Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-switch-conversion.c @ 137:d22083d7f10b
merge
| author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Thu, 08 Nov 2018 14:16:42 +0900 |
| parents | 84e7813d76e9 |
| children | 1830386684a0 |
| rev | line source |
|---|---|
| 111 | 1 /* Lower GIMPLE_SWITCH expressions to something more efficient than |
| 2 a jump table. | |
| 131 | 3 Copyright (C) 2006-2018 Free Software Foundation, Inc. |
| 0 | 4 |
| 5 This file is part of GCC. | |
| 6 | |
| 7 GCC is free software; you can redistribute it and/or modify it | |
| 8 under the terms of the GNU General Public License as published by the | |
| 9 Free Software Foundation; either version 3, or (at your option) any | |
| 10 later version. | |
| 11 | |
| 12 GCC is distributed in the hope that it will be useful, but WITHOUT | |
| 13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 15 for more details. | |
| 16 | |
| 17 You should have received a copy of the GNU General Public License | |
| 18 along with GCC; see the file COPYING3. If not, write to the Free | |
| 19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA | |
| 20 02110-1301, USA. */ | |
| 21 | |
| 111 | 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" | |
| 28 #include "backend.h" | |
| 29 #include "insn-codes.h" | |
| 30 #include "rtl.h" | |
| 31 #include "tree.h" | |
| 32 #include "gimple.h" | |
| 33 #include "cfghooks.h" | |
| 34 #include "tree-pass.h" | |
| 35 #include "ssa.h" | |
| 36 #include "optabs-tree.h" | |
| 37 #include "cgraph.h" | |
| 38 #include "gimple-pretty-print.h" | |
| 39 #include "params.h" | |
| 40 #include "fold-const.h" | |
| 41 #include "varasm.h" | |
| 42 #include "stor-layout.h" | |
| 43 #include "cfganal.h" | |
| 44 #include "gimplify.h" | |
| 45 #include "gimple-iterator.h" | |
| 46 #include "gimplify-me.h" | |
| 131 | 47 #include "gimple-fold.h" |
| 111 | 48 #include "tree-cfg.h" |
| 49 #include "cfgloop.h" | |
| 50 #include "alloc-pool.h" | |
| 51 #include "target.h" | |
| 52 #include "tree-into-ssa.h" | |
| 131 | 53 #include "omp-general.h" |
| 111 | 54 |
| 55 /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode | |
| 56 type in the GIMPLE type system that is language-independent? */ | |
| 57 #include "langhooks.h" | |
| 58 | |
| 131 | 59 #include "tree-switch-conversion.h" |
| 111 | 60 |
| 131 | 61 using namespace tree_switch_conversion; |
| 62 | |
| 63 /* Constructor. */ | |
| 64 | |
| 65 switch_conversion::switch_conversion (): m_final_bb (NULL), m_other_count (), | |
| 66 m_constructors (NULL), m_default_values (NULL), | |
| 67 m_arr_ref_first (NULL), m_arr_ref_last (NULL), | |
| 68 m_reason (NULL), m_default_case_nonstandard (false), m_cfg_altered (false) | |
| 111 | 69 { |
| 70 } | |
| 131 | 71 |
| 72 /* Collection information about SWTCH statement. */ | |
| 73 | |
| 74 void | |
| 75 switch_conversion::collect (gswitch *swtch) | |
| 111 | 76 { |
| 77 unsigned int branch_num = gimple_switch_num_labels (swtch); | |
| 78 tree min_case, max_case; | |
| 131 | 79 unsigned int i; |
| 111 | 80 edge e, e_default, e_first; |
| 81 edge_iterator ei; | |
| 131 | 82 |
| 83 m_switch = swtch; | |
| 111 | 84 |
| 85 /* The gimplifier has already sorted the cases by CASE_LOW and ensured there | |
| 86 is a default label which is the first in the vector. | |
| 87 Collect the bits we can deduce from the CFG. */ | |
| 131 | 88 m_index_expr = gimple_switch_index (swtch); |
| 89 m_switch_bb = gimple_bb (swtch); | |
| 90 e_default = gimple_switch_default_edge (cfun, swtch); | |
| 91 m_default_bb = e_default->dest; | |
| 92 m_default_prob = e_default->probability; | |
| 93 m_default_count = e_default->count (); | |
| 94 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) | |
| 111 | 95 if (e != e_default) |
| 131 | 96 m_other_count += e->count (); |
| 111 | 97 |
| 98 /* Get upper and lower bounds of case values, and the covered range. */ | |
| 99 min_case = gimple_switch_label (swtch, 1); | |
| 100 max_case = gimple_switch_label (swtch, branch_num - 1); | |
| 101 | |
| 131 | 102 m_range_min = CASE_LOW (min_case); |
| 111 | 103 if (CASE_HIGH (max_case) != NULL_TREE) |
| 131 | 104 m_range_max = CASE_HIGH (max_case); |
| 111 | 105 else |
| 131 | 106 m_range_max = CASE_LOW (max_case); |
| 107 | |
| 108 m_contiguous_range = true; | |
| 109 tree last = CASE_HIGH (min_case) ? CASE_HIGH (min_case) : m_range_min; | |
| 111 | 110 for (i = 2; i < branch_num; i++) |
| 111 { | |
| 112 tree elt = gimple_switch_label (swtch, i); | |
| 113 if (wi::to_wide (last) + 1 != wi::to_wide (CASE_LOW (elt))) | |
| 114 { | |
| 131 | 115 m_contiguous_range = false; |
| 111 | 116 break; |
| 117 } | |
| 118 last = CASE_HIGH (elt) ? CASE_HIGH (elt) : CASE_LOW (elt); | |
| 119 } | |
| 120 | |
| 131 | 121 if (m_contiguous_range) |
| 122 e_first = gimple_switch_edge (cfun, swtch, 1); | |
| 111 | 123 else |
| 131 | 124 e_first = e_default; |
| 111 | 125 |
| 126 /* See if there is one common successor block for all branch | |
| 127 targets. If it exists, record it in FINAL_BB. | |
| 128 Start with the destination of the first non-default case | |
| 129 if the range is contiguous and default case otherwise as | |
| 130 guess or its destination in case it is a forwarder block. */ | |
| 131 if (! single_pred_p (e_first->dest)) | |
| 131 | 132 m_final_bb = e_first->dest; |
| 111 | 133 else if (single_succ_p (e_first->dest) |
| 134 && ! single_pred_p (single_succ (e_first->dest))) | |
| 131 | 135 m_final_bb = single_succ (e_first->dest); |
| 111 | 136 /* Require that all switch destinations are either that common |
| 137 FINAL_BB or a forwarder to it, except for the default | |
| 138 case if contiguous range. */ | |
| 131 | 139 if (m_final_bb) |
| 140 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) | |
| 111 | 141 { |
| 131 | 142 if (e->dest == m_final_bb) |
| 111 | 143 continue; |
| 144 | |
| 145 if (single_pred_p (e->dest) | |
| 146 && single_succ_p (e->dest) | |
| 131 | 147 && single_succ (e->dest) == m_final_bb) |
| 111 | 148 continue; |
| 149 | |
| 131 | 150 if (e == e_default && m_contiguous_range) |
| 111 | 151 { |
| 131 | 152 m_default_case_nonstandard = true; |
| 111 | 153 continue; |
| 154 } | |
| 155 | |
| 131 | 156 m_final_bb = NULL; |
| 111 | 157 break; |
| 158 } | |
| 159 | |
| 131 | 160 m_range_size |
| 161 = int_const_binop (MINUS_EXPR, m_range_max, m_range_min); | |
| 111 | 162 |
| 163 /* Get a count of the number of case labels. Single-valued case labels | |
| 164 simply count as one, but a case range counts double, since it may | |
| 165 require two compares if it gets lowered as a branching tree. */ | |
| 131 | 166 m_count = 0; |
| 111 | 167 for (i = 1; i < branch_num; i++) |
| 168 { | |
| 169 tree elt = gimple_switch_label (swtch, i); | |
| 131 | 170 m_count++; |
| 111 | 171 if (CASE_HIGH (elt) |
| 172 && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt))) | |
| 131 | 173 m_count++; |
| 111 | 174 } |
| 131 | 175 |
| 111 | 176 /* Get the number of unique non-default targets out of the GIMPLE_SWITCH |
| 177 block. Assume a CFG cleanup would have already removed degenerate | |
| 178 switch statements, this allows us to just use EDGE_COUNT. */ | |
| 131 | 179 m_uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1; |
| 111 | 180 } |
| 0 | 181 |
| 131 | 182 /* Checks whether the range given by individual case statements of the switch |
| 0 | 183 switch statement isn't too big and whether the number of branches actually |
| 184 satisfies the size of the new array. */ | |
| 185 | |
| 131 | 186 bool |
| 187 switch_conversion::check_range () | |
| 0 | 188 { |
| 131 | 189 gcc_assert (m_range_size); |
| 190 if (!tree_fits_uhwi_p (m_range_size)) | |
| 0 | 191 { |
| 131 | 192 m_reason = "index range way too large or otherwise unusable"; |
| 0 | 193 return false; |
| 194 } | |
| 195 | |
| 131 | 196 if (tree_to_uhwi (m_range_size) |
| 197 > ((unsigned) m_count * SWITCH_CONVERSION_BRANCH_RATIO)) | |
| 0 | 198 { |
| 131 | 199 m_reason = "the maximum range-branch ratio exceeded"; |
| 0 | 200 return false; |
| 201 } | |
| 202 | |
| 203 return true; | |
| 204 } | |
| 205 | |
| 131 | 206 /* Checks whether all but the final BB basic blocks are empty. */ |
| 207 | |
| 208 bool | |
| 209 switch_conversion::check_all_empty_except_final () | |
| 0 | 210 { |
| 131 | 211 edge e, e_default = find_edge (m_switch_bb, m_default_bb); |
| 111 | 212 edge_iterator ei; |
| 213 | |
| 131 | 214 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
215 { |
| 131 | 216 if (e->dest == m_final_bb) |
| 111 | 217 continue; |
| 218 | |
| 219 if (!empty_block_p (e->dest)) | |
| 0 | 220 { |
| 131 | 221 if (m_contiguous_range && e == e_default) |
| 111 | 222 { |
| 131 | 223 m_default_case_nonstandard = true; |
| 111 | 224 continue; |
| 225 } | |
| 226 | |
| 131 | 227 m_reason = "bad case - a non-final BB not empty"; |
| 0 | 228 return false; |
| 229 } | |
| 230 } | |
| 231 | |
| 232 return true; | |
| 233 } | |
| 234 | |
| 235 /* This function checks whether all required values in phi nodes in final_bb | |
| 236 are constants. Required values are those that correspond to a basic block | |
| 237 which is a part of the examined switch statement. It returns true if the | |
| 238 phi nodes are OK, otherwise false. */ | |
| 239 | |
| 131 | 240 bool |
| 241 switch_conversion::check_final_bb () | |
| 0 | 242 { |
| 111 | 243 gphi_iterator gsi; |
| 244 | |
| 131 | 245 m_phi_count = 0; |
| 246 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 0 | 247 { |
| 111 | 248 gphi *phi = gsi.phi (); |
| 0 | 249 unsigned int i; |
| 250 | |
| 111 | 251 if (virtual_operand_p (gimple_phi_result (phi))) |
| 252 continue; | |
| 253 | |
| 131 | 254 m_phi_count++; |
| 0 | 255 |
| 256 for (i = 0; i < gimple_phi_num_args (phi); i++) | |
| 257 { | |
| 258 basic_block bb = gimple_phi_arg_edge (phi, i)->src; | |
| 259 | |
| 131 | 260 if (bb == m_switch_bb |
| 111 | 261 || (single_pred_p (bb) |
| 131 | 262 && single_pred (bb) == m_switch_bb |
| 263 && (!m_default_case_nonstandard | |
| 111 | 264 || empty_block_p (bb)))) |
| 0 | 265 { |
| 266 tree reloc, val; | |
| 111 | 267 const char *reason = NULL; |
| 0 | 268 |
| 269 val = gimple_phi_arg_def (phi, i); | |
| 270 if (!is_gimple_ip_invariant (val)) | |
| 111 | 271 reason = "non-invariant value from a case"; |
| 272 else | |
| 0 | 273 { |
| 111 | 274 reloc = initializer_constant_valid_p (val, TREE_TYPE (val)); |
| 275 if ((flag_pic && reloc != null_pointer_node) | |
| 276 || (!flag_pic && reloc == NULL_TREE)) | |
| 277 { | |
| 278 if (reloc) | |
| 279 reason | |
| 280 = "value from a case would need runtime relocations"; | |
| 281 else | |
| 282 reason | |
| 283 = "value from a case is not a valid initializer"; | |
| 284 } | |
| 0 | 285 } |
| 111 | 286 if (reason) |
| 0 | 287 { |
| 111 | 288 /* For contiguous range, we can allow non-constant |
| 289 or one that needs relocation, as long as it is | |
| 290 only reachable from the default case. */ | |
| 131 | 291 if (bb == m_switch_bb) |
| 292 bb = m_final_bb; | |
| 293 if (!m_contiguous_range || bb != m_default_bb) | |
| 111 | 294 { |
| 131 | 295 m_reason = reason; |
| 111 | 296 return false; |
| 297 } | |
| 298 | |
| 131 | 299 unsigned int branch_num = gimple_switch_num_labels (m_switch); |
| 111 | 300 for (unsigned int i = 1; i < branch_num; i++) |
| 301 { | |
| 131 | 302 if (gimple_switch_label_bb (cfun, m_switch, i) == bb) |
| 111 | 303 { |
| 131 | 304 m_reason = reason; |
| 111 | 305 return false; |
| 306 } | |
| 307 } | |
| 131 | 308 m_default_case_nonstandard = true; |
| 0 | 309 } |
| 310 } | |
| 311 } | |
| 312 } | |
| 313 | |
| 314 return true; | |
| 315 } | |
| 316 | |
| 317 /* The following function allocates default_values, target_{in,out}_names and | |
| 318 constructors arrays. The last one is also populated with pointers to | |
| 319 vectors that will become constructors of new arrays. */ | |
| 320 | |
| 131 | 321 void |
| 322 switch_conversion::create_temp_arrays () | |
| 0 | 323 { |
| 324 int i; | |
| 325 | |
| 131 | 326 m_default_values = XCNEWVEC (tree, m_phi_count * 3); |
| 111 | 327 /* ??? Macros do not support multi argument templates in their |
| 328 argument list. We create a typedef to work around that problem. */ | |
| 329 typedef vec<constructor_elt, va_gc> *vec_constructor_elt_gc; | |
| 131 | 330 m_constructors = XCNEWVEC (vec_constructor_elt_gc, m_phi_count); |
| 331 m_target_inbound_names = m_default_values + m_phi_count; | |
| 332 m_target_outbound_names = m_target_inbound_names + m_phi_count; | |
