Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-switch-conversion.c @ 158:494b0b89df80 default tip
...
| author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Mon, 25 May 2020 18:13:55 +0900 |
| parents | 1830386684a0 |
| children |
| rev | line source |
|---|---|
| 111 | 1 /* Lower GIMPLE_SWITCH expressions to something more efficient than |
| 2 a jump table. | |
| 145 | 3 Copyright (C) 2006-2020 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 "fold-const.h" | |
| 40 #include "varasm.h" | |
| 41 #include "stor-layout.h" | |
| 42 #include "cfganal.h" | |
| 43 #include "gimplify.h" | |
| 44 #include "gimple-iterator.h" | |
| 45 #include "gimplify-me.h" | |
| 131 | 46 #include "gimple-fold.h" |
| 111 | 47 #include "tree-cfg.h" |
| 48 #include "cfgloop.h" | |
| 49 #include "alloc-pool.h" | |
| 50 #include "target.h" | |
| 51 #include "tree-into-ssa.h" | |
| 131 | 52 #include "omp-general.h" |
| 111 | 53 |
| 54 /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode | |
| 55 type in the GIMPLE type system that is language-independent? */ | |
| 56 #include "langhooks.h" | |
| 57 | |
| 131 | 58 #include "tree-switch-conversion.h" |
| 111 | 59 |
| 131 | 60 using namespace tree_switch_conversion; |
| 61 | |
| 62 /* Constructor. */ | |
| 63 | |
| 145 | 64 switch_conversion::switch_conversion (): m_final_bb (NULL), |
| 131 | 65 m_constructors (NULL), m_default_values (NULL), |
| 66 m_arr_ref_first (NULL), m_arr_ref_last (NULL), | |
| 67 m_reason (NULL), m_default_case_nonstandard (false), m_cfg_altered (false) | |
| 111 | 68 { |
| 69 } | |
| 131 | 70 |
| 71 /* Collection information about SWTCH statement. */ | |
| 72 | |
| 73 void | |
| 74 switch_conversion::collect (gswitch *swtch) | |
| 111 | 75 { |
| 76 unsigned int branch_num = gimple_switch_num_labels (swtch); | |
| 77 tree min_case, max_case; | |
| 131 | 78 unsigned int i; |
| 111 | 79 edge e, e_default, e_first; |
| 80 edge_iterator ei; | |
| 131 | 81 |
| 82 m_switch = swtch; | |
| 111 | 83 |
| 84 /* The gimplifier has already sorted the cases by CASE_LOW and ensured there | |
| 85 is a default label which is the first in the vector. | |
| 86 Collect the bits we can deduce from the CFG. */ | |
| 131 | 87 m_index_expr = gimple_switch_index (swtch); |
| 88 m_switch_bb = gimple_bb (swtch); | |
| 89 e_default = gimple_switch_default_edge (cfun, swtch); | |
| 90 m_default_bb = e_default->dest; | |
| 91 m_default_prob = e_default->probability; | |
| 111 | 92 |
| 93 /* Get upper and lower bounds of case values, and the covered range. */ | |
| 94 min_case = gimple_switch_label (swtch, 1); | |
| 95 max_case = gimple_switch_label (swtch, branch_num - 1); | |
| 96 | |
| 131 | 97 m_range_min = CASE_LOW (min_case); |
| 111 | 98 if (CASE_HIGH (max_case) != NULL_TREE) |
| 131 | 99 m_range_max = CASE_HIGH (max_case); |
| 111 | 100 else |
| 131 | 101 m_range_max = CASE_LOW (max_case); |
| 102 | |
| 103 m_contiguous_range = true; | |
| 104 tree last = CASE_HIGH (min_case) ? CASE_HIGH (min_case) : m_range_min; | |
| 111 | 105 for (i = 2; i < branch_num; i++) |
| 106 { | |
| 107 tree elt = gimple_switch_label (swtch, i); | |
| 108 if (wi::to_wide (last) + 1 != wi::to_wide (CASE_LOW (elt))) | |
| 109 { | |
| 131 | 110 m_contiguous_range = false; |
| 111 | 111 break; |
| 112 } | |
| 113 last = CASE_HIGH (elt) ? CASE_HIGH (elt) : CASE_LOW (elt); | |
| 114 } | |
| 115 | |
| 131 | 116 if (m_contiguous_range) |
| 117 e_first = gimple_switch_edge (cfun, swtch, 1); | |
| 111 | 118 else |
| 131 | 119 e_first = e_default; |
| 111 | 120 |
| 121 /* See if there is one common successor block for all branch | |
| 122 targets. If it exists, record it in FINAL_BB. | |
| 123 Start with the destination of the first non-default case | |
| 124 if the range is contiguous and default case otherwise as | |
| 125 guess or its destination in case it is a forwarder block. */ | |
| 126 if (! single_pred_p (e_first->dest)) | |
| 131 | 127 m_final_bb = e_first->dest; |
| 111 | 128 else if (single_succ_p (e_first->dest) |
| 129 && ! single_pred_p (single_succ (e_first->dest))) | |
| 131 | 130 m_final_bb = single_succ (e_first->dest); |
| 111 | 131 /* Require that all switch destinations are either that common |
| 132 FINAL_BB or a forwarder to it, except for the default | |
| 133 case if contiguous range. */ | |
| 131 | 134 if (m_final_bb) |
| 135 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) | |
| 111 | 136 { |
| 131 | 137 if (e->dest == m_final_bb) |
| 111 | 138 continue; |
| 139 | |
| 140 if (single_pred_p (e->dest) | |
| 141 && single_succ_p (e->dest) | |
| 131 | 142 && single_succ (e->dest) == m_final_bb) |
| 111 | 143 continue; |
| 144 | |
| 131 | 145 if (e == e_default && m_contiguous_range) |
| 111 | 146 { |
| 131 | 147 m_default_case_nonstandard = true; |
| 111 | 148 continue; |
| 149 } | |
| 150 | |
| 131 | 151 m_final_bb = NULL; |
| 111 | 152 break; |
| 153 } | |
| 154 | |
| 131 | 155 m_range_size |
| 156 = int_const_binop (MINUS_EXPR, m_range_max, m_range_min); | |
| 111 | 157 |
| 158 /* Get a count of the number of case labels. Single-valued case labels | |
| 159 simply count as one, but a case range counts double, since it may | |
| 160 require two compares if it gets lowered as a branching tree. */ | |
| 131 | 161 m_count = 0; |
| 111 | 162 for (i = 1; i < branch_num; i++) |
| 163 { | |
| 164 tree elt = gimple_switch_label (swtch, i); | |
| 131 | 165 m_count++; |
| 111 | 166 if (CASE_HIGH (elt) |
| 167 && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt))) | |
| 131 | 168 m_count++; |
| 111 | 169 } |
| 131 | 170 |
| 111 | 171 /* Get the number of unique non-default targets out of the GIMPLE_SWITCH |
| 172 block. Assume a CFG cleanup would have already removed degenerate | |
| 173 switch statements, this allows us to just use EDGE_COUNT. */ | |
| 131 | 174 m_uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1; |
| 111 | 175 } |
| 0 | 176 |
| 131 | 177 /* Checks whether the range given by individual case statements of the switch |
| 0 | 178 switch statement isn't too big and whether the number of branches actually |
| 179 satisfies the size of the new array. */ | |
| 180 | |
| 131 | 181 bool |
| 182 switch_conversion::check_range () | |
| 0 | 183 { |
| 131 | 184 gcc_assert (m_range_size); |
| 185 if (!tree_fits_uhwi_p (m_range_size)) | |
| 0 | 186 { |
| 131 | 187 m_reason = "index range way too large or otherwise unusable"; |
| 0 | 188 return false; |
| 189 } | |
| 190 | |
| 131 | 191 if (tree_to_uhwi (m_range_size) |
| 145 | 192 > ((unsigned) m_count * param_switch_conversion_branch_ratio)) |
| 0 | 193 { |
| 131 | 194 m_reason = "the maximum range-branch ratio exceeded"; |
| 0 | 195 return false; |
| 196 } | |
| 197 | |
| 198 return true; | |
| 199 } | |
| 200 | |
| 131 | 201 /* Checks whether all but the final BB basic blocks are empty. */ |
| 202 | |
| 203 bool | |
| 204 switch_conversion::check_all_empty_except_final () | |
| 0 | 205 { |
| 131 | 206 edge e, e_default = find_edge (m_switch_bb, m_default_bb); |
| 111 | 207 edge_iterator ei; |
| 208 | |
| 131 | 209 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
|
210 { |
| 131 | 211 if (e->dest == m_final_bb) |
| 111 | 212 continue; |
| 213 | |
| 214 if (!empty_block_p (e->dest)) | |
| 0 | 215 { |
| 131 | 216 if (m_contiguous_range && e == e_default) |
| 111 | 217 { |
| 131 | 218 m_default_case_nonstandard = true; |
| 111 | 219 continue; |
| 220 } | |
| 221 | |
| 131 | 222 m_reason = "bad case - a non-final BB not empty"; |
| 0 | 223 return false; |
| 224 } | |
| 225 } | |
| 226 | |
| 227 return true; | |
| 228 } | |
| 229 | |
| 230 /* This function checks whether all required values in phi nodes in final_bb | |
| 231 are constants. Required values are those that correspond to a basic block | |
| 232 which is a part of the examined switch statement. It returns true if the | |
| 233 phi nodes are OK, otherwise false. */ | |
| 234 | |
| 131 | 235 bool |
| 236 switch_conversion::check_final_bb () | |
| 0 | 237 { |
| 111 | 238 gphi_iterator gsi; |
| 239 | |
| 131 | 240 m_phi_count = 0; |
| 241 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 0 | 242 { |
| 111 | 243 gphi *phi = gsi.phi (); |
| 0 | 244 unsigned int i; |
| 245 | |
| 111 | 246 if (virtual_operand_p (gimple_phi_result (phi))) |
| 247 continue; | |
| 248 | |
| 131 | 249 m_phi_count++; |
| 0 | 250 |
| 251 for (i = 0; i < gimple_phi_num_args (phi); i++) | |
| 252 { | |
| 253 basic_block bb = gimple_phi_arg_edge (phi, i)->src; | |
| 254 | |
| 131 | 255 if (bb == m_switch_bb |
| 111 | 256 || (single_pred_p (bb) |
| 131 | 257 && single_pred (bb) == m_switch_bb |
| 258 && (!m_default_case_nonstandard | |
| 111 | 259 || empty_block_p (bb)))) |
| 0 | 260 { |
| 261 tree reloc, val; | |
| 111 | 262 const char *reason = NULL; |
| 0 | 263 |
| 264 val = gimple_phi_arg_def (phi, i); | |
| 265 if (!is_gimple_ip_invariant (val)) | |
| 111 | 266 reason = "non-invariant value from a case"; |
| 267 else | |
| 0 | 268 { |
| 111 | 269 reloc = initializer_constant_valid_p (val, TREE_TYPE (val)); |
| 270 if ((flag_pic && reloc != null_pointer_node) | |
| 271 || (!flag_pic && reloc == NULL_TREE)) | |
| 272 { | |
| 273 if (reloc) | |
| 274 reason | |
| 275 = "value from a case would need runtime relocations"; | |
| 276 else | |
| 277 reason | |
| 278 = "value from a case is not a valid initializer"; | |
| 279 } | |
| 0 | 280 } |
| 111 | 281 if (reason) |
| 0 | 282 { |
| 111 | 283 /* For contiguous range, we can allow non-constant |
| 284 or one that needs relocation, as long as it is | |
| 285 only reachable from the default case. */ | |
| 131 | 286 if (bb == m_switch_bb) |
| 287 bb = m_final_bb; | |
| 288 if (!m_contiguous_range || bb != m_default_bb) | |
| 111 | 289 { |
| 131 | 290 m_reason = reason; |
| 111 | 291 return false; |
| 292 } | |
| 293 | |
| 131 | 294 unsigned int branch_num = gimple_switch_num_labels (m_switch); |
| 111 | 295 for (unsigned int i = 1; i < branch_num; i++) |
| 296 { | |
| 131 | 297 if (gimple_switch_label_bb (cfun, m_switch, i) == bb) |
| 111 | 298 { |
| 131 | 299 m_reason = reason; |
| 111 | 300 return false; |
| 301 } | |
| 302 } | |
| 131 | 303 m_default_case_nonstandard = true; |
| 0 | 304 } |
| 305 } | |
| 306 } | |
| 307 } | |
| 308 | |
| 309 return true; | |
| 310 } | |
| 311 | |
| 312 /* The following function allocates default_values, target_{in,out}_names and | |
| 313 constructors arrays. The last one is also populated with pointers to | |
| 314 vectors that will become constructors of new arrays. */ | |
| 315 | |
| 131 | 316 void |
| 317 switch_conversion::create_temp_arrays () | |
| 0 | 318 { |
| 319 int i; | |
| 320 | |
| 131 | 321 m_default_values = XCNEWVEC (tree, m_phi_count * 3); |
| 111 | 322 /* ??? Macros do not support multi argument templates in their |
| 323 argument list. We create a typedef to work around that problem. */ | |
| 324 typedef vec<constructor_elt, va_gc> *vec_constructor_elt_gc; | |
| 131 | 325 m_constructors = XCNEWVEC (vec_constructor_elt_gc, m_phi_count); |
| 326 m_target_inbound_names = m_default_values + m_phi_count; | |
| 327 m_target_outbound_names = m_target_inbound_names + m_phi_count; | |
| 328 for (i = 0; i < m_phi_count; i++) | |
| 329 vec_alloc (m_constructors[i], tree_to_uhwi (m_range_size) + 1); | |
| 0 | 330 } |
| 331 | |
| 332 /* Populate the array of default values in the order of phi nodes. | |
| 111 | 333 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch |
| 334 if the range is non-contiguous or the default case has standard | |
| 335 structure, otherwise it is the first non-default case instead. */ | |
| 0 | 336 |
| 131 | 337 void |
| 338 switch_conversion::gather_default_values (tree default_case) | |
