1 /* Callgraph transformations to handle inlining 2 Copyright (C) 2003, 2004, 2007, 2008, 2009, 2010, 2011 3 Free Software Foundation, Inc. 4 Contributed by Jan Hubicka 5 6 This file is part of GCC. 7 8 GCC is free software; you can redistribute it and/or modify it under 9 the terms of the GNU General Public License as published by the Free 10 Software Foundation; either version 3, or (at your option) any later 11 version. 12 13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 14 WARRANTY; without even the implied warranty of MERCHANTABILITY or 15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 16 for more details. 17 18 You should have received a copy of the GNU General Public License 19 along with GCC; see the file COPYING3. If not see 20 <http://www.gnu.org/licenses/>. */ 21 22 /* The inline decisions are stored in callgraph in "inline plan" and 23 applied later. 24 25 To mark given call inline, use inline_call function. 26 The function marks the edge inlinable and, if necessary, produces 27 virtual clone in the callgraph representing the new copy of callee's 28 function body. 29 30 The inline plan is applied on given function body by inline_transform. */ 31 32 #include "config.h" 33 #include "system.h" 34 #include "coretypes.h" 35 #include "tm.h" 36 #include "tree.h" 37 #include "langhooks.h" 38 #include "cgraph.h" 39 #include "timevar.h" 40 #include "output.h" 41 #include "intl.h" 42 #include "coverage.h" 43 #include "ggc.h" 44 #include "tree-flow.h" 45 #include "ipa-prop.h" 46 #include "ipa-inline.h" 47 #include "tree-inline.h" 48 #include "tree-pass.h" 49 50 int ncalls_inlined; 51 int nfunctions_inlined; 52 53 /* Scale frequency of NODE edges by FREQ_SCALE. */ 54 55 static void 56 update_noncloned_frequencies (struct cgraph_node *node, 57 int freq_scale) 58 { 59 struct cgraph_edge *e; 60 61 /* We do not want to ignore high loop nest after freq drops to 0. */ 62 if (!freq_scale) 63 freq_scale = 1; 64 for (e = node->callees; e; e = e->next_callee) 65 { 66 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE; 67 if (e->frequency > CGRAPH_FREQ_MAX) 68 e->frequency = CGRAPH_FREQ_MAX; 69 if (!e->inline_failed) 70 update_noncloned_frequencies (e->callee, freq_scale); 71 } 72 for (e = node->indirect_calls; e; e = e->next_callee) 73 { 74 e->frequency = e->frequency * (gcov_type) freq_scale / CGRAPH_FREQ_BASE; 75 if (e->frequency > CGRAPH_FREQ_MAX) 76 e->frequency = CGRAPH_FREQ_MAX; 77 } 78 } 79 80 /* We removed or are going to remove the last call to NODE. 81 Return true if we can and want proactively remove the NODE now. 82 This is important to do, since we want inliner to know when offline 83 copy of function was removed. */ 84 85 static bool 86 can_remove_node_now_p_1 (struct cgraph_node *node) 87 { 88 /* FIXME: When address is taken of DECL_EXTERNAL function we still 89 can remove its offline copy, but we would need to keep unanalyzed node in 90 the callgraph so references can point to it. */ 91 return (!node->address_taken 92 && !ipa_ref_has_aliases_p (&node->ref_list) 93 && cgraph_can_remove_if_no_direct_calls_p (node) 94 /* Inlining might enable more devirtualizing, so we want to remove 95 those only after all devirtualizable virtual calls are processed. 96 Lacking may edges in callgraph we just preserve them post 97 inlining. */ 98 && (!DECL_VIRTUAL_P (node->decl) 99 || (!DECL_COMDAT (node->decl) 100 && !DECL_EXTERNAL (node->decl))) 101 /* During early inlining some unanalyzed cgraph nodes might be in the 102 callgraph and they might reffer the function in question. */ 103 && !cgraph_new_nodes); 104 } 105 106 /* We are going to eliminate last direct call to NODE (or alias of it) via edge E. 107 Verify that the NODE can be removed from unit and if it is contained in comdat 108 group that the whole comdat group is removable. */ 109 110 static bool 111 can_remove_node_now_p (struct cgraph_node *node, struct cgraph_edge *e) 112 { 113 struct cgraph_node *next; 114 if (!can_remove_node_now_p_1 (node)) 115 return false; 116 117 /* When we see same comdat group, we need to be sure that all 118 items can be removed. */ 119 if (!node->same_comdat_group) 120 return true; 121 for (next = node->same_comdat_group; 122 next != node; next = next->same_comdat_group) 123 if ((next->callers && next->callers != e) 124 || !can_remove_node_now_p_1 (next)) 125 return false; 126 return true; 127 } 128 129 130 /* E is expected to be an edge being inlined. Clone destination node of 131 the edge and redirect it to the new clone. 