]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/tree-switch-conversion.c
lto-streamer-out.c (DFS::DFS_write_tree_body): Do not stream BINFO_BASE_ACCESSES...
[thirdparty/gcc.git] / gcc / tree-switch-conversion.c
CommitLineData
531b10fc
SB
1/* Lower GIMPLE_SWITCH expressions to something more efficient than
2 a jump table.
85ec4feb 3 Copyright (C) 2006-2018 Free Software Foundation, Inc.
b6e99746
MJ
4
5This file is part of GCC.
6
7GCC is free software; you can redistribute it and/or modify it
8under the terms of the GNU General Public License as published by the
9Free Software Foundation; either version 3, or (at your option) any
10later version.
11
12GCC is distributed in the hope that it will be useful, but WITHOUT
13ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15for more details.
16
17You should have received a copy of the GNU General Public License
18along with GCC; see the file COPYING3. If not, write to the Free
19Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
2002110-1301, USA. */
21
531b10fc
SB
22/* This file handles the lowering of GIMPLE_SWITCH to an indexed
23 load, or a series of bit-test-and-branch expressions. */
24
25#include "config.h"
26#include "system.h"
27#include "coretypes.h"
c7131fb2 28#include "backend.h"
957060b5
AM
29#include "insn-codes.h"
30#include "rtl.h"
c7131fb2
AM
31#include "tree.h"
32#include "gimple.h"
957060b5
AM
33#include "cfghooks.h"
34#include "tree-pass.h"
c7131fb2 35#include "ssa.h"
957060b5
AM
36#include "optabs-tree.h"
37#include "cgraph.h"
38#include "gimple-pretty-print.h"
531b10fc 39#include "params.h"
40e23961 40#include "fold-const.h"
d8a2d370
DN
41#include "varasm.h"
42#include "stor-layout.h"
60393bbc 43#include "cfganal.h"
45b0be94 44#include "gimplify.h"
5be5c238 45#include "gimple-iterator.h"
18f429e2 46#include "gimplify-me.h"
442b4905 47#include "tree-cfg.h"
a9e0d843 48#include "cfgloop.h"
9dc3d6a9
ML
49#include "alloc-pool.h"
50#include "target.h"
51#include "tree-into-ssa.h"
46dbeb40 52#include "omp-general.h"
7ee2468b
SB
53
54/* ??? For lang_hooks.types.type_for_mode, but is there a word_mode
55 type in the GIMPLE type system that is language-independent? */
531b10fc
SB
56#include "langhooks.h"
57
789410e4 58#include "tree-switch-conversion.h"
531b10fc 59\f
789410e4 60using namespace tree_switch_conversion;
531b10fc 61
789410e4 62/* Constructor. */
531b10fc 63
789410e4
ML
64switch_conversion::switch_conversion (): m_final_bb (NULL), m_other_count (),
65 m_constructors (NULL), m_default_values (NULL),
66 m_arr_ref_first (NULL), m_arr_ref_last (NULL),
67 m_reason (NULL), m_default_case_nonstandard (false), m_cfg_altered (false)
531b10fc 68{
531b10fc 69}
b6e99746 70
789410e4 71/* Collection information about SWTCH statement. */
886cd84f 72
789410e4
ML
73void
74switch_conversion::collect (gswitch *swtch)
b6e99746 75{
726a989a 76 unsigned int branch_num = gimple_switch_num_labels (swtch);
886cd84f 77 tree min_case, max_case;
789410e4 78 unsigned int i;
18bfe940 79 edge e, e_default, e_first;
886cd84f 80 edge_iterator ei;
18bfe940 81 basic_block first;
886cd84f 82
789410e4 83 m_switch = swtch;
b6e99746
MJ
84
85 /* The gimplifier has already sorted the cases by CASE_LOW and ensured there
fd8d363e
SB
86 is a default label which is the first in the vector.
87 Collect the bits we can deduce from the CFG. */
789410e4
ML
88 m_index_expr = gimple_switch_index (swtch);
89 m_switch_bb = gimple_bb (swtch);
90 m_default_bb
18bfe940 91 = label_to_block (CASE_LABEL (gimple_switch_default_label (swtch)));
789410e4
ML
92 e_default = find_edge (m_switch_bb, m_default_bb);
93 m_default_prob = e_default->probability;
94 m_default_count = e_default->count ();
95 FOR_EACH_EDGE (e, ei, m_switch_bb->succs)
886cd84f 96 if (e != e_default)
789410e4 97 m_other_count += e->count ();
b6e99746 98
18bfe940
JJ
99 /* Get upper and lower bounds of case values, and the covered range. */
100 min_case = gimple_switch_label (swtch, 1);
101 max_case = gimple_switch_label (swtch, branch_num - 1);
102
789410e4 103 m_range_min = CASE_LOW (min_case);
18bfe940 104 if (CASE_HIGH (max_case) != NULL_TREE)
789410e4 105 m_range_max = CASE_HIGH (max_case);
18bfe940 106 else
789410e4 107 m_range_max = CASE_LOW (max_case);
18bfe940 108
789410e4
ML
109 m_contiguous_range = true;
110 tree last = CASE_HIGH (min_case) ? CASE_HIGH (min_case) : m_range_min;
18bfe940
JJ
111 for (i = 2; i < branch_num; i++)
112 {
113 tree elt = gimple_switch_label (swtch, i);
8e6cdc90 114 if (wi::to_wide (last) + 1 != wi::to_wide (CASE_LOW (elt)))
18bfe940 115 {
789410e4 116 m_contiguous_range = false;
18bfe940
JJ
117 break;
118 }
119 last = CASE_HIGH (elt) ? CASE_HIGH (elt) : CASE_LOW (elt);
120 }
121
789410e4 122 if (m_contiguous_range)
18bfe940
JJ
123 {
124 first = label_to_block (CASE_LABEL (gimple_switch_label (swtch, 1)));
789410e4 125 e_first = find_edge (m_switch_bb, first);
18bfe940
JJ
126 }
127 else
128 {
789410e4 129 first = m_default_bb;
18bfe940
JJ
130 e_first = e_default;
131 }
132
886cd84f 133 /* See if there is one common successor block for all branch
866f20d6 134 targets. If it exists, record it in FINAL_BB.
18bfe940
JJ
135 Start with the destination of the first non-default case
136 if the range is contiguous and default case otherwise as
137 guess or its destination in case it is a forwarder block. */
138 if (! single_pred_p (e_first->dest))
789410e4 139 m_final_bb = e_first->dest;
18bfe940
JJ
140 else if (single_succ_p (e_first->dest)
141 && ! single_pred_p (single_succ (e_first->dest)))
789410e4 142 m_final_bb = single_succ (e_first->dest);
866f20d6 143 /* Require that all switch destinations are either that common
18bfe940
JJ
144 FINAL_BB or a forwarder to it, except for the default
145 case if contiguous range. */
789410e4
ML
146 if (m_final_bb)
147 FOR_EACH_EDGE (e, ei, m_switch_bb->succs)
886cd84f 148 {
789410e4 149 if (e->dest == m_final_bb)
886cd84f
SB
150 continue;
151
152 if (single_pred_p (e->dest)
153 && single_succ_p (e->dest)
789410e4 154 && single_succ (e->dest) == m_final_bb)
886cd84f
SB
155 continue;
156
789410e4 157 if (e == e_default && m_contiguous_range)
18bfe940 158 {
789410e4 159 m_default_case_nonstandard = true;
18bfe940
JJ
160 continue;
161 }
162
789410e4 163 m_final_bb = NULL;
886cd84f
SB
164 break;
165 }
166
789410e4
ML
167 m_range_size
168 = int_const_binop (MINUS_EXPR, m_range_max, m_range_min);
b6e99746 169
886cd84f
SB
170 /* Get a count of the number of case labels. Single-valued case labels
171 simply count as one, but a case range counts double, since it may
172 require two compares if it gets lowered as a branching tree. */
789410e4 173 m_count = 0;
886cd84f
SB
174 for (i = 1; i < branch_num; i++)
175 {
176 tree elt = gimple_switch_label (swtch, i);
789410e4 177 m_count++;
886cd84f
SB
178 if (CASE_HIGH (elt)
179 && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt)))
789410e4 180 m_count++;
886cd84f 181 }
dc223ad4
ML
182
183 /* Get the number of unique non-default targets out of the GIMPLE_SWITCH
184 block. Assume a CFG cleanup would have already removed degenerate
185 switch statements, this allows us to just use EDGE_COUNT. */
186 m_uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1;
886cd84f 187}
b6e99746 188
789410e4 189/* Checks whether the range given by individual case statements of the switch
886cd84f
SB
190 switch statement isn't too big and whether the number of branches actually
191 satisfies the size of the new array. */
b6e99746 192
789410e4
ML
193bool
194switch_conversion::check_range ()
886cd84f 195{
789410e4
ML
196 gcc_assert (m_range_size);
197 if (!tree_fits_uhwi_p (m_range_size))
b6e99746 198 {
789410e4 199 m_reason = "index range way too large or otherwise unusable";
b6e99746
MJ
200 return false;
201 }
202
789410e4
ML
203 if (tree_to_uhwi (m_range_size)
204 > ((unsigned) m_count * SWITCH_CONVERSION_BRANCH_RATIO))
b6e99746 205 {
789410e4 206 m_reason = "the maximum range-branch ratio exceeded";
b6e99746
MJ
207 return false;
208 }
209
210 return true;
211}
212
789410e4 213/* Checks whether all but the final BB basic blocks are empty. */
b6e99746 214
789410e4
ML
215bool
216switch_conversion::check_all_empty_except_final ()
b6e99746 217{
789410e4 218 edge e, e_default = find_edge (m_switch_bb, m_default_bb);
886cd84f 219 edge_iterator ei;
b6e99746 220
789410e4 221 FOR_EACH_EDGE (e, ei, m_switch_bb->succs)
b6e99746 222 {
789410e4 223 if (e->dest == m_final_bb)
886cd84f 224 continue;
b6e99746 225
886cd84f 226 if (!empty_block_p (e->dest))
b6e99746 227 {
789410e4 228 if (m_contiguous_range && e == e_default)
18bfe940 229 {
789410e4 230 m_default_case_nonstandard = true;
18bfe940
JJ
231 continue;
232 }
233
789410e4 234 m_reason = "bad case - a non-final BB not empty";
b6e99746
MJ
235 return false;
236 }
b6e99746
MJ
237 }
238
239 return true;
240}
241
242/* This function checks whether all required values in phi nodes in final_bb
243 are constants. Required values are those that correspond to a basic block
244 which is a part of the examined switch statement. It returns true if the
245 phi nodes are OK, otherwise false. */
246
789410e4
ML
247bool
248switch_conversion::check_final_bb ()
b6e99746 249{
538dd0b7 250 gphi_iterator gsi;
b6e99746 251
789410e4
ML
252 m_phi_count = 0;
253 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi))
b6e99746 254 {
538dd0b7 255 gphi *phi = gsi.phi ();
726a989a 256 unsigned int i;
b6e99746 257
18bfe940
JJ
258 if (virtual_operand_p (gimple_phi_result (phi)))
259 continue;
260
789410e4 261 m_phi_count++;
b6e99746 262
726a989a 263 for (i = 0; i < gimple_phi_num_args (phi); i++)
b6e99746 264 {
726a989a 265 basic_block bb = gimple_phi_arg_edge (phi, i)->src;
b6e99746 266
789410e4 267 if (bb == m_switch_bb
18bfe940 268 || (single_pred_p (bb)
789410e4
ML
269 && single_pred (bb) == m_switch_bb
270 && (!m_default_case_nonstandard
18bfe940 271 || empty_block_p (bb))))
b6e99746 272 {
f6e6e990 273 tree reloc, val;
18bfe940 274 const char *reason = NULL;
f6e6e990
JJ
275
276 val = gimple_phi_arg_def (phi, i);
277 if (!is_gimple_ip_invariant (val))
18bfe940
JJ
278 reason = "non-invariant value from a case";
279 else
f6e6e990 280 {
18bfe940
JJ
281 reloc = initializer_constant_valid_p (val, TREE_TYPE (val));
282 if ((flag_pic && reloc != null_pointer_node)
283 || (!flag_pic && reloc == NULL_TREE))
284 {
285 if (reloc)
286 reason
287 = "value from a case would need runtime relocations";
288 else
289 reason
290 = "value from a case is not a valid initializer";
291 }
f6e6e990 292 }
18bfe940 293 if (reason)
f6e6e990 294 {
18bfe940
JJ
295 /* For contiguous range, we can allow non-constant
296 or one that needs relocation, as long as it is
297 only reachable from the default case. */
789410e4
ML
298 if (bb == m_switch_bb)
299 bb = m_final_bb;
300 if (!m_contiguous_range || bb != m_default_bb)
18bfe940 301 {
789410e4 302 m_reason = reason;
18bfe940
JJ
303 return false;
304 }
305
789410e4 306 unsigned int branch_num = gimple_switch_num_labels (m_switch);
18bfe940
JJ
307 for (unsigned int i = 1; i < branch_num; i++)
308 {
789410e4 309 tree lab = CASE_LABEL (gimple_switch_label (m_switch, i));
18bfe940
JJ
310 if (label_to_block (lab) == bb)
311 {
789410e4 312 m_reason = reason;
18bfe940
JJ
313 return false;
314 }
315 }
789410e4 316 m_default_case_nonstandard = true;
f6e6e990 317 }
b6e99746
MJ
318 }
319 }
320 }
321
322 return true;
323}
324
325/* The following function allocates default_values, target_{in,out}_names and
326 constructors arrays. The last one is also populated with pointers to
327 vectors that will become constructors of new arrays. */
328
789410e4
ML
329void
330switch_conversion::create_temp_arrays ()
b6e99746
MJ
331{
332 int i;
333
789410e4 334 m_default_values = XCNEWVEC (tree, m_phi_count * 3);
9771b263
DN
335 /* ??? Macros do not support multi argument templates in their
336 argument list. We create a typedef to work around that problem. */
337 typedef vec<constructor_elt, va_gc> *vec_constructor_elt_gc;
789410e4
ML
338 m_constructors = XCNEWVEC (vec_constructor_elt_gc, m_phi_count);
339 m_target_inbound_names = m_default_values + m_phi_count;
340 m_target_outbound_names = m_target_inbound_names + m_phi_count;
341 for (i = 0; i < m_phi_count; i++)
342 vec_alloc (m_constructors[i], tree_to_uhwi (m_range_size) + 1);
b6e99746
MJ
343}
344
345/* Populate the array of default values in the order of phi nodes.
