| 1 | /* Gimple ranger SSA cache implementation. |
| 2 | Copyright (C) 2017-2025 Free Software Foundation, Inc. |
| 3 | Contributed by Andrew MacLeod <amacleod@redhat.com>. |
| 4 | |
| 5 | This file is part of GCC. |
| 6 | |
| 7 | GCC is free software; you can redistribute it and/or modify |
| 8 | it under the terms of the GNU General Public License as published by |
| 9 | the Free Software Foundation; either version 3, or (at your option) |
| 10 | any later version. |
| 11 | |
| 12 | GCC is distributed in the hope that it will be useful, |
| 13 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 15 | GNU General Public License 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 see |
| 19 | <http://www.gnu.org/licenses/>. */ |
| 20 | |
| 21 | #include "config.h" |
| 22 | #include "system.h" |
| 23 | #include "coretypes.h" |
| 24 | #include "backend.h" |
| 25 | #include "insn-codes.h" |
| 26 | #include "tree.h" |
| 27 | #include "gimple.h" |
| 28 | #include "ssa.h" |
| 29 | #include "gimple-pretty-print.h" |
| 30 | #include "gimple-range.h" |
| 31 | #include "value-range-storage.h" |
| 32 | #include "tree-cfg.h" |
| 33 | #include "target.h" |
| 34 | #include "attribs.h" |
| 35 | #include "gimple-iterator.h" |
| 36 | #include "gimple-walk.h" |
| 37 | #include "cfganal.h" |
| 38 | |
| 39 | #define DEBUG_RANGE_CACHE (dump_file \ |
| 40 | && (param_ranger_debug & RANGER_DEBUG_CACHE)) |
| 41 | |
| 42 | // This class represents the API into a cache of ranges for an SSA_NAME. |
| 43 | // Routines must be implemented to set, get, and query if a value is set. |
| 44 | |
| 45 | class ssa_block_ranges |
| 46 | { |
| 47 | public: |
| 48 | ssa_block_ranges (tree t) : m_type (t) { } |
| 49 | virtual bool set_bb_range (const_basic_block bb, const vrange &r) = 0; |
| 50 | virtual bool get_bb_range (vrange &r, const_basic_block bb) = 0; |
| 51 | virtual bool bb_range_p (const_basic_block bb) = 0; |
| 52 | |
| 53 | void dump(FILE *f); |
| 54 | private: |
| 55 | tree m_type; |
| 56 | }; |
| 57 | |
| 58 | // Print the list of known ranges for file F in a nice format. |
| 59 | |
| 60 | void |
| 61 | ssa_block_ranges::dump (FILE *f) |
| 62 | { |
| 63 | basic_block bb; |
| 64 | value_range r (m_type); |
| 65 | |
| 66 | FOR_EACH_BB_FN (bb, cfun) |
| 67 | if (get_bb_range (r, bb)) |
| 68 | { |
| 69 | fprintf (stream: f, format: "BB%d -> " , bb->index); |
| 70 | r.dump (f); |
| 71 | fprintf (stream: f, format: "\n" ); |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | // This class implements the range cache as a linear vector, indexed by BB. |
| 76 | // It caches a varying and undefined range which are used instead of |
| 77 | // allocating new ones each time. |
| 78 | |
| 79 | class sbr_vector : public ssa_block_ranges |
| 80 | { |
| 81 | public: |
| 82 | sbr_vector (tree t, vrange_allocator *allocator, bool zero_p = true); |
| 83 | |
| 84 | virtual bool set_bb_range (const_basic_block bb, const vrange &r) override; |
| 85 | virtual bool get_bb_range (vrange &r, const_basic_block bb) override; |
| 86 | virtual bool bb_range_p (const_basic_block bb) override; |
| 87 | protected: |
| 88 | vrange_storage **m_tab; // Non growing vector. |
| 89 | int m_tab_size; |
| 90 | vrange_storage *m_varying; |
| 91 | vrange_storage *m_undefined; |
| 92 | tree m_type; |
| 93 | vrange_allocator *m_range_allocator; |
| 94 | bool m_zero_p; |
| 95 | void grow (); |
| 96 | }; |
| 97 | |
| 98 | |
| 99 | // Initialize a block cache for an ssa_name of type T. |
| 100 | |
| 101 | sbr_vector::sbr_vector (tree t, vrange_allocator *allocator, bool zero_p) |
| 102 | : ssa_block_ranges (t) |
| 103 | { |
| 104 | gcc_checking_assert (TYPE_P (t)); |
| 105 | m_type = t; |
| 106 | m_zero_p = zero_p; |
| 107 | m_range_allocator = allocator; |
| 108 | m_tab_size = last_basic_block_for_fn (cfun) + 1; |
| 109 | m_tab = static_cast <vrange_storage **> |
| 110 | (allocator->alloc (size: m_tab_size * sizeof (vrange_storage *))); |
| 111 | if (zero_p) |
| 112 | memset (s: m_tab, c: 0, n: m_tab_size * sizeof (vrange *)); |
| 113 | |
| 114 | // Create the cached type range. |
| 115 | m_varying = m_range_allocator->clone_varying (type: t); |
| 116 | m_undefined = m_range_allocator->clone_undefined (type: t); |
| 117 | } |
| 118 | |
| 119 | // Grow the vector when the CFG has increased in size. |
| 120 | |
| 121 | void |
| 122 | sbr_vector::grow () |
| 123 | { |
| 124 | int curr_bb_size = last_basic_block_for_fn (cfun); |
| 125 | gcc_checking_assert (curr_bb_size > m_tab_size); |
| 126 | |
| 127 | // Increase the max of a)128, b)needed increase * 2, c)10% of current_size. |
| 128 | int inc = MAX ((curr_bb_size - m_tab_size) * 2, 128); |
| 129 | inc = MAX (inc, curr_bb_size / 10); |
| 130 | int new_size = inc + curr_bb_size; |
| 131 | |
| 132 | // Allocate new memory, copy the old vector and clear the new space. |
| 133 | vrange_storage **t = static_cast <vrange_storage **> |
| 134 | (m_range_allocator->alloc (size: new_size * sizeof (vrange_storage *))); |
| 135 | memcpy (dest: t, src: m_tab, n: m_tab_size * sizeof (vrange_storage *)); |
| 136 | if (m_zero_p) |
| 137 | memset (s: t + m_tab_size, c: 0, n: (new_size - m_tab_size) * sizeof (vrange_storage *)); |
| 138 | |
| 139 | m_tab = t; |
| 140 | m_tab_size = new_size; |
| 141 | } |
| 142 | |
| 143 | // Set the range for block BB to be R. |
| 144 | |
| 145 | bool |
| 146 | sbr_vector::set_bb_range (const_basic_block bb, const vrange &r) |
| 147 | { |
| 148 | vrange_storage *m; |
| 149 | if (bb->index >= m_tab_size) |
| 150 | grow (); |
| 151 | if (r.varying_p ()) |
| 152 | m = m_varying; |
| 153 | else if (r.undefined_p ()) |
| 154 | m = m_undefined; |
| 155 | else |
| 156 | m = m_range_allocator->clone (r); |
| 157 | m_tab[bb->index] = m; |
| 158 | return true; |
| 159 | } |
| 160 | |
| 161 | // Return the range associated with block BB in R. Return false if |
| 162 | // there is no range. |
| 163 | |
| 164 | bool |
| 165 | sbr_vector::get_bb_range (vrange &r, const_basic_block bb) |
| 166 | { |
| 167 | if (bb->index >= m_tab_size) |
| 168 | return false; |
| 169 | vrange_storage *m = m_tab[bb->index]; |
| 170 | if (m) |
| 171 | { |
| 172 | m->get_vrange (r, type: m_type); |
| 173 | return true; |
| 174 | } |
| 175 | return false; |
| 176 | } |
| 177 | |
| 178 | // Return true if a range is present. |
| 179 | |
| 180 | bool |
| 181 | sbr_vector::bb_range_p (const_basic_block bb) |
| 182 | { |
| 183 | if (bb->index < m_tab_size) |
| 184 | return m_tab[bb->index] != NULL; |
| 185 | return false; |
| 186 | } |
| 187 | |
| 188 | // Like an sbr_vector, except it uses a bitmap to manage whetehr vale is set |
| 189 | // or not rather than cleared memory. |
| 190 | |
| 191 | class sbr_lazy_vector : public sbr_vector |
| 192 | { |
| 193 | public: |
| 194 | sbr_lazy_vector (tree t, vrange_allocator *allocator, bitmap_obstack *bm); |
| 195 | |
| 196 | virtual bool set_bb_range (const_basic_block bb, const vrange &r) override; |
| 197 | virtual bool get_bb_range (vrange &r, const_basic_block bb) override; |
| 198 | virtual bool bb_range_p (const_basic_block bb) override; |
| 199 | protected: |
| 200 | bitmap m_has_value; |
| 201 | }; |
| 202 | |
| 203 | sbr_lazy_vector::sbr_lazy_vector (tree t, vrange_allocator *allocator, |
| 204 | bitmap_obstack *bm) |
| 205 | : sbr_vector (t, allocator, false) |
| 206 | { |
| 207 | m_has_value = BITMAP_ALLOC (obstack: bm); |
| 208 | } |
| 209 | |
| 210 | bool |
| 211 | sbr_lazy_vector::set_bb_range (const_basic_block bb, const vrange &r) |
| 212 | { |
| 213 | sbr_vector::set_bb_range (bb, r); |
| 214 | bitmap_set_bit (m_has_value, bb->index); |
| 215 | return true; |
| 216 | } |
| 217 | |
| 218 | bool |
| 219 | sbr_lazy_vector::get_bb_range (vrange &r, const_basic_block bb) |
| 220 | { |
| 221 | if (bitmap_bit_p (m_has_value, bb->index)) |
| 222 | return sbr_vector::get_bb_range (r, bb); |
| 223 | return false; |
| 224 | } |
| 225 | |
| 226 | bool |
| 227 | sbr_lazy_vector::bb_range_p (const_basic_block bb) |
| 228 | { |
| 229 | return bitmap_bit_p (m_has_value, bb->index); |
| 230 | } |
| 231 | |
| 232 | // This class implements the on entry cache via a sparse bitmap. |
| 233 | // It uses the quad bit routines to access 4 bits at a time. |
| 234 | // A value of 0 (the default) means there is no entry, and a value of |
| 235 | // 1 thru SBR_NUM represents an element in the m_range vector. |
| 236 | // Varying is given the first value (1) and pre-cached. |
| 237 | // SBR_NUM + 1 represents the value of UNDEFINED, and is never stored. |
| 238 | // SBR_NUM is the number of values that can be cached. |
| 239 | // Indexes are 1..SBR_NUM and are stored locally at m_range[0..SBR_NUM-1] |
| 240 | |
| 241 | #define SBR_NUM 14 |