| 0 | 339 { |
| 111 | 340 gphi_iterator gsi; |
| 131 | 341 basic_block bb = label_to_block (cfun, CASE_LABEL (default_case)); |
| 0 | 342 edge e; |
| 343 int i = 0; | |
| 344 | |
| 111 | 345 gcc_assert (CASE_LOW (default_case) == NULL_TREE |
| 131 | 346 || m_default_case_nonstandard); |
| 347 | |
| 348 if (bb == m_final_bb) | |
| 349 e = find_edge (m_switch_bb, bb); | |
| 0 | 350 else |
| 351 e = single_succ_edge (bb); | |
| 352 | |
| 131 | 353 for (gsi = gsi_start_phis (m_final_bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
| 0 | 354 { |
| 111 | 355 gphi *phi = gsi.phi (); |
| 356 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 357 continue; | |
| 0 | 358 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
| 359 gcc_assert (val); | |
| 131 | 360 m_default_values[i++] = val; |
| 0 | 361 } |
| 362 } | |
| 363 | |
| 364 /* The following function populates the vectors in the constructors array with | |
| 365 future contents of the static arrays. The vectors are populated in the | |
| 131 | 366 order of phi nodes. */ |
| 367 | |
| 368 void | |
| 369 switch_conversion::build_constructors () | |
| 0 | 370 { |
| 131 | 371 unsigned i, branch_num = gimple_switch_num_labels (m_switch); |
| 372 tree pos = m_range_min; | |
| 111 | 373 tree pos_one = build_int_cst (TREE_TYPE (pos), 1); |
| 0 | 374 |
| 375 for (i = 1; i < branch_num; i++) | |
| 376 { | |
| 131 | 377 tree cs = gimple_switch_label (m_switch, i); |
| 378 basic_block bb = label_to_block (cfun, CASE_LABEL (cs)); | |
| 0 | 379 edge e; |
| 380 tree high; | |
| 111 | 381 gphi_iterator gsi; |
| 0 | 382 int j; |
| 383 | |
| 131 | 384 if (bb == m_final_bb) |
| 385 e = find_edge (m_switch_bb, bb); | |
| 0 | 386 else |
| 387 e = single_succ_edge (bb); | |
| 388 gcc_assert (e); | |
| 389 | |
| 390 while (tree_int_cst_lt (pos, CASE_LOW (cs))) | |
| 391 { | |
| 392 int k; | |
| 131 | 393 for (k = 0; k < m_phi_count; k++) |
| 0 | 394 { |
| 111 | 395 constructor_elt elt; |
| 396 | |
| 131 | 397 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
| 111 | 398 elt.value |
| 131 | 399 = unshare_expr_without_location (m_default_values[k]); |
| 400 m_constructors[k]->quick_push (elt); | |
| 0 | 401 } |
| 402 | |
| 111 | 403 pos = int_const_binop (PLUS_EXPR, pos, pos_one); |
| 0 | 404 } |
| 405 gcc_assert (tree_int_cst_equal (pos, CASE_LOW (cs))); | |
| 406 | |
| 407 j = 0; | |
| 408 if (CASE_HIGH (cs)) | |
| 409 high = CASE_HIGH (cs); | |
| 410 else | |
| 411 high = CASE_LOW (cs); | |
| 131 | 412 for (gsi = gsi_start_phis (m_final_bb); |
| 0 | 413 !gsi_end_p (gsi); gsi_next (&gsi)) |
| 414 { | |
| 111 | 415 gphi *phi = gsi.phi (); |
| 416 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 417 continue; | |
| 0 | 418 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e); |
| 419 tree low = CASE_LOW (cs); | |
| 420 pos = CASE_LOW (cs); | |
| 421 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
422 do |
| 0 | 423 { |
| 111 | 424 constructor_elt elt; |
| 425 | |
| 131 | 426 elt.index = int_const_binop (MINUS_EXPR, pos, m_range_min); |
| 111 | 427 elt.value = unshare_expr_without_location (val); |
| 131 | 428 m_constructors[j]->quick_push (elt); |
| 111 | 429 |
| 430 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
|
431 } 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
|
432 && tree_int_cst_lt (low, pos)); |
| 0 | 433 j++; |
| 434 } | |
| 435 } | |
| 436 } | |
| 437 | |
| 131 | 438 /* If all values in the constructor vector are products of a linear function |
| 439 a * x + b, then return true. When true, COEFF_A and COEFF_B and | |
| 440 coefficients of the linear function. Note that equal values are special | |
| 441 case of a linear function with a and b equal to zero. */ | |
| 442 | |
| 443 bool | |
| 444 switch_conversion::contains_linear_function_p (vec<constructor_elt, va_gc> *vec, | |
| 445 wide_int *coeff_a, | |
| 446 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
|
447 { |
|
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
|
448 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
|
449 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
|
450 |
| 131 | 451 gcc_assert (vec->length () >= 2); |
| 452 | |
| 453 /* Let's try to find any linear function a * x + y that can apply to | |
| 454 given values. 'a' can be calculated as follows: | |
| 455 | |
| 456 a = (y2 - y1) / (x2 - x1) where x2 - x1 = 1 (consecutive case indices) | |
| 457 a = y2 - y1 | |
| 458 | |
| 459 and | |
| 460 | |
| 461 b = y2 - a * x2 | |
| 462 | |
| 463 */ | |
| 464 | |
| 465 tree elt0 = (*vec)[0].value; | |
| 466 tree elt1 = (*vec)[1].value; | |
| 467 | |
| 468 if (TREE_CODE (elt0) != INTEGER_CST || TREE_CODE (elt1) != INTEGER_CST) | |
| 469 return false; | |
| 470 | |
| 145 | 471 wide_int range_min |
| 472 = wide_int::from (wi::to_wide (m_range_min), | |
| 473 TYPE_PRECISION (TREE_TYPE (elt0)), | |
| 474 TYPE_SIGN (TREE_TYPE (m_range_min))); | |
| 131 | 475 wide_int y1 = wi::to_wide (elt0); |
| 476 wide_int y2 = wi::to_wide (elt1); | |
| 477 wide_int a = y2 - y1; | |
| 478 wide_int b = y2 - a * (range_min + 1); | |
| 479 | |
| 480 /* Verify that all values fulfill the linear function. */ | |
| 111 | 481 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
|
482 { |
| 131 | 483 if (TREE_CODE (elt->value) != INTEGER_CST) |
| 484 return false; | |
| 485 | |
| 486 wide_int value = wi::to_wide (elt->value); | |
| 487 if (a * range_min + b != value) | |
| 488 return false; | |
| 489 | |
| 490 ++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
|
491 } |
| 131 | 492 |
| 493 *coeff_a = a; | |
| 494 *coeff_b = b; | |
| 495 | |
| 496 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
|
497 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
498 |
| 111 | 499 /* Return type which should be used for array elements, either TYPE's |
| 500 main variant or, for integral types, some smaller integral type | |
| 501 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
|
502 |
| 131 | 503 tree |
| 504 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
|
505 { |
| 131 | 506 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
|
507 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
|
508 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
|
509 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
|
510 |
| 111 | 511 /* Types with alignments greater than their size can reach here, e.g. out of |
| 512 SRA. We couldn't use these as an array component type so get back to the | |
| 513 main variant first, which, for our purposes, is fine for other types as | |
| 514 well. */ | |
| 515 | |
| 516 type = TYPE_MAIN_VARIANT (type); | |
| 517 | |
|
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
|
518 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
|
519 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
|
520 |
| 111 | 521 scalar_int_mode type_mode = SCALAR_INT_TYPE_MODE (type); |
| 522 scalar_int_mode mode = get_narrowest_mode (type_mode); | |
| 523 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
|
524 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
|
525 |
| 131 | 526 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
|
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 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
|
530 { |
| 111 | 531 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
|
532 |
|
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 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
|
534 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
|
535 |
| 111 | 536 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
|
537 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
|
538 { |
|
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
|
539 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
|
540 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
|
541 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
|
542 |
| 111 | 543 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
|
544 { |
| 111 | 545 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
|
546 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
|
547 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
|
548 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
|
549 } |
| 111 | 550 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
|
551 { |
|
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 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
|
553 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
|
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 |
|
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 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
|
557 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
|
558 |
| 111 | 559 if (!GET_MODE_WIDER_MODE (mode).exists (&mode) |
| 560 || 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
|
561 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
|
562 } |
|
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
|
563 } |
|
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 |
|
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 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
|
566 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
|
567 smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0); |
| 111 | 568 if (GET_MODE_SIZE (type_mode) |
| 569 <= 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
|
570 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
|
571 |
|
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
|
572 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
|
573 } |
|
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 |
| 131 | 575 /* Create an appropriate array type and declaration and assemble a static |
| 576 array variable. Also create a load statement that initializes | |
| 577 the variable in question with a value from the static array. SWTCH is | |
| 578 the switch statement being converted, NUM is the index to | |
| 579 arrays of constructors, default values and target SSA names | |
| 580 for this particular array. ARR_INDEX_TYPE is the type of the index | |
| 581 of the new array, PHI is the phi node of the final BB that corresponds | |
| 582 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
|
583 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
|
584 new array. */ |
| 0 | 585 |
| 131 | 586 void |
| 587 switch_conversion::build_one_array (int num, tree arr_index_type, | |
| 588 gphi *phi, tree tidx) | |
| 0 | 589 { |
| 131 | 590 tree name; |
| 111 | 591 gimple *load; |
| 131 | 592 gimple_stmt_iterator gsi = gsi_for_stmt (m_switch); |
| 593 location_t loc = gimple_location (m_switch); | |
| 594 | |
| 595 gcc_assert (m_default_values[num]); | |
| 111 | 596 |
| 597 name = copy_ssa_name (PHI_RESULT (phi)); | |
| 131 | 598 m_target_inbound_names[num] = name; |
| 599 | |
| 145 | 600 vec<constructor_elt, va_gc> *constructor = m_constructors[num]; |
| 131 | 601 wide_int coeff_a, coeff_b; |
| 145 | 602 bool linear_p = contains_linear_function_p (constructor, &coeff_a, &coeff_b); |
| 603 tree type; | |
| 604 if (linear_p | |
| 605 && (type = range_check_type (TREE_TYPE ((*constructor)[0].value)))) | |
| 131 | 606 { |
| 607 if (dump_file && coeff_a.to_uhwi () > 0) | |
| 608 fprintf (dump_file, "Linear transformation with A = %" PRId64 | |
| 609 " and B = %" PRId64 "\n", coeff_a.to_shwi (), | |
| 610 coeff_b.to_shwi ()); | |
| 611 | |
| 145 | 612 /* We must use type of constructor values. */ |
| 131 | 613 gimple_seq seq = NULL; |
| 145 | 614 tree tmp = gimple_convert (&seq, type, m_index_expr); |
| 615 tree tmp2 = gimple_build (&seq, MULT_EXPR, type, | |
| 616 wide_int_to_tree (type, coeff_a), tmp); | |
