Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-threadedge.c @ 127:4c56639505ff
fix function.c and add CbC-example Makefile
| author | mir3636 |
|---|---|
| date | Wed, 11 Apr 2018 18:46:58 +0900 |
| parents | 04ced10e8804 |
| children | 84e7813d76e9 |
| rev | line source |
|---|---|
| 0 | 1 /* SSA Jump Threading |
| 111 | 2 Copyright (C) 2005-2017 Free Software Foundation, Inc. |
| 0 | 3 Contributed by Jeff Law <law@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" | |
| 111 | 24 #include "backend.h" |
| 0 | 25 #include "tree.h" |
| 111 | 26 #include "gimple.h" |
| 27 #include "predict.h" | |
| 28 #include "ssa.h" | |
| 29 #include "fold-const.h" | |
| 0 | 30 #include "cfgloop.h" |
| 111 | 31 #include "gimple-iterator.h" |
| 32 #include "tree-cfg.h" | |
| 33 #include "tree-ssa-threadupdate.h" | |
| 0 | 34 #include "params.h" |
| 111 | 35 #include "tree-ssa-scopedtables.h" |
| 36 #include "tree-ssa-threadedge.h" | |
| 37 #include "tree-ssa-dom.h" | |
| 38 #include "gimple-fold.h" | |
| 39 #include "cfganal.h" | |
| 0 | 40 |
| 41 /* To avoid code explosion due to jump threading, we limit the | |
| 42 number of statements we are going to copy. This variable | |
| 43 holds the number of statements currently seen that we'll have | |
| 44 to copy as part of the jump threading process. */ | |
| 45 static int stmt_count; | |
| 46 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
47 /* Array to record value-handles per SSA_NAME. */ |
| 111 | 48 vec<tree> ssa_name_values; |
| 49 | |
| 50 typedef tree (pfn_simplify) (gimple *, gimple *, | |
| 51 class avail_exprs_stack *, | |
| 52 basic_block); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
53 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
54 /* Set the value for the SSA name NAME to VALUE. */ |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
55 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
56 void |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
57 set_ssa_name_value (tree name, tree value) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
58 { |
| 111 | 59 if (SSA_NAME_VERSION (name) >= ssa_name_values.length ()) |
| 60 ssa_name_values.safe_grow_cleared (SSA_NAME_VERSION (name) + 1); | |
| 61 if (value && TREE_OVERFLOW_P (value)) | |
| 62 value = drop_tree_overflow (value); | |
| 63 ssa_name_values[SSA_NAME_VERSION (name)] = value; | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
64 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
65 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
66 /* Initialize the per SSA_NAME value-handles array. Returns it. */ |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
67 void |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
68 threadedge_initialize_values (void) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
69 { |
| 111 | 70 gcc_assert (!ssa_name_values.exists ()); |
| 71 ssa_name_values.create (num_ssa_names); | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
72 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
73 |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
74 /* Free the per SSA_NAME value-handle array. */ |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
75 void |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
76 threadedge_finalize_values (void) |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
77 { |
| 111 | 78 ssa_name_values.release (); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
79 } |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
80 |
| 0 | 81 /* Return TRUE if we may be able to thread an incoming edge into |
| 82 BB to an outgoing edge from BB. Return FALSE otherwise. */ | |
| 83 | |
| 84 bool | |
| 85 potentially_threadable_block (basic_block bb) | |
| 86 { | |