18bfe940
JJ
346 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch
347 if the range is non-contiguous or the default case has standard
348 structure, otherwise it is the first non-default case instead. */
b6e99746 349
789410e4
ML
350void
351switch_conversion::gather_default_values (tree default_case)
b6e99746 352{
538dd0b7 353 gphi_iterator gsi;
b6e99746
MJ
354 basic_block bb = label_to_block (CASE_LABEL (default_case));
355 edge e;
726a989a 356 int i = 0;
b6e99746 357
18bfe940 358 gcc_assert (CASE_LOW (default_case) == NULL_TREE
789410e4 359 || m_default_case_nonstandard);
b6e99746 360
789410e4
ML
361 if (bb == m_final_bb)
362 e = find_edge (m_switch_bb, bb);
b6e99746
MJ
363 else
364 e = single_succ_edge (bb);
365
789410e4 366 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi))
b6e99746 367 {
538dd0b7 368 gphi *phi = gsi.phi ();
18bfe940
JJ
369 if (virtual_operand_p (gimple_phi_result (phi)))
370 continue;
b6e99746
MJ
371 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e);
372 gcc_assert (val);
789410e4 373 m_default_values[i++] = val;
b6e99746
MJ
374 }
375}
376
377/* The following function populates the vectors in the constructors array with
378 future contents of the static arrays. The vectors are populated in the
789410e4 379 order of phi nodes. */
b6e99746 380
789410e4
ML
381void
382switch_conversion::build_constructors ()
b6e99746 383{
789410e4
ML
384 unsigned i, branch_num = gimple_switch_num_labels (m_switch);
385 tree pos = m_range_min;
18bfe940 386 tree pos_one = build_int_cst (TREE_TYPE (pos), 1);
b6e99746 387
726a989a 388 for (i = 1; i < branch_num; i++)
b6e99746 389 {
789410e4 390 tree cs = gimple_switch_label (m_switch, i);
b6e99746
MJ
391 basic_block bb = label_to_block (CASE_LABEL (cs));
392 edge e;
726a989a 393 tree high;
538dd0b7 394 gphi_iterator gsi;
b6e99746
MJ
395 int j;
396
789410e4
ML
397 if (bb == m_final_bb)
398 e = find_edge (m_switch_bb, bb);
b6e99746
MJ
399 else
400 e = single_succ_edge (bb);
401 gcc_assert (e);
402
403 while (tree_int_cst_lt (pos, CASE_LOW (cs)))
404 {
405 int k;
789410e4 406 for (k = 0; k < m_phi_count; k++)
b6e99746 407 {
f32682ca 408 constructor_elt elt;
b6e99746 409
789410e4 410 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min);
d1f98542 411 elt.value
789410e4
ML
412 = unshare_expr_without_location (m_default_values[k]);
413 m_constructors[k]->quick_push (elt);
b6e99746
MJ
414 }
415
18bfe940 416 pos = int_const_binop (PLUS_EXPR, pos, pos_one);
b6e99746 417 }
b1ae1681 418 gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs)));
b6e99746
MJ
419
420 j = 0;
421 if (CASE_HIGH (cs))
422 high = CASE_HIGH (cs);
423 else
b1ae1681 424 high = CASE_LOW (cs);
789410e4 425 for (gsi = gsi_start_phis (m_final_bb);
726a989a 426 !gsi_end_p (gsi); gsi_next (&gsi))
b6e99746 427 {
538dd0b7 428 gphi *phi = gsi.phi ();
18bfe940
JJ
429 if (virtual_operand_p (gimple_phi_result (phi)))
430 continue;
b6e99746 431 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e);
7f2a9982 432 tree low = CASE_LOW (cs);
b6e99746
MJ
433 pos = CASE_LOW (cs);
434
b8698a0f 435 do
b6e99746 436 {
f32682ca 437 constructor_elt elt;
b6e99746 438
789410e4 439 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min);
d1f98542 440 elt.value = unshare_expr_without_location (val);
789410e4 441 m_constructors[j]->quick_push (elt);
b6e99746 442
18bfe940 443 pos = int_const_binop (PLUS_EXPR, pos, pos_one);
7156c8ab
MJ
444 } while (!tree_int_cst_lt (high, pos)
445 && tree_int_cst_lt (low, pos));
b6e99746
MJ
446 j++;
447 }
448 }
449}
450
7156c8ab
MJ
451/* If all values in the constructor vector are the same, return the value.
452 Otherwise return NULL_TREE. Not supposed to be called for empty
453 vectors. */
454
789410e4
ML
455tree
456switch_conversion::contains_same_values_p (vec<constructor_elt, va_gc> *vec)
7156c8ab 457{
8e97bc2b 458 unsigned int i;
7156c8ab 459 tree prev = NULL_TREE;
8e97bc2b 460 constructor_elt *elt;
7156c8ab 461
9771b263 462 FOR_EACH_VEC_SAFE_ELT (vec, i, elt)
7156c8ab 463 {
7156c8ab
MJ
464 if (!prev)
465 prev = elt->value;
466 else if (!operand_equal_p (elt->value, prev, OEP_ONLY_CONST))
467 return NULL_TREE;
468 }
469 return prev;
470}
471
f1b0632a
OH
472/* Return type which should be used for array elements, either TYPE's
473 main variant or, for integral types, some smaller integral type
474 that can still hold all the constants. */
8e97bc2b 475
789410e4
ML
476tree
477switch_conversion::array_value_type (tree type, int num)
8e97bc2b 478{
789410e4 479 unsigned int i, len = vec_safe_length (m_constructors[num]);
8e97bc2b 480 constructor_elt *elt;
8e97bc2b
JJ
481 int sign = 0;
482 tree smaller_type;
483
f1b0632a
OH
484 /* Types with alignments greater than their size can reach here, e.g. out of
485 SRA. We couldn't use these as an array component type so get back to the
486 main variant first, which, for our purposes, is fine for other types as
487 well. */
488
489 type = TYPE_MAIN_VARIANT (type);
490
8e97bc2b
JJ
491 if (!INTEGRAL_TYPE_P (type))
492 return type;
493
7a504f33 494 scalar_int_mode type_mode = SCALAR_INT_TYPE_MODE (type);
095a2d76 495 scalar_int_mode mode = get_narrowest_mode (type_mode);
ec35d572 496 if (GET_MODE_SIZE (type_mode) <= GET_MODE_SIZE (mode))
8e97bc2b
JJ
497 return type;
498
789410e4 499 if (len < (optimize_bb_for_size_p (gimple_bb (m_switch)) ? 2 : 32))
8e97bc2b
JJ
500 return type;
501
789410e4 502 FOR_EACH_VEC_SAFE_ELT (m_constructors[num], i, elt)
8e97bc2b 503 {
807e902e 504 wide_int cst;
8e97bc2b
JJ
505
506 if (TREE_CODE (elt->value) != INTEGER_CST)
507 return type;
508
8e6cdc90 509 cst = wi::to_wide (elt->value);
8e97bc2b
JJ
510 while (1)
511 {
512 unsigned int prec = GET_MODE_BITSIZE (mode);
513 if (prec > HOST_BITS_PER_WIDE_INT)
514 return type;
515
807e902e 516 if (sign >= 0 && cst == wi::zext (cst, prec))
8e97bc2b 517 {
807e902e 518 if (sign == 0 && cst == wi::sext (cst, prec))
8e97bc2b
JJ
519 break;
520 sign = 1;
521 break;
522 }
807e902e 523 if (sign <= 0 && cst == wi::sext (cst, prec))
8e97bc2b
JJ
524 {
525 sign = -1;
526 break;
527 }
528
529 if (sign == 1)
530 sign = 0;
531
490d0f6c 532 if (!GET_MODE_WIDER_MODE (mode).exists (&mode)
ec35d572 533 || GET_MODE_SIZE (mode) >= GET_MODE_SIZE (type_mode))
8e97bc2b
JJ
534 return type;
535 }
536 }
537
538 if (sign == 0)
539 sign = TYPE_UNSIGNED (type) ? 1 : -1;
540 smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0);
7a504f33
RS
541 if (GET_MODE_SIZE (type_mode)
542 <= GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (smaller_type)))
8e97bc2b
JJ
543 return type;
544
545 return smaller_type;
546}
547
789410e4
ML
548/* Create an appropriate array type and declaration and assemble a static
549 array variable. Also create a load statement that initializes
550 the variable in question with a value from the static array. SWTCH is
551 the switch statement being converted, NUM is the index to
552 arrays of constructors, default values and target SSA names
553 for this particular array. ARR_INDEX_TYPE is the type of the index
554 of the new array, PHI is the phi node of the final BB that corresponds
555 to the value that will be loaded from the created array. TIDX
7156c8ab
MJ
556 is an ssa name of a temporary variable holding the index for loads from the
557 new array. */
b6e99746 558
789410e4
ML
559void
560switch_conversion::build_one_array (int num, tree arr_index_type,
561 gphi *phi, tree tidx)
b6e99746 562{
7156c8ab 563 tree name, cst;
355fe088 564 gimple *load;
789410e4
ML
565 gimple_stmt_iterator gsi = gsi_for_stmt (m_switch);
566 location_t loc = gimple_location (m_switch);
b6e99746 567
789410e4 568 gcc_assert (m_default_values[num]);
b6e99746 569
b731b390 570 name = copy_ssa_name (PHI_RESULT (phi));
789410e4 571 m_target_inbound_names[num] = name;
b6e99746 572
789410e4 573 cst = contains_same_values_p (m_constructors[num]);
7156c8ab
MJ
574 if (cst)
575 load = gimple_build_assign (name, cst);
576 else
577 {
8e97bc2b 578 tree array_type, ctor, decl, value_type, fetch, default_type;
7156c8ab 579
789410e4
ML
580 default_type = TREE_TYPE (m_default_values[num]);
581 value_type = array_value_type (default_type, num);
7156c8ab 582 array_type = build_array_type (value_type, arr_index_type);
8e97bc2b
JJ
583 if (default_type != value_type)
584 {
585 unsigned int i;
586 constructor_elt *elt;
587
789410e4 588 FOR_EACH_VEC_SAFE_ELT (m_constructors[num], i, elt)
8e97bc2b
JJ
589 elt->value = fold_convert (value_type, elt->value);
590 }
789410e4 591 ctor = build_constructor (array_type, m_constructors[num]);
7156c8ab 592 TREE_CONSTANT (ctor) = true;
5f7ae6b6 593 TREE_STATIC (ctor) = true;
7156c8ab 594
c2255bc4 595 decl = build_decl (loc, VAR_DECL, NULL_TREE, array_type);