| 617 tree tmp3 = gimple_build (&seq, PLUS_EXPR, type, tmp2, | |
| 618 wide_int_to_tree (type, coeff_b)); | |
| 131 | 619 tree tmp4 = gimple_convert (&seq, TREE_TYPE (name), tmp3); |
| 620 gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
| 621 load = gimple_build_assign (name, tmp4); | |
| 622 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
623 else |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
624 { |
|
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
|
625 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
|
626 |
| 131 | 627 default_type = TREE_TYPE (m_default_values[num]); |
| 628 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
|
629 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
|
630 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
|
631 { |
|
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 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
|
633 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
|
634 |
| 145 | 635 FOR_EACH_VEC_SAFE_ELT (constructor, 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
|
636 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
|
637 } |
| 145 | 638 ctor = build_constructor (array_type, constructor); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
639 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
|
640 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
|
641 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
642 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
|
643 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
|
644 DECL_INITIAL (decl) = ctor; |
| 0 | 645 |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
646 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
|
647 DECL_ARTIFICIAL (decl) = 1; |
| 111 | 648 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
|
649 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
|
650 TREE_READONLY (decl) = 1; |
| 111 | 651 DECL_IGNORED_P (decl) = 1; |
| 131 | 652 if (offloading_function_p (cfun->decl)) |
| 653 DECL_ATTRIBUTES (decl) | |
| 654 = tree_cons (get_identifier ("omp declare target"), NULL_TREE, | |
| 655 NULL_TREE); | |
| 111 | 656 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
|
657 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
658 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
|
659 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
|
660 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
|
661 { |
|
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 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
|
663 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
|
664 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
|
665 } |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
666 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
|
667 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
668 |
| 0 | 669 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
|
670 update_stmt (load); |
| 131 | 671 m_arr_ref_last = load; |
| 0 | 672 } |
| 673 | |
| 674 /* Builds and initializes static arrays initialized with values gathered from | |
| 131 | 675 the switch statement. Also creates statements that load values from |
| 0 | 676 them. */ |
| 677 | |
| 131 | 678 void |
| 679 switch_conversion::build_arrays () | |
| 0 | 680 { |
| 681 tree arr_index_type; | |
| 111 | 682 tree tidx, sub, utype; |
| 683 gimple *stmt; | |
| 0 | 684 gimple_stmt_iterator gsi; |
| 111 | 685 gphi_iterator gpi; |
| 0 | 686 int i; |
| 131 | 687 location_t loc = gimple_location (m_switch); |
| 688 | |
| 689 gsi = gsi_for_stmt (m_switch); | |
| 0 | 690 |
| 111 | 691 /* Make sure we do not generate arithmetics in a subrange. */ |
| 131 | 692 utype = TREE_TYPE (m_index_expr); |
| 111 | 693 if (TREE_TYPE (utype)) |
| 694 utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1); | |
| 695 else | |
| 696 utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1); | |
| 697 | |
| 131 | 698 arr_index_type = build_index_type (m_range_size); |
| 111 | 699 tidx = make_ssa_name (utype); |
| 700 sub = fold_build2_loc (loc, MINUS_EXPR, utype, | |
| 131 | 701 fold_convert_loc (loc, utype, m_index_expr), |
| 702 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
|
703 sub = force_gimple_operand_gsi (&gsi, sub, |
| 0 | 704 false, NULL, true, GSI_SAME_STMT); |
| 705 stmt = gimple_build_assign (tidx, sub); | |
| 706 | |
| 707 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
|
708 update_stmt (stmt); |
| 131 | 709 m_arr_ref_first = stmt; |
| 710 | |
| 711 for (gpi = gsi_start_phis (m_final_bb), i = 0; | |
| 111 | 712 !gsi_end_p (gpi); gsi_next (&gpi)) |
| 713 { | |
| 714 gphi *phi = gpi.phi (); | |
| 715 if (!virtual_operand_p (gimple_phi_result (phi))) | |
| 131 | 716 build_one_array (i++, arr_index_type, phi, tidx); |
| 111 | 717 else |
| 718 { | |
| 719 edge e; | |
| 720 edge_iterator ei; | |
| 131 | 721 FOR_EACH_EDGE (e, ei, m_switch_bb->succs) |
| 111 | 722 { |
| 131 | 723 if (e->dest == m_final_bb) |
| 111 | 724 break; |
| 131 | 725 if (!m_default_case_nonstandard |
| 726 || e->dest != m_default_bb) | |
| 111 | 727 { |
| 728 e = single_succ_edge (e->dest); | |
| 729 break; | |
| 730 } | |
| 731 } | |
| 131 | 732 gcc_assert (e && e->dest == m_final_bb); |
| 733 m_target_vop = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
| 111 | 734 } |
| 735 } | |
| 0 | 736 } |
| 737 | |
| 738 /* Generates and appropriately inserts loads of default values at the position | |
| 131 | 739 given by GSI. Returns the last inserted statement. */ |
| 740 | |
| 741 gassign * | |
| 742 switch_conversion::gen_def_assigns (gimple_stmt_iterator *gsi) | |
| 0 | 743 { |
| 744 int i; | |
| 111 | 745 gassign *assign = NULL; |
| 746 | |
| 131 | 747 for (i = 0; i < m_phi_count; i++) |
| 0 | 748 { |
| 131 | 749 tree name = copy_ssa_name (m_target_inbound_names[i]); |
| 750 m_target_outbound_names[i] = name; | |
| 751 assign = gimple_build_assign (name, m_default_values[i]); | |
| 0 | 752 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
|
753 update_stmt (assign); |
| 0 | 754 } |
| 755 return assign; | |
| 756 } | |
| 757 | |
| 758 /* Deletes the unused bbs and edges that now contain the switch statement and | |
| 131 | 759 its empty branch bbs. BBD is the now dead BB containing |
| 760 the original switch statement, FINAL is the last BB of the converted | |
| 761 switch statement (in terms of succession). */ | |
| 762 | |
| 763 void | |
| 764 switch_conversion::prune_bbs (basic_block bbd, basic_block final, | |
| 765 basic_block default_bb) | |
| 0 | 766 { |
| 767 edge_iterator ei; | |
| 768 edge e; | |
| 769 | |
| 770 for (ei = ei_start (bbd->succs); (e = ei_safe_edge (ei)); ) | |
| 771 { | |
| 772 basic_block bb; | |
| 773 bb = e->dest; | |
| 774 remove_edge (e); | |
| 111 | 775 if (bb != final && bb != default_bb) |
| 0 | 776 delete_basic_block (bb); |
| 777 } | |
| 778 delete_basic_block (bbd); | |
| 779 } | |
| 780 | |
| 781 /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge | |
| 782 from the basic block loading values from an array and E2F from the basic | |
| 783 block loading default values. BBF is the last switch basic block (see the | |
| 784 bbf description in the comment below). */ | |
| 785 | |
| 131 | 786 void |
| 787 switch_conversion::fix_phi_nodes (edge e1f, edge e2f, basic_block bbf) | |
| 0 | 788 { |
| 111 | 789 gphi_iterator gsi; |
| 0 | 790 int i; |
| 791 | |
| 792 for (gsi = gsi_start_phis (bbf), i = 0; | |
| 111 | 793 !gsi_end_p (gsi); gsi_next (&gsi)) |
| 0 | 794 { |
| 111 | 795 gphi *phi = gsi.phi (); |
| 796 tree inbound, outbound; | |
| 797 if (virtual_operand_p (gimple_phi_result (phi))) | |
| 131 | 798 inbound = outbound = m_target_vop; |
| 111 | 799 else |
| 800 { | |
| 131 | 801 inbound = m_target_inbound_names[i]; |
| 802 outbound = m_target_outbound_names[i++]; | |
| 111 | 803 } |
| 804 add_phi_arg (phi, inbound, e1f, UNKNOWN_LOCATION); | |
| 131 | 805 if (!m_default_case_nonstandard) |
| 111 | 806 add_phi_arg (phi, outbound, e2f, UNKNOWN_LOCATION); |
| 0 | 807 } |
| 808 } | |
| 809 | |
| 810 /* Creates a check whether the switch expression value actually falls into the | |
| 811 range given by all the cases. If it does not, the temporaries are loaded | |
| 131 | 812 with default values instead. */ |
| 813 | |
| 814 void | |
| 815 switch_conversion::gen_inbound_check () | |
| 0 | 816 { |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
817 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
|
818 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
|
819 tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION); |
| 111 | 820 glabel *label1, *label2, *label3; |
| 821 tree utype, tidx; | |
| 0 | 822 tree bound; |
| 823 | |
| 111 | 824 gcond *cond_stmt; |
| 825 | |
| 826 gassign *last_assign = NULL; | |
| 0 | 827 gimple_stmt_iterator gsi; |
| 828 basic_block bb0, bb1, bb2, bbf, bbd; | |
| 111 | 829 edge e01 = NULL, e02, e21, e1d, e1f, e2f; |
| 131 | 830 location_t loc = gimple_location (m_switch); |
| 831 | |
| 832 gcc_assert (m_default_values); | |
| 833 | |
| 834 bb0 = gimple_bb (m_switch); | |
| 835 | |
| 836 tidx = gimple_assign_lhs (m_arr_ref_first); | |
| 111 | 837 utype = TREE_TYPE (tidx); |
| 0 | 838 |
| 839 /* (end of) block 0 */ | |
| 131 | 840 gsi = gsi_for_stmt (m_arr_ref_first); |
| 111 | 841 gsi_next (&gsi); |
| 842 | |
| 131 | 843 bound = fold_convert_loc (loc, utype, m_range_size); |
| 111 | 844 cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE); |
| 0 | 845 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
|
846 update_stmt (cond_stmt); |
| 0 | 847 |
| 848 /* block 2 */ | |
| 131 | 849 if (!m_default_case_nonstandard) |
| 111 | 850 { |
| 851 label2 = gimple_build_label (label_decl2); | |
| 852 gsi_insert_before (&gsi, label2, GSI_SAME_STMT); | |
| 131 | 853 last_assign = gen_def_assigns (&gsi); |
| 111 | 854 } |
| 0 | 855 |
| 856 /* block 1 */ | |
| 857 label1 = gimple_build_label (label_decl1); | |
| 858 gsi_insert_before (&gsi, label1, GSI_SAME_STMT); | |
| 859 | |
| 860 /* block F */ | |
| 131 | 861 gsi = gsi_start_bb (m_final_bb); |
| 0 | 862 label3 = gimple_build_label (label_decl3); |
| 863 gsi_insert_before (&gsi, label3, GSI_SAME_STMT); | |
| 864 | |
| 865 /* cfg fix */ | |
| 866 e02 = split_block (bb0, cond_stmt); | |
| 867 bb2 = e02->dest; | |
| 868 | |
| 131 | 869 if (m_default_case_nonstandard) |
| 111 | 870 { |
| 871 bb1 = bb2; | |
| 131 | 872 bb2 = m_default_bb; |
| 111 | 873 e01 = e02; |
| 874 e01->flags = EDGE_TRUE_VALUE; | |
| 875 e02 = make_edge (bb0, bb2, EDGE_FALSE_VALUE); | |
| 876 edge e_default = find_edge (bb1, bb2); | |
| 877 for (gphi_iterator gsi = gsi_start_phis (bb2); | |
| 878 !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 879 { | |
| 880 gphi *phi = gsi.phi (); | |
| 881 tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e_default); | |
| 882 add_phi_arg (phi, arg, e02, | |
| 883 gimple_phi_arg_location_from_edge (phi, e_default)); | |
| 884 } | |
| 885 /* Partially fix the dominator tree, if it is available. */ | |
| 886 if (dom_info_available_p (CDI_DOMINATORS)) | |
| 887 redirect_immediate_dominators (CDI_DOMINATORS, bb1, bb0); | |
| 888 } | |
| 889 else | |
| 890 { | |
| 891 e21 = split_block (bb2, last_assign); | |
| 892 bb1 = e21->dest; | |
| 893 remove_edge (e21); | |
| 894 } | |
| 895 | |
| 131 | 896 e1d = split_block (bb1, m_arr_ref_last); |
| 0 | 897 bbd = e1d->dest; |
| 898 remove_edge (e1d); | |
| 899 | |
| 131 | 900 /* Flags and profiles of the edge for in-range values. */ |
| 901 if (!m_default_case_nonstandard) | |
| 111 | 902 e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE); |
