1 /* Induction variable canonicalization.
2 Copyright (C) 2004, 2005, 2007, 2008 Free Software Foundation, Inc.
4 This file is part of GCC.
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
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
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/>. */
20 /* This pass detects the loops that iterate a constant number of times,
21 adds a canonical induction variable (step -1, tested against 0)
22 and replaces the exit test. This enables the less powerful rtl
23 level analysis to use this information.
25 This might spoil the code in some cases (by increasing register pressure).
26 Note that in the case the new variable is not needed, ivopts will get rid
27 of it, so it might only be a problem when there are no other linear induction
28 variables. In that case the created optimization possibilities are likely
31 Additionally in case we detect that it is beneficial to unroll the
32 loop completely, we do it right here to expose the optimization
33 possibilities to the following passes. */
37 #include "coretypes.h"
42 #include "hard-reg-set.h"
43 #include "basic-block.h"
45 #include "diagnostic.h"
46 #include "tree-flow.h"
47 #include "tree-dump.h"
49 #include "tree-pass.h"
51 #include "tree-chrec.h"
52 #include "tree-scalar-evolution.h"
55 #include "tree-inline.h"
57 /* Specifies types of loops that may be unrolled. */
61 UL_SINGLE_ITER
, /* Only loops that exit immediately in the first
63 UL_NO_GROWTH
, /* Only loops whose unrolling will not cause increase
65 UL_ALL
/* All suitable loops. */
68 /* Adds a canonical induction variable to LOOP iterating NITER times. EXIT
69 is the exit edge whose condition is replaced. */
72 create_canonical_iv (struct loop
*loop
, edge exit
, tree niter
)
77 gimple_stmt_iterator incr_at
;
80 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
82 fprintf (dump_file
, "Added canonical iv to loop %d, ", loop
->num
);
83 print_generic_expr (dump_file
, niter
, TDF_SLIM
);
84 fprintf (dump_file
, " iterations.\n");
87 cond
= last_stmt (exit
->src
);
88 in
= EDGE_SUCC (exit
->src
, 0);
90 in
= EDGE_SUCC (exit
->src
, 1);
92 /* Note that we do not need to worry about overflows, since
93 type of niter is always unsigned and all comparisons are
94 just for equality/nonequality -- i.e. everything works
95 with a modulo arithmetics. */
97 type
= TREE_TYPE (niter
);
98 niter
= fold_build2 (PLUS_EXPR
, type
,
100 build_int_cst (type
, 1));
101 incr_at
= gsi_last_bb (in
->src
);
103 build_int_cst (type
, -1),
105 &incr_at
, false, NULL
, &var
);
107 cmp
= (exit
->flags
& EDGE_TRUE_VALUE
) ? EQ_EXPR
: NE_EXPR
;
108 gimple_cond_set_code (cond
, cmp
);
109 gimple_cond_set_lhs (cond
, var
);
110 gimple_cond_set_rhs (cond
, build_int_cst (type
, 0));
114 /* Computes an estimated number of insns in LOOP, weighted by WEIGHTS. */
117 tree_num_loop_insns (struct loop
*loop
, eni_weights
*weights
)
119 basic_block
*body
= get_loop_body (loop
);
120 gimple_stmt_iterator gsi
;
121 unsigned size
= 1, i
;
123 for (i
= 0; i
< loop
->num_nodes
; i
++)
124 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
); gsi_next (&gsi
))
125 size
+= estimate_num_insns (gsi_stmt (gsi
), weights
);
131 /* Estimate number of insns of completely unrolled loop. We assume
132 that the size of the unrolled loop is decreased in the
133 following way (the numbers of insns are based on what
134 estimate_num_insns returns for appropriate statements):
136 1) exit condition gets removed (2 insns)
137 2) increment of the control variable gets removed (2 insns)
138 3) All remaining statements are likely to get simplified
139 due to constant propagation. Hard to estimate; just
140 as a heuristics we decrease the rest by 1/3.
142 NINSNS is the number of insns in the loop before unrolling.
143 NUNROLL is the number of times the loop is unrolled. */
145 static unsigned HOST_WIDE_INT
146 estimated_unrolled_size (unsigned HOST_WIDE_INT ninsns
,
147 unsigned HOST_WIDE_INT nunroll
)
149 HOST_WIDE_INT unr_insns
= 2 * ((HOST_WIDE_INT
) ninsns
- 4) / 3;
152 unr_insns
*= (nunroll
+ 1);
157 /* Tries to unroll LOOP completely, i.e. NITER times.