| 333 for (i = 0; i < m_phi_count; i++) | |
| 334 vec_alloc (m_constructors[i], tree_to_uhwi (m_range_size) + 1); | |
| 0 | 335 } |
| 336 | |
| 337 /* Populate the array of default values in the order of phi nodes. | |
| 111 | 338 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch |
| 339 if the range is non-contiguous or the default case has standard | |
| 340 structure, otherwise it is the first non-default case instead. */ | |
| 0 | 341 |
| 131 | 342 void |
| 343 switch_conversion::gather_default_values (tree default_case) | |
| 0 | 344 { |
| 111 | 345 gphi_iterator gsi; |
| 131 | 346 basic_block bb = label_to_block (cfun, CASE_LABEL (default_case)); |
| 0 | 347 edge e; |
| 348 int i = 0; | |
| 349 | |
| 111 | 350 gcc_assert (CASE_LOW (default_case) == NULL_TREE |
| 131 | 351 || m_default_case_nonstandard); |
| 352 | |
| 353 if (bb == m_final_bb) | |
| 354 e = find_edge (m_switch_bb, bb); | |
| 0 | 355 else |
| 356 e = single_succ_edge (bb); | |
| 357 | |
| 131 | 358 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 0 | 359 { |
| 111 | 360 gphi *phi = gsi.phi (); |
| 361 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 362 continue; | |
| 0 | 363 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
| 364 gcc_assert (val); | |
| 131 | 365 m_default_values[i++] = val; |
| 0 | 366 } |
| 367 } | |
| 368 | |
| 369 /* The following function populates the vectors in the constructors array with | |
| 370 future contents of the static arrays. The vectors are populated in the | |
| 131 | 371 order of phi nodes. */ |
| 372 | |
| 373 void | |
| 374 switch_conversion::build_constructors () | |
| 0 | 375 { |
| 131 | 376 unsigned i, branch_num = gimple_switch_num_labels (m_switch); |
| 377 tree pos = m_range_min; | |
| 111 | 378 tree pos_one = build_int_cst (TREE_TYPE (pos), 1); |
| 0 | 379 |
| 380 for (i = 1; i < branch_num; i++) | |
| 381 { | |
| 131 | 382 tree cs = gimple_switch_label (m_switch, i); |
| 383 basic_block bb = label_to_block (cfun, CASE_LABEL (cs)); | |
| 0 | 384 edge e; |
| 385 tree high; | |
| 111 | 386 gphi_iterator gsi; |
| 0 | 387 int j; |
| 388 | |
| 131 | 389 if (bb == m_final_bb) |
| 390 e = find_edge (m_switch_bb, bb); | |
| 0 | 391 else |
| 392 e = single_succ_edge (bb); | |
| 393 gcc_assert (e); | |
| 394 | |
| 395 while (tree_int_cst_lt (pos, CASE_LOW (cs))) | |
| 396 { | |
| 397 int k; | |
| 131 | 398 for (k = 0; k < m_phi_count; k++) |
| 0 | 399 { |
| 111 | 400 constructor_elt elt; |
| 401 | |
| 131 | 402 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
| 111 | 403 elt.value |
| 131 | 404 = unshare_expr_without_location (m_default_values[k]); |
| 405 m_constructors[k]->quick_push (elt); | |
| 0 | 406 } |
| 407 | |
| 111 | 408 pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
| 0 | 409 } |
| 410 gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs))); | |
| 411 | |
| 412 j = 0; | |
| 413 if (CASE_HIGH (cs)) | |
| 414 high = CASE_HIGH (cs); | |
| 415 else | |
| 416 high = CASE_LOW (cs); | |
| 131 | 417 for (gsi = gsi_start_phis (m_final_bb); |
| 0 | 418 !gsi_end_p (gsi); gsi_next (&gsi)) |
| 419 { | |
| 111 | 420 gphi *phi = gsi.phi (); |
| 421 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 422 continue; | |
| 0 | 423 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
| 424 tree low = CASE_LOW (cs); | |
| 425 pos = CASE_LOW (cs); | |
| 426 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
427 do |
| 0 | 428 { |
| 111 | 429 constructor_elt elt; |
| 430 | |
| 131 | 431 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
| 111 | 432 elt.value = unshare_expr_without_location (val); |
| 131 | 433 m_constructors[j]->quick_push (elt); |
| 111 | 434 |
| 435 pos = int_const_binop (PLUS_EXPR, pos, pos_one); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
436 } while (!tree_int_cst_lt (high, pos) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
437 && tree_int_cst_lt (low, pos)); |
| 0 | 438 j++; |
| 439 } | |
| 440 } | |
| 441 } | |
| 442 | |
| 131 | 443 /* If all values in the constructor vector are products of a linear function |
| 444 a * x + b, then return true. When true, COEFF_A and COEFF_B and | |
| 445 coefficients of the linear function. Note that equal values are special | |
| 446 case of a linear function with a and b equal to zero. */ | |
| 447 | |
| 448 bool | |
| 449 switch_conversion::contains_linear_function_p (vec<constructor_elt, va_gc> *vec, | |
| 450 wide_int *coeff_a, | |
| 451 wide_int *coeff_b) | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
452 { |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
453 unsigned int i; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
454 constructor_elt *elt; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
455 |
| 131 | 456 gcc_assert (vec->length () >= 2); |
| 457 | |
| 458 /* Let's try to find any linear function a * x + y that can apply to | |
| 459 given values. 'a' can be calculated as follows: | |
| 460 | |
| 461 a = (y2 - y1) / (x2 - x1) where x2 - x1 = 1 (consecutive case indices) | |
| 462 a = y2 - y1 | |
| 463 | |
| 464 and | |
| 465 | |
| 466 b = y2 - a * x2 | |
| 467 | |
| 468 */ | |
| 469 | |
| 470 tree elt0 = (*vec)[0].value; | |
| 471 tree elt1 = (*vec)[1].value; | |
| 472 | |
| 473 if (TREE_CODE (elt0) != INTEGER_CST || TREE_CODE (elt1) != INTEGER_CST) | |
| 474 return false; | |
| 475 | |
| 476 wide_int range_min = wi::to_wide (fold_convert (TREE_TYPE (elt0), | |
| 477 m_range_min)); | |
| 478 wide_int y1 = wi::to_wide (elt0); | |
| 479 wide_int y2 = wi::to_wide (elt1); | |
| 480 wide_int a = y2 - y1; | |
| 481 wide_int b = y2 - a * (range_min + 1); | |
| 482 | |
| 483 /* Verify that all values fulfill the linear function. */ | |
| 111 | 484 FOR_EACH_VEC_SAFE_ELT (vec, i, elt) |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
485 { |
| 131 | 486 if (TREE_CODE (elt->value) != INTEGER_CST) |
| 487 return false; | |
| 488 | |
| 489 wide_int value = wi::to_wide (elt->value); | |
| 490 if (a * range_min + b != value) | |
| 491 return false; | |
| 492 | |
| 493 ++range_min; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
494 } |
| 131 | 495 |
| 496 *coeff_a = a; | |
| 497 *coeff_b = b; | |
| 498 | |
| 499 return true; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
500 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
501 |
| 111 | 502 /* Return type which should be used for array elements, either TYPE's |
| 503 main variant or, for integral types, some smaller integral type | |
| 504 that can still hold all the constants. */ | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
505 |
| 131 | 506 tree |
| 507 switch_conversion::array_value_type (tree type, int num) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
508 { |
| 131 | 509 unsigned int i, len = vec_safe_length (m_constructors[num]); |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
510 constructor_elt *elt; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
511 int sign = 0; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
512 tree smaller_type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
513 |
| 111 | 514 /* Types with alignments greater than their size can reach here, e.g. out of |
| 515 SRA. We couldn't use these as an array component type so get back to the | |
| 516 main variant first, which, for our purposes, is fine for other types as | |
| 517 well. */ | |
| 518 | |
| 519 type = TYPE_MAIN_VARIANT (type); | |
| 520 | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
521 if (!INTEGRAL_TYPE_P (type)) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
522 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
523 |
| 111 | 524 scalar_int_mode type_mode = SCALAR_INT_TYPE_MODE (type); |
| 525 scalar_int_mode mode = get_narrowest_mode (type_mode); | |
| 526 if (GET_MODE_SIZE (type_mode) <= GET_MODE_SIZE (mode)) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
527 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
528 |
| 131 | 529 if (len < (optimize_bb_for_size_p (gimple_bb (m_switch)) ? 2 : 32)) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
530 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
531 |
| 131 | 532 FOR_EACH_VEC_SAFE_ELT (m_constructors[num], i, elt) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
533 { |
| 111 | 534 wide_int cst; |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
535 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
536 if (TREE_CODE (elt->value) != INTEGER_CST) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
537 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
538 |
| 111 | 539 cst = wi::to_wide (elt->value); |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
540 while (1) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
541 { |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
542 unsigned int prec = GET_MODE_BITSIZE (mode); |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
543 if (prec > HOST_BITS_PER_WIDE_INT) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
544 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
545 |
| 111 | 546 if (sign >= 0 && cst == wi::zext (cst, prec)) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
547 { |
| 111 | 548 if (sign == 0 && cst == wi::sext (cst, prec)) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
549 break; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
550 sign = 1; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
551 break; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
552 } |
| 111 | 553 if (sign <= 0 && cst == wi::sext (cst, prec)) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
554 { |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
555 sign = -1; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
556 break; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
557 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
558 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
559 if (sign == 1) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
560 sign = 0; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
561 |
| 111 | 562 if (!GET_MODE_WIDER_MODE (mode).exists (&mode) |
| 563 || GET_MODE_SIZE (mode) >= GET_MODE_SIZE (type_mode)) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
564 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
565 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
566 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
567 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
568 if (sign == 0) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
569 sign = TYPE_UNSIGNED (type) ? 1 : -1; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
570 smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0); |
| 111 | 571 if (GET_MODE_SIZE (type_mode) |
| 572 <= GET_MODE_SIZE (SCALAR_INT_TYPE_MODE (smaller_type))) | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
573 return type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
574 |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
575 return smaller_type; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
576 } |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
577 |
| 131 | 578 /* Create an appropriate array type and declaration and assemble a static |
| 579 array variable. Also create a load statement that initializes | |
| 580 the variable in question with a value from the static array. SWTCH is | |
| 581 the switch statement being converted, NUM is the index to | |
| 582 arrays of constructors, default values and target SSA names | |
| 583 for this particular array. ARR_INDEX_TYPE is the type of the index | |
| 584 of the new array, PHI is the phi node of the final BB that corresponds | |
| 585 to the value that will be loaded from the created array. TIDX | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
586 is an ssa name of a temporary variable holding the index for loads from the |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
587 new array. */ |
| 0 | 588 |
| 131 | 589 void |
| 590 switch_conversion::build_one_array (int num, tree arr_index_type, | |
| 591 gphi *phi, tree tidx) | |
| 0 | 592 { |
| 131 | 593 tree name; |
| 111 | 594 gimple *load; |
| 131 | 595 gimple_stmt_iterator gsi = gsi_for_stmt (m_switch); |
| 596 location_t loc = gimple_location (m_switch); | |
| 597 | |
| 598 gcc_assert (m_default_values[num]); | |
| 111 | 599 |
| 600 name = copy_ssa_name (PHI_RESULT (phi)); | |
| 131 | 601 m_target_inbound_names[num] = name; |
| 602 | |
| 603 wide_int coeff_a, coeff_b; | |
| 604 bool linear_p = contains_linear_function_p (m_constructors[num], &coeff_a, | |