| 131 | 903 e01->probability = m_default_prob.invert (); |
| 904 | |
| 905 /* Flags and profiles of the edge taking care of out-of-range values. */ | |
| 0 | 906 e02->flags &= ~EDGE_FALLTHRU; |
| 907 e02->flags |= EDGE_FALSE_VALUE; | |
| 131 | 908 e02->probability = m_default_prob; |
| 909 | |
| 910 bbf = m_final_bb; | |
| 0 | 911 |
| 912 e1f = make_edge (bb1, bbf, EDGE_FALLTHRU); | |
| 111 | 913 e1f->probability = profile_probability::always (); |
| 914 | |
| 131 | 915 if (m_default_case_nonstandard) |
| 111 | 916 e2f = NULL; |
| 917 else | |
| 918 { | |
| 919 e2f = make_edge (bb2, bbf, EDGE_FALLTHRU); | |
| 920 e2f->probability = profile_probability::always (); | |
| 921 } | |
| 0 | 922 |
| 923 /* frequencies of the new BBs */ | |
| 131 | 924 bb1->count = e01->count (); |
| 925 bb2->count = e02->count (); | |
| 926 if (!m_default_case_nonstandard) | |
| 927 bbf->count = e1f->count () + e2f->count (); | |
| 111 | 928 |
| 929 /* Tidy blocks that have become unreachable. */ | |
| 131 | 930 prune_bbs (bbd, m_final_bb, |
| 931 m_default_case_nonstandard ? m_default_bb : NULL); | |
| 111 | 932 |
| 933 /* Fixup the PHI nodes in bbF. */ | |
| 131 | 934 fix_phi_nodes (e1f, e2f, bbf); |
| 111 | 935 |
| 936 /* Fix the dominator tree, if it is available. */ | |
| 937 if (dom_info_available_p (CDI_DOMINATORS)) | |
| 938 { | |
| 939 vec<basic_block> bbs_to_fix_dom; | |
| 940 | |
| 941 set_immediate_dominator (CDI_DOMINATORS, bb1, bb0); | |
| 131 | 942 if (!m_default_case_nonstandard) |
| 111 | 943 set_immediate_dominator (CDI_DOMINATORS, bb2, bb0); |
| 944 if (! get_immediate_dominator (CDI_DOMINATORS, bbf)) | |
| 945 /* If bbD was the immediate dominator ... */ | |
| 946 set_immediate_dominator (CDI_DOMINATORS, bbf, bb0); | |
| 947 | |
| 948 bbs_to_fix_dom.create (3 + (bb2 != bbf)); | |
| 949 bbs_to_fix_dom.quick_push (bb0); | |
| 950 bbs_to_fix_dom.quick_push (bb1); | |
| 951 if (bb2 != bbf) | |
| 952 bbs_to_fix_dom.quick_push (bb2); | |
| 953 bbs_to_fix_dom.quick_push (bbf); | |
| 954 | |
| 955 iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true); | |
| 956 bbs_to_fix_dom.release (); | |
| 957 } | |
| 0 | 958 } |
| 959 | |
| 131 | 960 /* The following function is invoked on every switch statement (the current |
| 961 one is given in SWTCH) and runs the individual phases of switch | |
| 962 conversion on it one after another until one fails or the conversion | |
| 963 is completed. On success, NULL is in m_reason, otherwise points | |
| 964 to a string with the reason why the conversion failed. */ | |
| 965 | |
| 966 void | |
| 967 switch_conversion::expand (gswitch *swtch) | |
| 0 | 968 { |
| 111 | 969 /* Group case labels so that we get the right results from the heuristics |
| 970 that decide on the code generation approach for this switch. */ | |
| 131 | 971 m_cfg_altered |= group_case_labels_stmt (swtch); |
| 111 | 972 |
| 973 /* If this switch is now a degenerate case with only a default label, | |
| 131 | 974 there is nothing left for us to do. */ |
| 111 | 975 if (gimple_switch_num_labels (swtch) < 2) |
| 131 | 976 { |
| 977 m_reason = "switch is a degenerate case"; | |
| 978 return; | |
| 979 } | |
| 980 | |
| 981 collect (swtch); | |
| 111 | 982 |
| 983 /* No error markers should reach here (they should be filtered out | |
| 984 during gimplification). */ | |
| 131 | 985 gcc_checking_assert (TREE_TYPE (m_index_expr) != error_mark_node); |
| 111 | 986 |
| 987 /* A switch on a constant should have been optimized in tree-cfg-cleanup. */ | |
| 131 | 988 gcc_checking_assert (!TREE_CONSTANT (m_index_expr)); |
| 989 | |
| 990 /* Prefer bit test if possible. */ | |
| 991 if (tree_fits_uhwi_p (m_range_size) | |
| 992 && bit_test_cluster::can_be_handled (tree_to_uhwi (m_range_size), m_uniq) | |
| 993 && bit_test_cluster::is_beneficial (m_count, m_uniq)) | |
| 0 | 994 { |
| 131 | 995 m_reason = "expanding as bit test is preferable"; |
| 996 return; | |
| 997 } | |
| 998 | |
| 999 if (m_uniq <= 2) | |
| 1000 { | |
| 1001 /* This will be expanded as a decision tree . */ | |
| 1002 m_reason = "expanding as jumps is preferable"; | |
| 1003 return; | |
| 0 | 1004 } |
| 1005 | |
| 111 | 1006 /* If there is no common successor, we cannot do the transformation. */ |
| 131 | 1007 if (!m_final_bb) |
| 1008 { | |
| 1009 m_reason = "no common successor to all case label target blocks found"; | |
| 1010 return; | |
| 1011 } | |
| 0 | 1012 |
| 1013 /* Check the case label values are within reasonable range: */ | |
| 131 | 1014 if (!check_range ()) |
| 111 | 1015 { |
| 131 | 1016 gcc_assert (m_reason); |
| 1017 return; | |
| 111 | 1018 } |
| 0 | 1019 |
| 1020 /* For all the cases, see whether they are empty, the assignments they | |
| 1021 represent constant and so on... */ | |
| 131 | 1022 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
|
1023 { |
| 131 | 1024 gcc_assert (m_reason); |
| 1025 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
|
1026 } |
| 131 | 1027 if (!check_final_bb ()) |
| 111 | 1028 { |
| 131 | 1029 gcc_assert (m_reason); |
| 1030 return; | |
| 111 | 1031 } |
| 0 | 1032 |
| 1033 /* At this point all checks have passed and we can proceed with the | |
| 1034 transformation. */ | |
| 1035 | |
| 131 | 1036 create_temp_arrays (); |
| 1037 gather_default_values (m_default_case_nonstandard | |
| 111 | 1038 ? gimple_switch_label (swtch, 1) |
| 131 | 1039 : gimple_switch_default_label (swtch)); |
| 1040 build_constructors (); | |
| 1041 | |
| 1042 build_arrays (); /* Build the static arrays and assignments. */ | |
| 1043 gen_inbound_check (); /* Build the bounds check. */ | |
| 1044 | |
| 1045 m_cfg_altered = true; | |
| 1046 } | |
| 1047 | |
| 1048 /* Destructor. */ | |
| 1049 | |
| 1050 switch_conversion::~switch_conversion () | |
| 1051 { | |
| 1052 XDELETEVEC (m_constructors); | |
| 1053 XDELETEVEC (m_default_values); | |
| 1054 } | |
| 1055 | |
| 1056 /* Constructor. */ | |
| 1057 | |
| 1058 group_cluster::group_cluster (vec<cluster *> &clusters, | |
| 1059 unsigned start, unsigned end) | |
| 1060 { | |
| 1061 gcc_checking_assert (end - start + 1 >= 1); | |
| 1062 m_prob = profile_probability::never (); | |
| 1063 m_cases.create (end - start + 1); | |
| 1064 for (unsigned i = start; i <= end; i++) | |
| 1065 { | |
| 1066 m_cases.quick_push (static_cast<simple_cluster *> (clusters[i])); | |
| 1067 m_prob += clusters[i]->m_prob; | |
| 1068 } | |
| 1069 m_subtree_prob = m_prob; | |
| 1070 } | |
| 1071 | |
| 1072 /* Destructor. */ | |
| 1073 | |
| 1074 group_cluster::~group_cluster () | |
| 1075 { | |
| 1076 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1077 delete m_cases[i]; | |
| 1078 | |
| 1079 m_cases.release (); | |
| 1080 } | |
| 1081 | |
| 1082 /* Dump content of a cluster. */ | |
| 1083 | |
| 1084 void | |
| 1085 group_cluster::dump (FILE *f, bool details) | |
| 1086 { | |
| 1087 unsigned total_values = 0; | |
| 1088 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1089 total_values += m_cases[i]->get_range (m_cases[i]->get_low (), | |
| 1090 m_cases[i]->get_high ()); | |
| 1091 | |
| 1092 unsigned comparison_count = 0; | |
| 1093 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1094 { | |
| 1095 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1096 comparison_count += sc->m_range_p ? 2 : 1; | |
| 1097 } | |
| 1098 | |
| 1099 unsigned HOST_WIDE_INT range = get_range (get_low (), get_high ()); | |
| 1100 fprintf (f, "%s", get_type () == JUMP_TABLE ? "JT" : "BT"); | |
| 1101 | |
| 1102 if (details) | |
| 1103 fprintf (f, "(values:%d comparisons:%d range:" HOST_WIDE_INT_PRINT_DEC | |
| 1104 " density: %.2f%%)", total_values, comparison_count, range, | |
| 1105 100.0f * comparison_count / range); | |
| 1106 | |
| 1107 fprintf (f, ":"); | |
| 1108 PRINT_CASE (f, get_low ()); | |
| 1109 fprintf (f, "-"); | |
| 1110 PRINT_CASE (f, get_high ()); | |
| 1111 fprintf (f, " "); | |
| 1112 } | |
| 1113 | |
| 1114 /* Emit GIMPLE code to handle the cluster. */ | |
| 1115 | |
| 1116 void | |
| 1117 jump_table_cluster::emit (tree index_expr, tree, | |
| 1118 tree default_label_expr, basic_block default_bb) | |
| 1119 { | |
| 1120 unsigned HOST_WIDE_INT range = get_range (get_low (), get_high ()); | |
| 1121 unsigned HOST_WIDE_INT nondefault_range = 0; | |
| 1122 | |
| 1123 /* For jump table we just emit a new gswitch statement that will | |
| 1124 be latter lowered to jump table. */ | |
| 1125 auto_vec <tree> labels; | |
| 1126 labels.create (m_cases.length ()); | |
| 1127 | |
| 1128 make_edge (m_case_bb, default_bb, 0); | |
| 1129 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1130 { | |
| 1131 labels.quick_push (unshare_expr (m_cases[i]->m_case_label_expr)); | |
| 1132 make_edge (m_case_bb, m_cases[i]->m_case_bb, 0); | |
| 1133 } | |
| 1134 | |
| 1135 gswitch *s = gimple_build_switch (index_expr, | |
| 1136 unshare_expr (default_label_expr), labels); | |
| 1137 gimple_stmt_iterator gsi = gsi_start_bb (m_case_bb); | |
| 1138 gsi_insert_after (&gsi, s, GSI_NEW_STMT); | |
| 1139 | |
| 1140 /* Set up even probabilities for all cases. */ | |
| 1141 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1142 { | |
| 1143 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1144 edge case_edge = find_edge (m_case_bb, sc->m_case_bb); | |
| 1145 unsigned HOST_WIDE_INT case_range | |
| 1146 = sc->get_range (sc->get_low (), sc->get_high ()); | |
| 1147 nondefault_range += case_range; | |
| 1148 | |
| 1149 /* case_edge->aux is number of values in a jump-table that are covered | |
| 1150 by the case_edge. */ | |
| 1151 case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + case_range); | |
| 1152 } | |
| 1153 | |
| 1154 edge default_edge = gimple_switch_default_edge (cfun, s); | |
| 1155 default_edge->probability = profile_probability::never (); | |
| 1156 | |
| 1157 for (unsigned i = 0; i < m_cases.length (); i++) | |
| 1158 { | |
| 1159 simple_cluster *sc = static_cast<simple_cluster *> (m_cases[i]); | |
| 1160 edge case_edge = find_edge (m_case_bb, sc->m_case_bb); | |
| 1161 case_edge->probability | |
| 1162 = profile_probability::always ().apply_scale ((intptr_t)case_edge->aux, | |
| 1163 range); | |
| 1164 } | |
| 1165 | |
| 1166 /* Number of non-default values is probability of default edge. */ | |
| 1167 default_edge->probability | |
| 1168 += profile_probability::always ().apply_scale (nondefault_range, | |
| 1169 range).invert (); | |
| 1170 | |
| 1171 switch_decision_tree::reset_out_edges_aux (s); | |
| 1172 } | |
| 1173 | |
| 1174 /* Find jump tables of given CLUSTERS, where all members of the vector | |
| 1175 are of type simple_cluster. New clusters are returned. */ | |
| 1176 | |
| 1177 vec<cluster *> | |
| 1178 jump_table_cluster::find_jump_tables (vec<cluster *> &clusters) | |
| 1179 { | |
| 1180 if (!is_enabled ()) | |
| 1181 return clusters.copy (); | |
| 1182 | |
| 1183 unsigned l = clusters.length (); | |
| 1184 auto_vec<min_cluster_item> min; | |
| 1185 min.reserve (l + 1); | |
| 1186 | |
| 1187 min.quick_push (min_cluster_item (0, 0, 0)); | |
| 1188 | |
| 1189 for (unsigned i = 1; i <= l; i++) | |
| 1190 { | |
| 1191 /* Set minimal # of clusters with i-th item to infinite. */ | |
| 1192 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); | |
| 1193 | |
| 1194 for (unsigned j = 0; j < i; j++) | |
| 1195 { | |
| 1196 unsigned HOST_WIDE_INT s = min[j].m_non_jt_cases; | |
| 1197 if (i - j < case_values_threshold ()) | |
| 1198 s += i - j; | |
| 1199 | |
| 1200 /* Prefer clusters with smaller number of numbers covered. */ | |
| 1201 if ((min[j].m_count + 1 < min[i].m_count | |
| 1202 || (min[j].m_count + 1 == min[i].m_count | |
| 1203 && s < min[i].m_non_jt_cases)) | |
| 1204 && can_be_handled (clusters, j, i - 1)) | |