158 UL determines which loops we are allowed to unroll.
159 EXIT is the exit of the loop that should be eliminated. */
162 try_unroll_loop_completely (struct loop
*loop
,
163 edge exit
, tree niter
,
164 enum unroll_level ul
)
166 unsigned HOST_WIDE_INT n_unroll
, ninsns
, max_unroll
, unr_insns
;
172 if (!host_integerp (niter
, 1))
174 n_unroll
= tree_low_cst (niter
, 1);
176 max_unroll
= PARAM_VALUE (PARAM_MAX_COMPLETELY_PEEL_TIMES
);
177 if (n_unroll
> max_unroll
)
182 if (ul
== UL_SINGLE_ITER
)
185 ninsns
= tree_num_loop_insns (loop
, &eni_size_weights
);
187 if (n_unroll
* ninsns
188 > (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS
))
191 unr_insns
= estimated_unrolled_size (ninsns
, n_unroll
);
192 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
194 fprintf (dump_file
, " Loop size: %d\n", (int) ninsns
);
195 fprintf (dump_file
, " Estimated size after unrolling: %d\n",
199 if (ul
== UL_NO_GROWTH
200 && unr_insns
> ninsns
)
202 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
203 fprintf (dump_file
, "Not unrolling loop %d.\n", loop
->num
);
213 VEC (edge
, heap
) *to_remove
= NULL
;
215 initialize_original_copy_tables ();
216 wont_exit
= sbitmap_alloc (n_unroll
+ 1);
217 sbitmap_ones (wont_exit
);
218 RESET_BIT (wont_exit
, 0);
220 if (!gimple_duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
223 DLTHE_FLAG_UPDATE_FREQ
224 | DLTHE_FLAG_COMPLETTE_PEEL
))
226 free_original_copy_tables ();
231 for (i
= 0; VEC_iterate (edge
, to_remove
, i
, e
); i
++)
233 bool ok
= remove_path (e
);
237 VEC_free (edge
, heap
, to_remove
);
239 free_original_copy_tables ();
242 cond
= last_stmt (exit
->src
);
243 if (exit
->flags
& EDGE_TRUE_VALUE
)
244 gimple_cond_make_true (cond
);
246 gimple_cond_make_false (cond
);
248 update_ssa (TODO_update_ssa
);
250 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
251 fprintf (dump_file
, "Unrolled loop %d completely.\n", loop
->num
);
256 /* Adds a canonical induction variable to LOOP if suitable.
257 CREATE_IV is true if we may create a new iv. UL determines
258 which loops we are allowed to completely unroll. If TRY_EVAL is true, we try
259 to determine the number of iterations of a loop by direct evaluation.
260 Returns true if cfg is changed. */
263 canonicalize_loop_induction_variables (struct loop
*loop
,
264 bool create_iv
, enum unroll_level ul
,
270 niter
= number_of_latch_executions (loop
);
271 if (TREE_CODE (niter
) == INTEGER_CST
)
273 exit
= single_exit (loop
);
274 if (!just_once_each_iteration_p (loop
, exit
->src
))
279 /* If the loop has more than one exit, try checking all of them
280 for # of iterations determinable through scev. */
281 if (!single_exit (loop
))
282 niter
= find_loop_niter (loop
, &exit
);
284 /* Finally if everything else fails, try brute force evaluation. */
286 && (chrec_contains_undetermined (niter
)
287 || TREE_CODE (niter
) != INTEGER_CST
))
288 niter
= find_loop_niter_by_eval (loop
, &exit
);
290 if (chrec_contains_undetermined (niter
)
291 || TREE_CODE (niter
) != INTEGER_CST
)
295 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
297 fprintf (dump_file
, "Loop %d iterates ", loop
->num
);
298 print_generic_expr (dump_file
, niter
, TDF_SLIM
);
299 fprintf (dump_file
, " times.\n");
302 if (try_unroll_loop_completely (loop
, exit
, niter
, ul
))
306 create_canonical_iv (loop
, exit
, niter
);
311 /* The main entry point of the pass. Adds canonical induction variables
312 to the suitable loops. */
315 canonicalize_induction_variables (void)
319 bool changed
= false;
321 FOR_EACH_LOOP (li
, loop
, 0)
323 changed
|= canonicalize_loop_induction_variables (loop
,
324 true, UL_SINGLE_ITER
,
328 /* Clean up the information about numbers of iterations, since brute force
329 evaluation could reveal new information. */
333 return TODO_cleanup_cfg
;
337 /* Unroll LOOPS completely if they iterate just few times. Unless
338 MAY_INCREASE_SIZE is true, perform the unrolling only if the
339 size of the code does not increase. */
342 tree_unroll_loops_completely (bool may_increase_size
, bool unroll_outer
)
347 enum unroll_level ul
;
353 FOR_EACH_LOOP (li
, loop
, LI_ONLY_INNERMOST
)
355 if (may_increase_size
&& maybe_hot_bb_p (loop
->header
)
356 /* Unroll outermost loops only if asked to do so or they do
357 not cause code growth. */
359 || loop_outer (loop_outer (loop
))))
363 changed
|= canonicalize_loop_induction_variables
364 (loop
, false, ul
, !flag_tree_loop_ivcanon
);
369 /* This will take care of removing completely unrolled loops
370 from the loop structures so we can continue unrolling now
372 if (cleanup_tree_cfg ())
373 update_ssa (TODO_update_ssa_only_virtuals
);
375 /* Clean up the information about numbers of iterations, since
376 complete unrolling might have invalidated it. */
385 /* Checks whether LOOP is empty. */
388 empty_loop_p (struct loop
*loop
)
391 struct tree_niter_desc niter
;
393 gimple_stmt_iterator gsi
;
396 /* If the loop has multiple exits, it is too hard for us to handle.