| 242 | #define SBR_UNDEF SBR_NUM + 1 |
| 243 | #define SBR_VARYING 1 |
| 244 | |
| 245 | class sbr_sparse_bitmap : public ssa_block_ranges |
| 246 | { |
| 247 | public: |
| 248 | sbr_sparse_bitmap (tree t, vrange_allocator *allocator, bitmap_obstack *bm); |
| 249 | virtual bool set_bb_range (const_basic_block bb, const vrange &r) override; |
| 250 | virtual bool get_bb_range (vrange &r, const_basic_block bb) override; |
| 251 | virtual bool bb_range_p (const_basic_block bb) override; |
| 252 | private: |
| 253 | void bitmap_set_quad (bitmap head, int quad, int quad_value); |
| 254 | int bitmap_get_quad (const_bitmap head, int quad); |
| 255 | vrange_allocator *m_range_allocator; |
| 256 | vrange_storage *m_range[SBR_NUM]; |
| 257 | bitmap_head bitvec; |
| 258 | tree m_type; |
| 259 | }; |
| 260 | |
| 261 | // Initialize a block cache for an ssa_name of type T. |
| 262 | |
| 263 | sbr_sparse_bitmap::sbr_sparse_bitmap (tree t, vrange_allocator *allocator, |
| 264 | bitmap_obstack *bm) |
| 265 | : ssa_block_ranges (t) |
| 266 | { |
| 267 | gcc_checking_assert (TYPE_P (t)); |
| 268 | m_type = t; |
| 269 | bitmap_initialize (head: &bitvec, obstack: bm); |
| 270 | bitmap_tree_view (&bitvec); |
| 271 | m_range_allocator = allocator; |
| 272 | // Pre-cache varying. |
| 273 | m_range[0] = m_range_allocator->clone_varying (type: t); |
| 274 | // Pre-cache zero and non-zero values for pointers. |
| 275 | if (POINTER_TYPE_P (t)) |
| 276 | { |
| 277 | prange nonzero; |
| 278 | nonzero.set_nonzero (t); |
| 279 | m_range[1] = m_range_allocator->clone (r: nonzero); |
| 280 | prange zero; |
| 281 | zero.set_zero (t); |
| 282 | m_range[2] = m_range_allocator->clone (r: zero); |
| 283 | } |
| 284 | else |
| 285 | m_range[1] = m_range[2] = NULL; |
| 286 | // Clear SBR_NUM entries. |
| 287 | for (int x = 3; x < SBR_NUM; x++) |
| 288 | m_range[x] = 0; |
| 289 | } |
| 290 | |
| 291 | // Set 4 bit values in a sparse bitmap. This allows a bitmap to |
| 292 | // function as a sparse array of 4 bit values. |
| 293 | // QUAD is the index, QUAD_VALUE is the 4 bit value to set. |
| 294 | |
| 295 | inline void |
| 296 | sbr_sparse_bitmap::bitmap_set_quad (bitmap head, int quad, int quad_value) |
| 297 | { |
| 298 | bitmap_set_aligned_chunk (head, quad, 4, (BITMAP_WORD) quad_value); |
| 299 | } |
| 300 | |
| 301 | // Get a 4 bit value from a sparse bitmap. This allows a bitmap to |
| 302 | // function as a sparse array of 4 bit values. |
| 303 | // QUAD is the index. |
| 304 | inline int |
| 305 | sbr_sparse_bitmap::bitmap_get_quad (const_bitmap head, int quad) |
| 306 | { |
| 307 | return (int) bitmap_get_aligned_chunk (head, quad, 4); |
| 308 | } |
| 309 | |
| 310 | // Set the range on entry to basic block BB to R. |
| 311 | |
| 312 | bool |
| 313 | sbr_sparse_bitmap::set_bb_range (const_basic_block bb, const vrange &r) |
| 314 | { |
| 315 | if (r.undefined_p ()) |
| 316 | { |
| 317 | bitmap_set_quad (head: &bitvec, quad: bb->index, SBR_UNDEF); |
| 318 | return true; |
| 319 | } |
| 320 | |
| 321 | // Loop thru the values to see if R is already present. |
| 322 | for (int x = 0; x < SBR_NUM; x++) |
| 323 | if (!m_range[x] || m_range[x]->equal_p (r)) |
| 324 | { |
| 325 | if (!m_range[x]) |
| 326 | m_range[x] = m_range_allocator->clone (r); |
| 327 | bitmap_set_quad (head: &bitvec, quad: bb->index, quad_value: x + 1); |
| 328 | return true; |
| 329 | } |
| 330 | // All values are taken, default to VARYING. |
| 331 | bitmap_set_quad (head: &bitvec, quad: bb->index, SBR_VARYING); |
| 332 | return false; |
| 333 | } |
| 334 | |
| 335 | // Return the range associated with block BB in R. Return false if |
| 336 | // there is no range. |
| 337 | |
| 338 | bool |
| 339 | sbr_sparse_bitmap::get_bb_range (vrange &r, const_basic_block bb) |
| 340 | { |
| 341 | int value = bitmap_get_quad (head: &bitvec, quad: bb->index); |
| 342 | |
| 343 | if (!value) |
| 344 | return false; |
| 345 | |
| 346 | gcc_checking_assert (value <= SBR_UNDEF); |
| 347 | if (value == SBR_UNDEF) |
| 348 | r.set_undefined (); |
| 349 | else |
| 350 | m_range[value - 1]->get_vrange (r, type: m_type); |
| 351 | return true; |
| 352 | } |
| 353 | |
| 354 | // Return true if a range is present. |
| 355 | |
| 356 | bool |
| 357 | sbr_sparse_bitmap::bb_range_p (const_basic_block bb) |
| 358 | { |
| 359 | return (bitmap_get_quad (head: &bitvec, quad: bb->index) != 0); |
| 360 | } |
| 361 | |
| 362 | // ------------------------------------------------------------------------- |
| 363 | |
| 364 | // Initialize the block cache. |
| 365 | |
| 366 | block_range_cache::block_range_cache () |
| 367 | { |
| 368 | bitmap_obstack_initialize (&m_bitmaps); |
| 369 | m_ssa_ranges.create (nelems: 0); |
| 370 | m_ssa_ranges.safe_grow_cleared (num_ssa_names); |
| 371 | m_range_allocator = new vrange_allocator; |
| 372 | } |
| 373 | |
| 374 | // Remove any m_block_caches which have been created. |
| 375 | |
| 376 | block_range_cache::~block_range_cache () |
| 377 | { |
| 378 | delete m_range_allocator; |
| 379 | // Release the vector itself. |
| 380 | m_ssa_ranges.release (); |
| 381 | bitmap_obstack_release (&m_bitmaps); |
| 382 | } |
| 383 | |
| 384 | // Set the range for NAME on entry to block BB to R. |
| 385 | // If it has not been accessed yet, allocate it first. |
| 386 | |
| 387 | bool |
| 388 | block_range_cache::set_bb_range (tree name, const_basic_block bb, |
| 389 | const vrange &r) |
| 390 | { |
| 391 | unsigned v = SSA_NAME_VERSION (name); |
| 392 | if (v >= m_ssa_ranges.length ()) |
| 393 | m_ssa_ranges.safe_grow_cleared (num_ssa_names); |
| 394 | |
| 395 | if (!m_ssa_ranges[v]) |
| 396 | { |
| 397 | // Use sparse bitmap representation if there are too many basic blocks. |
| 398 | if (last_basic_block_for_fn (cfun) > param_vrp_sparse_threshold) |
| 399 | { |
| 400 | void *r = m_range_allocator->alloc (size: sizeof (sbr_sparse_bitmap)); |
| 401 | m_ssa_ranges[v] = new (r) sbr_sparse_bitmap (TREE_TYPE (name), |
| 402 | m_range_allocator, |
| 403 | &m_bitmaps); |
| 404 | } |
| 405 | else if (last_basic_block_for_fn (cfun) < param_vrp_vector_threshold) |
| 406 | { |
| 407 | // For small CFGs use the basic vector implemntation. |
| 408 | void *r = m_range_allocator->alloc (size: sizeof (sbr_vector)); |
| 409 | m_ssa_ranges[v] = new (r) sbr_vector (TREE_TYPE (name), |
| 410 | m_range_allocator); |
| 411 | } |
| 412 | else |
| 413 | { |
| 414 | // Otherwise use the sparse vector implementation. |
| 415 | void *r = m_range_allocator->alloc (size: sizeof (sbr_lazy_vector)); |
| 416 | m_ssa_ranges[v] = new (r) sbr_lazy_vector (TREE_TYPE (name), |
| 417 | m_range_allocator, |
| 418 | &m_bitmaps); |
| 419 | } |
| 420 | } |
| 421 | return m_ssa_ranges[v]->set_bb_range (bb, r); |
| 422 | } |
| 423 | |
| 424 | |
| 425 | // Return a pointer to the ssa_block_cache for NAME. If it has not been |
| 426 | // accessed yet, return NULL. |
| 427 | |
| 428 | inline ssa_block_ranges * |
| 429 | block_range_cache::query_block_ranges (tree name) |
| 430 | { |
| 431 | unsigned v = SSA_NAME_VERSION (name); |
| 432 | if (v >= m_ssa_ranges.length () || !m_ssa_ranges[v]) |
| 433 | return NULL; |
| 434 | return m_ssa_ranges[v]; |
| 435 | } |
| 436 | |
| 437 | |
| 438 | |
| 439 | // Return the range for NAME on entry to BB in R. Return true if there |
| 440 | // is one. |
| 441 | |
| 442 | bool |
| 443 | block_range_cache::get_bb_range (vrange &r, tree name, const_basic_block bb) |
| 444 | { |
| 445 | ssa_block_ranges *ptr = query_block_ranges (name); |
| 446 | if (ptr) |
| 447 | return ptr->get_bb_range (r, bb); |
| 448 | return false; |
| 449 | } |
| 450 | |
| 451 | // Return true if NAME has a range set in block BB. |
| 452 | |
| 453 | bool |
| 454 | block_range_cache::bb_range_p (tree name, const_basic_block bb) |
| 455 | { |
| 456 | ssa_block_ranges *ptr = query_block_ranges (name); |
| 457 | if (ptr) |
| 458 | return ptr->bb_range_p (bb); |
| 459 | return false; |
| 460 | } |
| 461 | |
| 462 | // Print all known block caches to file F. |
| 463 | |
| 464 | void |
| 465 | block_range_cache::dump (FILE *f) |
| 466 | { |
| 467 | unsigned x; |
| 468 | for (x = 1; x < m_ssa_ranges.length (); ++x) |
| 469 | { |
| 470 | if (m_ssa_ranges[x]) |
| 471 | { |
| 472 | fprintf (stream: f, format: " Ranges for " ); |
| 473 | print_generic_expr (f, ssa_name (x), TDF_NONE); |
| 474 | fprintf (stream: f, format: ":\n" ); |
| 475 | m_ssa_ranges[x]->dump (f); |
| 476 | fprintf (stream: f, format: "\n" ); |
| 477 | } |
| 478 | } |
| 479 | } |
| 480 | |
| 481 | // Print all known ranges on entry to block BB to file F. |
| 482 | |
| 483 | void |
| 484 | block_range_cache::dump (FILE *f, basic_block bb, bool print_varying) |
| 485 | { |
| 486 | unsigned x; |
| 487 | bool summarize_varying = false; |