7156c8ab
MJ
596 TREE_STATIC (decl) = 1;
597 DECL_INITIAL (decl) = ctor;
598
599 DECL_NAME (decl) = create_tmp_var_name ("CSWTCH");
600 DECL_ARTIFICIAL (decl) = 1;
f8d851c6 601 DECL_IGNORED_P (decl) = 1;
7156c8ab 602 TREE_CONSTANT (decl) = 1;
2e3b4885 603 TREE_READONLY (decl) = 1;
d7438551 604 DECL_IGNORED_P (decl) = 1;
46dbeb40
TV
605 if (offloading_function_p (cfun->decl))
606 DECL_ATTRIBUTES (decl)
607 = tree_cons (get_identifier ("omp declare target"), NULL_TREE,
608 NULL_TREE);
9041d2e6 609 varpool_node::finalize_decl (decl);
7156c8ab
MJ
610
611 fetch = build4 (ARRAY_REF, value_type, decl, tidx, NULL_TREE,
612 NULL_TREE);
8e97bc2b
JJ
613 if (default_type != value_type)
614 {
615 fetch = fold_convert (default_type, fetch);
616 fetch = force_gimple_operand_gsi (&gsi, fetch, true, NULL_TREE,
617 true, GSI_SAME_STMT);
618 }
7156c8ab
MJ
619 load = gimple_build_assign (name, fetch);
620 }
b6e99746 621
726a989a 622 gsi_insert_before (&gsi, load, GSI_SAME_STMT);
7156c8ab 623 update_stmt (load);
789410e4 624 m_arr_ref_last = load;
b6e99746
MJ
625}
626
627/* Builds and initializes static arrays initialized with values gathered from
789410e4 628 the switch statement. Also creates statements that load values from
b6e99746
MJ
629 them. */
630
789410e4
ML
631void
632switch_conversion::build_arrays ()
b6e99746
MJ
633{
634 tree arr_index_type;
83d5977e 635 tree tidx, sub, utype;
355fe088 636 gimple *stmt;
726a989a 637 gimple_stmt_iterator gsi;
538dd0b7 638 gphi_iterator gpi;
b6e99746 639 int i;
789410e4 640 location_t loc = gimple_location (m_switch);
b6e99746 641
789410e4 642 gsi = gsi_for_stmt (m_switch);
04e78aa9 643
edb9b69e 644 /* Make sure we do not generate arithmetics in a subrange. */
789410e4 645 utype = TREE_TYPE (m_index_expr);
edb9b69e
JJ
646 if (TREE_TYPE (utype))
647 utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1);
648 else
649 utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1);
650
789410e4 651 arr_index_type = build_index_type (m_range_size);
b731b390 652 tidx = make_ssa_name (utype);
edb9b69e 653 sub = fold_build2_loc (loc, MINUS_EXPR, utype,
789410e4
ML
654 fold_convert_loc (loc, utype, m_index_expr),
655 fold_convert_loc (loc, utype, m_range_min));
fae1034e 656 sub = force_gimple_operand_gsi (&gsi, sub,
726a989a
RB
657 false, NULL, true, GSI_SAME_STMT);
658 stmt = gimple_build_assign (tidx, sub);
b6e99746 659
726a989a 660 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
7156c8ab 661 update_stmt (stmt);
789410e4 662 m_arr_ref_first = stmt;
b6e99746 663
789410e4 664 for (gpi = gsi_start_phis (m_final_bb), i = 0;
18bfe940
JJ
665 !gsi_end_p (gpi); gsi_next (&gpi))
666 {
667 gphi *phi = gpi.phi ();
668 if (!virtual_operand_p (gimple_phi_result (phi)))
789410e4 669 build_one_array (i++, arr_index_type, phi, tidx);
8dc6a926
JJ
670 else
671 {
672 edge e;
673 edge_iterator ei;
789410e4 674 FOR_EACH_EDGE (e, ei, m_switch_bb->succs)
8dc6a926 675 {
789410e4 676 if (e->dest == m_final_bb)
8dc6a926 677 break;
789410e4
ML
678 if (!m_default_case_nonstandard
679 || e->dest != m_default_bb)
8dc6a926
JJ
680 {
681 e = single_succ_edge (e->dest);
682 break;
683 }
684 }
789410e4
ML
685 gcc_assert (e && e->dest == m_final_bb);
686 m_target_vop = PHI_ARG_DEF_FROM_EDGE (phi, e);
8dc6a926 687 }
18bfe940 688 }
b6e99746
MJ
689}
690
691/* Generates and appropriately inserts loads of default values at the position
789410e4 692 given by GSI. Returns the last inserted statement. */
b6e99746 693
789410e4
ML
694gassign *
695switch_conversion::gen_def_assigns (gimple_stmt_iterator *gsi)
b6e99746
MJ
696{
697 int i;
538dd0b7 698 gassign *assign = NULL;
b6e99746 699
789410e4 700 for (i = 0; i < m_phi_count; i++)
b6e99746 701 {
789410e4
ML
702 tree name = copy_ssa_name (m_target_inbound_names[i]);
703 m_target_outbound_names[i] = name;
704 assign = gimple_build_assign (name, m_default_values[i]);
726a989a 705 gsi_insert_before (gsi, assign, GSI_SAME_STMT);
7156c8ab 706 update_stmt (assign);
b6e99746
MJ
707 }
708 return assign;
709}
710
711/* Deletes the unused bbs and edges that now contain the switch statement and
789410e4
ML
712 its empty branch bbs. BBD is the now dead BB containing
713 the original switch statement, FINAL is the last BB of the converted
714 switch statement (in terms of succession). */
b6e99746 715
789410e4
ML
716void
717switch_conversion::prune_bbs (basic_block bbd, basic_block final,
718 basic_block default_bb)
b6e99746
MJ
719{
720 edge_iterator ei;
721 edge e;
722
723 for (ei = ei_start (bbd->succs); (e = ei_safe_edge (ei)); )
724 {
725 basic_block bb;
726 bb = e->dest;
727 remove_edge (e);
18bfe940 728 if (bb != final && bb != default_bb)
b6e99746
MJ
729 delete_basic_block (bb);
730 }
731 delete_basic_block (bbd);
732}
733
734/* Add values to phi nodes in final_bb for the two new edges. E1F is the edge
735 from the basic block loading values from an array and E2F from the basic
736 block loading default values. BBF is the last switch basic block (see the
737 bbf description in the comment below). */
738
789410e4
ML
739void
740switch_conversion::fix_phi_nodes (edge e1f, edge e2f, basic_block bbf)
b6e99746 741{
538dd0b7 742 gphi_iterator gsi;
b6e99746
MJ
743 int i;
744
726a989a 745 for (gsi = gsi_start_phis (bbf), i = 0;
18bfe940 746 !gsi_end_p (gsi); gsi_next (&gsi))
b6e99746 747 {
538dd0b7 748 gphi *phi = gsi.phi ();
18bfe940
JJ
749 tree inbound, outbound;
750 if (virtual_operand_p (gimple_phi_result (phi)))
789410e4 751 inbound = outbound = m_target_vop;
18bfe940
JJ
752 else
753 {
789410e4
ML
754 inbound = m_target_inbound_names[i];
755 outbound = m_target_outbound_names[i++];
18bfe940
JJ
756 }
757 add_phi_arg (phi, inbound, e1f, UNKNOWN_LOCATION);
789410e4 758 if (!m_default_case_nonstandard)
18bfe940 759 add_phi_arg (phi, outbound, e2f, UNKNOWN_LOCATION);
b6e99746 760 }
b6e99746
MJ
761}
762
763/* Creates a check whether the switch expression value actually falls into the
764 range given by all the cases. If it does not, the temporaries are loaded
789410e4 765 with default values instead. */
b6e99746 766
789410e4
ML
767void
768switch_conversion::gen_inbound_check ()
b6e99746 769{
c2255bc4
AH
770 tree label_decl1 = create_artificial_label (UNKNOWN_LOCATION);
771 tree label_decl2 = create_artificial_label (UNKNOWN_LOCATION);
772 tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION);
538dd0b7 773 glabel *label1, *label2, *label3;
edb9b69e 774 tree utype, tidx;
b6e99746
MJ
775 tree bound;
776
538dd0b7 777 gcond *cond_stmt;
b6e99746 778
18bfe940 779 gassign *last_assign = NULL;
726a989a 780 gimple_stmt_iterator gsi;
b6e99746 781 basic_block bb0, bb1, bb2, bbf, bbd;
18bfe940 782 edge e01 = NULL, e02, e21, e1d, e1f, e2f;
789410e4 783 location_t loc = gimple_location (m_switch);
b6e99746 784
789410e4 785 gcc_assert (m_default_values);
6ab1ab14 786
789410e4 787 bb0 = gimple_bb (m_switch);
b6e99746 788
789410e4 789 tidx = gimple_assign_lhs (m_arr_ref_first);
edb9b69e 790 utype = TREE_TYPE (tidx);
145544ab 791
b6e99746 792 /* (end of) block 0 */
789410e4 793 gsi = gsi_for_stmt (m_arr_ref_first);
edb9b69e 794 gsi_next (&gsi);
b6e99746 795
789410e4 796 bound = fold_convert_loc (loc, utype, m_range_size);
edb9b69e 797 cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE);
726a989a 798 gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT);
7156c8ab 799 update_stmt (cond_stmt);
b6e99746
MJ
800
801 /* block 2 */
789410e4 802 if (!m_default_case_nonstandard)
18bfe940
JJ
803 {
804 label2 = gimple_build_label (label_decl2);
805 gsi_insert_before (&gsi, label2, GSI_SAME_STMT);
789410e4 806 last_assign = gen_def_assigns (&gsi);
18bfe940 807 }
b6e99746
MJ
808
809 /* block 1 */
726a989a
RB
810 label1 = gimple_build_label (label_decl1);
811 gsi_insert_before (&gsi, label1, GSI_SAME_STMT);
b6e99746
MJ
812
813 /* block F */
789410e4 814 gsi = gsi_start_bb (m_final_bb);
726a989a
RB
815 label3 = gimple_build_label (label_decl3);
816 gsi_insert_before (&gsi, label3, GSI_SAME_STMT);
b6e99746
MJ
817
818 /* cfg fix */
726a989a 819 e02 = split_block (bb0, cond_stmt);
b6e99746
MJ
820 bb2 = e02->dest;
821
789410e4 822 if (m_default_case_nonstandard)
18bfe940
JJ
823 {
824 bb1 = bb2;
789410e4 825 bb2 = m_default_bb;
18bfe940
JJ
826 e01 = e02;
827 e01->flags = EDGE_TRUE_VALUE;
828 e02 = make_edge (bb0, bb2, EDGE_FALSE_VALUE);
829 edge e_default = find_edge (bb1, bb2);
830 for (gphi_iterator gsi = gsi_start_phis (bb2);
831 !gsi_end_p (gsi); gsi_next (&gsi))
832 {
833 gphi *phi = gsi.phi ();
834 tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_default);
835 add_phi_arg (phi, arg, e02,
836 gimple_phi_arg_location_from_edge (phi, e_default));
837 }
838 /* Partially fix the dominator tree, if it is available. */