| 605 &coeff_b); | |
| 606 if (linear_p) | |
| 607 { | |
| 608 if (dump_file && coeff_a.to_uhwi () > 0) | |
| 609 fprintf (dump_file, "Linear transformation with A = %" PRId64 | |
| 610 " and B = %" PRId64 "\n", coeff_a.to_shwi (), | |
| 611 coeff_b.to_shwi ()); | |
| 612 | |
| 613 tree t = unsigned_type_for (TREE_TYPE (m_index_expr)); | |
| 614 gimple_seq seq = NULL; | |
| 615 tree tmp = gimple_convert (&seq, t, m_index_expr); | |
| 616 tree tmp2 = gimple_build (&seq, MULT_EXPR, t, | |
| 617 wide_int_to_tree (t, coeff_a), tmp); | |
| 618 tree tmp3 = gimple_build (&seq, PLUS_EXPR, t, tmp2, | |
| 619 wide_int_to_tree (t, coeff_b)); | |
| 620 tree tmp4 = gimple_convert (&seq, TREE_TYPE (name), tmp3); | |
| 621 gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
| 622 load = gimple_build_assign (name, tmp4); | |
| 623 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
624 else |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
625 { |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
626 tree array_type, ctor, decl, value_type, fetch, default_type; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
627 |
| 131 | 628 default_type = TREE_TYPE (m_default_values[num]); |
| 629 value_type = array_value_type (default_type, num); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
630 array_type = build_array_type (value_type, arr_index_type); |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
631 if (default_type != value_type) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
632 { |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
633 unsigned int i; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
634 constructor_elt *elt; |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
635 |
| 131 | 636 FOR_EACH_VEC_SAFE_ELT (m_constructors[num], i, elt) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
637 elt->value = fold_convert (value_type, elt->value); |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
638 } |
| 131 | 639 ctor = build_constructor (array_type, m_constructors[num]); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
640 TREE_CONSTANT (ctor) = true; |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
641 TREE_STATIC (ctor) = true; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
642 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
643 decl = build_decl (loc, VAR_DECL, NULL_TREE, array_type); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
644 TREE_STATIC (decl) = 1; |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
645 DECL_INITIAL (decl) = ctor; |
| 0 | 646 |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
647 DECL_NAME (decl) = create_tmp_var_name ("CSWTCH"); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
648 DECL_ARTIFICIAL (decl) = 1; |
| 111 | 649 DECL_IGNORED_P (decl) = 1; |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
650 TREE_CONSTANT (decl) = 1; |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
651 TREE_READONLY (decl) = 1; |
| 111 | 652 DECL_IGNORED_P (decl) = 1; |
| 131 | 653 if (offloading_function_p (cfun->decl)) |
| 654 DECL_ATTRIBUTES (decl) | |
| 655 = tree_cons (get_identifier ("omp declare target"), NULL_TREE, | |
| 656 NULL_TREE); | |
| 111 | 657 varpool_node::finalize_decl (decl); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
658 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
659 fetch = build4 (ARRAY_REF, value_type, decl, tidx, NULL_TREE, |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
660 NULL_TREE); |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
661 if (default_type != value_type) |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
662 { |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
663 fetch = fold_convert (default_type, fetch); |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
664 fetch = force_gimple_operand_gsi (&gsi, fetch, true, NULL_TREE, |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
665 true, GSI_SAME_STMT); |
|
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
666 } |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
667 load = gimple_build_assign (name, fetch); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
668 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
669 |
| 0 | 670 gsi_insert_before (&gsi, load, GSI_SAME_STMT); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
671 update_stmt (load); |
| 131 | 672 m_arr_ref_last = load; |
| 0 | 673 } |
| 674 | |
| 675 /* Builds and initializes static arrays initialized with values gathered from | |
| 131 | 676 the switch statement. Also creates statements that load values from |
| 0 | 677 them. */ |
| 678 | |
| 131 | 679 void |
| 680 switch_conversion::build_arrays () | |
| 0 | 681 { |
| 682 tree arr_index_type; | |
| 111 | 683 tree tidx, sub, utype; |
| 684 gimple *stmt; | |
| 0 | 685 gimple_stmt_iterator gsi; |
| 111 | 686 gphi_iterator gpi; |
| 0 | 687 int i; |
| 131 | 688 location_t loc = gimple_location (m_switch); |
| 689 | |
| 690 gsi = gsi_for_stmt (m_switch); | |
| 0 | 691 |
| 111 | 692 /* Make sure we do not generate arithmetics in a subrange. */ |
| 131 | 693 utype = TREE_TYPE (m_index_expr); |
| 111 | 694 if (TREE_TYPE (utype)) |
| 695 utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1); | |
| 696 else | |
| 697 utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1); | |
| 698 | |
| 131 | 699 arr_index_type = build_index_type (m_range_size); |
| 111 | 700 tidx = make_ssa_name (utype); |
| 701 sub = fold_build2_loc (loc, MINUS_EXPR, utype, | |
| 131 | 702 fold_convert_loc (loc, utype, m_index_expr), |
| 703 fold_convert_loc (loc, utype, m_range_min)); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
704 sub = force_gimple_operand_gsi (&gsi, sub, |
| 0 | 705 false, NULL, true, GSI_SAME_STMT); |
| 706 stmt = gimple_build_assign (tidx, sub); | |
| 707 | |
| 708 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
709 update_stmt (stmt); |
| 131 | 710 m_arr_ref_first = stmt; |
| 711 | |
| 712 for (gpi = gsi_start_phis (m_final_bb), i = 0; | |
| 111 | 713 !gsi_end_p (gpi); gsi_next (&gpi)) |
| 714 { | |
| 715 gphi *phi = gpi.phi (); | |
| 716 if (!virtual_operand_p (gimple_phi_result (phi))) | |
| 131 | 717 build_one_array (i++, arr_index_type, phi, tidx); |
| 111 | 718 else |
| 719 { | |
| 720 edge e; | |
| 721 edge_iterator ei; | |
| 131 | 722 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
| 111 | 723 { |
| 131 | 724 if (e->dest == m_final_bb) |
| 111 | 725 break; |
| 131 | 726 if (!m_default_case_nonstandard |
| 727 || e->dest != m_default_bb) | |
| 111 | 728 { |
| 729 e = single_succ_edge (e->dest); | |
| 730 break; | |
| 731 } | |
| 732 } | |
| 131 | 733 gcc_assert (e && e->dest == m_final_bb); |
| 734 m_target_vop = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
| 111 | 735 } |
| 736 } | |
| 0 | 737 } |
| 738 | |
| 739 /* Generates and appropriately inserts loads of default values at the position | |
| 131 | 740 given by GSI. Returns the last inserted statement. */ |
| 741 | |
| 742 gassign * | |
| 743 switch_conversion::gen_def_assigns (gimple_stmt_iterator *gsi) | |
| 0 | 744 { |
| 745 int i; | |
| 111 | 746 gassign *assign = NULL; |
| 747 | |
| 131 | 748 for (i = 0; i < m_phi_count; i++) |
| 0 | 749 { |
| 131 | 750 tree name = copy_ssa_name (m_target_inbound_names[i]); |
| 751 m_target_outbound_names[i] = name; | |
| 752 assign = gimple_build_assign (name, m_default_values[i]); | |
| 0 | 753 gsi_insert_before (gsi, assign, GSI_SAME_STMT); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
754 update_stmt (assign); |
| 0 | 755 } |
| 756 return assign; | |
| 757 } | |
| 758 | |
| 759 /* Deletes the unused bbs and edges that now contain the switch statement and | |
| 131 | 760 its empty branch bbs. BBD is the now dead BB containing |
| 761 the original switch statement, FINAL is the last BB of the converted | |
| 762 switch statement (in terms of succession). */ | |
| 763 | |
| 764 void | |
| 765 switch_conversion::prune_bbs (basic_block bbd, basic_block final, | |
| 766 basic_block default_bb) | |
| 0 | 767 { |
| 768 edge_iterator ei; | |
| 769 edge e; | |
| 770 | |
| 771 for (ei = ei_start (bbd->succs); (e = ei_safe_edge (ei)); ) | |
| 772 { | |
| 773 basic_block bb; | |
| 774 bb = e->dest; | |
| 775 remove_edge (e); | |
| 111 | 776 if (bb != final && bb != default_bb) |
| 0 | 777 delete_basic_block (bb); |
| 778 } | |
| 779 delete_basic_block (bbd); | |
| 780 } | |
| 781 | |
| 782 /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge | |
| 783 from the basic block loading values from an array and E2F from the basic | |
| 784 block loading default values. BBF is the last switch basic block (see the | |
| 785 bbf description in the comment below). */ | |
| 786 | |
| 131 | 787 void |
| 788 switch_conversion::fix_phi_nodes (edge e1f, edge e2f, basic_block bbf) | |
| 0 | 789 { |
| 111 | 790 gphi_iterator gsi; |
| 0 | 791 int i; |
| 792 | |
| 793 for (gsi = gsi_start_phis (bbf), i = 0; | |
| 111 | 794 !gsi_end_p (gsi); gsi_next (&gsi)) |
| 0 | 795 { |
| 111 | 796 gphi *phi = gsi.phi (); |
| 797 tree inbound, outbound; | |
| 798 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 131 | 799 inbound = outbound = m_target_vop; |
| 111 | 800 else |
| 801 { | |
| 131 | 802 inbound = m_target_inbound_names[i]; |
| 803 outbound = m_target_outbound_names[i++]; | |
| 111 | 804 } |
| 805 add_phi_arg (phi, inbound, e1f, UNKNOWN_LOCATION); | |
| 131 | 806 if (!m_default_case_nonstandard) |
| 111 | 807 add_phi_arg (phi, outbound, e2f, UNKNOWN_LOCATION); |
| 0 | 808 } |
| 809 } | |
| 810 | |
| 811 /* Creates a check whether the switch expression value actually falls into the | |
| 812 range given by all the cases. If it does not, the temporaries are loaded | |
| 131 | 813 with default values instead. */ |
| 814 | |
| 815 void | |
| 816 switch_conversion::gen_inbound_check () | |
| 0 | 817 { |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
818 tree label_decl1 = create_artificial_label (UNKNOWN_LOCATION); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
819 tree label_decl2 = create_artificial_label (UNKNOWN_LOCATION); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
820 tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION); |
| 111 | 821 glabel *label1, *label2, *label3; |
| 822 tree utype, tidx; | |
| 0 | 823 tree bound; |
| 824 | |
| 111 | 825 gcond *cond_stmt; |
| 826 | |
| 827 gassign *last_assign = NULL; | |
| 0 | 828 gimple_stmt_iterator gsi; |
| 829 basic_block bb0, bb1, bb2, bbf, bbd; | |
| 111 | 830 edge e01 = NULL, e02, e21, e1d, e1f, e2f; |
| 131 | 831 location_t loc = gimple_location (m_switch); |
| 832 | |
| 833 gcc_assert (m_default_values); | |
| 834 | |
| 835 bb0 = gimple_bb (m_switch); | |
| 836 | |
| 837 tidx = gimple_assign_lhs (m_arr_ref_first); | |
| 111 | 838 utype = TREE_TYPE (tidx); |
| 0 | 839 |
| 840 /* (end of) block 0 */ | |
| 131 | 841 gsi = gsi_for_stmt (m_arr_ref_first); |
| 111 | 842 gsi_next (&gsi); |
| 843 | |
| 131 | 844 bound = fold_convert_loc (loc, utype, m_range_size); |
| 111 | 845 cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE); |
| 0 | 846 gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
847 update_stmt (cond_stmt); |
| 0 | 848 |
| 849 /* block 2 */ | |
| 131 | 850 if (!m_default_case_nonstandard) |
| 111 | 851 { |
| 852 label2 = gimple_build_label (label_decl2); | |
| 853 gsi_insert_before (&gsi, label2, GSI_SAME_STMT); | |
| 131 | 854 last_assign = gen_def_assigns (&gsi); |
| 111 | 855 } |
| 0 | 856 |
| 857 /* block 1 */ | |
| 858 label1 = gimple_build_label (label_decl1); | |
| 859 gsi_insert_before (&gsi, label1, GSI_SAME_STMT); | |
| 860 | |
| 861 /* block F */ | |
| 131 | 862 gsi = gsi_start_bb (m_final_bb); |
| 0 | 863 label3 = gimple_build_label (label_decl3); |
| 864 gsi_insert_before (&gsi, label3, GSI_SAME_STMT); | |
| 865 | |
| 866 /* cfg fix */ | |
| 867 e02 = split_block (bb0, cond_stmt); | |
| 868 bb2 = e02->dest; | |
| 869 | |
| 131 | 870 if (m_default_case_nonstandard) |
| 111 | 871 { |
| 872 bb1 = bb2; | |
| 131 | 873 bb2 = m_default_bb; |
| 111 | 874 e01 = e02; |
| 875 e01->flags = EDGE_TRUE_VALUE; | |
| 876 e02 = make_edge (bb0, bb2, EDGE_FALSE_VALUE); | |
| 877 edge e_default = find_edge (bb1, bb2); | |
| 878 for (gphi_iterator gsi = gsi_start_phis (bb2); | |
| 879 !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 880 { | |
| 881 gphi *phi = gsi.phi (); | |
| 882 tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_default); | |