397 Similarly, if the exit is not dominating, we cannot determine
398 whether the loop is not infinite. */
399 exit
= single_dom_exit (loop
);
403 /* The loop must be finite. */
404 if (!number_of_iterations_exit (loop
, exit
, &niter
, false))
407 /* Values of all loop exit phi nodes must be invariants. */
408 for (gsi
= gsi_start(phi_nodes (exit
->dest
)); !gsi_end_p (gsi
); gsi_next (&gsi
))
410 gimple phi
= gsi_stmt (gsi
);
413 if (!is_gimple_reg (PHI_RESULT (phi
)))
416 def
= PHI_ARG_DEF_FROM_EDGE (phi
, exit
);
418 if (!expr_invariant_in_loop_p (loop
, def
))
422 /* And there should be no memory modifying or from other reasons
423 unremovable statements. */
424 body
= get_loop_body (loop
);
425 for (i
= 0; i
< loop
->num_nodes
; i
++)
427 /* Irreducible region might be infinite. */
428 if (body
[i
]->flags
& BB_IRREDUCIBLE_LOOP
)
434 for (gsi
= gsi_start_bb (body
[i
]); !gsi_end_p (gsi
); gsi_next (&gsi
))
436 gimple stmt
= gsi_stmt (gsi
);
438 if (!ZERO_SSA_OPERANDS (stmt
, SSA_OP_VIRTUAL_DEFS
)
439 || gimple_has_volatile_ops (stmt
))
445 /* Also, asm statements and calls may have side effects and we
446 cannot change the number of times they are executed. */
447 switch (gimple_code (stmt
))
450 if (gimple_has_side_effects (stmt
))
458 /* We cannot remove volatile assembler. */
459 if (gimple_asm_volatile_p (stmt
))
476 /* Remove LOOP by making it exit in the first iteration. */
479 remove_empty_loop (struct loop
*loop
)
481 edge exit
= single_dom_exit (loop
), non_exit
;
482 gimple cond_stmt
= last_stmt (exit
->src
);
484 unsigned n_before
, freq_in
, freq_h
;
485 gcov_type exit_count
= exit
->count
;
488 fprintf (dump_file
, "Removing empty loop %d\n", loop
->num
);
490 non_exit
= EDGE_SUCC (exit
->src
, 0);
491 if (non_exit
== exit
)
492 non_exit
= EDGE_SUCC (exit
->src
, 1);
494 if (exit
->flags
& EDGE_TRUE_VALUE
)
495 gimple_cond_make_true (cond_stmt
);
497 gimple_cond_make_false (cond_stmt
);
498 update_stmt (cond_stmt
);
500 /* Let us set the probabilities of the edges coming from the exit block. */
501 exit
->probability
= REG_BR_PROB_BASE
;
502 non_exit
->probability
= 0;
505 /* Update frequencies and counts. Everything before
506 the exit needs to be scaled FREQ_IN/FREQ_H times,
507 where FREQ_IN is the frequency of the entry edge
508 and FREQ_H is the frequency of the loop header.
509 Everything after the exit has zero frequency. */
510 freq_h
= loop
->header
->frequency
;
511 freq_in
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
514 body
= get_loop_body_in_dom_order (loop
);
515 for (n_before
= 1; n_before
<= loop
->num_nodes
; n_before
++)
516 if (body
[n_before
- 1] == exit
->src
)
518 scale_bbs_frequencies_int (body
, n_before
, freq_in
, freq_h
);
519 scale_bbs_frequencies_int (body
+ n_before
, loop
->num_nodes
- n_before
,
524 /* Number of executions of exit is not changed, thus we need to restore
525 the original value. */
526 exit
->count
= exit_count
;
529 /* Removes LOOP if it is empty. Returns true if LOOP is removed. CHANGED
530 is set to true if LOOP or any of its subloops is removed. */
533 try_remove_empty_loop (struct loop
*loop
, bool *changed
)
535 bool nonempty_subloop
= false;
538 /* First, all subloops must be removed. */
539 for (sub
= loop
->inner
; sub
; sub
= sub
->next
)
540 nonempty_subloop
|= !try_remove_empty_loop (sub
, changed
);
542 if (nonempty_subloop
|| !empty_loop_p (loop
))
545 remove_empty_loop (loop
);
550 /* Remove the empty loops. */
553 remove_empty_loops (void)
555 bool changed
= false;
558 for (loop
= current_loops
->tree_root
->inner
; loop
; loop
= loop
->next
)
559 try_remove_empty_loop (loop
, &changed
);
564 return TODO_cleanup_cfg
;