839 if (dom_info_available_p (CDI_DOMINATORS))
840 redirect_immediate_dominators (CDI_DOMINATORS, bb1, bb0);
841 }
842 else
843 {
844 e21 = split_block (bb2, last_assign);
845 bb1 = e21->dest;
846 remove_edge (e21);
847 }
b6e99746 848
789410e4 849 e1d = split_block (bb1, m_arr_ref_last);
b6e99746
MJ
850 bbd = e1d->dest;
851 remove_edge (e1d);
852
789410e4
ML
853 /* Flags and profiles of the edge for in-range values. */
854 if (!m_default_case_nonstandard)
18bfe940 855 e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE);
789410e4 856 e01->probability = m_default_prob.invert ();
b6e99746 857
789410e4 858 /* Flags and profiles of the edge taking care of out-of-range values. */
b6e99746
MJ
859 e02->flags &= ~EDGE_FALLTHRU;
860 e02->flags |= EDGE_FALSE_VALUE;
789410e4 861 e02->probability = m_default_prob;
b6e99746 862
789410e4 863 bbf = m_final_bb;
b6e99746
MJ
864
865 e1f = make_edge (bb1, bbf, EDGE_FALLTHRU);
357067f2 866 e1f->probability = profile_probability::always ();
b6e99746 867
789410e4 868 if (m_default_case_nonstandard)
18bfe940
JJ
869 e2f = NULL;
870 else
871 {
872 e2f = make_edge (bb2, bbf, EDGE_FALLTHRU);
357067f2 873 e2f->probability = profile_probability::always ();
18bfe940 874 }
b6e99746
MJ
875
876 /* frequencies of the new BBs */
e7a74006
JH
877 bb1->count = e01->count ();
878 bb2->count = e02->count ();
789410e4 879 if (!m_default_case_nonstandard)
e7a74006 880 bbf->count = e1f->count () + e2f->count ();
b6e99746 881
6ab1ab14 882 /* Tidy blocks that have become unreachable. */
789410e4
ML
883 prune_bbs (bbd, m_final_bb,
884 m_default_case_nonstandard ? m_default_bb : NULL);
b6e99746 885
6ab1ab14 886 /* Fixup the PHI nodes in bbF. */
789410e4 887 fix_phi_nodes (e1f, e2f, bbf);
b6e99746 888
6ab1ab14
SB
889 /* Fix the dominator tree, if it is available. */
890 if (dom_info_available_p (CDI_DOMINATORS))
891 {
9771b263 892 vec<basic_block> bbs_to_fix_dom;
6ab1ab14
SB
893
894 set_immediate_dominator (CDI_DOMINATORS, bb1, bb0);
789410e4 895 if (!m_default_case_nonstandard)
18bfe940 896 set_immediate_dominator (CDI_DOMINATORS, bb2, bb0);
531b10fc 897 if (! get_immediate_dominator (CDI_DOMINATORS, bbf))
6ab1ab14
SB
898 /* If bbD was the immediate dominator ... */
899 set_immediate_dominator (CDI_DOMINATORS, bbf, bb0);
900
18bfe940 901 bbs_to_fix_dom.create (3 + (bb2 != bbf));
9771b263
DN
902 bbs_to_fix_dom.quick_push (bb0);
903 bbs_to_fix_dom.quick_push (bb1);
18bfe940
JJ
904 if (bb2 != bbf)
905 bbs_to_fix_dom.quick_push (bb2);
9771b263 906 bbs_to_fix_dom.quick_push (bbf);
6ab1ab14
SB
907
908 iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true);
9771b263 909 bbs_to_fix_dom.release ();
6ab1ab14 910 }
b6e99746
MJ
911}
912
789410e4
ML
913/* The following function is invoked on every switch statement (the current
914 one is given in SWTCH) and runs the individual phases of switch
915 conversion on it one after another until one fails or the conversion
916 is completed. On success, NULL is in m_reason, otherwise points
917 to a string with the reason why the conversion failed. */
b6e99746 918
789410e4
ML
919void
920switch_conversion::expand (gswitch *swtch)
b6e99746 921{
238065a7
SB
922 /* Group case labels so that we get the right results from the heuristics
923 that decide on the code generation approach for this switch. */
789410e4 924 m_cfg_altered |= group_case_labels_stmt (swtch);
238065a7
SB
925
926 /* If this switch is now a degenerate case with only a default label,
789410e4 927 there is nothing left for us to do. */
238065a7 928 if (gimple_switch_num_labels (swtch) < 2)
789410e4
ML
929 {
930 m_reason = "switch is a degenerate case";
931 return;
932 }
886cd84f 933
789410e4 934 collect (swtch);
886cd84f
SB
935
936 /* No error markers should reach here (they should be filtered out
937 during gimplification). */
789410e4 938 gcc_checking_assert (TREE_TYPE (m_index_expr) != error_mark_node);
886cd84f 939
531b10fc 940 /* A switch on a constant should have been optimized in tree-cfg-cleanup. */
789410e4 941 gcc_checking_assert (!TREE_CONSTANT (m_index_expr));
886cd84f 942
dc223ad4
ML
943 /* Prefer bit test if possible. */
944 if (tree_fits_uhwi_p (m_range_size)
945 && bit_test_cluster::can_be_handled (tree_to_uhwi (m_range_size), m_uniq)
946 && bit_test_cluster::is_beneficial (m_count, m_uniq))
947 {
948 m_reason = "expanding as bit test is preferable";
949 return;
950 }
951
952 if (m_uniq <= 2)
953 {
954 /* This will be expanded as a decision tree . */
955 m_reason = "expanding as jumps is preferable";
956 return;
957 }
958
789410e4
ML
959 /* If there is no common successor, we cannot do the transformation. */
960 if (!m_final_bb)
886cd84f 961 {
789410e4
ML
962 m_reason = "no common successor to all case label target blocks found";
963 return;
886cd84f 964 }
b6e99746
MJ
965
966 /* Check the case label values are within reasonable range: */
789410e4 967 if (!check_range ())
fade902a 968 {
789410e4
ML
969 gcc_assert (m_reason);
970 return;
fade902a 971 }
b6e99746
MJ
972
973 /* For all the cases, see whether they are empty, the assignments they
974 represent constant and so on... */
789410e4 975 if (!check_all_empty_except_final ())
8e97bc2b 976 {
789410e4
ML
977 gcc_assert (m_reason);
978 return;
8e97bc2b 979 }
789410e4 980 if (!check_final_bb ())
fade902a 981 {
789410e4
ML
982 gcc_assert (m_reason);
983 return;
fade902a 984 }
b6e99746
MJ
985
986 /* At this point all checks have passed and we can proceed with the
987 transformation. */
988
789410e4
ML
989 create_temp_arrays ();
990 gather_default_values (m_default_case_nonstandard
18bfe940 991 ? gimple_switch_label (swtch, 1)
789410e4
ML
992 : gimple_switch_default_label (swtch));
993 build_constructors ();
b6e99746 994
789410e4
ML
995 build_arrays (); /* Build the static arrays and assignments. */
996 gen_inbound_check (); /* Build the bounds check. */
b6e99746 997
789410e4
ML
998 m_cfg_altered = true;
999}
1000
1001/* Destructor. */
1002
1003switch_conversion::~switch_conversion ()
1004{
1005 XDELETEVEC (m_constructors);
1006 XDELETEVEC (m_default_values);
b6e99746
MJ
1007}
1008
dc223ad4 1009/* Constructor. */
be55bfe6 1010
dc223ad4
ML
1011group_cluster::group_cluster (vec<cluster *> &clusters,
1012 unsigned start, unsigned end)
be55bfe6 1013{
dc223ad4
ML
1014 gcc_checking_assert (end - start + 1 >= 1);
1015 m_prob = profile_probability::never ();
1016 m_cases.create (end - start + 1);
1017 for (unsigned i = start; i <= end; i++)
1018 {
1019 m_cases.quick_push (static_cast<simple_cluster *> (clusters[i]));
1020 m_prob += clusters[i]->m_prob;
1021 }
1022 m_subtree_prob = m_prob;
1023}
be55bfe6 1024
dc223ad4
ML
1025/* Destructor. */
1026
1027group_cluster::~group_cluster ()
be55bfe6 1028{
dc223ad4
ML
1029 for (unsigned i = 0; i < m_cases.length (); i++)
1030 delete m_cases[i];
be55bfe6 1031
dc223ad4
ML
1032 m_cases.release ();
1033}
be55bfe6 1034
dc223ad4 1035/* Dump content of a cluster. */
be55bfe6 1036
dc223ad4
ML
1037void
1038group_cluster::dump (FILE *f, bool details)
b6e99746 1039{
dc223ad4
ML
1040 unsigned total_values = 0;
1041 for (unsigned i = 0; i < m_cases.length (); i++)
1042 total_values += m_cases[i]->get_range (m_cases[i]->get_low (),
1043 m_cases[i]->get_high ());
726a989a 1044
dc223ad4
ML
1045 unsigned comparison_count = 0;
1046 for (unsigned i = 0; i < m_cases.length (); i++)
1047 {
1048 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]);
1049 comparison_count += sc->m_range_p ? 2 : 1;
1050 }
b6e99746 1051
dc223ad4
ML
1052 unsigned HOST_WIDE_INT range = get_range (get_low (), get_high ());
1053 fprintf (f, "%s", get_type () == JUMP_TABLE ? "JT" : "BT");
531b10fc 1054
dc223ad4
ML
1055 if (details)
1056 fprintf (f, "(values:%d comparisons:%d range:" HOST_WIDE_INT_PRINT_DEC
1057 " density: %.2f%%)", total_values, comparison_count, range,
1058 100.0f * comparison_count / range);
b6e99746 1059
dc223ad4
ML
1060 fprintf (f, ":");
1061 PRINT_CASE (f, get_low ());
1062 fprintf (f, "-");
1063 PRINT_CASE (f, get_high ());
1064 fprintf (f, " ");
b6e99746
MJ
1065}
1066
dc223ad4 1067/* Emit GIMPLE code to handle the cluster. */
27a4cd48 1068
dc223ad4
ML
1069void
1070jump_table_cluster::emit (tree index_expr, tree,
1071 tree default_label_expr, basic_block default_bb)
27a4cd48 1072{
dc223ad4
ML
1073 /* For jump table we just emit a new gswitch statement that will
1074 be latter lowered to jump table. */
1075 auto_vec <tree> labels;
1076 labels.create (m_cases.length ());
1077
1078 make_edge (m_case_bb, default_bb, 0);
1079 for (unsigned i = 0; i < m_cases.length (); i++)
1080 {
1081 labels.quick_push (unshare_expr (m_cases[i]->m_case_label_expr));
1082 make_edge (m_case_bb, m_cases[i]->m_case_bb, 0);