| 1205 min[i] = min_cluster_item (min[j].m_count + 1, j, s); | |
| 1206 } | |
| 1207 | |
| 1208 gcc_checking_assert (min[i].m_count != INT_MAX); | |
| 1209 } | |
| 1210 | |
| 1211 /* No result. */ | |
| 145 | 1212 if (min[l].m_count == l) |
| 131 | 1213 return clusters.copy (); |
| 1214 | |
| 1215 vec<cluster *> output; | |
| 1216 output.create (4); | |
| 1217 | |
| 1218 /* Find and build the clusters. */ | |
| 145 | 1219 for (unsigned int end = l;;) |
| 131 | 1220 { |
| 1221 int start = min[end].m_start; | |
| 1222 | |
| 1223 /* Do not allow clusters with small number of cases. */ | |
| 1224 if (is_beneficial (clusters, start, end - 1)) | |
| 1225 output.safe_push (new jump_table_cluster (clusters, start, end - 1)); | |
| 1226 else | |
| 1227 for (int i = end - 1; i >= start; i--) | |
| 1228 output.safe_push (clusters[i]); | |
| 1229 | |
| 1230 end = start; | |
| 1231 | |
| 1232 if (start <= 0) | |
| 1233 break; | |
| 1234 } | |
| 1235 | |
| 1236 output.reverse (); | |
| 1237 return output; | |
| 1238 } | |
| 1239 | |
| 1240 /* Return true when cluster starting at START and ending at END (inclusive) | |
| 1241 can build a jump-table. */ | |
| 1242 | |
| 1243 bool | |
| 1244 jump_table_cluster::can_be_handled (const vec<cluster *> &clusters, | |
| 1245 unsigned start, unsigned end) | |
| 1246 { | |
| 1247 /* If the switch is relatively small such that the cost of one | |
| 1248 indirect jump on the target are higher than the cost of a | |
| 1249 decision tree, go with the decision tree. | |
| 1250 | |
| 1251 If range of values is much bigger than number of values, | |
| 1252 or if it is too large to represent in a HOST_WIDE_INT, | |
| 1253 make a sequence of conditional branches instead of a dispatch. | |
| 1254 | |
| 1255 The definition of "much bigger" depends on whether we are | |
| 145 | 1256 optimizing for size or for speed. |
| 1257 | |
| 1258 For algorithm correctness, jump table for a single case must return | |
| 131 | 1259 true. We bail out in is_beneficial if it's called just for |
| 1260 a single case. */ | |
| 1261 if (start == end) | |
| 1262 return true; | |
| 1263 | |
| 1264 unsigned HOST_WIDE_INT max_ratio | |
| 145 | 1265 = (optimize_insn_for_size_p () |
| 1266 ? param_jump_table_max_growth_ratio_for_size | |
| 1267 : param_jump_table_max_growth_ratio_for_speed); | |
| 131 | 1268 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (), |
| 1269 clusters[end]->get_high ()); | |
| 1270 /* Check overflow. */ | |
| 1271 if (range == 0) | |
| 1272 return false; | |
| 1273 | |
| 1274 unsigned HOST_WIDE_INT comparison_count = 0; | |
| 1275 for (unsigned i = start; i <= end; i++) | |
| 1276 { | |
| 1277 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1278 comparison_count += sc->m_range_p ? 2 : 1; | |
| 1279 } | |
| 1280 | |
| 145 | 1281 unsigned HOST_WIDE_INT lhs = 100 * range; |
| 1282 if (lhs < range) | |
| 1283 return false; | |
| 1284 | |
| 1285 return lhs <= max_ratio * comparison_count; | |
| 131 | 1286 } |
| 1287 | |
| 1288 /* Return true if cluster starting at START and ending at END (inclusive) | |
| 1289 is profitable transformation. */ | |
| 1290 | |
| 1291 bool | |
| 1292 jump_table_cluster::is_beneficial (const vec<cluster *> &, | |
| 1293 unsigned start, unsigned end) | |
| 1294 { | |
| 1295 /* Single case bail out. */ | |
| 1296 if (start == end) | |
| 1297 return false; | |
| 1298 | |
| 1299 return end - start + 1 >= case_values_threshold (); | |
| 1300 } | |
| 1301 | |
| 1302 /* Find bit tests of given CLUSTERS, where all members of the vector | |
| 1303 are of type simple_cluster. New clusters are returned. */ | |
| 1304 | |
| 1305 vec<cluster *> | |
| 1306 bit_test_cluster::find_bit_tests (vec<cluster *> &clusters) | |
| 1307 { | |
| 1308 unsigned l = clusters.length (); | |
| 1309 auto_vec<min_cluster_item> min; | |
| 1310 min.reserve (l + 1); | |
| 1311 | |
| 1312 min.quick_push (min_cluster_item (0, 0, 0)); | |
| 1313 | |
| 1314 for (unsigned i = 1; i <= l; i++) | |
| 1315 { | |
| 1316 /* Set minimal # of clusters with i-th item to infinite. */ | |
| 1317 min.quick_push (min_cluster_item (INT_MAX, INT_MAX, INT_MAX)); | |
| 1318 | |
| 1319 for (unsigned j = 0; j < i; j++) | |
| 1320 { | |
| 1321 if (min[j].m_count + 1 < min[i].m_count | |
| 1322 && can_be_handled (clusters, j, i - 1)) | |
| 1323 min[i] = min_cluster_item (min[j].m_count + 1, j, INT_MAX); | |
| 1324 } | |
| 1325 | |
| 1326 gcc_checking_assert (min[i].m_count != INT_MAX); | |
| 1327 } | |
| 1328 | |
| 1329 /* No result. */ | |
| 145 | 1330 if (min[l].m_count == l) |
| 131 | 1331 return clusters.copy (); |
| 1332 | |
| 145 | 1333 vec<cluster *> output; |
| 1334 output.create (4); | |
| 1335 | |
| 131 | 1336 /* Find and build the clusters. */ |
| 1337 for (unsigned end = l;;) | |
| 1338 { | |
| 1339 int start = min[end].m_start; | |
| 1340 | |
| 1341 if (is_beneficial (clusters, start, end - 1)) | |
| 1342 { | |
| 1343 bool entire = start == 0 && end == clusters.length (); | |
| 1344 output.safe_push (new bit_test_cluster (clusters, start, end - 1, | |
| 1345 entire)); | |
| 1346 } | |
| 1347 else | |
| 145 | 1348 for (int i = end - 1; i >= start; i--) |
| 131 | 1349 output.safe_push (clusters[i]); |
| 1350 | |
| 1351 end = start; | |
| 1352 | |
| 1353 if (start <= 0) | |
| 1354 break; | |
| 1355 } | |
| 1356 | |
| 1357 output.reverse (); | |
| 1358 return output; | |
| 1359 } | |
| 1360 | |
| 1361 /* Return true when RANGE of case values with UNIQ labels | |
| 1362 can build a bit test. */ | |
| 1363 | |
| 1364 bool | |
| 1365 bit_test_cluster::can_be_handled (unsigned HOST_WIDE_INT range, | |
| 1366 unsigned int uniq) | |
| 1367 { | |
| 1368 /* Check overflow. */ | |
| 1369 if (range == 0) | |
| 1370 return 0; | |
| 1371 | |
| 1372 if (range >= GET_MODE_BITSIZE (word_mode)) | |
| 1373 return false; | |
| 1374 | |
| 1375 return uniq <= 3; | |
| 1376 } | |
| 1377 | |
| 1378 /* Return true when cluster starting at START and ending at END (inclusive) | |
| 1379 can build a bit test. */ | |
| 1380 | |
| 1381 bool | |
| 1382 bit_test_cluster::can_be_handled (const vec<cluster *> &clusters, | |
| 1383 unsigned start, unsigned end) | |
| 1384 { | |
| 1385 /* For algorithm correctness, bit test for a single case must return | |
| 1386 true. We bail out in is_beneficial if it's called just for | |
| 1387 a single case. */ | |
| 1388 if (start == end) | |
| 1389 return true; | |
| 1390 | |
| 1391 unsigned HOST_WIDE_INT range = get_range (clusters[start]->get_low (), | |
| 1392 clusters[end]->get_high ()); | |
| 1393 auto_bitmap dest_bbs; | |
| 1394 | |
| 1395 for (unsigned i = start; i <= end; i++) | |
| 1396 { | |
| 1397 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1398 bitmap_set_bit (dest_bbs, sc->m_case_bb->index); | |
| 1399 } | |
| 1400 | |
| 1401 return can_be_handled (range, bitmap_count_bits (dest_bbs)); | |
| 1402 } | |
| 1403 | |
| 1404 /* Return true when COUNT of cases of UNIQ labels is beneficial for bit test | |
| 1405 transformation. */ | |
| 1406 | |
| 1407 bool | |
| 1408 bit_test_cluster::is_beneficial (unsigned count, unsigned uniq) | |
| 1409 { | |
| 1410 return (((uniq == 1 && count >= 3) | |
| 1411 || (uniq == 2 && count >= 5) | |
| 1412 || (uniq == 3 && count >= 6))); | |
| 1413 } | |
| 1414 | |
| 1415 /* Return true if cluster starting at START and ending at END (inclusive) | |
| 1416 is profitable transformation. */ | |
| 1417 | |
| 1418 bool | |
| 1419 bit_test_cluster::is_beneficial (const vec<cluster *> &clusters, | |
| 1420 unsigned start, unsigned end) | |
| 1421 { | |
| 1422 /* Single case bail out. */ | |
| 1423 if (start == end) | |
| 1424 return false; | |
| 1425 | |
| 1426 auto_bitmap dest_bbs; | |
| 1427 | |
| 1428 for (unsigned i = start; i <= end; i++) | |
| 1429 { | |
| 1430 simple_cluster *sc = static_cast<simple_cluster *> (clusters[i]); | |
| 1431 bitmap_set_bit (dest_bbs, sc->m_case_bb->index); | |
| 1432 } | |
| 1433 | |
| 1434 unsigned uniq = bitmap_count_bits (dest_bbs); | |
| 1435 unsigned count = end - start + 1; | |
| 1436 return is_beneficial (count, uniq); | |
| 1437 } | |
| 1438 | |
| 1439 /* Comparison function for qsort to order bit tests by decreasing | |
| 1440 probability of execution. */ | |
| 1441 | |
| 1442 int | |
| 1443 case_bit_test::cmp (const void *p1, const void *p2) | |
| 1444 { | |
| 145 | 1445 const case_bit_test *const d1 = (const case_bit_test *) p1; |
| 1446 const case_bit_test *const d2 = (const case_bit_test *) p2; | |
| 131 | 1447 |
| 1448 if (d2->bits != d1->bits) | |
| 1449 return d2->bits - d1->bits; | |
| 1450 | |
| 1451 /* Stabilize the sort. */ | |
| 1452 return (LABEL_DECL_UID (CASE_LABEL (d2->label)) | |
| 1453 - LABEL_DECL_UID (CASE_LABEL (d1->label))); | |
| 1454 } | |
| 1455 | |
| 1456 /* Expand a switch statement by a short sequence of bit-wise | |
| 1457 comparisons. "switch(x)" is effectively converted into | |
| 1458 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are | |
| 1459 integer constants. | |
| 1460 | |
| 1461 INDEX_EXPR is the value being switched on. | |
| 1462 | |
| 1463 MINVAL is the lowest case value of in the case nodes, | |
| 1464 and RANGE is highest value minus MINVAL. MINVAL and RANGE | |
| 1465 are not guaranteed to be of the same type as INDEX_EXPR | |
| 1466 (the gimplifier doesn't change the type of case label values, | |
| 1467 and MINVAL and RANGE are derived from those values). | |
| 1468 MAXVAL is MINVAL + RANGE. | |
| 1469 | |
| 1470 There *MUST* be max_case_bit_tests or less unique case | |
| 1471 node targets. */ | |
| 1472 | |
| 1473 void | |
| 1474 bit_test_cluster::emit (tree index_expr, tree index_type, | |
| 1475 tree, basic_block default_bb) | |
| 1476 { | |
| 145 | 1477 case_bit_test test[m_max_case_bit_tests] = { {} }; |
| 131 | 1478 unsigned int i, j, k; |
| 1479 unsigned int count; | |
| 1480 | |
| 145 | 1481 tree unsigned_index_type = range_check_type (index_type); |
| 131 | 1482 |
| 1483 gimple_stmt_iterator gsi; | |
| 1484 gassign *shift_stmt; | |
| 1485 | |
| 1486 tree idx, tmp, csui; | |
| 1487 tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1); | |
| 1488 tree word_mode_zero = fold_convert (word_type_node, integer_zero_node); | |
| 1489 tree word_mode_one = fold_convert (word_type_node, integer_one_node); | |
| 1490 int prec = TYPE_PRECISION (word_type_node); | |
| 1491 wide_int wone = wi::one (prec); | |
| 1492 | |
| 1493 tree minval = get_low (); | |
| 1494 tree maxval = get_high (); | |
| 1495 tree range = int_const_binop (MINUS_EXPR, maxval, minval); | |
| 1496 unsigned HOST_WIDE_INT bt_range = get_range (minval, maxval); | |
| 1497 | |
| 1498 /* Go through all case labels, and collect the case labels, profile | |
| 1499 counts, and other information we need to build the branch tests. */ | |
| 1500 count = 0; | |
| 1501 for (i = 0; i < m_cases.length (); i++) | |
| 1502 { | |
| 1503 unsigned int lo, hi; | |
| 1504 simple_cluster *n = static_cast<simple_cluster *> (m_cases[i]); | |
| 1505 for (k = 0; k < count; k++) | |
| 1506 if (n->m_case_bb == test[k].target_bb) | |
| 1507 break; | |
| 1508 | |
| 1509 if (k == count) | |
| 1510 { | |
| 1511 gcc_checking_assert (count < m_max_case_bit_tests); | |
| 1512 test[k].mask = wi::zero (prec); | |
| 1513 test[k].target_bb = n->m_case_bb; | |
| 1514 test[k].label = n->m_case_label_expr; | |
| 1515 test[k].bits = 0; | |
| 1516 count++; | |
| 1517 } | |
| 1518 | |
| 1519 test[k].bits += n->get_range (n->get_low (), n->get_high ()); | |
| 1520 | |
| 1521 lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_low (), minval)); | |
| 1522 if (n->get_high () == NULL_TREE) | |
| 1523 hi = lo; | |
| 1524 else | |