| 488 | for (x = 1; x < m_ssa_ranges.length (); ++x) |
| 489 | { |
| 490 | if (!m_ssa_ranges[x]) |
| 491 | continue; |
| 492 | |
| 493 | if (!gimple_range_ssa_p (ssa_name (x))) |
| 494 | continue; |
| 495 | |
| 496 | value_range r (TREE_TYPE (ssa_name (x))); |
| 497 | if (m_ssa_ranges[x]->get_bb_range (r, bb)) |
| 498 | { |
| 499 | if (!print_varying && r.varying_p ()) |
| 500 | { |
| 501 | summarize_varying = true; |
| 502 | continue; |
| 503 | } |
| 504 | print_generic_expr (f, ssa_name (x), TDF_NONE); |
| 505 | fprintf (stream: f, format: "\t" ); |
| 506 | r.dump(f); |
| 507 | fprintf (stream: f, format: "\n" ); |
| 508 | } |
| 509 | } |
| 510 | // If there were any varying entries, lump them all together. |
| 511 | if (summarize_varying) |
| 512 | { |
| 513 | fprintf (stream: f, format: "VARYING_P on entry : " ); |
| 514 | for (x = 1; x < m_ssa_ranges.length (); ++x) |
| 515 | { |
| 516 | if (!m_ssa_ranges[x]) |
| 517 | continue; |
| 518 | |
| 519 | if (!gimple_range_ssa_p (ssa_name (x))) |
| 520 | continue; |
| 521 | |
| 522 | value_range r (TREE_TYPE (ssa_name (x))); |
| 523 | if (m_ssa_ranges[x]->get_bb_range (r, bb)) |
| 524 | { |
| 525 | if (r.varying_p ()) |
| 526 | { |
| 527 | print_generic_expr (f, ssa_name (x), TDF_NONE); |
| 528 | fprintf (stream: f, format: " " ); |
| 529 | } |
| 530 | } |
| 531 | } |
| 532 | fprintf (stream: f, format: "\n" ); |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | // ------------------------------------------------------------------------- |
| 537 | |
| 538 | // Initialize an ssa cache. |
| 539 | |
| 540 | ssa_cache::ssa_cache () |
| 541 | { |
| 542 | m_tab.create (nelems: 0); |
| 543 | m_range_allocator = new vrange_allocator; |
| 544 | } |
| 545 | |
| 546 | // Deconstruct an ssa cache. |
| 547 | |
| 548 | ssa_cache::~ssa_cache () |
| 549 | { |
| 550 | m_tab.release (); |
| 551 | delete m_range_allocator; |
| 552 | } |
| 553 | |
| 554 | // Enable a query to evaluate staements/ramnges based on picking up ranges |
| 555 | // from just an ssa-cache. |
| 556 | |
| 557 | bool |
| 558 | ssa_cache::range_of_expr (vrange &r, tree expr, gimple *stmt) |
| 559 | { |
| 560 | if (!gimple_range_ssa_p (exp: expr)) |
| 561 | return get_tree_range (v&: r, expr, stmt); |
| 562 | |
| 563 | if (!get_range (r, name: expr)) |
| 564 | gimple_range_global (v&: r, name: expr, cfun); |
| 565 | return true; |
| 566 | } |
| 567 | |
| 568 | // Return TRUE if the global range of NAME has a cache entry. |
| 569 | |
| 570 | bool |
| 571 | ssa_cache::has_range (tree name) const |
| 572 | { |
| 573 | unsigned v = SSA_NAME_VERSION (name); |
| 574 | if (v >= m_tab.length ()) |
| 575 | return false; |
| 576 | return m_tab[v] != NULL; |
| 577 | } |
| 578 | |
| 579 | // Retrieve the global range of NAME from cache memory if it exists. |
| 580 | // Return the value in R. |
| 581 | |
| 582 | bool |
| 583 | ssa_cache::get_range (vrange &r, tree name) const |
| 584 | { |
| 585 | unsigned v = SSA_NAME_VERSION (name); |
| 586 | if (v >= m_tab.length ()) |
| 587 | return false; |
| 588 | |
| 589 | vrange_storage *stow = m_tab[v]; |
| 590 | if (!stow) |
| 591 | return false; |
| 592 | stow->get_vrange (r, TREE_TYPE (name)); |
| 593 | return true; |
| 594 | } |
| 595 | |
| 596 | // Set the range for NAME to R in the ssa cache. |
| 597 | // Return TRUE if there was already a range set, otherwise false. |
| 598 | |
| 599 | bool |
| 600 | ssa_cache::set_range (tree name, const vrange &r) |
| 601 | { |
| 602 | unsigned v = SSA_NAME_VERSION (name); |
| 603 | if (v >= m_tab.length ()) |
| 604 | m_tab.safe_grow_cleared (num_ssa_names + 1); |
| 605 | |
| 606 | vrange_storage *m = m_tab[v]; |
| 607 | if (m && m->fits_p (r)) |
| 608 | m->set_vrange (r); |
| 609 | else |
| 610 | m_tab[v] = m_range_allocator->clone (r); |
| 611 | return m != NULL; |
| 612 | } |
| 613 | |
| 614 | // If NAME has a range, intersect it with R, otherwise set it to R. |
| 615 | // Return TRUE if the range is new or changes. |
| 616 | |
| 617 | bool |
| 618 | ssa_cache::merge_range (tree name, const vrange &r) |
| 619 | { |
| 620 | unsigned v = SSA_NAME_VERSION (name); |
| 621 | if (v >= m_tab.length ()) |
| 622 | m_tab.safe_grow_cleared (num_ssa_names + 1); |
| 623 | |
| 624 | vrange_storage *m = m_tab[v]; |
| 625 | // Check if this is a new value. |
| 626 | if (!m) |
| 627 | m_tab[v] = m_range_allocator->clone (r); |
| 628 | else |
| 629 | { |
| 630 | value_range curr (TREE_TYPE (name)); |
| 631 | m->get_vrange (r&: curr, TREE_TYPE (name)); |
| 632 | // If there is no change, return false. |
| 633 | if (!curr.intersect (r)) |
| 634 | return false; |
| 635 | |
| 636 | if (m->fits_p (r: curr)) |
| 637 | m->set_vrange (curr); |
| 638 | else |
| 639 | m_tab[v] = m_range_allocator->clone (r: curr); |
| 640 | } |
| 641 | return true; |
| 642 | } |
| 643 | |
| 644 | // Set the range for NAME to R in the ssa cache. |
| 645 | |
| 646 | void |
| 647 | ssa_cache::clear_range (tree name) |
| 648 | { |
| 649 | unsigned v = SSA_NAME_VERSION (name); |
| 650 | if (v >= m_tab.length ()) |
| 651 | return; |
| 652 | m_tab[v] = NULL; |
| 653 | } |
| 654 | |
| 655 | // Clear the ssa cache. |
| 656 | |
| 657 | void |
| 658 | ssa_cache::clear () |
| 659 | { |
| 660 | if (m_tab.address ()) |
| 661 | memset (s: m_tab.address(), c: 0, n: m_tab.length () * sizeof (vrange *)); |
| 662 | } |
| 663 | |
| 664 | // Dump the contents of the ssa cache to F. |
| 665 | |
| 666 | void |
| 667 | ssa_cache::dump (FILE *f) |
| 668 | { |
| 669 | for (unsigned x = 1; x < num_ssa_names; x++) |
| 670 | { |
| 671 | if (!gimple_range_ssa_p (ssa_name (x))) |
| 672 | continue; |
| 673 | value_range r (TREE_TYPE (ssa_name (x))); |
| 674 | // Dump all non-varying ranges. |
| 675 | if (get_range (r, ssa_name (x)) && !r.varying_p ()) |
| 676 | { |
| 677 | print_generic_expr (f, ssa_name (x), TDF_NONE); |
| 678 | fprintf (stream: f, format: " : " ); |
| 679 | r.dump (f); |
| 680 | fprintf (stream: f, format: "\n" ); |
| 681 | } |
| 682 | } |
| 683 | |
| 684 | } |
| 685 | |
| 686 | // Construct an ssa_lazy_cache. If OB is specified, us it, otherwise use |
| 687 | // a local bitmap obstack. |
| 688 | |
| 689 | ssa_lazy_cache::ssa_lazy_cache (bitmap_obstack *ob) |
| 690 | { |
| 691 | if (!ob) |
| 692 | { |
| 693 | bitmap_obstack_initialize (&m_bitmaps); |
| 694 | m_ob = &m_bitmaps; |
| 695 | } |
| 696 | else |
| 697 | m_ob = ob; |
| 698 | active_p = BITMAP_ALLOC (obstack: m_ob); |
| 699 | } |
| 700 | |
| 701 | // Destruct an sa_lazy_cache. Free the bitmap if it came from a different |
| 702 | // obstack, or release the obstack if it was a local one. |
| 703 | |
| 704 | ssa_lazy_cache::~ssa_lazy_cache () |
| 705 | { |
| 706 | if (m_ob == &m_bitmaps) |
| 707 | bitmap_obstack_release (&m_bitmaps); |
| 708 | else |
| 709 | BITMAP_FREE (active_p); |
| 710 | } |
| 711 | |
| 712 | // Return true if NAME has an active range in the cache. |
| 713 | |
| 714 | bool |
| 715 | ssa_lazy_cache::has_range (tree name) const |
| 716 | { |
| 717 | return bitmap_bit_p (active_p, SSA_NAME_VERSION (name)); |
| 718 | } |
| 719 | |
| 720 | // Set range of NAME to R in a lazy cache. Return FALSE if it did not already |
| 721 | // have a range. |
| 722 | |
| 723 | bool |
| 724 | ssa_lazy_cache::set_range (tree name, const vrange &r) |
| 725 | { |
| 726 | unsigned v = SSA_NAME_VERSION (name); |
| 727 | if (!bitmap_set_bit (active_p, v)) |
| 728 | { |
| 729 | // There is already an entry, simply set it. |
| 730 | gcc_checking_assert (v < m_tab.length ()); |
| 731 | return ssa_cache::set_range (name, r); |
| 732 | } |
| 733 | if (v >= m_tab.length ()) |
| 734 | m_tab.safe_grow (num_ssa_names + 1); |
| 735 | m_tab[v] = m_range_allocator->clone (r); |
| 736 | return false; |
| 737 | } |
| 738 | |
| 739 | // If NAME has a range, intersect it with R, otherwise set it to R. |
| 740 | // Return TRUE if the range is new or changes. |
| 741 | |
| 742 | bool |
| 743 | ssa_lazy_cache::merge_range (tree name, const vrange &r) |
| 744 | { |
| 745 | unsigned v = SSA_NAME_VERSION (name); |
| 746 | if (!bitmap_set_bit (active_p, v)) |
| 747 | { |
| 748 | // There is already an entry, simply merge it. |
| 749 | gcc_checking_assert (v < m_tab.length ()); |
| 750 | return ssa_cache::merge_range (name, r); |
| 751 | } |
| 752 | if (v >= m_tab.length ()) |
| 753 | m_tab.safe_grow (num_ssa_names + 1); |
| 754 | m_tab[v] = m_range_allocator->clone (r); |
| 755 | return true; |
| 756 | } |
| 757 | |
| 758 | // Merge all elements of CACHE with this cache. |
| 759 | // Any names in CACHE that are not in this one are added. |
| 760 | // Any names in both are merged via merge_range.. |
| 761 | |
| 762 | void |
| 763 | ssa_lazy_cache::merge (const ssa_lazy_cache &cache) |