1083 }
1084
1085 gswitch *s = gimple_build_switch (index_expr,
1086 unshare_expr (default_label_expr), labels);
1087 gimple_stmt_iterator gsi = gsi_start_bb (m_case_bb);
1088 gsi_insert_after (&gsi, s, GSI_NEW_STMT);
27a4cd48 1089}
9dc3d6a9 1090
2f928c1b
ML
1091/* Find jump tables of given CLUSTERS, where all members of the vector
1092 are of type simple_cluster. New clusters are returned. */
1093
1094vec<cluster *>
1095jump_table_cluster::find_jump_tables (vec<cluster *> &clusters)
1096{
1097 unsigned l = clusters.length ();
1098 auto_vec<min_cluster_item> min;
1099 min.reserve (l + 1);
1100
1101 min.quick_push (min_cluster_item (0, 0, 0));
1102
1103 for (unsigned i = 1; i <= l; i++)
1104 {
1105 /* Set minimal # of clusters with i-th item to infinite. */
1106 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX));
1107
1108 for (unsigned j = 0; j < i; j++)
1109 {
1110 unsigned HOST_WIDE_INT s = min[j].m_non_jt_cases;
1111 if (i - j < case_values_threshold ())
1112 s += i - j;
1113
1114 /* Prefer clusters with smaller number of numbers covered. */
1115 if ((min[j].m_count + 1 < min[i].m_count
1116 || (min[j].m_count + 1 == min[i].m_count
1117 && s < min[i].m_non_jt_cases))
1118 && can_be_handled (clusters, j, i - 1))
1119 min[i] = min_cluster_item (min[j].m_count + 1, j, s);
1120 }
1121 }
1122
1123 /* No result. */
1124 if (min[l].m_count == INT_MAX)
1125 return clusters.copy ();
1126
1127 vec<cluster *> output;
1128 output.create (4);
1129
1130 /* Find and build the clusters. */
1131 for (int end = l;;)
1132 {
1133 int start = min[end].m_start;
1134
1135 /* Do not allow clusters with small number of cases. */
1136 if (is_beneficial (clusters, start, end - 1))
1137 output.safe_push (new jump_table_cluster (clusters, start, end - 1));
1138 else
1139 for (int i = end - 1; i >= start; i--)
1140 output.safe_push (clusters[i]);
1141
1142 end = start;
1143
1144 if (start <= 0)
1145 break;
1146 }
1147
1148 output.reverse ();
1149 return output;
1150}
1151
dc223ad4
ML
1152/* Return true when cluster starting at START and ending at END (inclusive)
1153 can build a jump-table. */
1154
1155bool
1156jump_table_cluster::can_be_handled (const vec<cluster *> &clusters,
1157 unsigned start, unsigned end)
9dc3d6a9 1158{
dc223ad4
ML
1159 /* If the switch is relatively small such that the cost of one
1160 indirect jump on the target are higher than the cost of a
1161 decision tree, go with the decision tree.
9dc3d6a9 1162
dc223ad4
ML
1163 If range of values is much bigger than number of values,
1164 or if it is too large to represent in a HOST_WIDE_INT,
1165 make a sequence of conditional branches instead of a dispatch.
9dc3d6a9 1166
dc223ad4 1167 The definition of "much bigger" depends on whether we are
de840bde 1168 optimizing for size or for speed. */
dc223ad4
ML
1169 if (!flag_jump_tables)
1170 return false;
9dc3d6a9 1171
de840bde 1172 unsigned HOST_WIDE_INT max_ratio = optimize_insn_for_size_p () ? 3 : 8;
9dc3d6a9 1173
dc223ad4
ML
1174 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (),
1175 clusters[end]->get_high ());
1176 /* Check overflow. */
1177 if (range == 0)
1178 return false;
9dc3d6a9 1179
dc223ad4
ML
1180 unsigned HOST_WIDE_INT comparison_count = 0;
1181 for (unsigned i = start; i <= end; i++)
1182 {
1183 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]);
1184 comparison_count += sc->m_range_p ? 2 : 1;
1185 }
9dc3d6a9 1186
dc223ad4 1187 return range <= max_ratio * comparison_count;
9dc3d6a9
ML
1188}
1189
dc223ad4
ML
1190/* Return true if cluster starting at START and ending at END (inclusive)
1191 is profitable transformation. */
9dc3d6a9 1192
dc223ad4
ML
1193bool
1194jump_table_cluster::is_beneficial (const vec<cluster *> &,
1195 unsigned start, unsigned end)
9dc3d6a9 1196{
dc223ad4 1197 return end - start + 1 >= case_values_threshold ();
9dc3d6a9
ML
1198}
1199
2f928c1b
ML
1200/* Find bit tests of given CLUSTERS, where all members of the vector
1201 are of type simple_cluster. New clusters are returned. */
1202
1203vec<cluster *>
1204bit_test_cluster::find_bit_tests (vec<cluster *> &clusters)
1205{
1206 vec<cluster *> output;
1207 output.create (4);
1208
1209 unsigned l = clusters.length ();
1210 auto_vec<min_cluster_item> min;
1211 min.reserve (l + 1);
1212
1213 min.quick_push (min_cluster_item (0, 0, 0));
1214
1215 for (unsigned i = 1; i <= l; i++)
1216 {
1217 /* Set minimal # of clusters with i-th item to infinite. */
1218 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX));
1219
1220 for (unsigned j = 0; j < i; j++)
1221 {
1222 if (min[j].m_count + 1 < min[i].m_count
1223 && can_be_handled (clusters, j, i - 1))
1224 min[i] = min_cluster_item (min[j].m_count + 1, j, INT_MAX);
1225 }
1226 }
1227
1228 /* No result. */
1229 if (min[l].m_count == INT_MAX)
1230 return clusters.copy ();
1231
1232 /* Find and build the clusters. */
1233 for (int end = l;;)
1234 {
1235 int start = min[end].m_start;
1236
1237 if (is_beneficial (clusters, start, end - 1))
1238 output.safe_push (new bit_test_cluster (clusters, start, end - 1));
1239 else
1240 for (int i = end - 1; i >= start; i--)
1241 output.safe_push (clusters[i]);
1242
1243 end = start;
1244
1245 if (start <= 0)
1246 break;
1247 }
1248
1249 output.reverse ();
1250 return output;
1251}
1252
dc223ad4
ML
1253/* Return true when RANGE of case values with UNIQ labels
1254 can build a bit test. */
9dc3d6a9 1255
dc223ad4
ML
1256bool
1257bit_test_cluster::can_be_handled (unsigned HOST_WIDE_INT range,
1258 unsigned int uniq)
9dc3d6a9 1259{
dc223ad4
ML
1260 /* Check overflow. */
1261 if (range == 0)
1262 return 0;
1263
1264 if (range >= GET_MODE_BITSIZE (word_mode))
1265 return false;
1266
1267 return uniq <= 3;
1268}
1269
1270/* Return true when cluster starting at START and ending at END (inclusive)
1271 can build a bit test. */
1272
1273bool
1274bit_test_cluster::can_be_handled (const vec<cluster *> &clusters,
1275 unsigned start, unsigned end)
1276{
1277 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (),
1278 clusters[end]->get_high ());
1279 auto_bitmap dest_bbs;
1280
1281 for (unsigned i = start; i <= end; i++)
9dc3d6a9 1282 {
dc223ad4
ML
1283 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]);
1284 bitmap_set_bit (dest_bbs, sc->m_case_bb->index);
9dc3d6a9 1285 }
9dc3d6a9 1286
dc223ad4
ML
1287 return can_be_handled (range, bitmap_count_bits (dest_bbs));
1288}
9dc3d6a9 1289
dc223ad4
ML
1290/* Return true when COUNT of cases of UNIQ labels is beneficial for bit test
1291 transformation. */
9dc3d6a9 1292
dc223ad4
ML
1293bool
1294bit_test_cluster::is_beneficial (unsigned count, unsigned uniq)
9dc3d6a9 1295{
dc223ad4
ML
1296 return (((uniq == 1 && count >= 3)
1297 || (uniq == 2 && count >= 5)
1298 || (uniq == 3 && count >= 6)));
9dc3d6a9
ML
1299}
1300
dc223ad4
ML
1301/* Return true if cluster starting at START and ending at END (inclusive)
1302 is profitable transformation. */
9dc3d6a9 1303
dc223ad4
ML
1304bool
1305bit_test_cluster::is_beneficial (const vec<cluster *> &clusters,
1306 unsigned start, unsigned end)
9dc3d6a9 1307{
dc223ad4 1308 auto_bitmap dest_bbs;
9dc3d6a9 1309
dc223ad4 1310 for (unsigned i = start; i <= end; i++)
9dc3d6a9 1311 {
dc223ad4
ML
1312 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]);
1313 bitmap_set_bit (dest_bbs, sc->m_case_bb->index);
9dc3d6a9 1314 }
9dc3d6a9 1315
dc223ad4
ML
1316 unsigned uniq = bitmap_count_bits (dest_bbs);
1317 unsigned count = end - start + 1;
1318 return is_beneficial (count, uniq);
9dc3d6a9
ML
1319}
1320
dc223ad4
ML
1321/* Comparison function for qsort to order bit tests by decreasing
1322 probability of execution. */
9dc3d6a9 1323
dc223ad4
ML
1324int
1325case_bit_test::cmp (const void *p1, const void *p2)
1326{
1327 const struct case_bit_test *const d1 = (const struct case_bit_test *) p1;
1328 const struct case_bit_test *const d2 = (const struct case_bit_test *) p2;
1329
1330 if (d2->bits != d1->bits)
1331 return d2->bits - d1->bits;
1332
1333 /* Stabilize the sort. */
1334 return (LABEL_DECL_UID (CASE_LABEL (d2->label))
1335 - LABEL_DECL_UID (CASE_LABEL (d1->label)));
1336}
1337
1338/* Expand a switch statement by a short sequence of bit-wise
1339 comparisons. "switch(x)" is effectively converted into
1340 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are
1341 integer constants.
1342
1343 INDEX_EXPR is the value being switched on.
1344
1345 MINVAL is the lowest case value of in the case nodes,
1346 and RANGE is highest value minus MINVAL. MINVAL and RANGE
1347 are not guaranteed to be of the same type as INDEX_EXPR
1348 (the gimplifier doesn't change the type of case label values,
1349 and MINVAL and RANGE are derived from those values).