| 87 gimple_stmt_iterator gsi; | |
| 88 | |
| 111 | 89 /* Special case. We can get blocks that are forwarders, but are |
| 90 not optimized away because they forward from outside a loop | |
| 91 to the loop header. We want to thread through them as we can | |
| 92 sometimes thread to the loop exit, which is obviously profitable. | |
| 93 the interesting case here is when the block has PHIs. */ | |
| 94 if (gsi_end_p (gsi_start_nondebug_bb (bb)) | |
| 95 && !gsi_end_p (gsi_start_phis (bb))) | |
| 96 return true; | |
| 97 | |
| 0 | 98 /* If BB has a single successor or a single predecessor, then |
| 99 there is no threading opportunity. */ | |
| 100 if (single_succ_p (bb) || single_pred_p (bb)) | |
| 101 return false; | |
| 102 | |
| 103 /* If BB does not end with a conditional, switch or computed goto, | |
| 104 then there is no threading opportunity. */ | |
| 105 gsi = gsi_last_bb (bb); | |
| 106 if (gsi_end_p (gsi) | |
| 107 || ! gsi_stmt (gsi) | |
| 108 || (gimple_code (gsi_stmt (gsi)) != GIMPLE_COND | |
| 109 && gimple_code (gsi_stmt (gsi)) != GIMPLE_GOTO | |
| 110 && gimple_code (gsi_stmt (gsi)) != GIMPLE_SWITCH)) | |
| 111 return false; | |
| 112 | |
| 113 return true; | |
| 114 } | |
| 115 | |
| 116 /* Record temporary equivalences created by PHIs at the target of the | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
117 edge E. Record unwind information for the equivalences onto STACK. |
| 0 | 118 |
| 119 If a PHI which prevents threading is encountered, then return FALSE | |
| 111 | 120 indicating we should not thread this edge, else return TRUE. |
| 121 | |
| 122 If SRC_MAP/DST_MAP exist, then mark the source and destination SSA_NAMEs | |
| 123 of any equivalences recorded. We use this to make invalidation after | |
| 124 traversing back edges less painful. */ | |
| 0 | 125 |
| 126 static bool | |
| 111 | 127 record_temporary_equivalences_from_phis (edge e, const_and_copies *const_and_copies) |
| 0 | 128 { |
| 111 | 129 gphi_iterator gsi; |
| 0 | 130 |
| 131 /* Each PHI creates a temporary equivalence, record them. | |
| 132 These are context sensitive equivalences and will be removed | |
| 133 later. */ | |
| 134 for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 135 { | |
| 111 | 136 gphi *phi = gsi.phi (); |
| 0 | 137 tree src = PHI_ARG_DEF_FROM_EDGE (phi, e); |
| 138 tree dst = gimple_phi_result (phi); | |
| 139 | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
140 /* If the desired argument is not the same as this PHI's result |
| 0 | 141 and it is set by a PHI in E->dest, then we can not thread |
| 142 through E->dest. */ | |
| 143 if (src != dst | |
| 144 && TREE_CODE (src) == SSA_NAME | |
| 145 && gimple_code (SSA_NAME_DEF_STMT (src)) == GIMPLE_PHI | |
| 146 && gimple_bb (SSA_NAME_DEF_STMT (src)) == e->dest) | |
| 147 return false; | |
| 148 | |
| 149 /* We consider any non-virtual PHI as a statement since it | |
| 150 count result in a constant assignment or copy operation. */ | |
| 111 | 151 if (!virtual_operand_p (dst)) |
| 0 | 152 stmt_count++; |
| 153 | |
| 111 | 154 const_and_copies->record_const_or_copy (dst, src); |
| 0 | 155 } |
| 156 return true; | |
| 157 } | |
| 158 | |
| 111 | 159 /* Valueize hook for gimple_fold_stmt_to_constant_1. */ |
| 0 | 160 |
| 161 static tree | |
| 111 | 162 threadedge_valueize (tree t) |
| 0 | 163 { |
| 111 | 164 if (TREE_CODE (t) == SSA_NAME) |
| 0 | 165 { |
| 111 | 166 tree tem = SSA_NAME_VALUE (t); |
| 167 if (tem) | |
| 168 return tem; | |
| 0 | 169 } |
| 111 | 170 return t; |
| 0 | 171 } |
| 172 | |
| 173 /* Try to simplify each statement in E->dest, ultimately leading to | |
| 174 a simplification of the COND_EXPR at the end of E->dest. | |
| 175 | |
| 176 Record unwind information for temporary equivalences onto STACK. | |