| 764 | { |
| 765 | unsigned x; |
| 766 | bitmap_iterator bi; |
| 767 | EXECUTE_IF_SET_IN_BITMAP (cache.active_p, 0, x, bi) |
| 768 | { |
| 769 | tree name = ssa_name (x); |
| 770 | value_range r(TREE_TYPE (name)); |
| 771 | cache.get_range (r, name); |
| 772 | merge_range (ssa_name (x), r); |
| 773 | } |
| 774 | } |
| 775 | |
| 776 | // Return TRUE if NAME has a range, and return it in R. |
| 777 | |
| 778 | bool |
| 779 | ssa_lazy_cache::get_range (vrange &r, tree name) const |
| 780 | { |
| 781 | if (!bitmap_bit_p (active_p, SSA_NAME_VERSION (name))) |
| 782 | return false; |
| 783 | return ssa_cache::get_range (r, name); |
| 784 | } |
| 785 | |
| 786 | // Remove NAME from the active range list. |
| 787 | |
| 788 | void |
| 789 | ssa_lazy_cache::clear_range (tree name) |
| 790 | { |
| 791 | bitmap_clear_bit (active_p, SSA_NAME_VERSION (name)); |
| 792 | } |
| 793 | |
| 794 | // Remove all ranges from the active range list. |
| 795 | |
| 796 | void |
| 797 | ssa_lazy_cache::clear () |
| 798 | { |
| 799 | bitmap_clear (active_p); |
| 800 | } |
| 801 | |
| 802 | // -------------------------------------------------------------------------- |
| 803 | |
| 804 | |
| 805 | // This class will manage the timestamps for each ssa_name. |
| 806 | // When a value is calculated, the timestamp is set to the current time. |
| 807 | // Current time is then incremented. Any dependencies will already have |
| 808 | // been calculated, and will thus have older timestamps. |
| 809 | // If one of those values is ever calculated again, it will get a newer |
| 810 | // timestamp, and the "current_p" check will fail. |
| 811 | |
| 812 | class temporal_cache |
| 813 | { |
| 814 | public: |
| 815 | temporal_cache (); |
| 816 | ~temporal_cache (); |
| 817 | bool current_p (tree name, tree dep1, tree dep2) const; |
| 818 | void set_timestamp (tree name); |
| 819 | void set_always_current (tree name, bool value); |
| 820 | bool always_current_p (tree name) const; |
| 821 | private: |
| 822 | int temporal_value (unsigned ssa) const; |
| 823 | int m_current_time; |
| 824 | vec <int> m_timestamp; |
| 825 | }; |
| 826 | |
| 827 | inline |
| 828 | temporal_cache::temporal_cache () |
| 829 | { |
| 830 | m_current_time = 1; |
| 831 | m_timestamp.create (nelems: 0); |
| 832 | m_timestamp.safe_grow_cleared (num_ssa_names); |
| 833 | } |
| 834 | |
| 835 | inline |
| 836 | temporal_cache::~temporal_cache () |
| 837 | { |
| 838 | m_timestamp.release (); |
| 839 | } |
| 840 | |
| 841 | // Return the timestamp value for SSA, or 0 if there isn't one. |
| 842 | |
| 843 | inline int |
| 844 | temporal_cache::temporal_value (unsigned ssa) const |
| 845 | { |
| 846 | if (ssa >= m_timestamp.length ()) |
| 847 | return 0; |
| 848 | return abs (x: m_timestamp[ssa]); |
| 849 | } |
| 850 | |
| 851 | // Return TRUE if the timestamp for NAME is newer than any of its dependents. |
| 852 | // Up to 2 dependencies can be checked. |
| 853 | |
| 854 | bool |
| 855 | temporal_cache::current_p (tree name, tree dep1, tree dep2) const |
| 856 | { |
| 857 | if (always_current_p (name)) |
| 858 | return true; |
| 859 | |
| 860 | // Any non-registered dependencies will have a value of 0 and thus be older. |
| 861 | // Return true if time is newer than either dependent. |
| 862 | int ts = temporal_value (SSA_NAME_VERSION (name)); |
| 863 | if (dep1 && ts < temporal_value (SSA_NAME_VERSION (dep1))) |
| 864 | return false; |
| 865 | if (dep2 && ts < temporal_value (SSA_NAME_VERSION (dep2))) |
| 866 | return false; |
| 867 | |
| 868 | return true; |
| 869 | } |
| 870 | |
| 871 | // This increments the global timer and sets the timestamp for NAME. |
| 872 | |
| 873 | inline void |
| 874 | temporal_cache::set_timestamp (tree name) |
| 875 | { |
| 876 | unsigned v = SSA_NAME_VERSION (name); |
| 877 | if (v >= m_timestamp.length ()) |
| 878 | m_timestamp.safe_grow_cleared (num_ssa_names + 20); |
| 879 | m_timestamp[v] = ++m_current_time; |
| 880 | } |
| 881 | |
| 882 | // Set the timestamp to 0, marking it as "always up to date". |
| 883 | |
| 884 | inline void |
| 885 | temporal_cache::set_always_current (tree name, bool value) |
| 886 | { |
| 887 | unsigned v = SSA_NAME_VERSION (name); |
| 888 | if (v >= m_timestamp.length ()) |
| 889 | m_timestamp.safe_grow_cleared (num_ssa_names + 20); |
| 890 | |
| 891 | int ts = abs (x: m_timestamp[v]); |
| 892 | // If this does not have a timestamp, create one. |
| 893 | if (ts == 0) |
| 894 | ts = ++m_current_time; |
| 895 | m_timestamp[v] = value ? -ts : ts; |
| 896 | } |
| 897 | |
| 898 | // Return true if NAME is always current. |
| 899 | |
| 900 | inline bool |
| 901 | temporal_cache::always_current_p (tree name) const |
| 902 | { |
| 903 | unsigned v = SSA_NAME_VERSION (name); |
| 904 | if (v >= m_timestamp.length ()) |
| 905 | return false; |
| 906 | return m_timestamp[v] <= 0; |
| 907 | } |
| 908 | |
| 909 | // -------------------------------------------------------------------------- |
| 910 | |
| 911 | // This class provides an abstraction of a list of blocks to be updated |
| 912 | // by the cache. It is currently a stack but could be changed. It also |
| 913 | // maintains a list of blocks which have failed propagation, and does not |
| 914 | // enter any of those blocks into the list. |
| 915 | |
| 916 | // A vector over the BBs is maintained, and an entry of 0 means it is not in |
| 917 | // a list. Otherwise, the entry is the next block in the list. -1 terminates |
| 918 | // the list. m_head points to the top of the list, -1 if the list is empty. |
| 919 | |
| 920 | class update_list |
| 921 | { |
| 922 | public: |
| 923 | update_list (); |
| 924 | ~update_list (); |
| 925 | void add (basic_block bb); |
| 926 | basic_block pop (); |
| 927 | inline bool empty_p () { return m_update_head == -1; } |
| 928 | inline void clear_failures () { bitmap_clear (m_propfail); } |
| 929 | inline void propagation_failed (basic_block bb) |
| 930 | { bitmap_set_bit (m_propfail, bb->index); } |
| 931 | private: |
| 932 | vec<int> m_update_list; |
| 933 | int m_update_head; |
| 934 | bitmap m_propfail; |
| 935 | bitmap_obstack m_bitmaps; |
| 936 | }; |
| 937 | |
| 938 | // Create an update list. |
| 939 | |
| 940 | update_list::update_list () |
| 941 | { |
| 942 | m_update_list.create (nelems: 0); |
| 943 | m_update_list.safe_grow_cleared (last_basic_block_for_fn (cfun) + 64); |
| 944 | m_update_head = -1; |
| 945 | bitmap_obstack_initialize (&m_bitmaps); |
| 946 | m_propfail = BITMAP_ALLOC (obstack: &m_bitmaps); |
| 947 | } |
| 948 | |
| 949 | // Destroy an update list. |
| 950 | |
| 951 | update_list::~update_list () |
| 952 | { |
| 953 | m_update_list.release (); |
| 954 | bitmap_obstack_release (&m_bitmaps); |
| 955 | } |
| 956 | |
| 957 | // Add BB to the list of blocks to update, unless it's already in the list. |
| 958 | |
| 959 | void |
| 960 | update_list::add (basic_block bb) |
| 961 | { |
| 962 | int i = bb->index; |
| 963 | // If propagation has failed for BB, or its already in the list, don't |
| 964 | // add it again. |
| 965 | if ((unsigned)i >= m_update_list.length ()) |
| 966 | m_update_list.safe_grow_cleared (len: i + 64); |
| 967 | if (!m_update_list[i] && !bitmap_bit_p (m_propfail, i)) |
| 968 | { |
| 969 | if (empty_p ()) |
| 970 | { |
| 971 | m_update_head = i; |
| 972 | m_update_list[i] = -1; |
| 973 | } |
| 974 | else |
| 975 | { |
| 976 | gcc_checking_assert (m_update_head > 0); |
| 977 | m_update_list[i] = m_update_head; |
| 978 | m_update_head = i; |
| 979 | } |
| 980 | } |
| 981 | } |
| 982 | |
| 983 | // Remove a block from the list. |
| 984 | |
| 985 | basic_block |
| 986 | update_list::pop () |
| 987 | { |
| 988 | gcc_checking_assert (!empty_p ()); |
| 989 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, m_update_head); |
| 990 | int pop = m_update_head; |
| 991 | m_update_head = m_update_list[pop]; |
| 992 | m_update_list[pop] = 0; |
| 993 | return bb; |
| 994 | } |
| 995 | |
| 996 | // -------------------------------------------------------------------------- |
| 997 | |
| 998 | ranger_cache::ranger_cache (int not_executable_flag, bool use_imm_uses) |
| 999 | { |
| 1000 | m_workback = vNULL; |
| 1001 | m_temporal = new temporal_cache; |
| 1002 | |
| 1003 | // If DOM info is available, spawn an oracle as well. |
| 1004 | create_relation_oracle (); |
| 1005 | // Create an infer oracle using this cache as the range query. The cache |
| 1006 | // version acts as a read-only query, and will spawn no additional lookups. |
| 1007 | // It just ues what is already known. |
| 1008 | create_infer_oracle (q: this, do_search: use_imm_uses); |
| 1009 | create_gori (not_executable_flag, param_vrp_switch_limit); |
| 1010 | |