1350 MAXVAL is MINVAL + RANGE.
9dc3d6a9 1351
dc223ad4
ML
1352 There *MUST* be max_case_bit_tests or less unique case
1353 node targets. */
1354
1355void
1356bit_test_cluster::emit (tree index_expr, tree index_type,
1357 tree, basic_block default_bb)
9dc3d6a9 1358{
dc223ad4
ML
1359 struct case_bit_test test[m_max_case_bit_tests] = { {} };
1360 unsigned int i, j, k;
1361 unsigned int count;
9dc3d6a9 1362
dc223ad4 1363 tree unsigned_index_type = unsigned_type_for (index_type);
9dc3d6a9 1364
dc223ad4
ML
1365 gimple_stmt_iterator gsi;
1366 gassign *shift_stmt;
9dc3d6a9 1367
dc223ad4
ML
1368 tree idx, tmp, csui;
1369 tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1);
1370 tree word_mode_zero = fold_convert (word_type_node, integer_zero_node);
1371 tree word_mode_one = fold_convert (word_type_node, integer_one_node);
1372 int prec = TYPE_PRECISION (word_type_node);
1373 wide_int wone = wi::one (prec);
9dc3d6a9 1374
dc223ad4
ML
1375 tree minval = get_low ();
1376 tree maxval = get_high ();
1377 tree range = int_const_binop (MINUS_EXPR, maxval, minval);
9dc3d6a9 1378
dc223ad4
ML
1379 /* Go through all case labels, and collect the case labels, profile
1380 counts, and other information we need to build the branch tests. */
1381 count = 0;
1382 for (i = 0; i < m_cases.length (); i++)
1383 {
1384 unsigned int lo, hi;
1385 simple_cluster *n = static_cast<simple_cluster *> (m_cases[i]);
1386 for (k = 0; k < count; k++)
1387 if (n->m_case_bb == test[k].target_bb)
1388 break;
1389
1390 if (k == count)
9dc3d6a9 1391 {
dc223ad4
ML
1392 gcc_checking_assert (count < m_max_case_bit_tests);
1393 test[k].mask = wi::zero (prec);
1394 test[k].target_bb = n->m_case_bb;
1395 test[k].label = n->m_case_label_expr;
1396 test[k].bits = 1;
1397 count++;
1398 }
1399 else
1400 test[k].bits++;
9dc3d6a9 1401
dc223ad4
ML
1402 lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_low (), minval));
1403 if (n->get_high () == NULL_TREE)
1404 hi = lo;
1405 else
1406 hi = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_high (),
1407 minval));
9dc3d6a9 1408
dc223ad4
ML
1409 for (j = lo; j <= hi; j++)
1410 test[k].mask |= wi::lshift (wone, j);
1411 }
1412
1413 qsort (test, count, sizeof (*test), case_bit_test::cmp);
1414
1415 /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of
1416 the minval subtractions, but it might make the mask constants more
1417 expensive. So, compare the costs. */
1418 if (compare_tree_int (minval, 0) > 0
1419 && compare_tree_int (maxval, GET_MODE_BITSIZE (word_mode)) < 0)
1420 {
1421 int cost_diff;
1422 HOST_WIDE_INT m = tree_to_uhwi (minval);
1423 rtx reg = gen_raw_REG (word_mode, 10000);
1424 bool speed_p = optimize_insn_for_speed_p ();
1425 cost_diff = set_rtx_cost (gen_rtx_PLUS (word_mode, reg,
1426 GEN_INT (-m)), speed_p);
1427 for (i = 0; i < count; i++)
1428 {
1429 rtx r = immed_wide_int_const (test[i].mask, word_mode);
1430 cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r),
1431 word_mode, speed_p);
1432 r = immed_wide_int_const (wi::lshift (test[i].mask, m), word_mode);
1433 cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r),
1434 word_mode, speed_p);
9dc3d6a9 1435 }
dc223ad4 1436 if (cost_diff > 0)
9dc3d6a9 1437 {
dc223ad4
ML
1438 for (i = 0; i < count; i++)
1439 test[i].mask = wi::lshift (test[i].mask, m);
1440 minval = build_zero_cst (TREE_TYPE (minval));
1441 range = maxval;
9dc3d6a9
ML
1442 }
1443 }
9dc3d6a9 1444
dc223ad4
ML
1445 /* Now build the test-and-branch code. */
1446
1447 gsi = gsi_last_bb (m_case_bb);
1448
1449 /* idx = (unsigned)x - minval. */
1450 idx = fold_convert (unsigned_index_type, index_expr);
1451 idx = fold_build2 (MINUS_EXPR, unsigned_index_type, idx,
1452 fold_convert (unsigned_index_type, minval));
1453 idx = force_gimple_operand_gsi (&gsi, idx,
1454 /*simple=*/true, NULL_TREE,
1455 /*before=*/true, GSI_SAME_STMT);
1456
1457 /* if (idx > range) goto default */
1458 range = force_gimple_operand_gsi (&gsi,
1459 fold_convert (unsigned_index_type, range),
1460 /*simple=*/true, NULL_TREE,
1461 /*before=*/true, GSI_SAME_STMT);
1462 tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range);
1463 basic_block new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, default_bb);
1464 gsi = gsi_last_bb (new_bb);
1465
1466 /* csui = (1 << (word_mode) idx) */
1467 csui = make_ssa_name (word_type_node);
1468 tmp = fold_build2 (LSHIFT_EXPR, word_type_node, word_mode_one,
1469 fold_convert (word_type_node, idx));
1470 tmp = force_gimple_operand_gsi (&gsi, tmp,
1471 /*simple=*/false, NULL_TREE,
1472 /*before=*/true, GSI_SAME_STMT);
1473 shift_stmt = gimple_build_assign (csui, tmp);
1474 gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT);
1475 update_stmt (shift_stmt);
1476
1477 /* for each unique set of cases:
1478 if (const & csui) goto target */
1479 for (k = 0; k < count; k++)
1480 {
1481 tmp = wide_int_to_tree (word_type_node, test[k].mask);
1482 tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp);
1483 tmp = force_gimple_operand_gsi (&gsi, tmp,
1484 /*simple=*/true, NULL_TREE,
1485 /*before=*/true, GSI_SAME_STMT);
1486 tmp = fold_build2 (NE_EXPR, boolean_type_node, tmp, word_mode_zero);
1487 new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, test[k].target_bb);
1488 gsi = gsi_last_bb (new_bb);
1489 }
9dc3d6a9 1490
dc223ad4
ML
1491 /* We should have removed all edges now. */
1492 gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0);
9dc3d6a9 1493
dc223ad4
ML
1494 /* If nothing matched, go to the default label. */
1495 make_edge (gsi_bb (gsi), default_bb, EDGE_FALLTHRU);
1496}
9dc3d6a9 1497
dc223ad4
ML
1498/* Split the basic block at the statement pointed to by GSIP, and insert
1499 a branch to the target basic block of E_TRUE conditional on tree
1500 expression COND.
9dc3d6a9 1501
dc223ad4
ML
1502 It is assumed that there is already an edge from the to-be-split
1503 basic block to E_TRUE->dest block. This edge is removed, and the
1504 profile information on the edge is re-used for the new conditional
1505 jump.
9dc3d6a9 1506
dc223ad4
ML
1507 The CFG is updated. The dominator tree will not be valid after
1508 this transformation, but the immediate dominators are updated if
1509 UPDATE_DOMINATORS is true.
9dc3d6a9 1510
dc223ad4 1511 Returns the newly created basic block. */
9dc3d6a9 1512
dc223ad4
ML
1513basic_block
1514bit_test_cluster::hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip,
1515 tree cond, basic_block case_bb)
9dc3d6a9 1516{
dc223ad4
ML
1517 tree tmp;
1518 gcond *cond_stmt;
1519 edge e_false;
1520 basic_block new_bb, split_bb = gsi_bb (*gsip);
9dc3d6a9 1521
dc223ad4
ML
1522 edge e_true = make_edge (split_bb, case_bb, EDGE_TRUE_VALUE);
1523 gcc_assert (e_true->src == split_bb);
9dc3d6a9 1524
dc223ad4
ML
1525 tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL,
1526 /*before=*/true, GSI_SAME_STMT);
1527 cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE);
1528 gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT);
9dc3d6a9 1529
dc223ad4
ML
1530 e_false = split_block (split_bb, cond_stmt);
1531 new_bb = e_false->dest;
1532 redirect_edge_pred (e_true, split_bb);
9dc3d6a9 1533
dc223ad4
ML
1534 e_false->flags &= ~EDGE_FALLTHRU;
1535 e_false->flags |= EDGE_FALSE_VALUE;
1536 e_false->probability = e_true->probability.invert ();
1537 new_bb->count = e_false->count ();
1538
1539 return new_bb;
9dc3d6a9
ML
1540}
1541
dc223ad4
ML
1542/* Compute the number of case labels that correspond to each outgoing edge of
1543 switch statement. Record this information in the aux field of the edge. */
9dc3d6a9 1544
dc223ad4
ML
1545void
1546switch_decision_tree::compute_cases_per_edge ()
1547{
1548 basic_block bb = gimple_bb (m_switch);
1549 reset_out_edges_aux ();
1550 int ncases = gimple_switch_num_labels (m_switch);
1551 for (int i = ncases - 1; i >= 1; --i)
1552 {
1553 tree elt = gimple_switch_label (m_switch, i);
1554 tree lab = CASE_LABEL (elt);
1555 basic_block case_bb = label_to_block_fn (cfun, lab);
1556 edge case_edge = find_edge (bb, case_bb);
1557 case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + 1);
1558 }
1559}
1560
1561/* Analyze switch statement and return true when the statement is expanded
1562 as decision tree. */
9dc3d6a9 1563
dc223ad4
ML
1564bool
1565switch_decision_tree::analyze_switch_statement ()
9dc3d6a9 1566{
dc223ad4
ML
1567 unsigned l = gimple_switch_num_labels (m_switch);
1568 basic_block bb = gimple_bb (m_switch);
1569 auto_vec<cluster *> clusters;
1570 clusters.create (l - 1);
1571
1572 tree default_label = CASE_LABEL (gimple_switch_default_label (m_switch));
1573 basic_block default_bb = label_to_block_fn (cfun, default_label);
1574 m_case_bbs.reserve (l);
1575 m_case_bbs.quick_push (default_bb);
1576
1577 compute_cases_per_edge ();
1578
1579 for (unsigned i = 1; i < l; i++)
1580 {
1581 tree elt = gimple_switch_label (m_switch, i);
1582 tree lab = CASE_LABEL (elt);
1583 basic_block case_bb = label_to_block_fn (cfun, lab);
1584 edge case_edge = find_edge (bb, case_bb);
1585 tree low = CASE_LOW (elt);
1586 tree high = CASE_HIGH (elt);
1587
1588 profile_probability p
