(basic_block bb)415 uncprop_dom_walker::before_dom_children (basic_block bb)
416 {
417   basic_block parent;
418   bool recorded = false;
419 
420   /* If this block is dominated by a single incoming edge and that edge
421      has an equivalency, then record the equivalency and push the
422      VALUE onto EQUIV_STACK.  Else push a NULL entry on EQUIV_STACK.  */
423   parent = get_immediate_dominator (CDI_DOMINATORS, bb);
424   if (parent)
425     {
426       edge e = single_pred_edge_ignoring_loop_edges (bb, false);
427 
428       if (e && e->src == parent && e->aux)
429 	{
430 	  struct edge_equivalency *equiv = (struct edge_equivalency *) e->aux;
431 
432 	  record_equiv (equiv->rhs, equiv->lhs);
433 	  m_equiv_stack.safe_push (equiv->rhs);
434 	  recorded = true;
435 	}
436     }
437 
438   if (!recorded)
439     m_equiv_stack.safe_push (NULL_TREE);
440 
441   uncprop_into_successor_phis (bb);
442   return NULL;
443 }
444 
445 namespace {
446 
447 const pass_data pass_data_uncprop =
448 {
449   GIMPLE_PASS, /* type */
450   "uncprop", /* name */
451   OPTGROUP_NONE, /* optinfo_flags */
452   TV_TREE_SSA_UNCPROP, /* tv_id */
453   ( PROP_cfg | PROP_ssa ), /* properties_required */
454   0, /* properties_provided */
455   0, /* properties_destroyed */
456   0, /* todo_flags_start */
457   0, /* todo_flags_finish */
458 };
459 
460 class pass_uncprop : public gimple_opt_pass
461 {
462 public:
pass_uncprop(gcc::context * ctxt)463   pass_uncprop (gcc::context *ctxt)
464     : gimple_opt_pass (pass_data_uncprop, ctxt)
465   {}
466 
467   /* opt_pass methods: */
clone()468   opt_pass * clone () { return new pass_uncprop (m_ctxt); }
gate(function *)469   virtual bool gate (function *) { return flag_tree_dom != 0; }
470   virtual unsigned int execute (function *);
471 
472 }; // class pass_uncprop
473 
474 unsigned int
execute(function * fun)475 pass_uncprop::execute (function *fun)
476 {
477   basic_block bb;
478 
479   associate_equivalences_with_edges ();
480 
481   /* Create our global data structures.  */
482   val_ssa_equiv = new val_ssa_equiv_t (1024);
483 
484   /* We're going to do a dominator walk, so ensure that we have
485      dominance information.  */
486   calculate_dominance_info (CDI_DOMINATORS);
487 
488   /* Recursively walk the dominator tree undoing unprofitable
489      constant/copy propagations.  */
490   uncprop_dom_walker (CDI_DOMINATORS).walk (fun->cfg->x_entry_block_ptr);
491 
492   /* we just need to empty elements out of the hash table, and cleanup the
493     AUX field on the edges.  */
494   delete val_ssa_equiv;
495   val_ssa_equiv = NULL;
496   FOR_EACH_BB_FN (bb, fun)
497     {
498       edge e;
499       edge_iterator ei;
500 
501       FOR_EACH_EDGE (e, ei, bb->succs)
502 	{
503 	  if (e->aux)
504 	    {
505 	      free (e->aux);
506 	      e->aux = NULL;
507 	    }
508 	}
509     }
510   return 0;
511 }
512 
513 } // anon namespace
514 
515 gimple_opt_pass *
make_pass_uncprop(gcc::context * ctxt)516 make_pass_uncprop (gcc::context *ctxt)
517 {
518   return new pass_uncprop (ctxt);
519 }
520