| 1011 | unsigned x, lim = last_basic_block_for_fn (cfun); |
| 1012 | // Calculate outgoing range info upfront. This will fully populate the |
| 1013 | // m_maybe_variant bitmap which will help eliminate processing of names |
| 1014 | // which never have their ranges adjusted. |
| 1015 | for (x = 0; x < lim ; x++) |
| 1016 | { |
| 1017 | basic_block bb = BASIC_BLOCK_FOR_FN (cfun, x); |
| 1018 | if (bb) |
| 1019 | gori_ssa ()->exports (bb); |
| 1020 | } |
| 1021 | m_update = new update_list (); |
| 1022 | } |
| 1023 | |
| 1024 | ranger_cache::~ranger_cache () |
| 1025 | { |
| 1026 | delete m_update; |
| 1027 | destroy_infer_oracle (); |
| 1028 | destroy_relation_oracle (); |
| 1029 | delete m_temporal; |
| 1030 | m_workback.release (); |
| 1031 | } |
| 1032 | |
| 1033 | // Dump the global caches to file F. if GORI_DUMP is true, dump the |
| 1034 | // gori map as well. |
| 1035 | |
| 1036 | void |
| 1037 | ranger_cache::dump (FILE *f) |
| 1038 | { |
| 1039 | fprintf (stream: f, format: "Non-varying global ranges:\n" ); |
| 1040 | fprintf (stream: f, format: "=========================:\n" ); |
| 1041 | m_globals.dump (f); |
| 1042 | fprintf (stream: f, format: "\n" ); |
| 1043 | } |
| 1044 | |
| 1045 | // Dump the caches for basic block BB to file F. |
| 1046 | |
| 1047 | void |
| 1048 | ranger_cache::dump_bb (FILE *f, basic_block bb) |
| 1049 | { |
| 1050 | gori_ssa ()->dump (f, bb, verbose: false); |
| 1051 | m_on_entry.dump (f, bb); |
| 1052 | m_relation->dump (f, bb); |
| 1053 | } |
| 1054 | |
| 1055 | // Get the global range for NAME, and return in R. Return false if the |
| 1056 | // global range is not set, and return the legacy global value in R. |
| 1057 | |
| 1058 | bool |
| 1059 | ranger_cache::get_global_range (vrange &r, tree name) const |
| 1060 | { |
| 1061 | if (m_globals.get_range (r, name)) |
| 1062 | return true; |
| 1063 | gimple_range_global (v&: r, name); |
| 1064 | return false; |
| 1065 | } |
| 1066 | |
| 1067 | // Get the global range for NAME, and return in R. Return false if the |
| 1068 | // global range is not set, and R will contain the legacy global value. |
| 1069 | // CURRENT_P is set to true if the value was in cache and not stale. |
| 1070 | // Otherwise, set CURRENT_P to false and mark as it always current. |
| 1071 | // If the global cache did not have a value, initialize it as well. |
| 1072 | // After this call, the global cache will have a value. |
| 1073 | |
| 1074 | bool |
| 1075 | ranger_cache::get_global_range (vrange &r, tree name, bool ¤t_p) |
| 1076 | { |
| 1077 | bool had_global = get_global_range (r, name); |
| 1078 | |
| 1079 | // If there was a global value, set current flag, otherwise set a value. |
| 1080 | current_p = false; |
| 1081 | if (had_global) |
| 1082 | current_p = r.singleton_p () |
| 1083 | || m_temporal->current_p (name, dep1: gori_ssa ()->depend1 (name), |
| 1084 | dep2: gori_ssa ()->depend2 (name)); |
| 1085 | else |
| 1086 | { |
| 1087 | // If no global value has been set and value is VARYING, fold the stmt |
| 1088 | // using just global ranges to get a better initial value. |
| 1089 | // After inlining we tend to decide some things are constant, so |
| 1090 | // so not do this evaluation after inlining. |
| 1091 | if (r.varying_p () && !cfun->after_inlining) |
| 1092 | { |
| 1093 | gimple *s = SSA_NAME_DEF_STMT (name); |
| 1094 | // Do not process PHIs as SCEV may be in use and it can |
| 1095 | // spawn cyclic lookups. |
| 1096 | if (gimple_get_lhs (s) == name && !is_a<gphi *> (p: s)) |
| 1097 | { |
| 1098 | if (!fold_range (r, s, q: get_global_range_query ())) |
| 1099 | gimple_range_global (v&: r, name); |
| 1100 | } |
| 1101 | } |
| 1102 | m_globals.set_range (name, r); |
| 1103 | } |
| 1104 | |
| 1105 | // If the existing value was not current, mark it as always current. |
| 1106 | if (!current_p) |
| 1107 | m_temporal->set_always_current (name, value: true); |
| 1108 | return had_global; |
| 1109 | } |
| 1110 | |
| 1111 | // Set the global range of NAME to R and give it a timestamp. |
| 1112 | |
| 1113 | void |
| 1114 | ranger_cache::set_global_range (tree name, const vrange &r, bool changed) |
| 1115 | { |
| 1116 | // Setting a range always clears the always_current flag. |
| 1117 | m_temporal->set_always_current (name, value: false); |
| 1118 | if (!changed) |
| 1119 | { |
| 1120 | // If there are dependencies, make sure this is not out of date. |
| 1121 | if (!m_temporal->current_p (name, dep1: gori_ssa ()->depend1 (name), |
| 1122 | dep2: gori_ssa ()->depend2 (name))) |
| 1123 | m_temporal->set_timestamp (name); |
| 1124 | return; |
| 1125 | } |
| 1126 | if (m_globals.set_range (name, r)) |
| 1127 | { |
| 1128 | // If there was already a range set, propagate the new value. |
| 1129 | basic_block bb = gimple_bb (SSA_NAME_DEF_STMT (name)); |
| 1130 | if (!bb) |
| 1131 | bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
| 1132 | |
| 1133 | if (DEBUG_RANGE_CACHE) |
| 1134 | fprintf (stream: dump_file, format: " GLOBAL :" ); |
| 1135 | |
| 1136 | propagate_updated_value (name, bb); |
| 1137 | } |
| 1138 | // Constants no longer need to tracked. Any further refinement has to be |
| 1139 | // undefined. Propagation works better with constants. PR 100512. |
| 1140 | // Pointers which resolve to non-zero also do not need |
| 1141 | // tracking in the cache as they will never change. See PR 98866. |
| 1142 | // Timestamp must always be updated, or dependent calculations may |
| 1143 | // not include this latest value. PR 100774. |
| 1144 | |
| 1145 | if (r.singleton_p () |
| 1146 | || (POINTER_TYPE_P (TREE_TYPE (name)) && r.nonzero_p ())) |
| 1147 | gori_ssa ()->set_range_invariant (name); |
| 1148 | m_temporal->set_timestamp (name); |
| 1149 | } |
| 1150 | |
| 1151 | // Provide lookup for the gori-computes class to access the best known range |
| 1152 | // of an ssa_name in any given basic block. Note, this does no additional |
| 1153 | // lookups, just accesses the data that is already known. |
| 1154 | |
| 1155 | // Get the range of NAME when the def occurs in block BB. If BB is NULL |
| 1156 | // get the best global value available. |
| 1157 | |
| 1158 | void |
| 1159 | ranger_cache::range_of_def (vrange &r, tree name, basic_block bb) |
| 1160 | { |
| 1161 | gcc_checking_assert (gimple_range_ssa_p (name)); |
| 1162 | gcc_checking_assert (!bb || bb == gimple_bb (SSA_NAME_DEF_STMT (name))); |
| 1163 | |
| 1164 | // Pick up the best global range available. |
| 1165 | if (!m_globals.get_range (r, name)) |
| 1166 | { |
| 1167 | // If that fails, try to calculate the range using just global values. |
| 1168 | gimple *s = SSA_NAME_DEF_STMT (name); |
| 1169 | if (gimple_get_lhs (s) == name) |
| 1170 | fold_range (r, s, q: get_global_range_query ()); |
| 1171 | else |
| 1172 | gimple_range_global (v&: r, name); |
| 1173 | } |
| 1174 | } |
| 1175 | |
| 1176 | // Get the range of NAME as it occurs on entry to block BB. Use MODE for |
| 1177 | // lookups. |
| 1178 | |
| 1179 | void |
| 1180 | ranger_cache::entry_range (vrange &r, tree name, basic_block bb, |
| 1181 | enum rfd_mode mode) |
| 1182 | { |
| 1183 | if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
| 1184 | { |
| 1185 | gimple_range_global (v&: r, name); |
| 1186 | return; |
| 1187 | } |
| 1188 | |
| 1189 | // If NAME is invariant, simply return the defining range. |
| 1190 | if (!gori ().has_edge_range_p (name)) |
| 1191 | { |
| 1192 | range_of_def (r, name); |
| 1193 | return; |
| 1194 | } |
| 1195 | |
| 1196 | // Look for the on-entry value of name in BB from the cache. |
| 1197 | // Otherwise pick up the best available global value. |
| 1198 | if (!m_on_entry.get_bb_range (r, name, bb)) |
| 1199 | if (!range_from_dom (r, name, bb, mode)) |
| 1200 | range_of_def (r, name); |
| 1201 | } |
| 1202 | |
| 1203 | // Get the range of NAME as it occurs on exit from block BB. Use MODE for |
| 1204 | // lookups. |
| 1205 | |
| 1206 | void |
| 1207 | ranger_cache::exit_range (vrange &r, tree name, basic_block bb, |
| 1208 | enum rfd_mode mode) |
| 1209 | { |
| 1210 | if (bb == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
| 1211 | { |
| 1212 | gimple_range_global (v&: r, name); |
| 1213 | return; |
| 1214 | } |
| 1215 | |
| 1216 | gimple *s = SSA_NAME_DEF_STMT (name); |
| 1217 | basic_block def_bb = gimple_bb (g: s); |
| 1218 | if (def_bb == bb) |
| 1219 | range_of_def (r, name, bb); |
| 1220 | else |
| 1221 | entry_range (r, name, bb, mode); |
| 1222 | } |
| 1223 | |
| 1224 | // Get the range of NAME on edge E using MODE, return the result in R. |
| 1225 | // Always returns a range and true. |
| 1226 | |
| 1227 | bool |
| 1228 | ranger_cache::edge_range (vrange &r, edge e, tree name, enum rfd_mode mode) |
| 1229 | { |