| 883 add_phi_arg (phi, arg, e02, | |
| 884 gimple_phi_arg_location_from_edge (phi, e_default)); | |
| 885 } | |
| 886 /* Partially fix the dominator tree, if it is available. */ | |
| 887 if (dom_info_available_p (CDI_DOMINATORS)) | |
| 888 redirect_immediate_dominators (CDI_DOMINATORS, bb1, bb0); | |
| 889 } | |
| 890 else | |
| 891 { | |
| 892 e21 = split_block (bb2, last_assign); | |
| 893 bb1 = e21->dest; | |
| 894 remove_edge (e21); | |
| 895 } | |
| 896 | |
| 131 | 897 e1d = split_block (bb1, m_arr_ref_last); |
| 0 | 898 bbd = e1d->dest; |
| 899 remove_edge (e1d); | |
| 900 | |
| 131 | 901 /* Flags and profiles of the edge for in-range values. */ |
| 902 if (!m_default_case_nonstandard) | |
| 111 | 903 e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE); |
| 131 | 904 e01->probability = m_default_prob.invert (); |
| 905 | |
| 906 /* Flags and profiles of the edge taking care of out-of-range values. */ | |
| 0 | 907 e02->flags &= ~EDGE_FALLTHRU; |
| 908 e02->flags |= EDGE_FALSE_VALUE; | |
| 131 | 909 e02->probability = m_default_prob; |
| 910 | |
| 911 bbf = m_final_bb; | |
| 0 | 912 |
| 913 e1f = make_edge (bb1, bbf, EDGE_FALLTHRU); | |
| 111 | 914 e1f->probability = profile_probability::always (); |
| 915 | |
| 131 | 916 if (m_default_case_nonstandard) |
| 111 | 917 e2f = NULL; |
| 918 else | |
| 919 { | |
| 920 e2f = make_edge (bb2, bbf, EDGE_FALLTHRU); | |
| 921 e2f->probability = profile_probability::always (); | |
| 922 } | |
| 0 | 923 |
| 924 /* frequencies of the new BBs */ | |
| 131 | 925 bb1->count = e01->count (); |
| 926 bb2->count = e02->count (); | |
| 927 if (!m_default_case_nonstandard) | |
| 928 bbf->count = e1f->count () + e2f->count (); | |
| 111 | 929 |
| 930 /* Tidy blocks that have become unreachable. */ | |
| 131 | 931 prune_bbs (bbd, m_final_bb, |
| 932 m_default_case_nonstandard ? m_default_bb : NULL); | |
| 111 | 933 |
| 934 /* Fixup the PHI nodes in bbF. */ | |
| 131 | 935 fix_phi_nodes (e1f, e2f, bbf); |
| 111 | 936 |
| 937 /* Fix the dominator tree, if it is available. */ | |
| 938 if (dom_info_available_p (CDI_DOMINATORS)) | |
| 939 { | |
| 940 vec<basic_block> bbs_to_fix_dom; | |
| 941 | |
| 942 set_immediate_dominator (CDI_DOMINATORS, bb1, bb0); | |
| 131 | 943 if (!m_default_case_nonstandard) |
| 111 | 944 set_immediate_dominator (CDI_DOMINATORS, bb2, bb0); |
| 945 if (! get_immediate_dominator (CDI_DOMINATORS, bbf)) | |
| 946 /* If bbD was the immediate dominator ... */ | |
| 947 set_immediate_dominator (CDI_DOMINATORS, bbf, bb0); | |
| 948 | |
| 949 bbs_to_fix_dom.create (3 + (bb2 != bbf)); | |
| 950 bbs_to_fix_dom.quick_push (bb0); | |
| 951 bbs_to_fix_dom.quick_push (bb1); | |
| 952 if (bb2 != bbf) | |
| 953 bbs_to_fix_dom.quick_push (bb2); | |
| 954 bbs_to_fix_dom.quick_push (bbf); | |
| 955 | |
| 956 iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true); | |
| 957 bbs_to_fix_dom.release (); | |
| 958 } | |
| 0 | 959 } |
| 960 | |
| 131 | 961 /* The following function is invoked on every switch statement (the current |
| 962 one is given in SWTCH) and runs the individual phases of switch | |
| 963 conversion on it one after another until one fails or the conversion | |
| 964 is completed. On success, NULL is in m_reason, otherwise points | |
| 965 to a string with the reason why the conversion failed. */ | |
| 966 | |
| 967 void | |
| 968 switch_conversion::expand (gswitch *swtch) | |
| 0 | 969 { |
| 111 | 970 /* Group case labels so that we get the right results from the heuristics |
| 971 that decide on the code generation approach for this switch. */ | |
| 131 | 972 m_cfg_altered |= group_case_labels_stmt (swtch); |
| 111 | 973 |
| 974 /* If this switch is now a degenerate case with only a default label, | |
| 131 | 975 there is nothing left for us to do. */ |
| 111 | 976 if (gimple_switch_num_labels (swtch) < 2) |
| 131 | 977 { |
| 978 m_reason = "switch is a degenerate case"; | |
| 979 return; | |
| 980 } | |
| 981 | |
| 982 collect (swtch); | |
| 111 | 983 |
| 984 /* No error markers should reach here (they should be filtered out | |
| 985 during gimplification). */ | |
| 131 | 986 gcc_checking_assert (TREE_TYPE (m_index_expr) != error_mark_node); |
| 111 | 987 |
| 988 /* A switch on a constant should have been optimized in tree-cfg-cleanup. */ | |
| 131 | 989 gcc_checking_assert (!TREE_CONSTANT (m_index_expr)); |
| 990 | |
| 991 /* Prefer bit test if possible. */ | |
| 992 if (tree_fits_uhwi_p (m_range_size) | |
| 993 && bit_test_cluster::can_be_handled (tree_to_uhwi (m_range_size), m_uniq) | |
| 994 && bit_test_cluster::is_beneficial (m_count, m_uniq)) | |
| 0 | 995 { |
| 131 | 996 m_reason = "expanding as bit test is preferable"; |
| 997 return; | |
| 998 } | |
| 999 | |
| 1000 if (m_uniq <= 2) | |
| 1001 { | |
| 1002 /* This will be expanded as a decision tree . */ | |
| 1003 m_reason = "expanding as jumps is preferable"; | |
| 1004 return; | |
| 0 | 1005 } |
| 1006 | |
| 111 | 1007 /* If there is no common successor, we cannot do the transformation. */ |
| 131 | 1008 if (!m_final_bb) |
| 1009 { | |
| 1010 m_reason = "no common successor to all case label target blocks found"; | |
| 1011 return; | |
| 1012 } | |
| 0 | 1013 |
| 1014 /* Check the case label values are within reasonable range: */ | |
| 131 | 1015 if (!check_range ()) |
| 111 | 1016 { |
| 131 | 1017 gcc_assert (m_reason); |
| 1018 return; | |
| 111 | 1019 } |
| 0 | 1020 |
| 1021 /* For all the cases, see whether they are empty, the assignments they | |
| 1022 represent constant and so on... */ | |
| 131 | 1023 if (!check_all_empty_except_final ()) |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1024 { |
| 131 | 1025 gcc_assert (m_reason); |
| 1026 return; | |
|
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1027 } |
| 131 | 1028 if (!check_final_bb ()) |
| 111 | 1029 { |
| 131 | 1030 gcc_assert (m_reason); |
| 1031 return; | |
| 111 | 1032 } |
| 0 | 1033 |
| 1034 /* At this point all checks have passed and we can proceed with the | |
| 1035 transformation. */ | |
| 1036 | |
| 131 | 1037 create_temp_arrays (); |
| 1038 gather_default_values (m_default_case_nonstandard | |
| 111 | 1039 ? gimple_switch_label (swtch, 1) |
| 131 | 1040 : gimple_switch_default_label (swtch)); |
| 1041 build_constructors (); | |
| 1042 | |
| 1043 build_arrays (); /* Build the static arrays and assignments. */ | |
| 1044 gen_inbound_check (); /* Build the bounds check. */ | |
| 1045 | |
| 1046 m_cfg_altered = true; | |
| 1047 } | |
| 1048 | |
| 1049 /* Destructor. */ | |
| 1050 | |
| 1051 switch_conversion::~switch_conversion () | |
| 1052 { | |
| 1053 XDELETEVEC (m_constructors); | |
| 1054 XDELETEVEC (m_default_values); | |
| 1055 } | |
| 1056 | |
| 1057 /* Constructor. */ | |
| 1058 | |
| 1059 group_cluster::group_cluster (vec<cluster *> &clusters, | |
| 1060 unsigned start, unsigned end) | |
| 1061 { | |
| 1062 gcc_checking_assert (end - start + 1 >= 1); | |
| 1063 m_prob = profile_probability::never (); | |
| 1064 m_cases.create (end - start + 1); | |
| 1065 for (unsigned i = start; i <= end; i++) | |
| 1066 { | |
| 1067 m_cases.quick_push (static_cast<simple_cluster *> (clusters[i])); | |
| 1068 m_prob += clusters[i]->m_prob; | |
| 1069 } | |
| 1070 m_subtree_prob = m_prob; | |
| 1071 } | |
| 1072 | |
| 1073 /* Destructor. */ | |
| 1074 | |
| 1075 group_cluster::~group_cluster () | |
| 1076 { | |
| 1077 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1078 delete m_cases[i]; | |
| 1079 | |
| 1080 m_cases.release (); | |
| 1081 } | |
| 1082 | |
| 1083 /* Dump content of a cluster. */ | |
| 1084 | |
| 1085 void | |
| 1086 group_cluster::dump (FILE *f, bool details) | |
| 1087 { | |
| 1088 unsigned total_values = 0; | |
| 1089 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1090 total_values += m_cases[i]->get_range (m_cases[i]->get_low (), | |
| 1091 m_cases[i]->get_high ()); | |
| 1092 | |
| 1093 unsigned comparison_count = 0; | |
| 1094 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1095 { | |
| 1096 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1097 comparison_count += sc->m_range_p ? 2 : 1; | |
| 1098 } | |
| 1099 | |
| 1100 unsigned HOST_WIDE_INT range = get_range (get_low (), get_high ()); | |
| 1101 fprintf (f, "%s", get_type () == JUMP_TABLE ? "JT" : "BT"); | |
| 1102 | |
| 1103 if (details) | |
| 1104 fprintf (f, "(values:%d comparisons:%d range:" HOST_WIDE_INT_PRINT_DEC | |
| 1105 " density: %.2f%%)", total_values, comparison_count, range, | |
| 1106 100.0f * comparison_count / range); | |
| 1107 | |
| 1108 fprintf (f, ":"); | |
| 1109 PRINT_CASE (f, get_low ()); | |
| 1110 fprintf (f, "-"); | |
| 1111 PRINT_CASE (f, get_high ()); | |
| 1112 fprintf (f, " "); | |
| 1113 } | |
| 1114 | |
| 1115 /* Emit GIMPLE code to handle the cluster. */ | |
| 1116 | |
| 1117 void | |
| 1118 jump_table_cluster::emit (tree index_expr, tree, | |
| 1119 tree default_label_expr, basic_block default_bb) | |
| 1120 { | |
| 1121 unsigned HOST_WIDE_INT range = get_range (get_low (), get_high ()); | |
| 1122 unsigned HOST_WIDE_INT nondefault_range = 0; | |
| 1123 | |
| 1124 /* For jump table we just emit a new gswitch statement that will | |
| 1125 be latter lowered to jump table. */ | |
| 1126 auto_vec <tree> labels; | |
| 1127 labels.create (m_cases.length ()); | |
| 1128 | |
| 1129 make_edge (m_case_bb, default_bb, 0); | |
| 1130 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1131 { | |
| 1132 labels.quick_push (unshare_expr (m_cases[i]->m_case_label_expr)); | |
| 1133 make_edge (m_case_bb, m_cases[i]->m_case_bb, 0); | |
| 1134 } | |
| 1135 | |
| 1136 gswitch *s = gimple_build_switch (index_expr, | |
| 1137 unshare_expr (default_label_expr), labels); | |
| 1138 gimple_stmt_iterator gsi = gsi_start_bb (m_case_bb); | |
| 1139 gsi_insert_after (&gsi, s, GSI_NEW_STMT); | |
| 1140 | |
| 1141 /* Set up even probabilities for all cases. */ | |
| 1142 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1143 { | |
| 1144 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1145 edge case_edge = find_edge (m_case_bb, sc->m_case_bb); | |
| 1146 unsigned HOST_WIDE_INT case_range | |
| 1147 = sc->get_range (sc->get_low (), sc->get_high ()); | |
| 1148 nondefault_range += case_range; | |
| 1149 | |
| 1150 /* case_edge->aux is number of values in a jump-table that are covered | |
| 1151 by the case_edge. */ | |
| 1152 case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + case_range); | |
| 1153 } | |
| 1154 | |
| 1155 edge default_edge = gimple_switch_default_edge (cfun, s); | |
| 1156 default_edge->probability = profile_probability::never (); | |
| 1157 | |
| 1158 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1159 { | |
| 1160 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1161 edge case_edge = find_edge (m_case_bb, sc->m_case_bb); | |
| 1162 case_edge->probability | |
| 1163 = profile_probability::always ().apply_scale ((intptr_t)case_edge->aux, | |
| 1164 range); | |
| 1165 } | |
| 1166 | |
| 1167 /* Number of non-default values is probability of default edge. */ | |
| 1168 default_edge->probability | |
| 1169 += profile_probability::always ().apply_scale (nondefault_range, | |
| 1170 range).invert (); | |
| 1171 | |
| 1172 switch_decision_tree::reset_out_edges_aux (s); | |
| 1173 } | |
| 1174 | |
| 1175 /* Find jump tables of given CLUSTERS, where all members of the vector | |
| 1176 are of type simple_cluster. New clusters are returned. */ | |
| 1177 | |
| 1178 vec<cluster *> | |
| 1179 jump_table_cluster::find_jump_tables (vec<cluster *> &clusters) | |
| 1180 { | |
| 1181 if (!is_enabled ()) | |
| 1182 return clusters.copy (); | |
| 1183 | |
| 1184 unsigned l = clusters.length (); | |
| 1185 auto_vec<min_cluster_item> min; | |
| 1186 min.reserve (l + 1); | |
| 1187 | |
| 1188 min.quick_push (min_cluster_item (0, 0, 0)); | |
| 1189 | |
| 1190 for (unsigned i = 1; i <= l; i++) | |
| 1191 { | |
| 1192 /* Set minimal # of clusters with i-th item to infinite. */ | |
| 1193 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); | |
| 1194 | |
| 1195 for (unsigned j = 0; j < i; j++) | |
| 1196 { | |
| 1197 unsigned HOST_WIDE_INT s = min[j].m_non_jt_cases; | |
| 1198 if (i - j < case_values_threshold ()) | |
| 1199 s += i - j; | |
| 1200 | |
| 1201 /* Prefer clusters with smaller number of numbers covered. */ | |
| 1202 if ((min[j].m_count + 1 < min[i].m_count | |
| 1203 || (min[j].m_count + 1 == min[i].m_count | |
| 1204 && s < min[i].m_non_jt_cases)) | |
| 1205 && can_be_handled (clusters, j, i - 1)) | |