132 DUPLICATE is used for bookkeeping on whether we are actually creating new 133 clones or re-using node originally representing out-of-line function call. 134 */ 135 136 void 137 clone_inlined_nodes (struct cgraph_edge *e, bool duplicate, 138 bool update_original, int *overall_size) 139 { 140 if (duplicate) 141 { 142 /* We may eliminate the need for out-of-line copy to be output. 143 In that case just go ahead and re-use it. This is not just an 144 memory optimization. Making offline copy of fuction disappear 145 from the program will improve future decisions on inlining. */ 146 if (!e->callee->callers->next_caller 147 /* Recursive inlining never wants the master clone to 148 be overwritten. */ 149 && update_original 150 && can_remove_node_now_p (e->callee, e)) 151 { 152 /* TODO: When callee is in a comdat group, we could remove all of it, 153 including all inline clones inlined into it. That would however 154 need small function inlining to register edge removal hook to 155 maintain the priority queue. 156 157 For now we keep the ohter functions in the group in program until 158 cgraph_remove_unreachable_functions gets rid of them. */ 159 gcc_assert (!e->callee->global.inlined_to); 160 if (e->callee->analyzed && !DECL_EXTERNAL (e->callee->decl)) 161 { 162 if (overall_size) 163 *overall_size -= inline_summary (e->callee)->size; 164 nfunctions_inlined++; 165 } 166 duplicate = false; 167 e->callee->local.externally_visible = false; 168 update_noncloned_frequencies (e->callee, e->frequency); 169 } 170 else 171 { 172 struct cgraph_node *n; 173 n = cgraph_clone_node (e->callee, e->callee->decl, 174 e->count, e->frequency, 175 update_original, NULL, true); 176 cgraph_redirect_edge_callee (e, n); 177 } 178 } 179 180 if (e->caller->global.inlined_to) 181 e->callee->global.inlined_to = e->caller->global.inlined_to; 182 else 183 e->callee->global.inlined_to = e->caller; 184 185 /* Recursively clone all bodies. */ 186 for (e = e->callee->callees; e; e = e->next_callee) 187 if (!e->inline_failed) 188 clone_inlined_nodes (e, duplicate, update_original, overall_size); 189 } 190 191 192 /* Mark edge E as inlined and update callgraph accordingly. UPDATE_ORIGINAL 193 specify whether profile of original function should be updated. If any new 194 indirect edges are discovered in the process, add them to NEW_EDGES, unless 195 it is NULL. Return true iff any new callgraph edges were discovered as a 196 result of inlining. */ 197 198 bool 199 inline_call (struct cgraph_edge *e, bool update_original, 200 VEC (cgraph_edge_p, heap) **new_edges, 201 int *overall_size) 202 { 203 int old_size = 0, new_size = 0; 204 struct cgraph_node *to = NULL; 205 struct cgraph_edge *curr = e; 206 struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL); 207 208 /* Don't inline inlined edges. */ 209 gcc_assert (e->inline_failed); 210 /* Don't even think of inlining inline clone. */ 211 gcc_assert (!callee->global.inlined_to); 212 213 e->inline_failed = CIF_OK; 214 DECL_POSSIBLY_INLINED (callee->decl) = true; 215 216 to = e->caller; 217 if (to->global.inlined_to) 218 to = to->global.inlined_to; 219 220 /* If aliases are involved, redirect edge to the actual destination and 221 possibly remove the aliases. */ 222 if (e->callee != callee) 223 { 224 struct cgraph_node *alias = e->callee, *next_alias; 225 cgraph_redirect_edge_callee (e, callee); 226 while (alias && alias != callee) 227 { 228 if (!alias->callers 229 && can_remove_node_now_p (alias, e)) 230 { 231 next_alias = cgraph_alias_aliased_node (alias); 232 cgraph_remove_node (alias); 233 alias = next_alias; 234 } 235 else 236 break; 237 } 238 } 239 240 clone_inlined_nodes (e, true, update_original, overall_size); 241 242 gcc_assert (curr->callee->global.inlined_to == to); 243 244 old_size = inline_summary (to)->size; 245 inline_merge_summary (e); 246 new_size = inline_summary (to)->size; 247 if (overall_size) 248 *overall_size += new_size - old_size; 249 ncalls_inlined++; 250 251 /* This must happen after inline_merge_summary that rely on jump 252 