| 1230 | exit_range (r, name, bb: e->src, mode); |
| 1231 | // If this is not an abnormal edge, check for inferred ranges on exit. |
| 1232 | if ((e->flags & (EDGE_EH | EDGE_ABNORMAL)) == 0) |
| 1233 | infer_oracle ().maybe_adjust_range (r, name, e->src); |
| 1234 | value_range er (TREE_TYPE (name)); |
| 1235 | if (gori ().edge_range_p (er, e, name, *this)) |
| 1236 | r.intersect (er); |
| 1237 | return true; |
| 1238 | } |
| 1239 | |
| 1240 | |
| 1241 | |
| 1242 | // Implement range_of_expr. |
| 1243 | |
| 1244 | bool |
| 1245 | ranger_cache::range_of_expr (vrange &r, tree name, gimple *stmt) |
| 1246 | { |
| 1247 | if (!gimple_range_ssa_p (exp: name)) |
| 1248 | { |
| 1249 | get_tree_range (v&: r, expr: name, stmt); |
| 1250 | return true; |
| 1251 | } |
| 1252 | |
| 1253 | basic_block bb = gimple_bb (g: stmt); |
| 1254 | gimple *def_stmt = SSA_NAME_DEF_STMT (name); |
| 1255 | basic_block def_bb = gimple_bb (g: def_stmt); |
| 1256 | |
| 1257 | if (bb == def_bb) |
| 1258 | range_of_def (r, name, bb); |
| 1259 | else |
| 1260 | entry_range (r, name, bb, mode: RFD_NONE); |
| 1261 | return true; |
| 1262 | } |
| 1263 | |
| 1264 | |
| 1265 | // Implement range_on_edge. Always return the best available range using |
| 1266 | // the current cache values. |
| 1267 | |
| 1268 | bool |
| 1269 | ranger_cache::range_on_edge (vrange &r, edge e, tree expr) |
| 1270 | { |
| 1271 | if (gimple_range_ssa_p (exp: expr)) |
| 1272 | return edge_range (r, e, name: expr, mode: RFD_NONE); |
| 1273 | return get_tree_range (v&: r, expr, NULL); |
| 1274 | } |
| 1275 | |
| 1276 | // Return a static range for NAME on entry to basic block BB in R. If |
| 1277 | // calc is true, fill any cache entries required between BB and the |
| 1278 | // def block for NAME. Otherwise, return false if the cache is empty. |
| 1279 | |
| 1280 | bool |
| 1281 | ranger_cache::block_range (vrange &r, basic_block bb, tree name, bool calc) |
| 1282 | { |
| 1283 | gcc_checking_assert (gimple_range_ssa_p (name)); |
| 1284 | |
| 1285 | // If there are no range calculations anywhere in the IL, global range |
| 1286 | // applies everywhere, so don't bother caching it. |
| 1287 | if (!gori ().has_edge_range_p (name)) |
| 1288 | return false; |
| 1289 | |
| 1290 | if (calc) |
| 1291 | { |
| 1292 | gimple *def_stmt = SSA_NAME_DEF_STMT (name); |
| 1293 | basic_block def_bb = NULL; |
| 1294 | if (def_stmt) |
| 1295 | def_bb = gimple_bb (g: def_stmt); |
| 1296 | if (!def_bb) |
| 1297 | { |
| 1298 | // If we get to the entry block, this better be a default def |
| 1299 | // or range_on_entry was called for a block not dominated by |
| 1300 | // the def. But it could be also SSA_NAME defined by a statement |
| 1301 | // not yet in the IL (such as queued edge insertion), in that case |
| 1302 | // just punt. |
| 1303 | if (!SSA_NAME_IS_DEFAULT_DEF (name)) |
| 1304 | return false; |
| 1305 | def_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
| 1306 | } |
| 1307 | |
| 1308 | // There is no range on entry for the definition block. |
| 1309 | if (def_bb == bb) |
| 1310 | return false; |
| 1311 | |
| 1312 | // Otherwise, go figure out what is known in predecessor blocks. |
| 1313 | fill_block_cache (name, bb, def_bb); |
| 1314 | gcc_checking_assert (m_on_entry.bb_range_p (name, bb)); |
| 1315 | } |
| 1316 | return m_on_entry.get_bb_range (r, name, bb); |
| 1317 | } |
| 1318 | |
| 1319 | // If there is anything in the propagation update_list, continue |
| 1320 | // processing NAME until the list of blocks is empty. |
| 1321 | |
| 1322 | void |
| 1323 | ranger_cache::propagate_cache (tree name) |
| 1324 | { |
| 1325 | basic_block bb; |
| 1326 | edge_iterator ei; |
| 1327 | edge e; |
| 1328 | tree type = TREE_TYPE (name); |
| 1329 | value_range new_range (type); |
| 1330 | value_range current_range (type); |
| 1331 | value_range e_range (type); |
| 1332 | |
| 1333 | // Process each block by seeing if its calculated range on entry is |
| 1334 | // the same as its cached value. If there is a difference, update |
| 1335 | // the cache to reflect the new value, and check to see if any |
| 1336 | // successors have cache entries which may need to be checked for |
| 1337 | // updates. |
| 1338 | |
| 1339 | while (!m_update->empty_p ()) |
| 1340 | { |
| 1341 | bb = m_update->pop (); |
| 1342 | gcc_checking_assert (m_on_entry.bb_range_p (name, bb)); |
| 1343 | m_on_entry.get_bb_range (r&: current_range, name, bb); |
| 1344 | |
| 1345 | if (DEBUG_RANGE_CACHE) |
| 1346 | { |
| 1347 | fprintf (stream: dump_file, format: "FWD visiting block %d for " , bb->index); |
| 1348 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1349 | fprintf (stream: dump_file, format: " starting range : " ); |
| 1350 | current_range.dump (dump_file); |
| 1351 | fprintf (stream: dump_file, format: "\n" ); |
| 1352 | } |
| 1353 | |
| 1354 | // Calculate the "new" range on entry by unioning the pred edges. |
| 1355 | new_range.set_undefined (); |
| 1356 | FOR_EACH_EDGE (e, ei, bb->preds) |
| 1357 | { |
| 1358 | edge_range (r&: e_range, e, name, mode: RFD_READ_ONLY); |
| 1359 | if (DEBUG_RANGE_CACHE) |
| 1360 | { |
| 1361 | fprintf (stream: dump_file, format: " edge %d->%d :" , e->src->index, bb->index); |
| 1362 | e_range.dump (dump_file); |
| 1363 | fprintf (stream: dump_file, format: "\n" ); |
| 1364 | } |
| 1365 | new_range.union_ (r: e_range); |
| 1366 | if (new_range.varying_p ()) |
| 1367 | break; |
| 1368 | } |
| 1369 | |
| 1370 | // If the range on entry has changed, update it. |
| 1371 | if (new_range != current_range) |
| 1372 | { |
| 1373 | bool ok_p = m_on_entry.set_bb_range (name, bb, r: new_range); |
| 1374 | // If the cache couldn't set the value, mark it as failed. |
| 1375 | if (!ok_p) |
| 1376 | m_update->propagation_failed (bb); |
| 1377 | if (DEBUG_RANGE_CACHE) |
| 1378 | { |
| 1379 | if (!ok_p) |
| 1380 | { |
| 1381 | fprintf (stream: dump_file, format: " Cache failure to store value:" ); |
| 1382 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1383 | fprintf (stream: dump_file, format: " " ); |
| 1384 | } |
| 1385 | else |
| 1386 | { |
| 1387 | fprintf (stream: dump_file, format: " Updating range to " ); |
| 1388 | new_range.dump (dump_file); |
| 1389 | } |
| 1390 | fprintf (stream: dump_file, format: "\n Updating blocks :" ); |
| 1391 | } |
| 1392 | // Mark each successor that has a range to re-check its range |
| 1393 | FOR_EACH_EDGE (e, ei, bb->succs) |
| 1394 | if (m_on_entry.bb_range_p (name, bb: e->dest)) |
| 1395 | { |
| 1396 | if (DEBUG_RANGE_CACHE) |
| 1397 | fprintf (stream: dump_file, format: " bb%d" ,e->dest->index); |
| 1398 | m_update->add (bb: e->dest); |
| 1399 | } |
| 1400 | if (DEBUG_RANGE_CACHE) |
| 1401 | fprintf (stream: dump_file, format: "\n" ); |
| 1402 | } |
| 1403 | } |
| 1404 | if (DEBUG_RANGE_CACHE) |
| 1405 | { |
| 1406 | fprintf (stream: dump_file, format: "DONE visiting blocks for " ); |
| 1407 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1408 | fprintf (stream: dump_file, format: "\n" ); |
| 1409 | } |
| 1410 | m_update->clear_failures (); |
| 1411 | } |
| 1412 | |
| 1413 | // Check to see if an update to the value for NAME in BB has any effect |
| 1414 | // on values already in the on-entry cache for successor blocks. |
| 1415 | // If it does, update them. Don't visit any blocks which don't have a cache |
| 1416 | // entry. |
| 1417 | |
| 1418 | void |
| 1419 | ranger_cache::propagate_updated_value (tree name, basic_block bb) |
| 1420 | { |
| 1421 | edge e; |
| 1422 | edge_iterator ei; |
| 1423 | |
| 1424 | // The update work list should be empty at this point. |
| 1425 | gcc_checking_assert (m_update->empty_p ()); |
| 1426 | gcc_checking_assert (bb); |
| 1427 | |
| 1428 | if (DEBUG_RANGE_CACHE) |
| 1429 | { |
| 1430 | fprintf (stream: dump_file, format: " UPDATE cache for " ); |
| 1431 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1432 | fprintf (stream: dump_file, format: " in BB %d : successors : " , bb->index); |
| 1433 | } |
| 1434 | FOR_EACH_EDGE (e, ei, bb->succs) |
| 1435 | { |
| 1436 | // Only update active cache entries. |
| 1437 | if (m_on_entry.bb_range_p (name, bb: e->dest)) |
| 1438 | { |
| 1439 | m_update->add (bb: e->dest); |
| 1440 | if (DEBUG_RANGE_CACHE) |
| 1441 | fprintf (stream: dump_file, format: " UPDATE: bb%d" , e->dest->index); |
| 1442 | } |
| 1443 | } |
| 1444 | if (!m_update->empty_p ()) |
| 1445 | { |
| 1446 | if (DEBUG_RANGE_CACHE) |
| 1447 | fprintf (stream: dump_file, format: "\n" ); |
| 1448 | propagate_cache (name); |
| 1449 | } |
| 1450 | else |
| 1451 | { |
| 1452 | if (DEBUG_RANGE_CACHE) |
| 1453 | fprintf (stream: dump_file, format: " : No updates!\n" ); |
| 1454 | } |
| 1455 | } |
| 1456 | |
| 1457 | // Make sure that the range-on-entry cache for NAME is set for block BB. |