| 1525 hi = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_high (), | |
| 1526 minval)); | |
| 1527 | |
| 1528 for (j = lo; j <= hi; j++) | |
| 1529 test[k].mask |= wi::lshift (wone, j); | |
| 1530 } | |
| 1531 | |
| 1532 qsort (test, count, sizeof (*test), case_bit_test::cmp); | |
| 1533 | |
| 1534 /* If all values are in the 0 .. BITS_PER_WORD-1 range, we can get rid of | |
| 1535 the minval subtractions, but it might make the mask constants more | |
| 1536 expensive. So, compare the costs. */ | |
| 1537 if (compare_tree_int (minval, 0) > 0 | |
| 1538 && compare_tree_int (maxval, GET_MODE_BITSIZE (word_mode)) < 0) | |
| 1539 { | |
| 1540 int cost_diff; | |
| 1541 HOST_WIDE_INT m = tree_to_uhwi (minval); | |
| 1542 rtx reg = gen_raw_REG (word_mode, 10000); | |
| 1543 bool speed_p = optimize_insn_for_speed_p (); | |
| 1544 cost_diff = set_rtx_cost (gen_rtx_PLUS (word_mode, reg, | |
| 1545 GEN_INT (-m)), speed_p); | |
| 1546 for (i = 0; i < count; i++) | |
| 1547 { | |
| 1548 rtx r = immed_wide_int_const (test[i].mask, word_mode); | |
| 1549 cost_diff += set_src_cost (gen_rtx_AND (word_mode, reg, r), | |
| 1550 word_mode, speed_p); | |
| 1551 r = immed_wide_int_const (wi::lshift (test[i].mask, m), word_mode); | |
| 1552 cost_diff -= set_src_cost (gen_rtx_AND (word_mode, reg, r), | |
| 1553 word_mode, speed_p); | |
| 1554 } | |
| 1555 if (cost_diff > 0) | |
| 1556 { | |
| 1557 for (i = 0; i < count; i++) | |
| 1558 test[i].mask = wi::lshift (test[i].mask, m); | |
| 1559 minval = build_zero_cst (TREE_TYPE (minval)); | |
| 1560 range = maxval; | |
| 1561 } | |
| 1562 } | |
| 1563 | |
| 1564 /* Now build the test-and-branch code. */ | |
| 1565 | |
| 1566 gsi = gsi_last_bb (m_case_bb); | |
| 1567 | |
| 1568 /* idx = (unsigned)x - minval. */ | |
| 1569 idx = fold_convert (unsigned_index_type, index_expr); | |
| 1570 idx = fold_build2 (MINUS_EXPR, unsigned_index_type, idx, | |
| 1571 fold_convert (unsigned_index_type, minval)); | |
| 1572 idx = force_gimple_operand_gsi (&gsi, idx, | |
| 1573 /*simple=*/true, NULL_TREE, | |
| 1574 /*before=*/true, GSI_SAME_STMT); | |
| 1575 | |
| 1576 if (m_handles_entire_switch) | |
| 1577 { | |
| 1578 /* if (idx > range) goto default */ | |
| 1579 range | |
| 1580 = force_gimple_operand_gsi (&gsi, | |
| 1581 fold_convert (unsigned_index_type, range), | |
| 1582 /*simple=*/true, NULL_TREE, | |
| 1583 /*before=*/true, GSI_SAME_STMT); | |
| 1584 tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range); | |
| 1585 basic_block new_bb | |
| 1586 = hoist_edge_and_branch_if_true (&gsi, tmp, default_bb, | |
| 1587 profile_probability::unlikely ()); | |
| 1588 gsi = gsi_last_bb (new_bb); | |
| 1589 } | |
| 1590 | |
| 1591 /* csui = (1 << (word_mode) idx) */ | |
| 1592 csui = make_ssa_name (word_type_node); | |
| 1593 tmp = fold_build2 (LSHIFT_EXPR, word_type_node, word_mode_one, | |
| 1594 fold_convert (word_type_node, idx)); | |
| 1595 tmp = force_gimple_operand_gsi (&gsi, tmp, | |
| 1596 /*simple=*/false, NULL_TREE, | |
| 1597 /*before=*/true, GSI_SAME_STMT); | |
| 1598 shift_stmt = gimple_build_assign (csui, tmp); | |
| 1599 gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT); | |
| 1600 update_stmt (shift_stmt); | |
| 1601 | |
| 1602 profile_probability prob = profile_probability::always (); | |
| 1603 | |
| 1604 /* for each unique set of cases: | |
| 1605 if (const & csui) goto target */ | |
| 1606 for (k = 0; k < count; k++) | |
| 1607 { | |
| 1608 prob = profile_probability::always ().apply_scale (test[k].bits, | |
| 1609 bt_range); | |
| 1610 bt_range -= test[k].bits; | |
| 1611 tmp = wide_int_to_tree (word_type_node, test[k].mask); | |
| 1612 tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp); | |
| 1613 tmp = force_gimple_operand_gsi (&gsi, tmp, | |
| 1614 /*simple=*/true, NULL_TREE, | |
| 1615 /*before=*/true, GSI_SAME_STMT); | |
| 1616 tmp = fold_build2 (NE_EXPR, boolean_type_node, tmp, word_mode_zero); | |
| 1617 basic_block new_bb | |
| 1618 = hoist_edge_and_branch_if_true (&gsi, tmp, test[k].target_bb, prob); | |
| 1619 gsi = gsi_last_bb (new_bb); | |
| 1620 } | |
| 1621 | |
| 1622 /* We should have removed all edges now. */ | |
| 1623 gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0); | |
| 1624 | |
| 1625 /* If nothing matched, go to the default label. */ | |
| 1626 edge e = make_edge (gsi_bb (gsi), default_bb, EDGE_FALLTHRU); | |
| 1627 e->probability = profile_probability::always (); | |
| 1628 } | |
| 1629 | |
| 1630 /* Split the basic block at the statement pointed to by GSIP, and insert | |
| 1631 a branch to the target basic block of E_TRUE conditional on tree | |
| 1632 expression COND. | |
| 1633 | |
| 1634 It is assumed that there is already an edge from the to-be-split | |
| 1635 basic block to E_TRUE->dest block. This edge is removed, and the | |
| 1636 profile information on the edge is re-used for the new conditional | |
| 1637 jump. | |
| 1638 | |
| 1639 The CFG is updated. The dominator tree will not be valid after | |
| 1640 this transformation, but the immediate dominators are updated if | |
| 1641 UPDATE_DOMINATORS is true. | |
| 1642 | |
| 1643 Returns the newly created basic block. */ | |
| 1644 | |
| 1645 basic_block | |
| 1646 bit_test_cluster::hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip, | |
| 1647 tree cond, basic_block case_bb, | |
| 1648 profile_probability prob) | |
| 1649 { | |
| 1650 tree tmp; | |
| 1651 gcond *cond_stmt; | |
| 1652 edge e_false; | |
| 1653 basic_block new_bb, split_bb = gsi_bb (*gsip); | |
| 1654 | |
| 1655 edge e_true = make_edge (split_bb, case_bb, EDGE_TRUE_VALUE); | |
| 1656 e_true->probability = prob; | |
| 1657 gcc_assert (e_true->src == split_bb); | |
| 1658 | |
| 1659 tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL, | |
| 1660 /*before=*/true, GSI_SAME_STMT); | |
| 1661 cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE); | |
| 1662 gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT); | |
| 1663 | |
| 1664 e_false = split_block (split_bb, cond_stmt); | |
| 1665 new_bb = e_false->dest; | |
| 1666 redirect_edge_pred (e_true, split_bb); | |
| 1667 | |
| 1668 e_false->flags &= ~EDGE_FALLTHRU; | |
| 1669 e_false->flags |= EDGE_FALSE_VALUE; | |
| 1670 e_false->probability = e_true->probability.invert (); | |
| 1671 new_bb->count = e_false->count (); | |
| 1672 | |
| 1673 return new_bb; | |
| 1674 } | |
| 1675 | |
| 1676 /* Compute the number of case labels that correspond to each outgoing edge of | |
| 1677 switch statement. Record this information in the aux field of the edge. */ | |
| 1678 | |
| 1679 void | |
| 1680 switch_decision_tree::compute_cases_per_edge () | |
| 1681 { | |
| 1682 reset_out_edges_aux (m_switch); | |
| 1683 int ncases = gimple_switch_num_labels (m_switch); | |
| 1684 for (int i = ncases - 1; i >= 1; --i) | |
| 1685 { | |
| 1686 edge case_edge = gimple_switch_edge (cfun, m_switch, i); | |
| 1687 case_edge->aux = (void *) ((intptr_t) (case_edge->aux) + 1); | |
| 1688 } | |
| 1689 } | |
| 1690 | |
| 1691 /* Analyze switch statement and return true when the statement is expanded | |
| 1692 as decision tree. */ | |
| 1693 | |
| 1694 bool | |
| 1695 switch_decision_tree::analyze_switch_statement () | |
| 1696 { | |
| 1697 unsigned l = gimple_switch_num_labels (m_switch); | |
| 1698 basic_block bb = gimple_bb (m_switch); | |
| 1699 auto_vec<cluster *> clusters; | |
| 1700 clusters.create (l - 1); | |
| 1701 | |
| 1702 basic_block default_bb = gimple_switch_default_bb (cfun, m_switch); | |
| 1703 m_case_bbs.reserve (l); | |
| 1704 m_case_bbs.quick_push (default_bb); | |
| 1705 | |
| 1706 compute_cases_per_edge (); | |
| 1707 | |
| 1708 for (unsigned i = 1; i < l; i++) | |
| 1709 { | |
| 1710 tree elt = gimple_switch_label (m_switch, i); | |
| 1711 tree lab = CASE_LABEL (elt); | |
| 1712 basic_block case_bb = label_to_block (cfun, lab); | |
| 1713 edge case_edge = find_edge (bb, case_bb); | |
| 1714 tree low = CASE_LOW (elt); | |
| 1715 tree high = CASE_HIGH (elt); | |
| 1716 | |
| 1717 profile_probability p | |
| 1718 = case_edge->probability.apply_scale (1, (intptr_t) (case_edge->aux)); | |
| 1719 clusters.quick_push (new simple_cluster (low, high, elt, case_edge->dest, | |
| 1720 p)); | |
| 1721 m_case_bbs.quick_push (case_edge->dest); | |
| 1722 } | |
| 1723 | |
| 1724 reset_out_edges_aux (m_switch); | |
| 1725 | |
| 1726 /* Find jump table clusters. */ | |
| 1727 vec<cluster *> output = jump_table_cluster::find_jump_tables (clusters); | |
| 1728 | |
| 1729 /* Find bit test clusters. */ | |
| 1730 vec<cluster *> output2; | |
| 1731 auto_vec<cluster *> tmp; | |
| 1732 output2.create (1); | |
| 1733 tmp.create (1); | |
| 1734 | |
| 1735 for (unsigned i = 0; i < output.length (); i++) | |
| 1736 { | |
| 1737 cluster *c = output[i]; | |
| 1738 if (c->get_type () != SIMPLE_CASE) | |
| 1739 { | |
| 1740 if (!tmp.is_empty ()) | |
| 1741 { | |
| 1742 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp); | |
| 1743 output2.safe_splice (n); | |
| 1744 n.release (); | |
| 1745 tmp.truncate (0); | |
| 1746 } | |
| 1747 output2.safe_push (c); | |
| 1748 } | |
| 1749 else | |
| 1750 tmp.safe_push (c); | |
| 1751 } | |
| 1752 | |
| 1753 /* We still can have a temporary vector to test. */ | |
| 1754 if (!tmp.is_empty ()) | |
| 1755 { | |
| 1756 vec<cluster *> n = bit_test_cluster::find_bit_tests (tmp); | |
| 1757 output2.safe_splice (n); | |
| 1758 n.release (); | |
| 1759 } | |
| 1760 | |
| 1761 if (dump_file) | |
| 1762 { | |
| 1763 fprintf (dump_file, ";; GIMPLE switch case clusters: "); | |
| 1764 for (unsigned i = 0; i < output2.length (); i++) | |
| 1765 output2[i]->dump (dump_file, dump_flags & TDF_DETAILS); | |
| 1766 fprintf (dump_file, "\n"); | |
| 1767 } | |
| 1768 | |
| 1769 output.release (); | |
| 1770 | |
| 1771 bool expanded = try_switch_expansion (output2); | |
| 1772 | |
| 1773 for (unsigned i = 0; i < output2.length (); i++) | |
| 1774 delete output2[i]; | |
| 1775 | |
| 1776 output2.release (); | |
| 1777 | |
| 1778 return expanded; | |
| 1779 } | |
| 1780 | |
| 1781 /* Attempt to expand CLUSTERS as a decision tree. Return true when | |
| 1782 expanded. */ | |
| 1783 | |
| 1784 bool | |
| 1785 switch_decision_tree::try_switch_expansion (vec<cluster *> &clusters) | |
| 1786 { | |
| 1787 tree index_expr = gimple_switch_index (m_switch); | |
| 1788 tree index_type = TREE_TYPE (index_expr); | |
| 1789 basic_block bb = gimple_bb (m_switch); | |
| 1790 | |
| 145 | 1791 if (gimple_switch_num_labels (m_switch) == 1 |
| 1792 || range_check_type (index_type) == NULL_TREE) | |
| 131 | 1793 return false; |
| 1794 | |
| 1795 /* Find the default case target label. */ | |
| 1796 edge default_edge = gimple_switch_default_edge (cfun, m_switch); | |
| 1797 m_default_bb = default_edge->dest; | |
| 1798 | |
| 1799 /* Do the insertion of a case label into m_case_list. The labels are | |
| 1800 fed to us in descending order from the sorted vector of case labels used | |
| 1801 in the tree part of the middle end. So the list we construct is | |
| 1802 sorted in ascending order. */ | |
| 1803 | |
| 1804 for (int i = clusters.length () - 1; i >= 0; i--) | |
| 1805 { | |
| 1806 case_tree_node *r = m_case_list; | |
| 1807 m_case_list = m_case_node_pool.allocate (); | |
| 1808 m_case_list->m_right = r; | |
| 1809 m_case_list->m_c = clusters[i]; | |
| 1810 } | |
| 1811 | |
| 1812 record_phi_operand_mapping (); | |
| 1813 | |
| 1814 /* Split basic block that contains the gswitch statement. */ | |
| 1815 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 1816 edge e; | |
| 1817 if (gsi_end_p (gsi)) | |
| 1818 e = split_block_after_labels (bb); | |
| 1819 else | |
| 1820 { | |
| 1821 gsi_prev (&gsi); | |