| 177 | |
| 178 Use SIMPLIFY (a pointer to a callback function) to further simplify | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
179 statements using pass specific information. |
| 0 | 180 |
| 181 We might consider marking just those statements which ultimately | |
| 182 feed the COND_EXPR. It's not clear if the overhead of bookkeeping | |
| 183 would be recovered by trying to simplify fewer statements. | |
| 184 | |
| 185 If we are able to simplify a statement into the form | |
| 186 SSA_NAME = (SSA_NAME | gimple invariant), then we can record | |
| 187 a context sensitive equivalence which may help us simplify | |
| 188 later statements in E->dest. */ | |
| 189 | |
| 111 | 190 static gimple * |
| 0 | 191 record_temporary_equivalences_from_stmts_at_dest (edge e, |
| 111 | 192 const_and_copies *const_and_copies, |
| 193 avail_exprs_stack *avail_exprs_stack, | |
| 194 pfn_simplify simplify) | |
| 0 | 195 { |
| 111 | 196 gimple *stmt = NULL; |
| 0 | 197 gimple_stmt_iterator gsi; |
| 198 int max_stmt_count; | |
| 199 | |
| 200 max_stmt_count = PARAM_VALUE (PARAM_MAX_JUMP_THREAD_DUPLICATION_STMTS); | |
| 201 | |
| 202 /* Walk through each statement in the block recording equivalences | |
| 203 we discover. Note any equivalences we discover are context | |
| 204 sensitive (ie, are dependent on traversing E) and must be unwound | |
| 205 when we're finished processing E. */ | |
| 206 for (gsi = gsi_start_bb (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 207 { | |
| 208 tree cached_lhs = NULL; | |
| 209 | |
| 210 stmt = gsi_stmt (gsi); | |
| 211 | |
| 212 /* Ignore empty statements and labels. */ | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
213 if (gimple_code (stmt) == GIMPLE_NOP |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
214 || gimple_code (stmt) == GIMPLE_LABEL |
|
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
215 || is_gimple_debug (stmt)) |
| 0 | 216 continue; |
| 217 | |
| 218 /* If the statement has volatile operands, then we assume we | |
| 219 can not thread through this block. This is overly | |
| 220 conservative in some ways. */ | |
| 111 | 221 if (gimple_code (stmt) == GIMPLE_ASM |
| 222 && gimple_asm_volatile_p (as_a <gasm *> (stmt))) | |
| 223 return NULL; | |
| 224 | |
| 225 /* If the statement is a unique builtin, we can not thread | |
| 226 through here. */ | |
| 227 if (gimple_code (stmt) == GIMPLE_CALL | |
| 228 && gimple_call_internal_p (stmt) | |
| 229 && gimple_call_internal_unique_p (stmt)) | |
| 0 | 230 return NULL; |
| 231 | |
| 232 /* If duplicating this block is going to cause too much code | |
| 233 expansion, then do not thread through this block. */ | |
| 234 stmt_count++; | |
| 235 if (stmt_count > max_stmt_count) | |
| 236 return NULL; | |
| 237 | |
| 238 /* If this is not a statement that sets an SSA_NAME to a new | |
| 239 value, then do not try to simplify this statement as it will | |
| 240 not simplify in any way that is helpful for jump threading. */ | |
| 241 if ((gimple_code (stmt) != GIMPLE_ASSIGN | |
| 242 || TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) | |
| 243 && (gimple_code (stmt) != GIMPLE_CALL | |
| 244 || gimple_call_lhs (stmt) == NULL_TREE | |
| 245 || TREE_CODE (gimple_call_lhs (stmt)) != SSA_NAME)) | |
| 246 continue; | |
| 247 | |
| 248 /* The result of __builtin_object_size depends on all the arguments | |
| 249 of a phi node. Temporarily using only one edge produces invalid | |
| 250 results. For example | |
| 251 | |
| 252 if (x < 6) | |
| 253 goto l; | |
| 254 else | |
| 255 goto l; | |
| 256 | |
| 257 l: | |
| 258 r = PHI <&w[2].a[1](2), &a.a[6](3)> | |
| 259 __builtin_object_size (r, 0) | |
| 260 | |
| 261 The result of __builtin_object_size is defined to be the maximum of | |