| 1458 | // Work back through the CFG to DEF_BB ensuring the range is calculated |
| 1459 | // on the block/edges leading back to that point. |
| 1460 | |
| 1461 | void |
| 1462 | ranger_cache::fill_block_cache (tree name, basic_block bb, basic_block def_bb) |
| 1463 | { |
| 1464 | edge_iterator ei; |
| 1465 | edge e; |
| 1466 | tree type = TREE_TYPE (name); |
| 1467 | value_range block_result (type); |
| 1468 | value_range undefined (type); |
| 1469 | |
| 1470 | // At this point we shouldn't be looking at the def, entry block. |
| 1471 | gcc_checking_assert (bb != def_bb && bb != ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
| 1472 | unsigned start_length = m_workback.length (); |
| 1473 | |
| 1474 | // If the block cache is set, then we've already visited this block. |
| 1475 | if (m_on_entry.bb_range_p (name, bb)) |
| 1476 | return; |
| 1477 | |
| 1478 | if (DEBUG_RANGE_CACHE) |
| 1479 | { |
| 1480 | fprintf (stream: dump_file, format: "\n" ); |
| 1481 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1482 | fprintf (stream: dump_file, format: " : " ); |
| 1483 | } |
| 1484 | |
| 1485 | // Check if a dominators can supply the range. |
| 1486 | if (range_from_dom (r&: block_result, name, bb, RFD_FILL)) |
| 1487 | { |
| 1488 | if (DEBUG_RANGE_CACHE) |
| 1489 | { |
| 1490 | fprintf (stream: dump_file, format: "Filled from dominator! : " ); |
| 1491 | block_result.dump (dump_file); |
| 1492 | fprintf (stream: dump_file, format: "\n" ); |
| 1493 | } |
| 1494 | // See if any equivalences can refine it. |
| 1495 | // PR 109462, like 108139 below, a one way equivalence introduced |
| 1496 | // by a PHI node can also be through the definition side. Disallow it. |
| 1497 | tree equiv_name; |
| 1498 | relation_kind rel; |
| 1499 | int prec = TYPE_PRECISION (type); |
| 1500 | // If there are too many basic blocks, do not attempt to process |
| 1501 | // equivalencies. |
| 1502 | if (last_basic_block_for_fn (cfun) > param_vrp_sparse_threshold) |
| 1503 | { |
| 1504 | m_on_entry.set_bb_range (name, bb, r: block_result); |
| 1505 | gcc_checking_assert (m_workback.length () == start_length); |
| 1506 | return; |
| 1507 | } |
| 1508 | FOR_EACH_PARTIAL_AND_FULL_EQUIV (m_relation, bb, name, equiv_name, rel) |
| 1509 | { |
| 1510 | basic_block equiv_bb = gimple_bb (SSA_NAME_DEF_STMT (equiv_name)); |
| 1511 | |
| 1512 | // Ignore partial equivs that are smaller than this object. |
| 1513 | if (rel != VREL_EQ && prec > pe_to_bits (t: rel)) |
| 1514 | continue; |
| 1515 | |
| 1516 | // Check if the equiv has any ranges calculated. |
| 1517 | if (!gori ().has_edge_range_p (equiv_name)) |
| 1518 | continue; |
| 1519 | |
| 1520 | // Check if the equiv definition dominates this block |
| 1521 | if (equiv_bb == bb || |
| 1522 | (equiv_bb && !dominated_by_p (CDI_DOMINATORS, bb, equiv_bb))) |
| 1523 | continue; |
| 1524 | |
| 1525 | if (DEBUG_RANGE_CACHE) |
| 1526 | { |
| 1527 | if (rel == VREL_EQ) |
| 1528 | fprintf (stream: dump_file, format: "Checking Equivalence (" ); |
| 1529 | else |
| 1530 | fprintf (stream: dump_file, format: "Checking Partial equiv (" ); |
| 1531 | print_relation (f: dump_file, rel); |
| 1532 | fprintf (stream: dump_file, format: ") " ); |
| 1533 | print_generic_expr (dump_file, equiv_name, TDF_SLIM); |
| 1534 | fprintf (stream: dump_file, format: "\n" ); |
| 1535 | } |
| 1536 | value_range equiv_range (TREE_TYPE (equiv_name)); |
| 1537 | if (range_from_dom (r&: equiv_range, name: equiv_name, bb, RFD_READ_ONLY)) |
| 1538 | { |
| 1539 | if (rel != VREL_EQ) |
| 1540 | range_cast (r&: equiv_range, type); |
| 1541 | else |
| 1542 | adjust_equivalence_range (range&: equiv_range); |
| 1543 | |
| 1544 | if (block_result.intersect (r: equiv_range)) |
| 1545 | { |
| 1546 | if (DEBUG_RANGE_CACHE) |
| 1547 | { |
| 1548 | if (rel == VREL_EQ) |
| 1549 | fprintf (stream: dump_file, format: "Equivalence update! : " ); |
| 1550 | else |
| 1551 | fprintf (stream: dump_file, format: "Partial equiv update! : " ); |
| 1552 | print_generic_expr (dump_file, equiv_name, TDF_SLIM); |
| 1553 | fprintf (stream: dump_file, format: " has range : " ); |
| 1554 | equiv_range.dump (dump_file); |
| 1555 | fprintf (stream: dump_file, format: " refining range to :" ); |
| 1556 | block_result.dump (dump_file); |
| 1557 | fprintf (stream: dump_file, format: "\n" ); |
| 1558 | } |
| 1559 | } |
| 1560 | } |
| 1561 | } |
| 1562 | |
| 1563 | m_on_entry.set_bb_range (name, bb, r: block_result); |
| 1564 | gcc_checking_assert (m_workback.length () == start_length); |
| 1565 | return; |
| 1566 | } |
| 1567 | |
| 1568 | // Visit each block back to the DEF. Initialize each one to UNDEFINED. |
| 1569 | // m_visited at the end will contain all the blocks that we needed to set |
| 1570 | // the range_on_entry cache for. |
| 1571 | m_workback.safe_push (obj: bb); |
| 1572 | undefined.set_undefined (); |
| 1573 | m_on_entry.set_bb_range (name, bb, r: undefined); |
| 1574 | gcc_checking_assert (m_update->empty_p ()); |
| 1575 | |
| 1576 | while (m_workback.length () > start_length) |
| 1577 | { |
| 1578 | basic_block node = m_workback.pop (); |
| 1579 | if (DEBUG_RANGE_CACHE) |
| 1580 | { |
| 1581 | fprintf (stream: dump_file, format: "BACK visiting block %d for " , node->index); |
| 1582 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1583 | fprintf (stream: dump_file, format: "\n" ); |
| 1584 | } |
| 1585 | |
| 1586 | FOR_EACH_EDGE (e, ei, node->preds) |
| 1587 | { |
| 1588 | basic_block pred = e->src; |
| 1589 | value_range r (TREE_TYPE (name)); |
| 1590 | |
| 1591 | if (DEBUG_RANGE_CACHE) |
| 1592 | fprintf (stream: dump_file, format: " %d->%d " ,e->src->index, e->dest->index); |
| 1593 | |
| 1594 | // If the pred block is the def block add this BB to update list. |
| 1595 | if (pred == def_bb) |
| 1596 | { |
| 1597 | m_update->add (bb: node); |
| 1598 | continue; |
| 1599 | } |
| 1600 | |
| 1601 | // If the pred is entry but NOT def, then it is used before |
| 1602 | // defined, it'll get set to [] and no need to update it. |
| 1603 | if (pred == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
| 1604 | { |
| 1605 | if (DEBUG_RANGE_CACHE) |
| 1606 | fprintf (stream: dump_file, format: "entry: bail." ); |
| 1607 | continue; |
| 1608 | } |
| 1609 | |
| 1610 | // Regardless of whether we have visited pred or not, if the |
| 1611 | // pred has inferred ranges, revisit this block. |
| 1612 | // Don't search the DOM tree. |
| 1613 | if (infer_oracle ().has_range_p (pred, name)) |
| 1614 | { |
| 1615 | if (DEBUG_RANGE_CACHE) |
| 1616 | fprintf (stream: dump_file, format: "Inferred range: update " ); |
| 1617 | m_update->add (bb: node); |
| 1618 | } |
| 1619 | |
| 1620 | // If the pred block already has a range, or if it can contribute |
| 1621 | // something new. Ie, the edge generates a range of some sort. |
| 1622 | if (m_on_entry.get_bb_range (r, name, bb: pred)) |
| 1623 | { |
| 1624 | if (DEBUG_RANGE_CACHE) |
| 1625 | { |
| 1626 | fprintf (stream: dump_file, format: "has cache, " ); |
| 1627 | r.dump (dump_file); |
| 1628 | fprintf (stream: dump_file, format: ", " ); |
| 1629 | } |
| 1630 | if (!r.undefined_p () || gori ().has_edge_range_p (name, e)) |
| 1631 | { |
| 1632 | m_update->add (bb: node); |
| 1633 | if (DEBUG_RANGE_CACHE) |
| 1634 | fprintf (stream: dump_file, format: "update. " ); |
| 1635 | } |
| 1636 | continue; |
| 1637 | } |
| 1638 | |
| 1639 | if (DEBUG_RANGE_CACHE) |
| 1640 | fprintf (stream: dump_file, format: "pushing undefined pred block.\n" ); |
| 1641 | // If the pred hasn't been visited (has no range), add it to |
| 1642 | // the list. |
| 1643 | gcc_checking_assert (!m_on_entry.bb_range_p (name, pred)); |
| 1644 | m_on_entry.set_bb_range (name, bb: pred, r: undefined); |
| 1645 | m_workback.safe_push (obj: pred); |
| 1646 | } |
| 1647 | } |
| 1648 | |
| 1649 | if (DEBUG_RANGE_CACHE) |
| 1650 | fprintf (stream: dump_file, format: "\n" ); |
| 1651 | |
| 1652 | // Now fill in the marked blocks with values. |
| 1653 | propagate_cache (name); |
| 1654 | if (DEBUG_RANGE_CACHE) |
| 1655 | fprintf (stream: dump_file, format: " Propagation update done.\n" ); |
| 1656 | } |
| 1657 | |
| 1658 | // Resolve the range of BB if the dominators range is R by calculating incoming |
| 1659 | // edges to this block. All lead back to the dominator so should be cheap. |
| 1660 | // The range for BB is set and returned in R. |
| 1661 | |
| 1662 | void |
| 1663 | ranger_cache::resolve_dom (vrange &r, tree name, basic_block bb) |
| 1664 | { |
| 1665 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (name)); |
| 1666 | basic_block dom_bb = get_immediate_dominator (CDI_DOMINATORS, bb); |
| 1667 | |
| 1668 | // if it doesn't already have a value, store the incoming range. |