| 1206 min[i] = min_cluster_item (min[j].m_count + 1, j, s); | |
| 1207 } | |
| 1208 | |
| 1209 gcc_checking_assert (min[i].m_count != INT_MAX); | |
| 1210 } | |
| 1211 | |
| 1212 /* No result. */ | |
| 1213 if (min[l].m_count == INT_MAX) | |
| 1214 return clusters.copy (); | |
| 1215 | |
| 1216 vec<cluster *> output; | |
| 1217 output.create (4); | |
| 1218 | |
| 1219 /* Find and build the clusters. */ | |
| 1220 for (int end = l;;) | |
| 1221 { | |
| 1222 int start = min[end].m_start; | |
| 1223 | |
| 1224 /* Do not allow clusters with small number of cases. */ | |
| 1225 if (is_beneficial (clusters, start, end - 1)) | |
| 1226 output.safe_push (new jump_table_cluster (clusters, start, end - 1)); | |
| 1227 else | |
| 1228 for (int i = end - 1; i >= start; i--) | |
| 1229 output.safe_push (clusters[i]); | |
| 1230 | |
| 1231 end = start; | |
| 1232 | |
| 1233 if (start <= 0) | |
| 1234 break; | |
| 1235 } | |
| 1236 | |
| 1237 output.reverse (); | |
| 1238 return output; | |
| 1239 } | |
| 1240 | |
| 1241 /* Return true when cluster starting at START and ending at END (inclusive) | |
| 1242 can build a jump-table. */ | |
| 1243 | |
| 1244 bool | |
| 1245 jump_table_cluster::can_be_handled (const vec<cluster *> &clusters, | |
| 1246 unsigned start, unsigned end) | |
| 1247 { | |
| 1248 /* If the switch is relatively small such that the cost of one | |
| 1249 indirect jump on the target are higher than the cost of a | |
| 1250 decision tree, go with the decision tree. | |
| 1251 | |
| 1252 If range of values is much bigger than number of values, | |
| 1253 or if it is too large to represent in a HOST_WIDE_INT, | |
| 1254 make a sequence of conditional branches instead of a dispatch. | |
| 1255 | |
| 1256 The definition of "much bigger" depends on whether we are | |
| 1257 optimizing for size or for speed. */ | |
| 1258 if (!flag_jump_tables) | |
| 1259 return false; | |
| 1260 | |
| 1261 /* For algorithm correctness, jump table for a single case must return | |
| 1262 true. We bail out in is_beneficial if it's called just for | |
| 1263 a single case. */ | |
| 1264 if (start == end) | |
| 1265 return true; | |
| 1266 | |
| 1267 unsigned HOST_WIDE_INT max_ratio | |
| 1268 = optimize_insn_for_size_p () ? max_ratio_for_size : max_ratio_for_speed; | |
| 1269 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (), | |
| 1270 clusters[end]->get_high ()); | |
| 1271 /* Check overflow. */ | |
| 1272 if (range == 0) | |
| 1273 return false; | |
| 1274 | |
| 1275 unsigned HOST_WIDE_INT comparison_count = 0; | |
| 1276 for (unsigned i = start; i <= end; i++) | |
| 1277 { | |
| 1278 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1279 comparison_count += sc->m_range_p ? 2 : 1; | |
| 1280 } | |
| 1281 | |
| 1282 return range <= max_ratio * comparison_count; | |
| 1283 } | |
| 1284 | |
| 1285 /* Return true if cluster starting at START and ending at END (inclusive) | |
| 1286 is profitable transformation. */ | |
| 1287 | |
| 1288 bool | |
| 1289 jump_table_cluster::is_beneficial (const vec<cluster *> &, | |
| 1290 unsigned start, unsigned end) | |
| 1291 { | |
| 1292 /* Single case bail out. */ | |
| 1293 if (start == end) | |
| 1294 return false; | |
| 1295 | |
| 1296 return end - start + 1 >= case_values_threshold (); | |
| 1297 } | |
| 1298 | |
| 1299 /* Definition of jump_table_cluster constants. */ | |
| 1300 | |
| 1301 const unsigned HOST_WIDE_INT jump_table_cluster::max_ratio_for_size; | |
| 1302 const unsigned HOST_WIDE_INT jump_table_cluster::max_ratio_for_speed; | |
| 1303 | |
| 1304 /* Find bit tests of given CLUSTERS, where all members of the vector | |
| 1305 are of type simple_cluster. New clusters are returned. */ | |
| 1306 | |
| 1307 vec<cluster *> | |
| 1308 bit_test_cluster::find_bit_tests (vec<cluster *> &clusters) | |
| 1309 { | |
| 1310 vec<cluster *> output; | |
| 1311 output.create (4); | |
| 1312 | |
| 1313 unsigned l = clusters.length (); | |
| 1314 auto_vec<min_cluster_item> min; | |
| 1315 min.reserve (l + 1); | |
| 1316 | |
| 1317 min.quick_push (min_cluster_item (0, 0, 0)); | |
| 1318 | |
| 1319 for (unsigned i = 1; i <= l; i++) | |
| 1320 { | |
| 1321 /* Set minimal # of clusters with i-th item to infinite. */ | |
| 1322 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); | |
| 1323 | |
| 1324 for (unsigned j = 0; j < i; j++) | |
| 1325 { | |
| 1326 if (min[j].m_count + 1 < min[i].m_count | |
| 1327 && can_be_handled (clusters, j, i - 1)) | |
| 1328 min[i] = min_cluster_item (min[j].m_count + 1, j, INT_MAX); | |
| 1329 } | |
| 1330 | |
| 1331 gcc_checking_assert (min[i].m_count != INT_MAX); | |
| 1332 } | |
| 1333 | |
| 1334 /* No result. */ | |
| 1335 if (min[l].m_count == INT_MAX) | |
| 1336 return clusters.copy (); | |
| 1337 | |
| 1338 /* Find and build the clusters. */ | |
| 1339 for (unsigned end = l;;) | |
| 1340 { | |
| 1341 int start = min[end].m_start; | |
| 1342 | |
| 1343 if (is_beneficial (clusters, start, end - 1)) | |
| 1344 { | |
| 1345 bool entire = start == 0 && end == clusters.length (); | |
| 1346 output.safe_push (new bit_test_cluster (clusters, start, end - 1, | |
| 1347 entire)); | |
| 1348 } | |
| 1349 else | |
| 1350 for (int i = end - 1; i >= start; i--) | |
| 1351 output.safe_push (clusters[i]); | |
| 1352 | |
| 1353 end = start; | |
| 1354 | |
| 1355 if (start <= 0) | |
| 1356 break; | |
| 1357 } | |
| 1358 | |
| 1359 output.reverse (); | |
| 1360 return output; | |
| 1361 } | |
| 1362 | |
| 1363 /* Return true when RANGE of case values with UNIQ labels | |
| 1364 can build a bit test. */ | |
| 1365 | |
| 1366 bool | |
| 1367 bit_test_cluster::can_be_handled (unsigned HOST_WIDE_INT range, | |
| 1368 unsigned int uniq) | |
| 1369 { | |
| 1370 /* Check overflow. */ | |
| 1371 if (range == 0) | |
| 1372 return 0; | |
| 1373 | |
| 1374 if (range >= GET_MODE_BITSIZE (word_mode)) | |
| 1375 return false; | |
| 1376 | |
| 1377 return uniq <= 3; | |
| 1378 } | |
| 1379 | |
| 1380 /* Return true when cluster starting at START and ending at END (inclusive) | |
| 1381 can build a bit test. */ | |
| 1382 | |
| 1383 bool | |
| 1384 bit_test_cluster::can_be_handled (const vec<cluster *> &clusters, | |
| 1385 unsigned start, unsigned end) | |
| 1386 { | |
| 1387 /* For algorithm correctness, bit test for a single case must return | |
| 1388 true. We bail out in is_beneficial if it's called just for | |
| 1389 a single case. */ | |
| 1390 if (start == end) | |
| 1391 return true; | |
| 1392 | |
| 1393 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (), | |
| 1394 clusters[end]->get_high ()); | |
| 1395 auto_bitmap dest_bbs; | |
| 1396 | |
| 1397 for (unsigned i = start; i <= end; i++) | |
| 1398 { | |
| 1399 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1400 bitmap_set_bit (dest_bbs, sc->m_case_bb->index); | |
| 1401 } | |
| 1402 | |
| 1403 return can_be_handled (range, bitmap_count_bits (dest_bbs)); | |
| 1404 } | |
| 1405 | |
| 1406 /* Return true when COUNT of cases of UNIQ labels is beneficial for bit test | |
| 1407 transformation. */ | |
| 1408 | |
| 1409 bool | |
| 1410 bit_test_cluster::is_beneficial (unsigned count, unsigned uniq) | |
| 1411 { | |
| 1412 return (((uniq == 1 && count >= 3) | |
| 1413 || (uniq == 2 && count >= 5) | |
| 1414 || (uniq == 3 && count >= 6))); | |
| 1415 } | |
| 1416 | |
| 1417 /* Return true if cluster starting at START and ending at END (inclusive) | |
| 1418 is profitable transformation. */ | |
| 1419 | |
| 1420 bool | |
| 1421 bit_test_cluster::is_beneficial (const vec<cluster *> &clusters, | |
| 1422 unsigned start, unsigned end) | |
| 1423 { | |
| 1424 /* Single case bail out. */ | |
| 1425 if (start == end) | |
| 1426 return false; | |
| 1427 | |
| 1428 auto_bitmap dest_bbs; | |
| 1429 | |
| 1430 for (unsigned i = start; i <= end; i++) | |
| 1431 { | |
| 1432 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1433 bitmap_set_bit (dest_bbs, sc->m_case_bb->index); | |
| 1434 } | |
| 1435 | |
| 1436 unsigned uniq = bitmap_count_bits (dest_bbs); | |
| 1437 unsigned count = end - start + 1; | |
| 1438 return is_beneficial (count, uniq); | |
| 1439 } | |
| 1440 | |
| 1441 /* Comparison function for qsort to order bit tests by decreasing | |
| 1442 probability of execution. */ | |
| 1443 | |
| 1444 int | |
| 1445 case_bit_test::cmp (const void *p1, const void *p2) | |
| 1446 { | |
| 1447 const struct case_bit_test *const d1 = (const struct case_bit_test *) p1; | |
| 1448 const struct case_bit_test *const d2 = (const struct case_bit_test *) p2; | |
| 1449 | |
| 1450 if (d2->bits != d1->bits) | |
| 1451 return d2->bits - d1->bits; | |
| 1452 | |
| 1453 /* Stabilize the sort. */ | |
| 1454 return (LABEL_DECL_UID (CASE_LABEL (d2->label)) | |
| 1455 - LABEL_DECL_UID (CASE_LABEL (d1->label))); | |
| 1456 } | |
| 1457 | |
| 1458 /* Expand a switch statement by a short sequence of bit-wise | |
| 1459 comparisons. "switch(x)" is effectively converted into | |
| 1460 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are | |
| 1461 integer constants. | |
| 1462 | |
| 1463 INDEX_EXPR is the value being switched on. | |
| 1464 | |
| 1465 MINVAL is the lowest case value of in the case nodes, | |
| 1466 and RANGE is highest value minus MINVAL. MINVAL and RANGE | |
| 1467 are not guaranteed to be of the same type as INDEX_EXPR | |
| 1468 (the gimplifier doesn't change the type of case label values, | |
| 1469 and MINVAL and RANGE are derived from those values). | |
| 1470 MAXVAL is MINVAL + RANGE. | |
| 1471 | |
| 1472 There *MUST* be max_case_bit_tests or less unique case | |
| 1473 node targets. */ | |
| 1474 | |
| 1475 void | |
| 1476 bit_test_cluster::emit (tree index_expr, tree index_type, | |
| 1477 tree, basic_block default_bb) | |
| 1478 { | |
| 1479 struct case_bit_test test[m_max_case_bit_tests] = { {} }; | |
| 1480 unsigned int i, j, k; | |
| 1481 unsigned int count; | |
| 1482 | |
| 1483 tree unsigned_index_type = unsigned_type_for (index_type); | |
| 1484 | |
| 1485 gimple_stmt_iterator gsi; | |
| 1486 gassign *shift_stmt; | |
| 1487 | |
| 1488 tree idx, tmp, csui; | |
| 1489 tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1); | |
| 1490 tree word_mode_zero = fold_convert (word_type_node, integer_zero_node); | |
| 1491 tree word_mode_one = fold_convert (word_type_node, integer_one_node); | |
| 1492 int prec = TYPE_PRECISION (word_type_node); | |
| 1493 wide_int wone = wi::one (prec); | |
| 1494 | |
| 1495 tree minval = get_low (); | |
| 1496 tree maxval = get_high (); | |
| 1497 tree range = int_const_binop (MINUS_EXPR, maxval, minval); | |
| 1498 unsigned HOST_WIDE_INT bt_range = get_range (minval, maxval); | |
| 1499 | |
| 1500 /* Go through all case labels, and collect the case labels, profile | |
| 1501 counts, and other information we need to build the branch tests. */ | |
| 1502 count = 0; | |
| 1503 for (i = 0; i < m_cases.length (); i++) | |
| 1504 { | |
| 1505 unsigned int lo, hi; | |
| 1506 simple_cluster *n = static_cast<simple_cluster *> (m_cases[i]); | |
| 1507 for (k = 0; k < count; k++) | |
| 1508 if (n->m_case_bb == test[k].target_bb) | |
| 1509 break; | |
| 1510 | |
| 1511 if (k == count) | |
| 1512 { | |
| 1513 gcc_checking_assert (count < m_max_case_bit_tests); | |
| 1514 test[k].mask = wi::zero (prec); | |
| 1515 test[k].target_bb = n->m_case_bb; | |
| 1516 test[k].label = n->m_case_label_expr; | |
| 1517 test[k].bits = 0; | |
| 1518 count++; | |
| 1519 } | |
| 1520 | |
| 1521 test[k].bits += n->get_range (n->get_low (), n->get_high ()); | |
| 1522 | |
| 1523 lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_low (), minval)); | |
| 1524 if (n->get_high () == NULL_TREE) | |
| 1525 hi = lo; | |
| 1526 else | |
| 1527 hi = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_high (), | |
| 1528 minval)); | |
| 1529 | |
| 1530 for (j = lo; j <= hi; j++) | |
| 1531 test[k].mask |= wi::lshift (wone, j); | |
| 1532 } | |
| 1533 | |
| 1534 qsort (test, count, sizeof (*test), case_bit_test::cmp); | |
| 1535 | |
| 1536 /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of | |
| 1537 the minval subtractions, but it might make the mask constants more | |
| 1538 expensive. So, compare the costs. */ | |
| 1539 if (compare_tree_int (minval, 0) > 0 | |
| 1540 && compare_tree_int (maxval, GET_MODE_BITSIZE (word_mode)) < 0) | |
| 1541 { | |
| 1542 int cost_diff; | |
| 1543 HOST_WIDE_INT m = tree_to_uhwi (minval); | |
| 1544 rtx reg = gen_raw_REG (word_mode, 10000); | |
| 1545 bool speed_p = optimize_insn_for_speed_p (); | |
| 1546 cost_diff = set_rtx_cost (gen_rtx_PLUS (word_mode, reg, | |