| 1822 e = split_block (bb, gsi_stmt (gsi)); | |
| 1823 } | |
| 1824 bb = split_edge (e); | |
| 1825 | |
| 1826 /* Create new basic blocks for non-case clusters where specific expansion | |
| 1827 needs to happen. */ | |
| 1828 for (unsigned i = 0; i < clusters.length (); i++) | |
| 1829 if (clusters[i]->get_type () != SIMPLE_CASE) | |
| 1830 { | |
| 1831 clusters[i]->m_case_bb = create_empty_bb (bb); | |
| 145 | 1832 clusters[i]->m_case_bb->count = bb->count; |
| 131 | 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 | |
| 145 | 1945 bb = emit_case_nodes (bb, index_expr, m_case_list, default_prob, index_type, |
| 1946 gimple_location (m_switch)); | |
| 131 | 1947 |
| 1948 if (bb) | |
| 1949 emit_jump (bb, m_default_bb); | |
| 1950 | |
| 1951 /* Remove all edges and do just an edge that will reach default_bb. */ | |
| 1952 bb = gimple_bb (m_switch); | |
| 1953 gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
| 1954 gsi_remove (&gsi, true); | |
| 1955 | |
| 1956 delete_basic_block (bb); | |
| 1957 } | |
| 1958 | |
| 1959 /* Take an ordered list of case nodes | |
| 1960 and transform them into a near optimal binary tree, | |
| 1961 on the assumption that any target code selection value is as | |
| 1962 likely as any other. | |
| 1963 | |
| 1964 The transformation is performed by splitting the ordered | |
| 1965 list into two equal sections plus a pivot. The parts are | |
| 1966 then attached to the pivot as left and right branches. Each | |
| 1967 branch is then transformed recursively. */ | |
| 1968 | |
| 1969 void | |
| 1970 switch_decision_tree::balance_case_nodes (case_tree_node **head, | |
| 1971 case_tree_node *parent) | |
| 1972 { | |
| 1973 case_tree_node *np; | |
| 1974 | |
| 1975 np = *head; | |
| 1976 if (np) | |
| 1977 { | |
| 1978 int i = 0; | |
| 1979 int ranges = 0; | |
| 1980 case_tree_node **npp; | |
| 1981 case_tree_node *left; | |
| 1982 profile_probability prob = profile_probability::never (); | |
| 1983 | |
| 1984 /* Count the number of entries on branch. Also count the ranges. */ | |
| 1985 | |
| 1986 while (np) | |
| 1987 { | |
| 1988 if (!tree_int_cst_equal (np->m_c->get_low (), np->m_c->get_high ())) | |
| 1989 ranges++; | |
| 1990 | |
| 1991 i++; | |
| 1992 prob += np->m_c->m_prob; | |
| 1993 np = np->m_right; | |
| 1994 } | |
| 1995 | |
| 1996 if (i > 2) | |
| 1997 { | |
| 1998 /* Split this list if it is long enough for that to help. */ | |
| 1999 npp = head; | |
| 2000 left = *npp; | |
| 2001 profile_probability pivot_prob = prob.apply_scale (1, 2); | |
| 2002 | |
| 2003 /* Find the place in the list that bisects the list's total cost, | |
| 2004 where ranges count as 2. */ | |
| 2005 while (1) | |
| 2006 { | |
| 2007 /* Skip nodes while their probability does not reach | |
| 2008 that amount. */ | |
| 2009 prob -= (*npp)->m_c->m_prob; | |
| 2010 if ((prob.initialized_p () && prob < pivot_prob) | |
| 2011 || ! (*npp)->m_right) | |
| 2012 break; | |
| 2013 npp = &(*npp)->m_right; | |
| 2014 } | |
| 2015 | |
| 2016 np = *npp; | |
| 2017 *npp = 0; | |
| 2018 *head = np; | |
| 2019 np->m_parent = parent; | |
| 2020 np->m_left = left == np ? NULL : left; | |
| 2021 | |
| 2022 /* Optimize each of the two split parts. */ | |
| 2023 balance_case_nodes (&np->m_left, np); | |
| 2024 balance_case_nodes (&np->m_right, np); | |
| 2025 np->m_c->m_subtree_prob = np->m_c->m_prob; | |
| 2026 if (np->m_left) | |
| 2027 np->m_c->m_subtree_prob += np->m_left->m_c->m_subtree_prob; | |
| 2028 if (np->m_right) | |
| 2029 np->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; | |
| 2030 } | |
| 2031 else | |
| 2032 { | |
| 2033 /* Else leave this branch as one level, | |
| 2034 but fill in `parent' fields. */ | |
| 2035 np = *head; | |
| 2036 np->m_parent = parent; | |
| 2037 np->m_c->m_subtree_prob = np->m_c->m_prob; | |
| 2038 for (; np->m_right; np = np->m_right) | |
| 2039 { | |
| 2040 np->m_right->m_parent = np; | |
| 2041 (*head)->m_c->m_subtree_prob += np->m_right->m_c->m_subtree_prob; | |
| 2042 } | |
| 2043 } | |
| 2044 } | |
| 2045 } | |
| 2046 | |
| 2047 /* Dump ROOT, a list or tree of case nodes, to file. */ | |
| 2048 | |
| 2049 void | |
| 2050 switch_decision_tree::dump_case_nodes (FILE *f, case_tree_node *root, | |
| 2051 int indent_step, int indent_level) | |
| 2052 { | |
| 2053 if (root == 0) | |
| 2054 return; | |
| 2055 indent_level++; | |
| 2056 | |
| 2057 dump_case_nodes (f, root->m_left, indent_step, indent_level); | |
| 2058 | |
| 2059 fputs (";; ", f); | |
| 2060 fprintf (f, "%*s", indent_step * indent_level, ""); | |
| 2061 root->m_c->dump (f); | |
| 2062 root->m_c->m_prob.dump (f); | |
| 2063 fputs (" subtree: ", f); | |
| 2064 root->m_c->m_subtree_prob.dump (f); | |
| 2065 fputs (")\n", f); | |
| 2066 | |
| 2067 dump_case_nodes (f, root->m_right, indent_step, indent_level); | |
| 2068 } | |
| 2069 | |
| 2070 | |
| 2071 /* Add an unconditional jump to CASE_BB that happens in basic block BB. */ | |
| 2072 | |
| 2073 void | |
| 2074 switch_decision_tree::emit_jump (basic_block bb, basic_block case_bb) | |
| 2075 { | |
| 2076 edge e = single_succ_edge (bb); | |
| 2077 redirect_edge_succ (e, case_bb); | |
| 2078 } | |
| 2079 | |
| 2080 /* Generate code to compare OP0 with OP1 so that the condition codes are | |
| 2081 set and to jump to LABEL_BB if the condition is true. | |
| 2082 COMPARISON is the GIMPLE comparison (EQ, NE, GT, etc.). | |
| 2083 PROB is the probability of jumping to LABEL_BB. */ | |
| 2084 | |
| 2085 basic_block | |
| 2086 switch_decision_tree::emit_cmp_and_jump_insns (basic_block bb, tree op0, | |
| 2087 tree op1, tree_code comparison, | |
| 2088 basic_block label_bb, | |
| 145 | 2089 profile_probability prob, |
| 2090 location_t loc) | |
| 131 | 2091 { |
| 2092 // TODO: it's once called with lhs != index. | |
| 2093 op1 = fold_convert (TREE_TYPE (op0), op1); | |
| 2094 | |
| 2095 gcond *cond = gimple_build_cond (comparison, op0, op1, NULL_TREE, NULL_TREE); | |
| 145 | 2096 gimple_set_location (cond, loc); |
| 131 | 2097 gimple_stmt_iterator gsi = gsi_last_bb (bb); |
| 2098 gsi_insert_after (&gsi, cond, GSI_NEW_STMT); | |
| 2099 | |
| 2100 gcc_assert (single_succ_p (bb)); | |
| 2101 | |
| 2102 /* Make a new basic block where false branch will take place. */ | |
| 2103 edge false_edge = split_block (bb, cond); | |
| 2104 false_edge->flags = EDGE_FALSE_VALUE; | |
| 2105 false_edge->probability = prob.invert (); | |
| 2106 | |
| 2107 edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
| 2108 true_edge->probability = prob; | |
| 2109 | |
| 2110 return false_edge->dest; | |
| 2111 } | |
| 2112 | |
| 2113 /* Generate code to jump to LABEL if OP0 and OP1 are equal. | |
| 2114 PROB is the probability of jumping to LABEL_BB. | |
| 2115 BB is a basic block where the new condition will be placed. */ | |
| 2116 | |
| 2117 basic_block | |
| 2118 switch_decision_tree::do_jump_if_equal (basic_block bb, tree op0, tree op1, | |
| 2119 basic_block label_bb, | |
| 145 | 2120 profile_probability prob, |
| 2121 location_t loc) | |
| 131 | 2122 { |
| 2123 op1 = fold_convert (TREE_TYPE (op0), op1); | |
| 2124 | |
| 2125 gcond *cond = gimple_build_cond (EQ_EXPR, op0, op1, NULL_TREE, NULL_TREE); | |
| 145 | 2126 gimple_set_location (cond, loc); |
| 131 | 2127 gimple_stmt_iterator gsi = gsi_last_bb (bb); |
| 2128 gsi_insert_before (&gsi, cond, GSI_SAME_STMT); | |
| 2129 | |
| 2130 gcc_assert (single_succ_p (bb)); | |
| 2131 | |
| 2132 /* Make a new basic block where false branch will take place. */ | |
| 2133 edge false_edge = split_block (bb, cond); | |
| 2134 false_edge->flags = EDGE_FALSE_VALUE; | |
| 2135 false_edge->probability = prob.invert (); | |
| 2136 | |
| 2137 edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE); | |
| 2138 true_edge->probability = prob; | |
| 2139 | |
| 2140 return false_edge->dest; | |
| 2141 } | |
| 2142 | |
| 2143 /* Emit step-by-step code to select a case for the value of INDEX. | |
| 2144 The thus generated decision tree follows the form of the | |
| 2145 case-node binary tree NODE, whose nodes represent test conditions. | |
| 2146 DEFAULT_PROB is probability of cases leading to default BB. | |
| 2147 INDEX_TYPE is the type of the index of the switch. */ | |
| 2148 | |
| 2149 basic_block | |
| 2150 switch_decision_tree::emit_case_nodes (basic_block bb, tree index, | |
| 2151 case_tree_node *node, | |
| 2152 profile_probability default_prob, | |
| 145 | 2153 tree index_type, location_t loc) |
| 131 | 2154 { |
| 2155 profile_probability p; | |
| 2156 | |
| 2157 /* If node is null, we are done. */ | |
| 2158 if (node == NULL) | |
| 2159 return bb; | |
| 2160 | |
| 2161 /* Single value case. */ | |
| 2162 if (node->m_c->is_single_value_p ()) | |
| 2163 { | |
| 2164 /* Node is single valued. First see if the index expression matches | |
| 2165 this node and then check our children, if any. */ | |
| 2166 p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2167 bb = do_jump_if_equal (bb, index, node->m_c->get_low (), | |
| 145 | 2168 node->m_c->m_case_bb, p, loc); |
| 131 | 2169 /* Since this case is taken at this point, reduce its weight from |
| 2170 subtree_weight. */ | |
| 2171 node->m_c->m_subtree_prob -= p; | |
| 2172 | |
| 2173 if (node->m_left != NULL && node->m_right != NULL) | |
| 2174 { | |
| 2175 /* 1) the node has both children | |
| 2176 | |
| 2177 If both children are single-valued cases with no | |
| 2178 children, finish up all the work. This way, we can save | |
| 2179 one ordered comparison. */ | |
| 2180 | |
| 2181 if (!node->m_left->has_child () | |
| 2182 && node->m_left->m_c->is_single_value_p () | |
| 2183 && !node->m_right->has_child () | |
| 2184 && node->m_right->m_c->is_single_value_p ()) | |
| 2185 { | |
| 2186 p = (node->m_right->m_c->m_prob | |
| 2187 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2188 bb = do_jump_if_equal (bb, index, node->m_right->m_c->get_low (), | |
| 145 | 2189 node->m_right->m_c->m_case_bb, p, loc); |
| 131 | 2190 |
| 2191 p = (node->m_left->m_c->m_prob | |
| 2192 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2193 bb = do_jump_if_equal (bb, index, node->m_left->m_c->get_low (), | |
| 145 | 2194 node->m_left->m_c->m_case_bb, p, loc); |
| 131 | 2195 } |
| 2196 else | |
| 2197 { | |
| 2198 /* Branch to a label where we will handle it later. */ | |
| 2199 basic_block test_bb = split_edge (single_succ_edge (bb)); | |
| 2200 redirect_edge_succ (single_pred_edge (test_bb), | |
| 2201 single_succ_edge (bb)->dest); | |
| 2202 | |
| 2203 p = ((node->m_right->m_c->m_subtree_prob | |
| 2204 + default_prob.apply_scale (1, 2)) | |
| 2205 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2206 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 145 | 2207 GT_EXPR, test_bb, p, loc); |
| 131 | 2208 default_prob = default_prob.apply_scale (1, 2); |
| 2209 | |
| 2210 /* Handle the left-hand subtree. */ | |
| 2211 bb = emit_case_nodes (bb, index, node->m_left, | |
| 145 | 2212 default_prob, index_type, loc); |
| 131 | 2213 |
| 2214 /* If the left-hand subtree fell through, | |
| 2215 don't let it fall into the right-hand subtree. */ | |
| 2216 if (bb && m_default_bb) | |
| 2217 emit_jump (bb, m_default_bb); | |
| 2218 | |
| 2219 bb = emit_case_nodes (test_bb, index, node->m_right, | |
| 145 | 2220 default_prob, index_type, loc); |
| 131 | 2221 } |
| 2222 } | |
| 2223 else if (node->m_left == NULL && node->m_right != NULL) | |
| 2224 { | |
| 2225 /* 2) the node has only right child. */ | |
| 2226 | |
| 2227 /* Here we have a right child but no left so we issue a conditional | |
| 2228 branch to default and process the right child. | |
| 2229 | |
| 2230 Omit the conditional branch to default if the right child | |