functions of callee to not be updated. */ 253 if (optimize) 254 return ipa_propagate_indirect_call_infos (curr, new_edges); 255 else 256 return false; 257 } 258 259 260 /* Copy function body of NODE and redirect all inline clones to it. 261 This is done before inline plan is applied to NODE when there are 262 still some inline clones if it. 263 264 This is neccesary because inline decisions are not really transitive 265 and the other inline clones may have different bodies. */ 266 267 static struct cgraph_node * 268 save_inline_function_body (struct cgraph_node *node) 269 { 270 struct cgraph_node *first_clone, *n; 271 272 if (dump_file) 273 fprintf (dump_file, "\nSaving body of %s for later reuse\n", 274 cgraph_node_name (node)); 275 276 gcc_assert (node == cgraph_get_node (node->decl)); 277 278 /* first_clone will be turned into real function. */ 279 first_clone = node->clones; 280 first_clone->decl = copy_node (node->decl); 281 cgraph_insert_node_to_hashtable (first_clone); 282 gcc_assert (first_clone == cgraph_get_node (first_clone->decl)); 283 284 /* Now reshape the clone tree, so all other clones descends from 285 first_clone. */ 286 if (first_clone->next_sibling_clone) 287 { 288 for (n = first_clone->next_sibling_clone; n->next_sibling_clone; n = n->next_sibling_clone) 289 n->clone_of = first_clone; 290 n->clone_of = first_clone; 291 n->next_sibling_clone = first_clone->clones; 292 if (first_clone->clones) 293 first_clone->clones->prev_sibling_clone = n; 294 first_clone->clones = first_clone->next_sibling_clone; 295 first_clone->next_sibling_clone->prev_sibling_clone = NULL; 296 first_clone->next_sibling_clone = NULL; 297 gcc_assert (!first_clone->prev_sibling_clone); 298 } 299 first_clone->clone_of = NULL; 300 301 /* Now node in question has no clones. */ 302 node->clones = NULL; 303 304 /* Inline clones share decl with the function they are cloned 305 from. Walk the whole clone tree and redirect them all to the 306 new decl. */ 307 if (first_clone->clones) 308 for (n = first_clone->clones; n != first_clone;) 309 { 310 gcc_assert (n->decl == node->decl); 311 n->decl = first_clone->decl; 312 if (n->clones) 313 n = n->clones; 314 else if (n->next_sibling_clone) 315 n = n->next_sibling_clone; 316 else 317 { 318 while (n != first_clone && !n->next_sibling_clone) 319 n = n->clone_of; 320 if (n != first_clone) 321 n = n->next_sibling_clone; 322 } 323 } 324 325 /* Copy the OLD_VERSION_NODE function tree to the new version. */ 326 tree_function_versioning (node->decl, first_clone->decl, NULL, true, NULL, 327 false, NULL, NULL); 328 329 /* The function will be short lived and removed after we inline all the clones, 330 but make it internal so we won't confuse ourself. */ 331 DECL_EXTERNAL (first_clone->decl) = 0; 332 DECL_COMDAT_GROUP (first_clone->decl) = NULL_TREE; 333 TREE_PUBLIC (first_clone->decl) = 0; 334 DECL_COMDAT (first_clone->decl) = 0; 335 VEC_free (ipa_opt_pass, heap, 336 first_clone->ipa_transforms_to_apply); 337 first_clone->ipa_transforms_to_apply = NULL; 338 339 #ifdef ENABLE_CHECKING 340 verify_cgraph_node (first_clone); 341 #endif 342 return first_clone; 343 } 344 345 346 /* Apply inline plan to function. */ 347 348 unsigned int 349 inline_transform (struct cgraph_node *node) 350 { 351 unsigned int todo = 0; 352 struct cgraph_edge *e; 353 354 /* FIXME: Currently the pass manager is adding inline transform more than 355 once to some clones. This needs revisiting after WPA cleanups. */ 356 if (cfun->after_inlining) 357 return 0; 358 359 /* We might need the body of this function so that we can expand 360 it inline somewhere else. */ 361 if (cgraph_preserve_function_body_p (node)) 362 save_inline_function_body (node); 363 364 for (e = node->callees; e; e = e->next_callee) 365 cgraph_redirect_edge_call_stmt_to_callee (e); 366 367 timevar_push (TV_INTEGRATION); 368 if (node->callees) 369 todo = optimize_inline_calls (current_function_decl); 370 timevar_pop (TV_INTEGRATION); 371 372 cfun->always_inline_functions_inlined = true; 373 cfun->after_inlining = true; 374 todo |= execute_fixup_cfg (); 375 376 if (!(todo & TODO_update_ssa_any)) 377 /* Redirecting edges might lead to a need for vops to be recomputed. */ 378 todo |= TODO_update_ssa_only_virtuals; 379 380 return todo; 381 } 382