| 1547 GEN_INT (-m)), speed_p); | |
| 1548 for (i = 0; i < count; i++) | |
| 1549 { | |
| 1550 rtx r = immed_wide_int_const (test[i].mask, word_mode); | |
| 1551 cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r), | |
| 1552 word_mode, speed_p); | |
| 1553 r = immed_wide_int_const (wi::lshift (test[i].mask, m), word_mode); | |
| 1554 cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r), | |
| 1555 word_mode, speed_p); | |
| 1556 } | |
| 1557 if (cost_diff > 0) | |
| 1558 { | |
| 1559 for (i = 0; i < count; i++) | |
| 1560 test[i].mask = wi::lshift (test[i].mask, m); | |
| 1561 minval = build_zero_cst (TREE_TYPE (minval)); | |
| 1562 range = maxval; | |
| 1563 } | |
| 1564 } | |
| 1565 | |
| 1566 /* Now build the test-and-branch code. */ | |
| 1567 | |
| 1568 gsi = gsi_last_bb (m_case_bb); | |
| 1569 | |
| 1570 /* idx = (unsigned)x - minval. */ | |
| 1571 idx = fold_convert (unsigned_index_type, index_expr); | |
| 1572 idx = fold_build2 (MINUS_EXPR, unsigned_index_type, idx, | |
| 1573 fold_convert (unsigned_index_type, minval)); | |
| 1574 idx = force_gimple_operand_gsi (&gsi, idx, | |
| 1575 /*simple=*/true, NULL_TREE, | |
| 1576 /*before=*/true, GSI_SAME_STMT); | |
| 1577 | |
| 1578 if (m_handles_entire_switch) | |
| 1579 { | |
| 1580 /* if (idx > range) goto default */ | |
| 1581 range | |
| 1582 = force_gimple_operand_gsi (&gsi, | |
| 1583 fold_convert (unsigned_index_type, range), | |
| 1584 /*simple=*/true, NULL_TREE, | |
| 1585 /*before=*/true, GSI_SAME_STMT); | |
| 1586 tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range); | |
| 1587 basic_block new_bb | |
| 1588 = hoist_edge_and_branch_if_true (&gsi, tmp, default_bb, | |
| 1589 profile_probability::unlikely ()); | |
| 1590 gsi = gsi_last_bb (new_bb); | |
| 1591 } | |
| 1592 | |
| 1593 /* csui = (1 << (word_mode) idx) */ | |
| 1594 csui = make_ssa_name (word_type_node); | |
| 1595 tmp = fold_build2 (LSHIFT_EXPR, word_type_node, word_mode_one, | |
| 1596 fold_convert (word_type_node, idx)); | |
| 1597 tmp = force_gimple_operand_gsi (&gsi, tmp, | |
| 1598 /*simple=*/false, NULL_TREE, | |
| 1599 /*before=*/true, GSI_SAME_STMT); | |
| 1600 shift_stmt = gimple_build_assign (csui, tmp); | |
| 1601 gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT); | |
| 1602 update_stmt (shift_stmt); | |
| 1603 | |
| 1604 profile_probability prob = profile_probability::always (); | |
| 1605 | |
| 1606 /* for each unique set of cases: | |
| 1607 if (const & csui) goto target */ | |
| 1608 for (k = 0; k < count; k++) | |
| 1609 { | |
| 1610 prob = profile_probability::always ().apply_scale (test[k].bits, | |
| 1611 bt_range); | |
| 1612 bt_range -= test[k].bits; | |
| 1613 tmp = wide_int_to_tree (word_type_node, test[k].mask); | |
| 1614 tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp); | |
| 1615 tmp = force_gimple_operand_gsi (&gsi, tmp, | |
| 1616 /*simple=*/true, NULL_TREE, | |
| 1617 /*before=*/true, GSI_SAME_STMT); | |
| 1618 tmp = fold_build2 (NE_EXPR, boolean_type_node, tmp, word_mode_zero); | |
| 1619 basic_block new_bb | |
| 1620 = hoist_edge_and_branch_if_true (&gsi, tmp, test[k].target_bb, prob); | |
| 1621 gsi = gsi_last_bb (new_bb); | |
| 1622 } | |
| 1623 | |
| 1624 /* We should have removed all edges now. */ | |
| 1625 gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0); | |
| 1626 | |
| 1627 /* If nothing matched, go to the default label. */ | |
| 1628 edge e = make_edge (gsi_bb (gsi), default_bb, EDGE_FALLTHRU); | |
| 1629 e->probability = profile_probability::always (); | |
| 1630 } | |
| 1631 | |
| 1632 /* Split the basic block at the statement pointed to by GSIP, and insert | |
| 1633 a branch to the target basic block of E_TRUE conditional on tree | |
| 1634 expression COND. | |
| 1635 | |
| 1636 It is assumed that there is already an edge from the to-be-split | |
| 1637 basic block to E_TRUE->dest block. This edge is removed, and the | |
| 1638 profile information on the edge is re-used for the new conditional | |
| 1639 jump. | |
| 1640 | |
| 1641 The CFG is updated. The dominator tree will not be valid after | |
| 1642 this transformation, but the immediate dominators are updated if | |
| 1643 UPDATE_DOMINATORS is true. | |
| 1644 | |
| 1645 Returns the newly created basic block. */ | |
| 1646 | |
| 1647 basic_block | |
| 1648 bit_test_cluster::hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip, | |
| 1649 tree cond, basic_block case_bb, | |
| 1650 profile_probability prob) | |
| 1651 { | |
| 1652 tree tmp; | |
| 1653 gcond *cond_stmt; | |
| 1654 edge e_false; | |
| 1655 basic_block new_bb, split_bb = gsi_bb (*gsip); | |
| 1656 | |
| 1657 edge e_true = make_edge (split_bb, case_bb, EDGE_TRUE_VALUE); | |
| 1658 e_true->probability = prob; | |
| 1659 gcc_assert (e_true->src == split_bb); | |
| 1660 | |
| 1661 tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL, | |
| 1662 /*before=*/true, GSI_SAME_STMT); | |
| 1663 cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE); | |
| 1664 gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT); | |
| 1665 | |
| 1666 e_false = split_block (split_bb, cond_stmt); | |
| 1667 new_bb = e_false->dest; | |
| 1668 redirect_edge_pred (e_true, split_bb); | |
| 1669 | |
| 1670 e_false->flags &= ~EDGE_FALLTHRU; | |
| 1671 e_false->flags |= EDGE_FALSE_VALUE; | |
| 1672 e_false->probability = e_true->probability.invert (); | |
| 1673 new_bb->count = e_false->count (); | |
| 1674 | |
| 1675 return new_bb; | |
| 1676 } | |
| 1677 | |
| 1678 /* Compute the number of case labels that correspond to each outgoing edge of | |
| 1679 switch statement. Record this information in the aux field of the edge. */ | |
| 1680 | |
| 1681 void | |
| 1682 switch_decision_tree::compute_cases_per_edge () | |
| 1683 { | |
| 1684 reset_out_edges_aux (m_switch); | |
| 1685 int ncases = gimple_switch_num_labels (m_switch); | |
| 1686 for (int i = ncases - 1; i >= 1; --i) | |
| 1687 { | |
| 1688 edge case_edge = gimple_switch_edge (cfun, m_switch, i); | |
| 1689 case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + 1); | |
| 1690 } | |
| 1691 } | |
| 1692 | |
| 1693 /* Analyze switch statement and return true when the statement is expanded | |
| 1694 as decision tree. */ | |
| 1695 | |
| 1696 bool | |
| 1697 switch_decision_tree::analyze_switch_statement () | |
| 1698 { | |
| 1699 unsigned l = gimple_switch_num_labels (m_switch); | |
| 1700 basic_block bb = gimple_bb (m_switch); | |
| 1701 auto_vec<cluster *> clusters; | |
| 1702 clusters.create (l - 1); | |
| 1703 | |
| 1704 basic_block default_bb = gimple_switch_default_bb (cfun, m_switch); | |
| 1705 m_case_bbs.reserve (l); | |
| 1706 m_case_bbs.quick_push (default_bb); | |
| 1707 | |
| 1708 compute_cases_per_edge (); | |
| 1709 | |
| 1710 for (unsigned i = 1; i < l; i++) | |
| 1711 { | |
| 1712 tree elt = gimple_switch_label (m_switch, i); | |
| 1713 tree lab = CASE_LABEL (elt); | |
| 1714 basic_block case_bb = label_to_block (cfun, lab); | |
| 1715 edge case_edge = find_edge (bb, case_bb); | |
| 1716 tree low = CASE_LOW (elt); | |
| 1717 tree high = CASE_HIGH (elt); | |
| 1718 | |
| 1719 profile_probability p | |
| 1720 = case_edge->probability.apply_scale (1, (intptr_t) (case_edge->aux)); | |
| 1721 clusters.quick_push (new simple_cluster (low, high, elt, case_edge->dest, | |
| 1722 p)); | |
| 1723 m_case_bbs.quick_push (case_edge->dest); | |
| 1724 } | |
| 1725 | |
| 1726 reset_out_edges_aux (m_switch); | |
| 1727 | |
| 1728 /* Find jump table clusters. */ | |
| 1729 vec<cluster *> output = jump_table_cluster::find_jump_tables (clusters); | |
| 1730 | |
| 1731 /* Find bit test clusters. */ | |
| 1732 vec<cluster *> output2; | |
| 1733 auto_vec<cluster *> tmp; | |
| 1734 output2.create (1); | |
| 1735 tmp.create (1); | |
| 1736 | |
| 1737 for (unsigned i = 0; i < output.length (); i++) | |
| 1738 { | |
| 1739 cluster *c = output[i]; | |
| 1740 if (c->get_type () != SIMPLE_CASE) | |
| 1741 { | |
| 1742 if (!tmp.is_empty ()) | |
| 1743 { | |
| 1744 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp); | |
| 1745 output2.safe_splice (n); | |
| 1746 n.release (); | |
| 1747 tmp.truncate (0); | |
| 1748 } | |
| 1749 output2.safe_push (c); | |
| 1750 } | |
| 1751 else | |
| 1752 tmp.safe_push (c); | |
| 1753 } | |
| 1754 | |
| 1755 /* We still can have a temporary vector to test. */ | |
| 1756 if (!tmp.is_empty ()) | |
| 1757 { | |
| 1758 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp); | |
| 1759 output2.safe_splice (n); | |
| 1760 n.release (); | |
| 1761 } | |
| 1762 | |
| 1763 if (dump_file) | |
| 1764 { | |
| 1765 fprintf (dump_file, ";; GIMPLE switch case clusters: "); | |
| 1766 for (unsigned i = 0; i < output2.length (); i++) | |
| 1767 output2[i]->dump (dump_file, dump_flags & TDF_DETAILS); | |
| 1768 fprintf (dump_file, "\n"); | |
| 1769 } | |
| 1770 | |
| 1771 output.release (); | |
| 1772 | |
| 1773 bool expanded = try_switch_expansion (output2); | |
| 1774 | |
| 1775 for (unsigned i = 0; i < output2.length (); i++) | |
| 1776 delete output2[i]; | |
| 1777 | |
| 1778 output2.release (); | |
| 1779 | |
| 1780 return expanded; | |
| 1781 } | |
| 1782 | |
| 1783 /* Attempt to expand CLUSTERS as a decision tree. Return true when | |
| 1784 expanded. */ | |
| 1785 | |
| 1786 bool | |
| 1787 switch_decision_tree::try_switch_expansion (vec<cluster *> &clusters) | |
| 1788 { | |
| 1789 tree index_expr = gimple_switch_index (m_switch); | |
| 1790 tree index_type = TREE_TYPE (index_expr); | |
| 1791 basic_block bb = gimple_bb (m_switch); | |
| 1792 | |
| 1793 if (gimple_switch_num_labels (m_switch) == 1) | |
| 1794 return false; | |
| 1795 | |
| 1796 /* Find the default case target label. */ | |
| 1797 edge default_edge = gimple_switch_default_edge (cfun, m_switch); | |
| 1798 m_default_bb = default_edge->dest; | |
| 1799 | |
| 1800 /* Do the insertion of a case label into m_case_list. The labels are | |
| 1801 fed to us in descending order from the sorted vector of case labels used | |
| 1802 in the tree part of the middle end. So the list we construct is | |
| 1803 sorted in ascending order. */ | |
| 1804 | |
| 1805 for (int i = clusters.length () - 1; i >= 0; i--) | |
| 1806 { | |
| 1807 case_tree_node *r = m_case_list; | |
| 1808 m_case_list = m_case_node_pool.allocate (); | |
| 1809 m_case_list->m_right = r; | |
| 1810 m_case_list->m_c = clusters[i]; | |
| 1811 } | |
| 1812 | |
| 1813 record_phi_operand_mapping (); | |
| 1814 | |
| 1815 /* Split basic block that contains the gswitch statement. */ | |
| 1816 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 1817 edge e; | |
| 1818 if (gsi_end_p (gsi)) | |
| 1819 e = split_block_after_labels (bb); | |
| 1820 else | |
| 1821 { | |
| 1822 gsi_prev (&gsi); | |
| 1823 e = split_block (bb, gsi_stmt (gsi)); | |
| 1824 } | |
| 1825 bb = split_edge (e); | |
| 1826 | |
| 1827 /* Create new basic blocks for non-case clusters where specific expansion | |
| 1828 needs to happen. */ | |
| 1829 for (unsigned i = 0; i < clusters.length (); i++) | |
| 1830 if (clusters[i]->get_type () != SIMPLE_CASE) | |
| 1831 { | |
| 1832 clusters[i]->m_case_bb = create_empty_bb (bb); | |
| 1833 clusters[i]->m_case_bb->loop_father = bb->loop_father; | |
| 1834 } | |
| 1835 | |
| 1836 /* Do not do an extra work for a single cluster. */ | |
| 1837 if (clusters.length () == 1 | |
| 1838 && clusters[0]->get_type () != SIMPLE_CASE) | |
| 1839 { | |
| 1840 cluster *c = clusters[0]; | |
| 1841 c->emit (index_expr, index_type, | |
| 1842 gimple_switch_default_label (m_switch), m_default_bb); | |
| 1843 redirect_edge_succ (single_succ_edge (bb), c->m_case_bb); | |
| 1844 } | |
| 1845 else | |
| 1846 { | |
| 1847 emit (bb, index_expr, default_edge->probability, index_type); | |
| 1848 | |
| 1849 /* Emit cluster-specific switch handling. */ | |
| 1850 for (unsigned i = 0; i < clusters.length (); i++) | |
| 1851 if (clusters[i]->get_type () != SIMPLE_CASE) | |
| 1852 clusters[i]->emit (index_expr, index_type, | |
| 1853 gimple_switch_default_label (m_switch), | |
| 1854 m_default_bb); | |
| 1855 } | |
| 1856 | |
| 1857 fix_phi_operands_for_edges (); | |
| 1858 | |
| 1859 return true; | |
| 1860 } | |
| 1861 | |
| 1862 /* Before switch transformation, record all SSA_NAMEs defined in switch BB | |
| 1863 and used in a label basic block. */ | |
| 1864 | |
| 1865 void | |
| 1866 switch_decision_tree::record_phi_operand_mapping () | |
| 1867 { | |
| 1868 basic_block switch_bb = gimple_bb (m_switch); | |
| 1869 /* Record all PHI nodes that have to be fixed after conversion. */ | |