1589 = case_edge->probability.apply_scale (1, (intptr_t) (case_edge->aux));
1590 clusters.quick_push (new simple_cluster (low, high, elt, case_bb, p));
1591 m_case_bbs.quick_push (case_bb);
1592 }
1593
1594 reset_out_edges_aux ();
1595
2f928c1b
ML
1596 /* Find jump table clusters. */
1597 vec<cluster *> output = jump_table_cluster::find_jump_tables (clusters);
1598
1599 /* Find jump table clusters. */
1600 vec<cluster *> output2;
1601 auto_vec<cluster *> tmp;
1602 output2.create (1);
1603 tmp.create (1);
1604
1605 for (unsigned i = 0; i < output.length (); i++)
1606 {
1607 cluster *c = output[i];
1608 if (c->get_type () != SIMPLE_CASE)
1609 {
1610 if (!tmp.is_empty ())
1611 {
1612 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp);
1613 output2.safe_splice (n);
1614 n.release ();
1615 tmp.truncate (0);
1616 }
1617 output2.safe_push (c);
1618 }
1619 else
1620 tmp.safe_push (c);
1621 }
1622
1623 /* We still can have a temporary vector to test. */
1624 if (!tmp.is_empty ())
1625 {
1626 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp);
1627 output2.safe_splice (n);
1628 n.release ();
1629 }
9dc3d6a9
ML
1630
1631 if (dump_file)
9dc3d6a9 1632 {
dc223ad4 1633 fprintf (dump_file, ";; GIMPLE switch case clusters: ");
2f928c1b
ML
1634 for (unsigned i = 0; i < output2.length (); i++)
1635 output2[i]->dump (dump_file, dump_flags & TDF_DETAILS);
dc223ad4
ML
1636 fprintf (dump_file, "\n");
1637 }
1638
2f928c1b 1639 output.release ();
dc223ad4 1640
2f928c1b
ML
1641 bool expanded = try_switch_expansion (output2);
1642
1643 for (unsigned i = 0; i < output2.length (); i++)
1644 delete output2[i];
1645
1646 output2.release ();
dc223ad4
ML
1647
1648 return expanded;
1649}
1650
1651/* Attempt to expand CLUSTERS as a decision tree. Return true when
1652 expanded. */
1653
1654bool
1655switch_decision_tree::try_switch_expansion (vec<cluster *> &clusters)
1656{
1657 tree index_expr = gimple_switch_index (m_switch);
1658 tree index_type = TREE_TYPE (index_expr);
1659 basic_block bb = gimple_bb (m_switch);
1660
1661 if (gimple_switch_num_labels (m_switch) == 1)
1662 return false;
1663
1664 /* Find the default case target label. */
1665 tree default_label_expr = CASE_LABEL (gimple_switch_default_label (m_switch));
1666 m_default_bb = label_to_block_fn (cfun, default_label_expr);
1667 edge default_edge = find_edge (bb, m_default_bb);
1668
1669 /* Do the insertion of a case label into m_case_list. The labels are
1670 fed to us in descending order from the sorted vector of case labels used
1671 in the tree part of the middle end. So the list we construct is
1672 sorted in ascending order. */
1673
1674 for (int i = clusters.length () - 1; i >= 0; i--)
1675 {
1676 case_tree_node *r = m_case_list;
1677 m_case_list = m_case_node_pool.allocate ();
1678 m_case_list->m_right = r;
1679 m_case_list->m_c = clusters[i];
9dc3d6a9
ML
1680 }
1681
dc223ad4
ML
1682 record_phi_operand_mapping ();
1683
1684 /* Split basic block that contains the gswitch statement. */
9dc3d6a9
ML
1685 gimple_stmt_iterator gsi = gsi_last_bb (bb);
1686 edge e;
1687 if (gsi_end_p (gsi))
1688 e = split_block_after_labels (bb);
1689 else
1690 {
1691 gsi_prev (&gsi);
1692 e = split_block (bb, gsi_stmt (gsi));
1693 }
1694 bb = split_edge (e);
1695
dc223ad4
ML
1696 /* Create new basic blocks for non-case clusters where specific expansion
1697 needs to happen. */
1698 for (unsigned i = 0; i < clusters.length (); i++)
1699 if (clusters[i]->get_type () != SIMPLE_CASE)
1700 {
1701 clusters[i]->m_case_bb = create_empty_bb (bb);
1702 clusters[i]->m_case_bb->loop_father = bb->loop_father;
1703 }
9dc3d6a9 1704
dc223ad4
ML
1705 /* Do not do an extra work for a single cluster. */
1706 if (clusters.length () == 1
1707 && clusters[0]->get_type () != SIMPLE_CASE)
1708 clusters[0]->emit (index_expr, index_type,
1709 gimple_switch_default_label (m_switch), m_default_bb);
1710 else
1711 {
1712 emit (bb, index_expr, default_edge->probability, index_type);
1713
1714 /* Emit cluster-specific switch handling. */
1715 for (unsigned i = 0; i < clusters.length (); i++)
1716 if (clusters[i]->get_type () != SIMPLE_CASE)
1717 clusters[i]->emit (index_expr, index_type,
1718 gimple_switch_default_label (m_switch),
1719 m_default_bb);
1720 }
9dc3d6a9 1721
dc223ad4
ML
1722 fix_phi_operands_for_edges ();
1723
1724 return true;
9dc3d6a9
ML
1725}
1726
dc223ad4
ML
1727/* Before switch transformation, record all SSA_NAMEs defined in switch BB
1728 and used in a label basic block. */
1729
1730void
1731switch_decision_tree::record_phi_operand_mapping ()
9dc3d6a9 1732{
dc223ad4 1733 basic_block switch_bb = gimple_bb (m_switch);
9dc3d6a9 1734 /* Record all PHI nodes that have to be fixed after conversion. */
dc223ad4 1735 for (unsigned i = 0; i < m_case_bbs.length (); i++)
9dc3d6a9 1736 {
9dc3d6a9 1737 gphi_iterator gsi;
dc223ad4 1738 basic_block bb = m_case_bbs[i];
9dc3d6a9
ML
1739 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1740 {
1741 gphi *phi = gsi.phi ();
1742
1743 for (unsigned i = 0; i < gimple_phi_num_args (phi); i++)
1744 {
1745 basic_block phi_src_bb = gimple_phi_arg_edge (phi, i)->src;
1746 if (phi_src_bb == switch_bb)
1747 {
1748 tree def = gimple_phi_arg_def (phi, i);
1749 tree result = gimple_phi_result (phi);
dc223ad4 1750 m_phi_mapping.put (result, def);
9dc3d6a9
ML
1751 break;
1752 }
1753 }
1754 }
1755 }
1756}
1757
dc223ad4
ML
1758/* Append new operands to PHI statements that were introduced due to
1759 addition of new edges to case labels. */
9dc3d6a9 1760
dc223ad4
ML
1761void
1762switch_decision_tree::fix_phi_operands_for_edges ()
9dc3d6a9 1763{
dc223ad4 1764 gphi_iterator gsi;
9dc3d6a9 1765
dc223ad4
ML
1766 for (unsigned i = 0; i < m_case_bbs.length (); i++)
1767 {
1768 basic_block bb = m_case_bbs[i];
1769 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1770 {
1771 gphi *phi = gsi.phi ();
1772 for (unsigned j = 0; j < gimple_phi_num_args (phi); j++)
1773 {
1774 tree def = gimple_phi_arg_def (phi, j);
1775 if (def == NULL_TREE)
1776 {
1777 edge e = gimple_phi_arg_edge (phi, j);
1778 tree *definition
1779 = m_phi_mapping.get (gimple_phi_result (phi));
1780 gcc_assert (definition);
1781 add_phi_arg (phi, *definition, e, UNKNOWN_LOCATION);
1782 }
1783 }
1784 }
1785 }
1786}
9dc3d6a9 1787
dc223ad4
ML
1788/* Generate a decision tree, switching on INDEX_EXPR and jumping to
1789 one of the labels in CASE_LIST or to the DEFAULT_LABEL.
9dc3d6a9 1790
dc223ad4
ML
1791 We generate a binary decision tree to select the appropriate target
1792 code. */
9dc3d6a9 1793
dc223ad4
ML
1794void
1795switch_decision_tree::emit (basic_block bb, tree index_expr,
1796 profile_probability default_prob, tree index_type)
1797{
1798 balance_case_nodes (&m_case_list, NULL);
9dc3d6a9 1799
dc223ad4
ML
1800 if (dump_file)
1801 dump_function_to_file (current_function_decl, dump_file, dump_flags);
1802 if (dump_file && (dump_flags & TDF_DETAILS))
1803 {
1804 int indent_step = ceil_log2 (TYPE_PRECISION (index_type)) + 2;
1805 fprintf (dump_file, ";; Expanding GIMPLE switch as decision tree:\n");
1806 gcc_assert (m_case_list != NULL);
1807 dump_case_nodes (dump_file, m_case_list, indent_step, 0);
1808 }
9dc3d6a9 1809
dc223ad4 1810 bb = emit_case_nodes (bb, index_expr, m_case_list, default_prob, index_type);
9dc3d6a9 1811
dc223ad4
ML
1812 if (bb)
1813 emit_jump (bb, m_default_bb);
9dc3d6a9 1814
dc223ad4
ML
1815 /* Remove all edges and do just an edge that will reach default_bb. */
1816 bb = gimple_bb (m_switch);
1817 gimple_stmt_iterator gsi = gsi_last_bb (bb);
1818 gsi_remove (&gsi, true);
9dc3d6a9 1819
dc223ad4
ML
1820 delete_basic_block (bb);
1821}
1822
1823/* Take an ordered list of case nodes
1824 and transform them into a near optimal binary tree,
1825 on the assumption that any target code selection value is as
1826 likely as any other.
1827
1828 The transformation is performed by splitting the ordered
1829 list into two equal sections plus a pivot. The parts are
1830 then attached to the pivot as left and right branches. Each
1831 branch is then transformed recursively. */
1832
1833void
1834switch_decision_tree::balance_case_nodes (case_tree_node **head,
1835 case_tree_node *parent)
1836{
1837 case_tree_node *np;
1838
1839 np = *head;
1840 if (np)
9dc3d6a9 1841 {
dc223ad4
ML
1842 int i = 0;
1843 int ranges = 0;
1844 case_tree_node **npp;
1845 case_tree_node *left;
9dc3d6a9 1846
dc223ad4 1847 /* Count the number of entries on branch. Also count the ranges. */
9dc3d6a9 1848
dc223ad4
ML
1849 while (np)
1850 {
1851 if (!tree_int_cst_equal (np->m_c->get_low (), np->m_c->get_high ()))
1852 ranges++;
9dc3d6a9 1853
dc223ad4
ML
1854 i++;
1855 np = np->m_right;
1856 }
9dc3d6a9 1857
dc223ad4
ML
1858 if (i > 2)
1859 {
1860 /* Split this list if it is long enough for that to help. */
1861 npp = head;
1862 left = *npp;
9dc3d6a9 1863
dc223ad4
ML
1864 /* If there are just three nodes, split at the middle one. */
1865 if (i == 3)
1866 npp = &(*npp)->m_right;
1867 else
1868 {
1869 /* Find the place in the list that bisects the list's total cost,
1870 where ranges count as 2.