| 262 remaining bytes. If we use only one edge on the phi, the result will | |
| 263 change to be the remaining bytes for the corresponding phi argument. | |
| 264 | |
| 265 Similarly for __builtin_constant_p: | |
| 266 | |
| 267 r = PHI <1(2), 2(3)> | |
| 268 __builtin_constant_p (r) | |
| 269 | |
| 270 Both PHI arguments are constant, but x ? 1 : 2 is still not | |
| 271 constant. */ | |
| 272 | |
| 273 if (is_gimple_call (stmt)) | |
| 274 { | |
| 275 tree fndecl = gimple_call_fndecl (stmt); | |
| 276 if (fndecl | |
| 277 && (DECL_FUNCTION_CODE (fndecl) == BUILT_IN_OBJECT_SIZE | |
| 278 || DECL_FUNCTION_CODE (fndecl) == BUILT_IN_CONSTANT_P)) | |
| 279 continue; | |
| 280 } | |
| 281 | |
| 282 /* At this point we have a statement which assigns an RHS to an | |
| 283 SSA_VAR on the LHS. We want to try and simplify this statement | |
| 284 to expose more context sensitive equivalences which in turn may | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
285 allow us to simplify the condition at the end of the loop. |
| 0 | 286 |
| 287 Handle simple copy operations as well as implied copies from | |
| 288 ASSERT_EXPRs. */ | |
| 289 if (gimple_assign_single_p (stmt) | |
| 290 && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME) | |
| 291 cached_lhs = gimple_assign_rhs1 (stmt); | |
| 292 else if (gimple_assign_single_p (stmt) | |
| 293 && TREE_CODE (gimple_assign_rhs1 (stmt)) == ASSERT_EXPR) | |
| 294 cached_lhs = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0); | |
| 295 else | |
| 296 { | |
| 297 /* A statement that is not a trivial copy or ASSERT_EXPR. | |
| 111 | 298 Try to fold the new expression. Inserting the |
| 299 expression into the hash table is unlikely to help. */ | |
| 300 /* ??? The DOM callback below can be changed to setting | |
| 301 the mprts_hook around the call to thread_across_edge, | |
| 302 avoiding the use substitution. The VRP hook should be | |
| 303 changed to properly valueize operands itself using | |
| 304 SSA_NAME_VALUE in addition to its own lattice. */ | |
| 305 cached_lhs = gimple_fold_stmt_to_constant_1 (stmt, | |
| 306 threadedge_valueize); | |
| 307 if (NUM_SSA_OPERANDS (stmt, SSA_OP_ALL_USES) != 0 | |
| 308 && (!cached_lhs | |
| 309 || (TREE_CODE (cached_lhs) != SSA_NAME | |
| 310 && !is_gimple_min_invariant (cached_lhs)))) | |
| 0 | 311 { |
| 111 | 312 /* We're going to temporarily copy propagate the operands |
| 313 and see if that allows us to simplify this statement. */ | |
| 314 tree *copy; | |
| 315 ssa_op_iter iter; | |
| 316 use_operand_p use_p; | |
| 317 unsigned int num, i = 0; | |
| 0 | 318 |
| 111 | 319 num = NUM_SSA_OPERANDS (stmt, SSA_OP_ALL_USES); |
| 320 copy = XALLOCAVEC (tree, num); | |
| 321 | |
| 322 /* Make a copy of the uses & vuses into USES_COPY, then cprop into | |
| 323 the operands. */ | |
| 324 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
| 325 { | |
| 326 tree tmp = NULL; | |
| 327 tree use = USE_FROM_PTR (use_p); | |
| 0 | 328 |
| 111 | 329 copy[i++] = use; |
| 330 if (TREE_CODE (use) == SSA_NAME) | |
| 331 tmp = SSA_NAME_VALUE (use); | |
| 332 if (tmp) | |
| 333 SET_USE (use_p, tmp); | |
| 334 } | |
| 0 | 335 |
| 111 | 336 cached_lhs = (*simplify) (stmt, stmt, avail_exprs_stack, e->src); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
337 |
| 111 | 338 /* Restore the statement's original uses/defs. */ |
| 339 i = 0; | |
| 340 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_ALL_USES) | |
| 341 SET_USE (use_p, copy[i++]); | |
| 342 } | |
| 0 | 343 } |
| 344 | |
| 345 /* Record the context sensitive equivalence if we were able | |
| 346 to simplify this statement. */ | |
| 347 if (cached_lhs | |
| 348 && (TREE_CODE (cached_lhs) == SSA_NAME | |
| 349 || is_gimple_min_invariant (cached_lhs))) | |