| 1870 for (unsigned i = 0; i < m_case_bbs.length (); i++) | |
| 1871 { | |
| 1872 gphi_iterator gsi; | |
| 1873 basic_block bb = m_case_bbs[i]; | |
| 1874 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 1875 { | |
| 1876 gphi *phi = gsi.phi (); | |
| 1877 | |
| 1878 for (unsigned i = 0; i < gimple_phi_num_args (phi); i++) | |
| 1879 { | |
| 1880 basic_block phi_src_bb = gimple_phi_arg_edge (phi, i)->src; | |
| 1881 if (phi_src_bb == switch_bb) | |
| 1882 { | |
| 1883 tree def = gimple_phi_arg_def (phi, i); | |
| 1884 tree result = gimple_phi_result (phi); | |
| 1885 m_phi_mapping.put (result, def); | |
| 1886 break; | |
| 1887 } | |
| 1888 } | |
| 1889 } | |
| 1890 } | |
| 1891 } | |
| 1892 | |
| 1893 /* Append new operands to PHI statements that were introduced due to | |
| 1894 addition of new edges to case labels. */ | |
| 1895 | |
| 1896 void | |
| 1897 switch_decision_tree::fix_phi_operands_for_edges () | |
| 1898 { | |
| 1899 gphi_iterator gsi; | |
| 1900 | |
| 1901 for (unsigned i = 0; i < m_case_bbs.length (); i++) | |
| 1902 { | |
| 1903 basic_block bb = m_case_bbs[i]; | |
| 1904 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 1905 { | |
| 1906 gphi *phi = gsi.phi (); | |
| 1907 for (unsigned j = 0; j < gimple_phi_num_args (phi); j++) | |
| 1908 { | |
| 1909 tree def = gimple_phi_arg_def (phi, j); | |
| 1910 if (def == NULL_TREE) | |
| 1911 { | |
| 1912 edge e = gimple_phi_arg_edge (phi, j); | |
| 1913 tree *definition | |
| 1914 = m_phi_mapping.get (gimple_phi_result (phi)); | |
| 1915 gcc_assert (definition); | |
| 1916 add_phi_arg (phi, *definition, e, UNKNOWN_LOCATION); | |
| 1917 } | |
| 1918 } | |
| 1919 } | |
| 1920 } | |
| 1921 } | |
| 1922 | |
| 1923 /* Generate a decision tree, switching on INDEX_EXPR and jumping to | |
| 1924 one of the labels in CASE_LIST or to the DEFAULT_LABEL. | |
| 1925 | |
| 1926 We generate a binary decision tree to select the appropriate target | |
| 1927 code. */ | |
| 1928 | |
| 1929 void | |
| 1930 switch_decision_tree::emit (basic_block bb, tree index_expr, | |
| 1931 profile_probability default_prob, tree index_type) | |
| 1932 { | |
| 1933 balance_case_nodes (&m_case_list, NULL); | |
| 1934 | |
| 1935 if (dump_file) | |
| 1936 dump_function_to_file (current_function_decl, dump_file, dump_flags); | |
| 1937 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1938 { | |
| 1939 int indent_step = ceil_log2 (TYPE_PRECISION (index_type)) + 2; | |
| 1940 fprintf (dump_file, ";; Expanding GIMPLE switch as decision tree:\n"); | |
| 1941 gcc_assert (m_case_list != NULL); | |
| 1942 dump_case_nodes (dump_file, m_case_list, indent_step, 0); | |
| 1943 } | |
| 1944 | |
| 1945 bb = emit_case_nodes (bb, index_expr, m_case_list, default_prob, index_type); | |
| 1946 | |
| 1947 if (bb) | |
| 1948 emit_jump (bb, m_default_bb); | |
| 1949 | |
| 1950 /* Remove all edges and do just an edge that will reach default_bb. */ | |
| 1951 bb = gimple_bb (m_switch); | |
| 1952 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 1953 gsi_remove (&gsi, true); | |
| 1954 | |
| 1955 delete_basic_block (bb); | |
| 1956 } | |
| 1957 | |
| 1958 /* Take an ordered list of case nodes | |
| 1959 and transform them into a near optimal binary tree, | |
| 1960 on the assumption that any target code selection value is as | |
| 1961 likely as any other. | |
| 1962 | |
| 1963 The transformation is performed by splitting the ordered | |
| 1964 list into two equal sections plus a pivot. The parts are | |
| 1965 then attached to the pivot as left and right branches. Each | |
| 1966 branch is then transformed recursively. */ | |
| 1967 | |
| 1968 void | |
| 1969 switch_decision_tree::balance_case_nodes (case_tree_node **head, | |
| 1970 case_tree_node *parent) | |
| 1971 { | |
| 1972 case_tree_node *np; | |
| 1973 | |
| 1974 np = *head; | |
| 1975 if (np) | |
| 1976 { | |
| 1977 int i = 0; | |
| 1978 int ranges = 0; | |
| 1979 case_tree_node **npp; | |
| 1980 case_tree_node *left; | |
| 1981 profile_probability prob = profile_probability::never (); | |
| 1982 | |
| 1983 /* Count the number of entries on branch. Also count the ranges. */ | |
| 1984 | |
| 1985 while (np) | |
| 1986 { | |
| 1987 if (!tree_int_cst_equal (np->m_c->get_low (), np->m_c->get_high ())) | |
| 1988 ranges++; | |
| 1989 | |
| 1990 i++; | |
| 1991 prob += np->m_c->m_prob; | |
| 1992 np = np->m_right; | |
| 1993 } | |
| 1994 | |
| 1995 if (i > 2) | |
| 1996 { | |
| 1997 /* Split this list if it is long enough for that to help. */ | |
| 1998 npp = head; | |
| 1999 left = *npp; | |
| 2000 profile_probability pivot_prob = prob.apply_scale (1, 2); | |
| 2001 | |
| 2002 /* Find the place in the list that bisects the list's total cost, | |
| 2003 where ranges count as 2. */ | |
| 2004 while (1) | |
| 2005 { | |
| 2006 /* Skip nodes while their probability does not reach | |
| 2007 that amount. */ | |
| 2008 prob -= (*npp)->m_c->m_prob; | |
| 2009 if ((prob.initialized_p () && prob < pivot_prob) | |
| 2010 || ! (*npp)->m_right) | |
| 2011 break; | |
| 2012 npp = &(*npp)->m_right; | |
| 2013 } | |
| 2014 | |
| 2015 np = *npp; | |
| 2016 *npp = 0; | |
| 2017 *head = np; | |
| 2018 np->m_parent = parent; | |
| 2019 np->m_left = left == np ? NULL : left; | |
| 2020 | |
| 2021 /* Optimize each of the two split parts. */ | |
| 2022 balance_case_nodes (&np->m_left, np); | |
| 2023 balance_case_nodes (&np->m_right, np); | |
| 2024 np->m_c->m_subtree_prob = np->m_c->m_prob; | |
| 2025 if (np->m_left) | |
| 2026 np->m_c->m_subtree_prob += np->m_left->m_c->m_subtree_prob; | |
| 2027 if (np->m_right) | |
| 2028 np->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; | |
| 2029 } | |
| 2030 else | |
| 2031 { | |
| 2032 /* Else leave this branch as one level, | |
| 2033 but fill in `parent' fields. */ | |
| 2034 np = *head; | |
| 2035 np->m_parent = parent; | |
| 2036 np->m_c->m_subtree_prob = np->m_c->m_prob; | |
| 2037 for (; np->m_right; np = np->m_right) | |
| 2038 { | |
| 2039 np->m_right->m_parent = np; | |
| 2040 (*head)->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; | |
| 2041 } | |
| 2042 } | |
| 2043 } | |
| 2044 } | |
| 2045 | |
| 2046 /* Dump ROOT, a list or tree of case nodes, to file. */ | |
| 2047 | |
| 2048 void | |
| 2049 switch_decision_tree::dump_case_nodes (FILE *f, case_tree_node *root, | |
| 2050 int indent_step, int indent_level) | |
| 2051 { | |
| 2052 if (root == 0) | |
| 2053 return; | |
| 2054 indent_level++; | |
| 2055 | |
| 2056 dump_case_nodes (f, root->m_left, indent_step, indent_level); | |
| 2057 | |
| 2058 fputs (";; ", f); | |
| 2059 fprintf (f, "%*s", indent_step * indent_level, ""); | |
| 2060 root->m_c->dump (f); | |
| 2061 root->m_c->m_prob.dump (f); | |
| 2062 fputs (" subtree: ", f); | |
| 2063 root->m_c->m_subtree_prob.dump (f); | |
| 2064 fputs (")\n", f); | |
| 2065 | |
| 2066 dump_case_nodes (f, root->m_right, indent_step, indent_level); | |
| 2067 } | |
| 2068 | |
| 2069 | |
| 2070 /* Add an unconditional jump to CASE_BB that happens in basic block BB. */ | |
| 2071 | |
| 2072 void | |
| 2073 switch_decision_tree::emit_jump (basic_block bb, basic_block case_bb) | |
| 2074 { | |
| 2075 edge e = single_succ_edge (bb); | |
| 2076 redirect_edge_succ (e, case_bb); | |
| 2077 } | |
| 2078 | |
| 2079 /* Generate code to compare OP0 with OP1 so that the condition codes are | |
| 2080 set and to jump to LABEL_BB if the condition is true. | |
| 2081 COMPARISON is the GIMPLE comparison (EQ, NE, GT, etc.). | |
| 2082 PROB is the probability of jumping to LABEL_BB. */ | |
| 2083 | |
| 2084 basic_block | |
| 2085 switch_decision_tree::emit_cmp_and_jump_insns (basic_block bb, tree op0, | |
| 2086 tree op1, tree_code comparison, | |
| 2087 basic_block label_bb, | |
| 2088 profile_probability prob) | |
| 2089 { | |
| 2090 // TODO: it's once called with lhs != index. | |
| 2091 op1 = fold_convert (TREE_TYPE (op0), op1); | |
| 2092 | |
| 2093 gcond *cond = gimple_build_cond (comparison, op0, op1, NULL_TREE, NULL_TREE); | |
| 2094 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 2095 gsi_insert_after (&gsi, cond, GSI_NEW_STMT); | |
| 2096 | |
| 2097 gcc_assert (single_succ_p (bb)); | |
| 2098 | |
| 2099 /* Make a new basic block where false branch will take place. */ | |
| 2100 edge false_edge = split_block (bb, cond); | |
| 2101 false_edge->flags = EDGE_FALSE_VALUE; | |
| 2102 false_edge->probability = prob.invert (); | |
| 2103 | |
| 2104 edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
| 2105 true_edge->probability = prob; | |
| 2106 | |
| 2107 return false_edge->dest; | |
| 2108 } | |
| 2109 | |
| 2110 /* Generate code to jump to LABEL if OP0 and OP1 are equal. | |
| 2111 PROB is the probability of jumping to LABEL_BB. | |
| 2112 BB is a basic block where the new condition will be placed. */ | |
| 2113 | |
| 2114 basic_block | |
| 2115 switch_decision_tree::do_jump_if_equal (basic_block bb, tree op0, tree op1, | |
| 2116 basic_block label_bb, | |
| 2117 profile_probability prob) | |
| 2118 { | |
| 2119 op1 = fold_convert (TREE_TYPE (op0), op1); | |
| 2120 | |
| 2121 gcond *cond = gimple_build_cond (EQ_EXPR, op0, op1, NULL_TREE, NULL_TREE); | |
| 2122 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 2123 gsi_insert_before (&gsi, cond, GSI_SAME_STMT); | |
| 2124 | |
| 2125 gcc_assert (single_succ_p (bb)); | |
| 2126 | |
| 2127 /* Make a new basic block where false branch will take place. */ | |
| 2128 edge false_edge = split_block (bb, cond); | |
| 2129 false_edge->flags = EDGE_FALSE_VALUE; | |
| 2130 false_edge->probability = prob.invert (); | |
| 2131 | |
| 2132 edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
| 2133 true_edge->probability = prob; | |
| 2134 | |
| 2135 return false_edge->dest; | |
| 2136 } | |
| 2137 | |
| 2138 /* Emit step-by-step code to select a case for the value of INDEX. | |
| 2139 The thus generated decision tree follows the form of the | |
| 2140 case-node binary tree NODE, whose nodes represent test conditions. | |
| 2141 DEFAULT_PROB is probability of cases leading to default BB. | |
| 2142 INDEX_TYPE is the type of the index of the switch. */ | |
| 2143 | |
| 2144 basic_block | |
| 2145 switch_decision_tree::emit_case_nodes (basic_block bb, tree index, | |
| 2146 case_tree_node *node, | |
| 2147 profile_probability default_prob, | |
| 2148 tree index_type) | |
| 2149 { | |
| 2150 profile_probability p; | |
| 2151 | |
| 2152 /* If node is null, we are done. */ | |
| 2153 if (node == NULL) | |
| 2154 return bb; | |
| 2155 | |
| 2156 /* Single value case. */ | |
| 2157 if (node->m_c->is_single_value_p ()) | |
| 2158 { | |
| 2159 /* Node is single valued. First see if the index expression matches | |
| 2160 this node and then check our children, if any. */ | |
| 2161 p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2162 bb = do_jump_if_equal (bb, index, node->m_c->get_low (), | |
| 2163 node->m_c->m_case_bb, p); | |
| 2164 /* Since this case is taken at this point, reduce its weight from | |
| 2165 subtree_weight. */ | |
| 2166 node->m_c->m_subtree_prob -= p; | |
| 2167 | |
| 2168 if (node->m_left != NULL && node->m_right != NULL) | |
| 2169 { | |
| 2170 /* 1) the node has both children | |
| 2171 | |
| 2172 If both children are single-valued cases with no | |
| 2173 children, finish up all the work. This way, we can save | |
| 2174 one ordered comparison. */ | |
| 2175 | |
| 2176 if (!node->m_left->has_child () | |
| 2177 && node->m_left->m_c->is_single_value_p () | |
| 2178 && !node->m_right->has_child () | |
| 2179 && node->m_right->m_c->is_single_value_p ()) | |
| 2180 { | |
| 2181 p = (node->m_right->m_c->m_prob | |
| 2182 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2183 bb = do_jump_if_equal (bb, index, node->m_right->m_c->get_low (), | |
| 2184 node->m_right->m_c->m_case_bb, p); | |
| 2185 | |
| 2186 p = (node->m_left->m_c->m_prob | |
| 2187 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2188 bb = do_jump_if_equal (bb, index, node->m_left->m_c->get_low (), | |
| 2189 node->m_left->m_c->m_case_bb, p); | |
| 2190 } | |
| 2191 else | |
| 2192 { | |
| 2193 /* Branch to a label where we will handle it later. */ | |
| 2194 basic_block test_bb = split_edge (single_succ_edge (bb)); | |
| 2195 redirect_edge_succ (single_pred_edge (test_bb), | |
| 2196 single_succ_edge (bb)->dest); | |
| 2197 | |
| 2198 p = ((node->m_right->m_c->m_subtree_prob | |
| 2199 + default_prob.apply_scale (1, 2)) | |