| 1669 | if (!m_on_entry.bb_range_p (name, bb: dom_bb) && def_bb != dom_bb) |
| 1670 | { |
| 1671 | // If the range can't be store, don't try to accumulate |
| 1672 | // the range in PREV_BB due to excessive recalculations. |
| 1673 | if (!m_on_entry.set_bb_range (name, bb: dom_bb, r)) |
| 1674 | return; |
| 1675 | } |
| 1676 | // With the dominator set, we should be able to cheaply query |
| 1677 | // each incoming edge now and accumulate the results. |
| 1678 | r.set_undefined (); |
| 1679 | edge e; |
| 1680 | edge_iterator ei; |
| 1681 | value_range er (TREE_TYPE (name)); |
| 1682 | FOR_EACH_EDGE (e, ei, bb->preds) |
| 1683 | { |
| 1684 | // If the predecessor is dominated by this block, then there is a back |
| 1685 | // edge, and won't provide anything useful. We'll actually end up with |
| 1686 | // VARYING as we will not resolve this node. |
| 1687 | if (dominated_by_p (CDI_DOMINATORS, e->src, bb)) |
| 1688 | continue; |
| 1689 | edge_range (r&: er, e, name, mode: RFD_READ_ONLY); |
| 1690 | r.union_ (er); |
| 1691 | } |
| 1692 | // Set the cache in PREV_BB so it is not calculated again. |
| 1693 | m_on_entry.set_bb_range (name, bb, r); |
| 1694 | } |
| 1695 | |
| 1696 | // Get the range of NAME from dominators of BB and return it in R. Search the |
| 1697 | // dominator tree based on MODE. |
| 1698 | |
| 1699 | bool |
| 1700 | ranger_cache::range_from_dom (vrange &r, tree name, basic_block start_bb, |
| 1701 | enum rfd_mode mode) |
| 1702 | { |
| 1703 | if (mode == RFD_NONE || !dom_info_available_p (CDI_DOMINATORS)) |
| 1704 | return false; |
| 1705 | |
| 1706 | // Search back to the definition block or entry block. |
| 1707 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (name)); |
| 1708 | if (def_bb == NULL) |
| 1709 | def_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
| 1710 | |
| 1711 | basic_block bb; |
| 1712 | basic_block prev_bb = start_bb; |
| 1713 | |
| 1714 | // Track any inferred ranges seen. |
| 1715 | value_range infer (TREE_TYPE (name)); |
| 1716 | infer.set_varying (TREE_TYPE (name)); |
| 1717 | |
| 1718 | // Range on entry to the DEF block should not be queried. |
| 1719 | gcc_checking_assert (start_bb != def_bb); |
| 1720 | unsigned start_limit = m_workback.length (); |
| 1721 | |
| 1722 | // Default value is global range. |
| 1723 | get_global_range (r, name); |
| 1724 | |
| 1725 | // The dominator of EXIT_BLOCK doesn't seem to be set, so at least handle |
| 1726 | // the common single exit cases. |
| 1727 | if (start_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) && single_pred_p (bb: start_bb)) |
| 1728 | bb = single_pred_edge (bb: start_bb)->src; |
| 1729 | else |
| 1730 | bb = get_immediate_dominator (CDI_DOMINATORS, start_bb); |
| 1731 | |
| 1732 | // Search until a value is found, pushing blocks which may need calculating. |
| 1733 | for ( ; bb; prev_bb = bb, bb = get_immediate_dominator (CDI_DOMINATORS, bb)) |
| 1734 | { |
| 1735 | // Accumulate any block exit inferred ranges. |
| 1736 | infer_oracle ().maybe_adjust_range (infer, name, bb); |
| 1737 | |
| 1738 | // This block has an outgoing range. |
| 1739 | if (gori ().has_edge_range_p (name, bb)) |
| 1740 | m_workback.safe_push (obj: prev_bb); |
| 1741 | else |
| 1742 | { |
| 1743 | // Normally join blocks don't carry any new range information on |
| 1744 | // incoming edges. If the first incoming edge to this block does |
| 1745 | // generate a range, calculate the ranges if all incoming edges |
| 1746 | // are also dominated by the dominator. (Avoids backedges which |
| 1747 | // will break the rule of moving only upward in the dominator tree). |
| 1748 | // If the first pred does not generate a range, then we will be |
| 1749 | // using the dominator range anyway, so that's all the check needed. |
| 1750 | if (EDGE_COUNT (prev_bb->preds) > 1 |
| 1751 | && gori ().has_edge_range_p (name, EDGE_PRED (prev_bb, 0)->src)) |
| 1752 | { |
| 1753 | edge e; |
| 1754 | edge_iterator ei; |
| 1755 | bool all_dom = true; |
| 1756 | FOR_EACH_EDGE (e, ei, prev_bb->preds) |
| 1757 | if (e->src != bb |
| 1758 | && !dominated_by_p (CDI_DOMINATORS, e->src, bb)) |
| 1759 | { |
| 1760 | all_dom = false; |
| 1761 | break; |
| 1762 | } |
| 1763 | if (all_dom) |
| 1764 | m_workback.safe_push (obj: prev_bb); |
| 1765 | } |
| 1766 | } |
| 1767 | |
| 1768 | if (def_bb == bb) |
| 1769 | break; |
| 1770 | |
| 1771 | if (m_on_entry.get_bb_range (r, name, bb)) |
| 1772 | break; |
| 1773 | } |
| 1774 | |
| 1775 | if (DEBUG_RANGE_CACHE) |
| 1776 | { |
| 1777 | fprintf (stream: dump_file, format: "CACHE: BB %d DOM query for " , start_bb->index); |
| 1778 | print_generic_expr (dump_file, name, TDF_SLIM); |
| 1779 | fprintf (stream: dump_file, format: ", found " ); |
| 1780 | r.dump (dump_file); |
| 1781 | if (bb) |
| 1782 | fprintf (stream: dump_file, format: " at BB%d\n" , bb->index); |
| 1783 | else |
| 1784 | fprintf (stream: dump_file, format: " at function top\n" ); |
| 1785 | } |
| 1786 | |
| 1787 | // Now process any blocks wit incoming edges that nay have adjustments. |
| 1788 | while (m_workback.length () > start_limit) |
| 1789 | { |
| 1790 | value_range er (TREE_TYPE (name)); |
| 1791 | prev_bb = m_workback.pop (); |
| 1792 | if (!single_pred_p (bb: prev_bb)) |
| 1793 | { |
| 1794 | // Non single pred means we need to cache a value in the dominator |
| 1795 | // so we can cheaply calculate incoming edges to this block, and |
| 1796 | // then store the resulting value. If processing mode is not |
| 1797 | // RFD_FILL, then the cache cant be stored to, so don't try. |
| 1798 | // Otherwise this becomes a quadratic timed calculation. |
| 1799 | if (mode == RFD_FILL) |
| 1800 | resolve_dom (r, name, bb: prev_bb); |
| 1801 | continue; |
| 1802 | } |
| 1803 | |
| 1804 | edge e = single_pred_edge (bb: prev_bb); |
| 1805 | bb = e->src; |
| 1806 | if (gori ().edge_range_p (er, e, name, *this)) |
| 1807 | { |
| 1808 | r.intersect (er); |
| 1809 | // If this is a normal edge, apply any inferred ranges. |
| 1810 | if ((e->flags & (EDGE_EH | EDGE_ABNORMAL)) == 0) |
| 1811 | infer_oracle ().maybe_adjust_range (r, name, bb); |
| 1812 | |
| 1813 | if (DEBUG_RANGE_CACHE) |
| 1814 | { |
| 1815 | fprintf (stream: dump_file, format: "CACHE: Adjusted edge range for %d->%d : " , |
| 1816 | bb->index, prev_bb->index); |
| 1817 | r.dump (dump_file); |
| 1818 | fprintf (stream: dump_file, format: "\n" ); |
| 1819 | } |
| 1820 | } |
| 1821 | } |
| 1822 | |
| 1823 | // Apply non-null if appropriate. |
| 1824 | if (!has_abnormal_call_or_eh_pred_edge_p (bb: start_bb)) |
| 1825 | r.intersect (infer); |
| 1826 | |
| 1827 | if (DEBUG_RANGE_CACHE) |
| 1828 | { |
| 1829 | fprintf (stream: dump_file, format: "CACHE: Range for DOM returns : " ); |
| 1830 | r.dump (dump_file); |
| 1831 | fprintf (stream: dump_file, format: "\n" ); |
| 1832 | } |
| 1833 | return true; |
| 1834 | } |
| 1835 | |
| 1836 | // This routine will register an inferred value in block BB, and possibly |
| 1837 | // update the on-entry cache if appropriate. |
| 1838 | |
| 1839 | void |
| 1840 | ranger_cache::register_inferred_value (const vrange &ir, tree name, |
| 1841 | basic_block bb) |
| 1842 | { |
| 1843 | value_range r (TREE_TYPE (name)); |
| 1844 | if (!m_on_entry.get_bb_range (r, name, bb)) |
| 1845 | exit_range (r, name, bb, mode: RFD_READ_ONLY); |
| 1846 | if (r.intersect (r: ir)) |
| 1847 | { |
| 1848 | m_on_entry.set_bb_range (name, bb, r); |
| 1849 | // If this range was invariant before, remove invariant. |
| 1850 | if (!gori ().has_edge_range_p (name)) |
| 1851 | gori_ssa ()->set_range_invariant (name, invariant: false); |
| 1852 | } |
| 1853 | } |
| 1854 | |
| 1855 | // This routine is used during a block walk to adjust any inferred ranges |
| 1856 | // of operands on stmt S. |
| 1857 | |
| 1858 | void |
| 1859 | ranger_cache::apply_inferred_ranges (gimple *s) |
| 1860 | { |
| 1861 | bool update = true; |
| 1862 | |
| 1863 | basic_block bb = gimple_bb (g: s); |
| 1864 | gimple_infer_range infer(s, this); |
| 1865 | if (infer.num () == 0) |
| 1866 | return; |
| 1867 | |
| 1868 | // Do not update the on-entry cache for block ending stmts. |
| 1869 | if (stmt_ends_bb_p (s)) |
| 1870 | { |
| 1871 | edge_iterator ei; |
| 1872 | edge e; |
| 1873 | FOR_EACH_EDGE (e, ei, gimple_bb (s)->succs) |
| 1874 | if (!(e->flags & (EDGE_ABNORMAL|EDGE_EH))) |
| 1875 | break; |
| 1876 | if (e == NULL) |
| 1877 | update = false; |
| 1878 | } |
| 1879 | |
| 1880 | infer_oracle ().add_ranges (s, infer); |
| 1881 | if (update) |
| 1882 | for (unsigned x = 0; x < infer.num (); x++) |
| 1883 | register_inferred_value (ir: infer.range (index: x), name: infer.name (index: x), bb); |
| 1884 | } |
| 1885 | |