1871 Here I gets half the total cost. */
1872 i = (i + ranges + 1) / 2;
1873 while (1)
1874 {
1875 /* Skip nodes while their cost does not reach that amount. */
1876 if (!tree_int_cst_equal ((*npp)->m_c->get_low (),
1877 (*npp)->m_c->get_high ()))
1878 i--;
1879 i--;
1880 if (i <= 0)
1881 break;
1882 npp = &(*npp)->m_right;
1883 }
1884 }
1885 *head = np = *npp;
1886 *npp = 0;
1887 np->m_parent = parent;
1888 np->m_left = left;
9dc3d6a9 1889
dc223ad4
ML
1890 /* Optimize each of the two split parts. */
1891 balance_case_nodes (&np->m_left, np);
1892 balance_case_nodes (&np->m_right, np);
1893 np->m_c->m_subtree_prob = np->m_c->m_prob;
1894 np->m_c->m_subtree_prob += np->m_left->m_c->m_subtree_prob;
1895 np->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob;
1896 }
1897 else
1898 {
1899 /* Else leave this branch as one level,
1900 but fill in `parent' fields. */
1901 np = *head;
1902 np->m_parent = parent;
1903 np->m_c->m_subtree_prob = np->m_c->m_prob;
1904 for (; np->m_right; np = np->m_right)
1905 {
1906 np->m_right->m_parent = np;
1907 (*head)->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob;
1908 }
1909 }
9dc3d6a9 1910 }
dc223ad4
ML
1911}
1912
1913/* Dump ROOT, a list or tree of case nodes, to file. */
9dc3d6a9 1914
dc223ad4
ML
1915void
1916switch_decision_tree::dump_case_nodes (FILE *f, case_tree_node *root,
1917 int indent_step, int indent_level)
1918{
1919 if (root == 0)
1920 return;
1921 indent_level++;
1922
1923 dump_case_nodes (f, root->m_left, indent_step, indent_level);
1924
1925 fputs (";; ", f);
1926 fprintf (f, "%*s", indent_step * indent_level, "");
1927 root->m_c->dump (f);
1928 root->m_c->m_prob.dump (f);
1929 fputs ("\n", f);
1930
1931 dump_case_nodes (f, root->m_right, indent_step, indent_level);
1932}
1933
1934
1935/* Add an unconditional jump to CASE_BB that happens in basic block BB. */
1936
1937void
1938switch_decision_tree::emit_jump (basic_block bb, basic_block case_bb)
1939{
1940 edge e = single_succ_edge (bb);
1941 redirect_edge_succ (e, case_bb);
1942}
1943
1944/* Generate code to compare OP0 with OP1 so that the condition codes are
1945 set and to jump to LABEL_BB if the condition is true.
1946 COMPARISON is the GIMPLE comparison (EQ, NE, GT, etc.).
1947 PROB is the probability of jumping to LABEL_BB. */
1948
1949basic_block
1950switch_decision_tree::emit_cmp_and_jump_insns (basic_block bb, tree op0,
1951 tree op1, tree_code comparison,
1952 basic_block label_bb,
1953 profile_probability prob)
1954{
1955 // TODO: it's once called with lhs != index.
1956 op1 = fold_convert (TREE_TYPE (op0), op1);
1957
1958 gcond *cond = gimple_build_cond (comparison, op0, op1, NULL_TREE, NULL_TREE);
1959 gimple_stmt_iterator gsi = gsi_last_bb (bb);
1960 gsi_insert_after (&gsi, cond, GSI_NEW_STMT);
1961
1962 gcc_assert (single_succ_p (bb));
1963
1964 /* Make a new basic block where false branch will take place. */
1965 edge false_edge = split_block (bb, cond);
1966 false_edge->flags = EDGE_FALSE_VALUE;
1967 false_edge->probability = prob.invert ();
1968
1969 edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE);
1970 true_edge->probability = prob;
1971
1972 return false_edge->dest;
1973}
1974
1975/* Emit step-by-step code to select a case for the value of INDEX.
1976 The thus generated decision tree follows the form of the
1977 case-node binary tree NODE, whose nodes represent test conditions.
1978 DEFAULT_PROB is probability of cases leading to default BB.
1979 INDEX_TYPE is the type of the index of the switch. */
1980
1981basic_block
1982switch_decision_tree::emit_case_nodes (basic_block bb, tree index,
1983 case_tree_node *node,
1984 profile_probability default_prob,
1985 tree index_type)
1986{
1987 /* If node is null, we are done. */
1988 if (node == NULL)
1989 return bb;
1990
1991 /* Branch to a label where we will handle it later. */
1992 basic_block test_bb = split_edge (single_succ_edge (bb));
1993 redirect_edge_succ (single_pred_edge (test_bb),
1994 single_succ_edge (bb)->dest);
1995
1996 profile_probability probability
1997 = (node->m_right
1998 ? node->m_right->m_c->m_subtree_prob : profile_probability::never ());
1999 probability = ((probability + default_prob.apply_scale (1, 2))
2000 / (node->m_c->m_subtree_prob + default_prob));
2001 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), GT_EXPR,
2002 test_bb, probability);
2003 default_prob = default_prob.apply_scale (1, 2);
2004
2005 /* Value belongs to this node or to the left-hand subtree. */
2006 probability = node->m_c->m_prob /
2007 (node->m_c->m_subtree_prob + default_prob);
2008 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_low (), GE_EXPR,
2009 node->m_c->m_case_bb, probability);
2010
2011 /* Handle the left-hand subtree. */
2012 bb = emit_case_nodes (bb, index, node->m_left,
2013 default_prob, index_type);
2014
2015 /* If the left-hand subtree fell through,
2016 don't let it fall into the right-hand subtree. */
2017 if (m_default_bb)
2018 emit_jump (bb, m_default_bb);
2019
2020 bb = emit_case_nodes (test_bb, index, node->m_right,
2021 default_prob, index_type);
2022
2023 return bb;
2024}
2025
2026/* The main function of the pass scans statements for switches and invokes
2027 process_switch on them. */
2028
2029namespace {
2030
2031const pass_data pass_data_convert_switch =
2032{
2033 GIMPLE_PASS, /* type */
2034 "switchconv", /* name */
2035 OPTGROUP_NONE, /* optinfo_flags */
2036 TV_TREE_SWITCH_CONVERSION, /* tv_id */
2037 ( PROP_cfg | PROP_ssa ), /* properties_required */
2038 0, /* properties_provided */
2039 0, /* properties_destroyed */
2040 0, /* todo_flags_start */
2041 TODO_update_ssa, /* todo_flags_finish */
2042};
2043
2044class pass_convert_switch : public gimple_opt_pass
2045{
2046public:
2047 pass_convert_switch (gcc::context *ctxt)
2048 : gimple_opt_pass (pass_data_convert_switch, ctxt)
2049 {}
2050
2051 /* opt_pass methods: */
2052 virtual bool gate (function *) { return flag_tree_switch_conversion != 0; }
2053 virtual unsigned int execute (function *);
2054
2055}; // class pass_convert_switch
2056
2057unsigned int
2058pass_convert_switch::execute (function *fun)
2059{
2060 basic_block bb;
2061 bool cfg_altered = false;
2062
2063 FOR_EACH_BB_FN (bb, fun)
2064 {
2065 gimple *stmt = last_stmt (bb);
2066 if (stmt && gimple_code (stmt) == GIMPLE_SWITCH)
2067 {
2068 if (dump_file)
2069 {
2070 expanded_location loc = expand_location (gimple_location (stmt));
2071
2072 fprintf (dump_file, "beginning to process the following "
2073 "SWITCH statement (%s:%d) : ------- \n",
2074 loc.file, loc.line);
2075 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2076 putc ('\n', dump_file);
2077 }
2078
2079 switch_conversion sconv;
2080 sconv.expand (as_a <gswitch *> (stmt));
2081 cfg_altered |= sconv.m_cfg_altered;
2082 if (!sconv.m_reason)
2083 {
2084 if (dump_file)
2085 {
2086 fputs ("Switch converted\n", dump_file);
2087 fputs ("--------------------------------\n", dump_file);
2088 }
2089
2090 /* Make no effort to update the post-dominator tree.
2091 It is actually not that hard for the transformations
2092 we have performed, but it is not supported
2093 by iterate_fix_dominators. */
2094 free_dominance_info (CDI_POST_DOMINATORS);
2095 }
2096 else
2097 {
2098 if (dump_file)
2099 {
2100 fputs ("Bailing out - ", dump_file);
2101 fputs (sconv.m_reason, dump_file);
2102 fputs ("\n--------------------------------\n", dump_file);
2103 }
2104 }
2105 }
2106 }
2107
2108 return cfg_altered ? TODO_cleanup_cfg : 0;;
2109}
2110
2111} // anon namespace
2112
2113gimple_opt_pass *
2114make_pass_convert_switch (gcc::context *ctxt)
2115{
2116 return new pass_convert_switch (ctxt);
9dc3d6a9
ML
2117}
2118
2119/* The main function of the pass scans statements for switches and invokes
2120 process_switch on them. */
2121
2122namespace {
2123
eb63c01f 2124template <bool O0> class pass_lower_switch: public gimple_opt_pass
9dc3d6a9 2125{
eb63c01f
ML
2126public:
2127 pass_lower_switch (gcc::context *ctxt) : gimple_opt_pass (data, ctxt) {}
2128
2129 static const pass_data data;
2130 opt_pass *
2131 clone ()
2132 {
2133 return new pass_lower_switch<O0> (m_ctxt);
2134 }
2135
2136 virtual bool
2137 gate (function *)
2138 {
2139 return !O0 || !optimize;
2140 }
2141
2142 virtual unsigned int execute (function *fun);
2143}; // class pass_lower_switch
2144
2145template <bool O0>
2146const pass_data pass_lower_switch<O0>::data = {
2147 GIMPLE_PASS, /* type */
2148 O0 ? "switchlower_O0" : "switchlower", /* name */
9dc3d6a9
ML
2149 OPTGROUP_NONE, /* optinfo_flags */
2150 TV_TREE_SWITCH_LOWERING, /* tv_id */
2151 ( PROP_cfg | PROP_ssa ), /* properties_required */
2152 0, /* properties_provided */
2153 0, /* properties_destroyed */
2154 0, /* todo_flags_start */
2155 TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */
2156};
2157
eb63c01f 2158template <bool O0>
9dc3d6a9 2159unsigned int
eb63c01f 2160pass_lower_switch<O0>::execute (function *fun)
9dc3d6a9
ML
2161{
2162 basic_block bb;
2163 bool expanded = false;
2164
dc223ad4
ML
2165 auto_vec<gimple *> switch_statements;
2166 switch_statements.create (1);
2167
9dc3d6a9
ML
2168 FOR_EACH_BB_FN (bb, fun)
2169 {
2170 gimple *stmt = last_stmt (bb);
2171 if (stmt && gimple_code (stmt) == GIMPLE_SWITCH)
dc223ad4
ML
2172 switch_statements.safe_push (stmt);
2173 }
2174
2175 for (unsigned i = 0; i < switch_statements.length (); i++)
2176 {
2177 gimple *stmt = switch_statements[i];
2178 if (dump_file)
9dc3d6a9 2179 {
dc223ad4 2180 expanded_location loc = expand_location (gimple_location (stmt));
9dc3d6a9 2181
dc223ad4
ML
2182 fprintf (dump_file, "beginning to process the following "
2183 "SWITCH statement (%s:%d) : ------- \n",
2184 loc.file, loc.line);
2185 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
2186 putc ('\n', dump_file);
2187 }
9dc3d6a9 2188
dc223ad4
ML
2189 gswitch *swtch = dyn_cast<gswitch *> (stmt);
2190 if (swtch)
2191 {
2192 switch_decision_tree dt (swtch);
2193 expanded |= dt.analyze_switch_statement ();
9dc3d6a9
ML
2194 }
2195 }
2196
2197 if (expanded)
2198 {
2199 free_dominance_info (CDI_DOMINATORS);
2200 free_dominance_info (CDI_POST_DOMINATORS);
2201 mark_virtual_operands_for_renaming (cfun);
2202 }
2203
2204 return 0;
2205}
2206
2207} // anon namespace
2208
2209gimple_opt_pass *
eb63c01f 2210make_pass_lower_switch_O0 (gcc::context *ctxt)
9dc3d6a9 2211{
eb63c01f 2212 return new pass_lower_switch<true> (ctxt);
9dc3d6a9 2213}
eb63c01f
ML
2214gimple_opt_pass *
2215make_pass_lower_switch (gcc::context *ctxt)
9dc3d6a9 2216{
eb63c01f 2217 return new pass_lower_switch<false> (ctxt);
9dc3d6a9
ML
2218}
2219
9dc3d6a9 2220