| 2200 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2201 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 2202 GT_EXPR, test_bb, p); | |
| 2203 default_prob = default_prob.apply_scale (1, 2); | |
| 2204 | |
| 2205 /* Handle the left-hand subtree. */ | |
| 2206 bb = emit_case_nodes (bb, index, node->m_left, | |
| 2207 default_prob, index_type); | |
| 2208 | |
| 2209 /* If the left-hand subtree fell through, | |
| 2210 don't let it fall into the right-hand subtree. */ | |
| 2211 if (bb && m_default_bb) | |
| 2212 emit_jump (bb, m_default_bb); | |
| 2213 | |
| 2214 bb = emit_case_nodes (test_bb, index, node->m_right, | |
| 2215 default_prob, index_type); | |
| 2216 } | |
| 2217 } | |
| 2218 else if (node->m_left == NULL && node->m_right != NULL) | |
| 2219 { | |
| 2220 /* 2) the node has only right child. */ | |
| 2221 | |
| 2222 /* Here we have a right child but no left so we issue a conditional | |
| 2223 branch to default and process the right child. | |
| 2224 | |
| 2225 Omit the conditional branch to default if the right child | |
| 2226 does not have any children and is single valued; it would | |
| 2227 cost too much space to save so little time. */ | |
| 2228 | |
| 2229 if (node->m_right->has_child () | |
| 2230 || !node->m_right->m_c->is_single_value_p ()) | |
| 2231 { | |
| 2232 p = (default_prob.apply_scale (1, 2) | |
| 2233 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2234 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_low (), | |
| 2235 LT_EXPR, m_default_bb, p); | |
| 2236 default_prob = default_prob.apply_scale (1, 2); | |
| 2237 | |
| 2238 bb = emit_case_nodes (bb, index, node->m_right, default_prob, | |
| 2239 index_type); | |
| 2240 } | |
| 2241 else | |
| 2242 { | |
| 2243 /* We cannot process node->right normally | |
| 2244 since we haven't ruled out the numbers less than | |
| 2245 this node's value. So handle node->right explicitly. */ | |
| 2246 p = (node->m_right->m_c->m_subtree_prob | |
| 2247 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2248 bb = do_jump_if_equal (bb, index, node->m_right->m_c->get_low (), | |
| 2249 node->m_right->m_c->m_case_bb, p); | |
| 2250 } | |
| 2251 } | |
| 2252 else if (node->m_left != NULL && node->m_right == NULL) | |
| 2253 { | |
| 2254 /* 3) just one subtree, on the left. Similar case as previous. */ | |
| 2255 | |
| 2256 if (node->m_left->has_child () | |
| 2257 || !node->m_left->m_c->is_single_value_p ()) | |
| 2258 { | |
| 2259 p = (default_prob.apply_scale (1, 2) | |
| 2260 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2261 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 2262 GT_EXPR, m_default_bb, p); | |
| 2263 default_prob = default_prob.apply_scale (1, 2); | |
| 2264 | |
| 2265 bb = emit_case_nodes (bb, index, node->m_left, default_prob, | |
| 2266 index_type); | |
| 2267 } | |
| 2268 else | |
| 2269 { | |
| 2270 /* We cannot process node->left normally | |
| 2271 since we haven't ruled out the numbers less than | |
| 2272 this node's value. So handle node->left explicitly. */ | |
| 2273 p = (node->m_left->m_c->m_subtree_prob | |
| 2274 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2275 bb = do_jump_if_equal (bb, index, node->m_left->m_c->get_low (), | |
| 2276 node->m_left->m_c->m_case_bb, p); | |
| 2277 } | |
| 2278 } | |
| 2279 } | |
| 2280 else | |
| 2281 { | |
| 2282 /* Node is a range. These cases are very similar to those for a single | |
| 2283 value, except that we do not start by testing whether this node | |
| 2284 is the one to branch to. */ | |
| 2285 if (node->has_child () || node->m_c->get_type () != SIMPLE_CASE) | |
| 2286 { | |
| 2287 /* Branch to a label where we will handle it later. */ | |
| 2288 basic_block test_bb = split_edge (single_succ_edge (bb)); | |
| 2289 redirect_edge_succ (single_pred_edge (test_bb), | |
| 2290 single_succ_edge (bb)->dest); | |
| 2291 | |
| 2292 | |
| 2293 profile_probability right_prob = profile_probability::never (); | |
| 2294 if (node->m_right) | |
| 2295 right_prob = node->m_right->m_c->m_subtree_prob; | |
| 2296 p = ((right_prob + default_prob.apply_scale (1, 2)) | |
| 2297 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2298 | |
| 2299 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 2300 GT_EXPR, test_bb, p); | |
| 2301 default_prob = default_prob.apply_scale (1, 2); | |
| 2302 | |
| 2303 /* Value belongs to this node or to the left-hand subtree. */ | |
| 2304 p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2305 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_low (), | |
| 2306 GE_EXPR, node->m_c->m_case_bb, p); | |
| 2307 | |
| 2308 /* Handle the left-hand subtree. */ | |
| 2309 bb = emit_case_nodes (bb, index, node->m_left, | |
| 2310 default_prob, index_type); | |
| 2311 | |
| 2312 /* If the left-hand subtree fell through, | |
| 2313 don't let it fall into the right-hand subtree. */ | |
| 2314 if (bb && m_default_bb) | |
| 2315 emit_jump (bb, m_default_bb); | |
| 2316 | |
| 2317 bb = emit_case_nodes (test_bb, index, node->m_right, | |
| 2318 default_prob, index_type); | |
| 2319 } | |
| 2320 else | |
| 2321 { | |
| 2322 /* Node has no children so we check low and high bounds to remove | |
| 2323 redundant tests. Only one of the bounds can exist, | |
| 2324 since otherwise this node is bounded--a case tested already. */ | |
| 2325 tree lhs, rhs; | |
| 2326 generate_range_test (bb, index, node->m_c->get_low (), | |
| 2327 node->m_c->get_high (), &lhs, &rhs); | |
| 2328 p = default_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2329 | |
| 2330 bb = emit_cmp_and_jump_insns (bb, lhs, rhs, GT_EXPR, | |
| 2331 m_default_bb, p); | |
| 2332 | |
| 2333 emit_jump (bb, node->m_c->m_case_bb); | |
| 2334 return NULL; | |
| 2335 } | |
| 2336 } | |
| 2337 | |
| 2338 return bb; | |
| 0 | 2339 } |
| 2340 | |
| 2341 /* The main function of the pass scans statements for switches and invokes | |
| 2342 process_switch on them. */ | |
| 2343 | |
| 111 | 2344 namespace { |
| 2345 | |
| 2346 const pass_data pass_data_convert_switch = | |
| 2347 { | |
| 2348 GIMPLE_PASS, /* type */ | |
| 2349 "switchconv", /* name */ | |
| 2350 OPTGROUP_NONE, /* optinfo_flags */ | |
| 2351 TV_TREE_SWITCH_CONVERSION, /* tv_id */ | |
| 2352 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 2353 0, /* properties_provided */ | |
| 2354 0, /* properties_destroyed */ | |
| 2355 0, /* todo_flags_start */ | |
| 2356 TODO_update_ssa, /* todo_flags_finish */ | |
| 2357 }; | |
| 2358 | |
| 2359 class pass_convert_switch : public gimple_opt_pass | |
| 2360 { | |
| 2361 public: | |
| 2362 pass_convert_switch (gcc::context *ctxt) | |
| 2363 : gimple_opt_pass (pass_data_convert_switch, ctxt) | |
| 2364 {} | |
| 2365 | |
| 2366 /* opt_pass methods: */ | |
| 2367 virtual bool gate (function *) { return flag_tree_switch_conversion != 0; } | |
| 2368 virtual unsigned int execute (function *); | |
| 2369 | |
| 2370 }; // class pass_convert_switch | |
| 2371 | |
| 2372 unsigned int | |
| 2373 pass_convert_switch::execute (function *fun) | |
| 0 | 2374 { |
| 2375 basic_block bb; | |
| 131 | 2376 bool cfg_altered = false; |
| 0 | 2377 |
| 111 | 2378 FOR_EACH_BB_FN (bb, fun) |
| 0 | 2379 { |
| 111 | 2380 gimple *stmt = last_stmt (bb); |
| 0 | 2381 if (stmt && gimple_code (stmt) == GIMPLE_SWITCH) |
| 2382 { | |
| 2383 if (dump_file) | |
| 2384 { | |
| 2385 expanded_location loc = expand_location (gimple_location (stmt)); | |
| 2386 | |
| 2387 fprintf (dump_file, "beginning to process the following " | |
| 2388 "SWITCH statement (%s:%d) : ------- \n", | |
| 2389 loc.file, loc.line); | |
| 2390 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2391 putc ('\n', dump_file); |
| 0 | 2392 } |
| 2393 | |
| 131 | 2394 switch_conversion sconv; |
| 2395 sconv.expand (as_a <gswitch *> (stmt)); | |
| 2396 cfg_altered |= sconv.m_cfg_altered; | |
| 2397 if (!sconv.m_reason) | |
| 0 | 2398 { |
| 2399 if (dump_file) | |
| 2400 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2401 fputs ("Switch converted\n", dump_file); |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2402 fputs ("--------------------------------\n", dump_file); |
| 0 | 2403 } |
| 111 | 2404 |
| 131 | 2405 /* Make no effort to update the post-dominator tree. |
| 2406 It is actually not that hard for the transformations | |
| 2407 we have performed, but it is not supported | |
| 2408 by iterate_fix_dominators. */ | |
| 111 | 2409 free_dominance_info (CDI_POST_DOMINATORS); |
| 0 | 2410 } |
| 2411 else | |
| 2412 { | |
| 2413 if (dump_file) | |
| 2414 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2415 fputs ("Bailing out - ", dump_file); |
| 131 | 2416 fputs (sconv.m_reason, dump_file); |
| 111 | 2417 fputs ("\n--------------------------------\n", dump_file); |
| 0 | 2418 } |
| 2419 } | |
| 2420 } | |
| 2421 } | |
| 2422 | |
| 131 | 2423 return cfg_altered ? TODO_cleanup_cfg : 0;; |
| 0 | 2424 } |
| 2425 | |
| 111 | 2426 } // anon namespace |
| 2427 | |
| 2428 gimple_opt_pass * | |
| 2429 make_pass_convert_switch (gcc::context *ctxt) | |
| 2430 { | |
| 2431 return new pass_convert_switch (ctxt); | |
| 2432 } | |
| 2433 | |
| 2434 /* The main function of the pass scans statements for switches and invokes | |
| 2435 process_switch on them. */ | |
| 2436 | |
| 2437 namespace { | |
| 2438 | |
| 131 | 2439 template <bool O0> class pass_lower_switch: public gimple_opt_pass |
| 111 | 2440 { |
| 131 | 2441 public: |
| 2442 pass_lower_switch (gcc::context *ctxt) : gimple_opt_pass (data, ctxt) {} | |
| 2443 | |
| 2444 static const pass_data data; | |
| 2445 opt_pass * | |
| 2446 clone () | |
| 2447 { | |
| 2448 return new pass_lower_switch<O0> (m_ctxt); | |
| 2449 } | |
| 2450 | |
| 2451 virtual bool | |
| 2452 gate (function *) | |
| 2453 { | |
| 2454 return !O0 || !optimize; | |
| 2455 } | |
| 2456 | |
| 2457 virtual unsigned int execute (function *fun); | |
| 2458 }; // class pass_lower_switch | |
| 2459 | |
| 2460 template <bool O0> | |
| 2461 const pass_data pass_lower_switch<O0>::data = { | |
| 2462 GIMPLE_PASS, /* type */ | |
| 2463 O0 ? "switchlower_O0" : "switchlower", /* name */ | |
| 111 | 2464 OPTGROUP_NONE, /* optinfo_flags */ |
| 2465 TV_TREE_SWITCH_LOWERING, /* tv_id */ | |
| 2466 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 2467 0, /* properties_provided */ | |
| 2468 0, /* properties_destroyed */ | |
| 2469 0, /* todo_flags_start */ | |
| 2470 TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ | |
| 2471 }; | |
| 2472 | |
| 131 | 2473 template <bool O0> |
| 111 | 2474 unsigned int |
| 131 | 2475 pass_lower_switch<O0>::execute (function *fun) |
| 111 | 2476 { |
| 2477 basic_block bb; | |
| 2478 bool expanded = false; | |
| 2479 | |
| 131 | 2480 auto_vec<gimple *> switch_statements; |
| 2481 switch_statements.create (1); | |
| 2482 | |
| 111 | 2483 FOR_EACH_BB_FN (bb, fun) |
| 2484 { | |
| 2485 gimple *stmt = last_stmt (bb); | |
| 131 | 2486 gswitch *swtch; |
| 2487 if (stmt && (swtch = dyn_cast<gswitch *> (stmt))) | |
| 2488 { | |
| 2489 if (!O0) | |
| 2490 group_case_labels_stmt (swtch); | |
| 2491 switch_statements.safe_push (swtch); | |
| 2492 } | |
| 2493 } | |
| 2494 | |
| 2495 for (unsigned i = 0; i < switch_statements.length (); i++) | |
| 2496 { | |
| 2497 gimple *stmt = switch_statements[i]; | |
| 2498 if (dump_file) | |
| 111 | 2499 { |
| 131 | 2500 expanded_location loc = expand_location (gimple_location (stmt)); |
| 2501 | |
| 2502 fprintf (dump_file, "beginning to process the following " | |
| 2503 "SWITCH statement (%s:%d) : ------- \n", | |
| 2504 loc.file, loc.line); | |
| 2505 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
| 2506 putc ('\n', dump_file); | |
| 2507 } | |
| 2508 | |
| 2509 gswitch *swtch = dyn_cast<gswitch *> (stmt); | |
| 2510 if (swtch) | |
| 2511 { | |
| 2512 switch_decision_tree dt (swtch); | |
| 2513 expanded |= dt.analyze_switch_statement (); | |
| 111 | 2514 } |
| 2515 } | |
| 2516 | |
| 2517 if (expanded) | |
| 2518 { | |
| 2519 free_dominance_info (CDI_DOMINATORS); | |
| 2520 free_dominance_info (CDI_POST_DOMINATORS); | |
| 2521 mark_virtual_operands_for_renaming (cfun); | |
| 2522 } | |
| 2523 | |
| 2524 return 0; | |
| 2525 } | |
| 2526 | |
| 2527 } // anon namespace | |
| 2528 | |
| 2529 gimple_opt_pass * | |
| 131 | 2530 make_pass_lower_switch_O0 (gcc::context *ctxt) |
| 2531 { | |
| 2532 return new pass_lower_switch<true> (ctxt); | |
| 2533 } | |
| 2534 gimple_opt_pass * | |
| 111 | 2535 make_pass_lower_switch (gcc::context *ctxt) |
| 2536 { | |
| 131 | 2537 return new pass_lower_switch<false> (ctxt); |
| 111 | 2538 } |
| 2539 | |
| 131 | 2540 |