| 2231 does not have any children and is single valued; it would | |
| 2232 cost too much space to save so little time. */ | |
| 2233 | |
| 2234 if (node->m_right->has_child () | |
| 2235 || !node->m_right->m_c->is_single_value_p ()) | |
| 2236 { | |
| 2237 p = (default_prob.apply_scale (1, 2) | |
| 2238 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2239 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_low (), | |
| 145 | 2240 LT_EXPR, m_default_bb, p, loc); |
| 131 | 2241 default_prob = default_prob.apply_scale (1, 2); |
| 2242 | |
| 2243 bb = emit_case_nodes (bb, index, node->m_right, default_prob, | |
| 145 | 2244 index_type, loc); |
| 131 | 2245 } |
| 2246 else | |
| 2247 { | |
| 2248 /* We cannot process node->right normally | |
| 2249 since we haven't ruled out the numbers less than | |
| 2250 this node's value. So handle node->right explicitly. */ | |
| 2251 p = (node->m_right->m_c->m_subtree_prob | |
| 2252 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2253 bb = do_jump_if_equal (bb, index, node->m_right->m_c->get_low (), | |
| 145 | 2254 node->m_right->m_c->m_case_bb, p, loc); |
| 131 | 2255 } |
| 2256 } | |
| 2257 else if (node->m_left != NULL && node->m_right == NULL) | |
| 2258 { | |
| 2259 /* 3) just one subtree, on the left. Similar case as previous. */ | |
| 2260 | |
| 2261 if (node->m_left->has_child () | |
| 2262 || !node->m_left->m_c->is_single_value_p ()) | |
| 2263 { | |
| 2264 p = (default_prob.apply_scale (1, 2) | |
| 2265 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2266 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 145 | 2267 GT_EXPR, m_default_bb, p, loc); |
| 131 | 2268 default_prob = default_prob.apply_scale (1, 2); |
| 2269 | |
| 2270 bb = emit_case_nodes (bb, index, node->m_left, default_prob, | |
| 145 | 2271 index_type, loc); |
| 131 | 2272 } |
| 2273 else | |
| 2274 { | |
| 2275 /* We cannot process node->left normally | |
| 2276 since we haven't ruled out the numbers less than | |
| 2277 this node's value. So handle node->left explicitly. */ | |
| 2278 p = (node->m_left->m_c->m_subtree_prob | |
| 2279 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2280 bb = do_jump_if_equal (bb, index, node->m_left->m_c->get_low (), | |
| 145 | 2281 node->m_left->m_c->m_case_bb, p, loc); |
| 131 | 2282 } |
| 2283 } | |
| 2284 } | |
| 2285 else | |
| 2286 { | |
| 2287 /* Node is a range. These cases are very similar to those for a single | |
| 2288 value, except that we do not start by testing whether this node | |
| 2289 is the one to branch to. */ | |
| 2290 if (node->has_child () || node->m_c->get_type () != SIMPLE_CASE) | |
| 2291 { | |
| 2292 /* Branch to a label where we will handle it later. */ | |
| 2293 basic_block test_bb = split_edge (single_succ_edge (bb)); | |
| 2294 redirect_edge_succ (single_pred_edge (test_bb), | |
| 2295 single_succ_edge (bb)->dest); | |
| 2296 | |
| 2297 | |
| 2298 profile_probability right_prob = profile_probability::never (); | |
| 2299 if (node->m_right) | |
| 2300 right_prob = node->m_right->m_c->m_subtree_prob; | |
| 2301 p = ((right_prob + default_prob.apply_scale (1, 2)) | |
| 2302 / (node->m_c->m_subtree_prob + default_prob)); | |
| 2303 | |
| 2304 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (), | |
| 145 | 2305 GT_EXPR, test_bb, p, loc); |
| 131 | 2306 default_prob = default_prob.apply_scale (1, 2); |
| 2307 | |
| 2308 /* Value belongs to this node or to the left-hand subtree. */ | |
| 2309 p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2310 bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_low (), | |
| 145 | 2311 GE_EXPR, node->m_c->m_case_bb, p, loc); |
| 131 | 2312 |
| 2313 /* Handle the left-hand subtree. */ | |
| 2314 bb = emit_case_nodes (bb, index, node->m_left, | |
| 145 | 2315 default_prob, index_type, loc); |
| 131 | 2316 |
| 2317 /* If the left-hand subtree fell through, | |
| 2318 don't let it fall into the right-hand subtree. */ | |
| 2319 if (bb && m_default_bb) | |
| 2320 emit_jump (bb, m_default_bb); | |
| 2321 | |
| 2322 bb = emit_case_nodes (test_bb, index, node->m_right, | |
| 145 | 2323 default_prob, index_type, loc); |
| 131 | 2324 } |
| 2325 else | |
| 2326 { | |
| 2327 /* Node has no children so we check low and high bounds to remove | |
| 2328 redundant tests. Only one of the bounds can exist, | |
| 2329 since otherwise this node is bounded--a case tested already. */ | |
| 2330 tree lhs, rhs; | |
| 2331 generate_range_test (bb, index, node->m_c->get_low (), | |
| 2332 node->m_c->get_high (), &lhs, &rhs); | |
| 2333 p = default_prob / (node->m_c->m_subtree_prob + default_prob); | |
| 2334 | |
| 2335 bb = emit_cmp_and_jump_insns (bb, lhs, rhs, GT_EXPR, | |
| 145 | 2336 m_default_bb, p, loc); |
| 131 | 2337 |
| 2338 emit_jump (bb, node->m_c->m_case_bb); | |
| 2339 return NULL; | |
| 2340 } | |
| 2341 } | |
| 2342 | |
| 2343 return bb; | |
| 0 | 2344 } |
| 2345 | |
| 2346 /* The main function of the pass scans statements for switches and invokes | |
| 2347 process_switch on them. */ | |
| 2348 | |
| 111 | 2349 namespace { |
| 2350 | |
| 2351 const pass_data pass_data_convert_switch = | |
| 2352 { | |
| 2353 GIMPLE_PASS, /* type */ | |
| 2354 "switchconv", /* name */ | |
| 2355 OPTGROUP_NONE, /* optinfo_flags */ | |
| 2356 TV_TREE_SWITCH_CONVERSION, /* tv_id */ | |
| 2357 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 2358 0, /* properties_provided */ | |
| 2359 0, /* properties_destroyed */ | |
| 2360 0, /* todo_flags_start */ | |
| 2361 TODO_update_ssa, /* todo_flags_finish */ | |
| 2362 }; | |
| 2363 | |
| 2364 class pass_convert_switch : public gimple_opt_pass | |
| 2365 { | |
| 2366 public: | |
| 2367 pass_convert_switch (gcc::context *ctxt) | |
| 2368 : gimple_opt_pass (pass_data_convert_switch, ctxt) | |
| 2369 {} | |
| 2370 | |
| 2371 /* opt_pass methods: */ | |
| 2372 virtual bool gate (function *) { return flag_tree_switch_conversion != 0; } | |
| 2373 virtual unsigned int execute (function *); | |
| 2374 | |
| 2375 }; // class pass_convert_switch | |
| 2376 | |
| 2377 unsigned int | |
| 2378 pass_convert_switch::execute (function *fun) | |
| 0 | 2379 { |
| 2380 basic_block bb; | |
| 131 | 2381 bool cfg_altered = false; |
| 0 | 2382 |
| 111 | 2383 FOR_EACH_BB_FN (bb, fun) |
| 0 | 2384 { |
| 111 | 2385 gimple *stmt = last_stmt (bb); |
| 0 | 2386 if (stmt && gimple_code (stmt) == GIMPLE_SWITCH) |
| 2387 { | |
| 2388 if (dump_file) | |
| 2389 { | |
| 2390 expanded_location loc = expand_location (gimple_location (stmt)); | |
| 2391 | |
| 2392 fprintf (dump_file, "beginning to process the following " | |
| 2393 "SWITCH statement (%s:%d) : ------- \n", | |
| 2394 loc.file, loc.line); | |
| 2395 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
|
2396 putc ('\n', dump_file); |
| 0 | 2397 } |
| 2398 | |
| 131 | 2399 switch_conversion sconv; |
| 2400 sconv.expand (as_a <gswitch *> (stmt)); | |
| 2401 cfg_altered |= sconv.m_cfg_altered; | |
| 2402 if (!sconv.m_reason) | |
| 0 | 2403 { |
| 2404 if (dump_file) | |
| 2405 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2406 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
|
2407 fputs ("--------------------------------\n", dump_file); |
| 0 | 2408 } |
| 111 | 2409 |
| 131 | 2410 /* Make no effort to update the post-dominator tree. |
| 2411 It is actually not that hard for the transformations | |
| 2412 we have performed, but it is not supported | |
| 2413 by iterate_fix_dominators. */ | |
| 111 | 2414 free_dominance_info (CDI_POST_DOMINATORS); |
| 0 | 2415 } |
| 2416 else | |
| 2417 { | |
| 2418 if (dump_file) | |
| 2419 { | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
2420 fputs ("Bailing out - ", dump_file); |
| 131 | 2421 fputs (sconv.m_reason, dump_file); |
| 111 | 2422 fputs ("\n--------------------------------\n", dump_file); |
| 0 | 2423 } |
| 2424 } | |
| 2425 } | |
| 2426 } | |
| 2427 | |
| 131 | 2428 return cfg_altered ? TODO_cleanup_cfg : 0;; |
| 0 | 2429 } |
| 2430 | |
| 111 | 2431 } // anon namespace |
| 2432 | |
| 2433 gimple_opt_pass * | |
| 2434 make_pass_convert_switch (gcc::context *ctxt) | |
| 2435 { | |
| 2436 return new pass_convert_switch (ctxt); | |
| 2437 } | |
| 2438 | |
| 2439 /* The main function of the pass scans statements for switches and invokes | |
| 2440 process_switch on them. */ | |
| 2441 | |
| 2442 namespace { | |
| 2443 | |
| 131 | 2444 template <bool O0> class pass_lower_switch: public gimple_opt_pass |
| 111 | 2445 { |
| 131 | 2446 public: |
| 2447 pass_lower_switch (gcc::context *ctxt) : gimple_opt_pass (data, ctxt) {} | |
| 2448 | |
| 2449 static const pass_data data; | |
| 2450 opt_pass * | |
| 2451 clone () | |
| 2452 { | |
| 2453 return new pass_lower_switch<O0> (m_ctxt); | |
| 2454 } | |
| 2455 | |
| 2456 virtual bool | |
| 2457 gate (function *) | |
| 2458 { | |
| 2459 return !O0 || !optimize; | |
| 2460 } | |
| 2461 | |
| 2462 virtual unsigned int execute (function *fun); | |
| 2463 }; // class pass_lower_switch | |
| 2464 | |
| 2465 template <bool O0> | |
| 2466 const pass_data pass_lower_switch<O0>::data = { | |
| 2467 GIMPLE_PASS, /* type */ | |
| 2468 O0 ? "switchlower_O0" : "switchlower", /* name */ | |
| 111 | 2469 OPTGROUP_NONE, /* optinfo_flags */ |
| 2470 TV_TREE_SWITCH_LOWERING, /* tv_id */ | |
| 2471 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
| 2472 0, /* properties_provided */ | |
| 2473 0, /* properties_destroyed */ | |
| 2474 0, /* todo_flags_start */ | |
| 2475 TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ | |
| 2476 }; | |
| 2477 | |
| 131 | 2478 template <bool O0> |
| 111 | 2479 unsigned int |
| 131 | 2480 pass_lower_switch<O0>::execute (function *fun) |
| 111 | 2481 { |
| 2482 basic_block bb; | |
| 2483 bool expanded = false; | |
| 2484 | |
| 131 | 2485 auto_vec<gimple *> switch_statements; |
| 2486 switch_statements.create (1); | |
| 2487 | |
| 111 | 2488 FOR_EACH_BB_FN (bb, fun) |
| 2489 { | |
| 2490 gimple *stmt = last_stmt (bb); | |
| 131 | 2491 gswitch *swtch; |
| 2492 if (stmt && (swtch = dyn_cast<gswitch *> (stmt))) | |
| 2493 { | |
| 2494 if (!O0) | |
| 2495 group_case_labels_stmt (swtch); | |
| 2496 switch_statements.safe_push (swtch); | |
| 2497 } | |
| 2498 } | |
| 2499 | |
| 2500 for (unsigned i = 0; i < switch_statements.length (); i++) | |
| 2501 { | |
| 2502 gimple *stmt = switch_statements[i]; | |
| 2503 if (dump_file) | |
| 111 | 2504 { |
| 131 | 2505 expanded_location loc = expand_location (gimple_location (stmt)); |
| 2506 | |
| 2507 fprintf (dump_file, "beginning to process the following " | |
| 2508 "SWITCH statement (%s:%d) : ------- \n", | |
| 2509 loc.file, loc.line); | |
| 2510 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
| 2511 putc ('\n', dump_file); | |
| 2512 } | |
| 2513 | |
| 2514 gswitch *swtch = dyn_cast<gswitch *> (stmt); | |
| 2515 if (swtch) | |
| 2516 { | |
| 2517 switch_decision_tree dt (swtch); | |
| 2518 expanded |= dt.analyze_switch_statement (); | |
| 111 | 2519 } |
| 2520 } | |
| 2521 | |
| 2522 if (expanded) | |
| 2523 { | |
| 2524 free_dominance_info (CDI_DOMINATORS); | |
| 2525 free_dominance_info (CDI_POST_DOMINATORS); | |
| 2526 mark_virtual_operands_for_renaming (cfun); | |
| 2527 } | |
| 2528 | |
| 2529 return 0; | |
| 2530 } | |
| 2531 | |
| 2532 } // anon namespace | |
| 2533 | |
| 2534 gimple_opt_pass * | |
| 131 | 2535 make_pass_lower_switch_O0 (gcc::context *ctxt) |
| 2536 { | |
| 2537 return new pass_lower_switch<true> (ctxt); | |
| 2538 } | |
| 2539 gimple_opt_pass * | |
| 111 | 2540 make_pass_lower_switch (gcc::context *ctxt) |
| 2541 { | |
| 131 | 2542 return new pass_lower_switch<false> (ctxt); |
| 111 | 2543 } |
| 2544 | |
| 131 | 2545 |