| 111 | 350 const_and_copies->record_const_or_copy (gimple_get_lhs (stmt), |
| 351 cached_lhs); | |
| 0 | 352 } |
| 353 return stmt; | |
| 354 } | |
| 355 | |
| 111 | 356 static tree simplify_control_stmt_condition_1 (edge, gimple *, |
| 357 class avail_exprs_stack *, | |
| 358 tree, enum tree_code, tree, | |
| 359 gcond *, pfn_simplify, | |
| 360 unsigned); | |
| 361 | |
| 0 | 362 /* Simplify the control statement at the end of the block E->dest. |
| 363 | |
| 364 To avoid allocating memory unnecessarily, a scratch GIMPLE_COND | |
| 365 is available to use/clobber in DUMMY_COND. | |
| 366 | |
| 367 Use SIMPLIFY (a pointer to a callback function) to further simplify | |
| 368 a condition using pass specific information. | |
| 369 | |
| 370 Return the simplified condition or NULL if simplification could | |
| 111 | 371 not be performed. When simplifying a GIMPLE_SWITCH, we may return |
| 372 the CASE_LABEL_EXPR that will be taken. | |
| 373 | |
| 374 The available expression table is referenced via AVAIL_EXPRS_STACK. */ | |
| 0 | 375 |
| 376 static tree | |
| 377 simplify_control_stmt_condition (edge e, | |
| 111 | 378 gimple *stmt, |
| 379 class avail_exprs_stack *avail_exprs_stack, | |
| 380 gcond *dummy_cond, | |
| 381 pfn_simplify simplify) | |
| 0 | 382 { |
| 383 tree cond, cached_lhs; | |
| 384 enum gimple_code code = gimple_code (stmt); | |
| 385 | |
| 386 /* For comparisons, we have to update both operands, then try | |
| 387 to simplify the comparison. */ | |
| 388 if (code == GIMPLE_COND) | |
| 389 { | |
| 390 tree op0, op1; | |
| 391 enum tree_code cond_code; | |
| 392 | |
| 393 op0 = gimple_cond_lhs (stmt); | |
| 394 op1 = gimple_cond_rhs (stmt); | |
| 395 cond_code = gimple_cond_code (stmt); | |
| 396 | |
| 397 /* Get the current value of both operands. */ | |
| 398 if (TREE_CODE (op0) == SSA_NAME) | |
| 399 { | |
| 111 | 400 for (int i = 0; i < 2; i++) |
| 401 { | |
| 402 if (TREE_CODE (op0) == SSA_NAME | |
| 403 && SSA_NAME_VALUE (op0)) | |
| 404 op0 = SSA_NAME_VALUE (op0); | |
| 405 else | |
| 406 break; | |
| 407 } | |
| 0 | 408 } |
| 409 | |
| 410 if (TREE_CODE (op1) == SSA_NAME) | |
| 411 { | |
| 111 | 412 for (int i = 0; i < 2; i++) |
| 413 { | |
| 414 if (TREE_CODE (op1) == SSA_NAME | |
| 415 && SSA_NAME_VALUE (op1)) | |
| 416 op1 = SSA_NAME_VALUE (op1); | |
| 417 else | |
| 418 break; | |
| 419 } | |
| 0 | 420 } |
| 421 | |
| 111 | 422 const unsigned recursion_limit = 4; |
| 0 | 423 |
| 111 | 424 cached_lhs |
| 425 = simplify_control_stmt_condition_1 (e, stmt, avail_exprs_stack, | |
| 426 op0, cond_code, op1, | |
| 427 dummy_cond, simplify, | |
| 428 recursion_limit); | |
| 429 | |
| 430 /* If we were testing an integer/pointer against a constant, then | |
| 431 we can use the FSM code to trace the value of the SSA_NAME. If | |
| 432 a value is found, then the condition will collapse to a constant. | |
| 0 | 433 |
| 111 | 434 Return the SSA_NAME we want to trace back rather than the full |
| 435 expression and give the FSM threader a chance to find its value. */ | |
| 436 if (cached_lhs == NULL) | |
| 437 { | |
| 438 /* Recover the original operands. They may have been simplified | |
| 439 using context sensitive equivalences. Those context sensitive | |
| 440 equivalences may not be valid on paths found by the FSM optimizer. */ | |
| 441 tree op0 = gimple_cond_lhs (stmt); | |
| 442 tree op1 = gimple_cond_rhs (stmt); | |
| 0 | 443 |
| 111 | 444 if ((INTEGRAL_TYPE_P (TREE_TYPE (op0)) |
| 445 || POINTER_TYPE_P (TREE_TYPE (op0))) | |
| 446 && TREE_CODE (op0) == SSA_NAME | |
| 447 && TREE_CODE (op1) == INTEGER_CST) | |
| 448 return op0; | |
| 449 } | |
| 0 | 450 |
| 451 return cached_lhs; | |
| 452 } | |
| 453 | |
| 454 if (code == GIMPLE_SWITCH) | |
| 111 | 455 cond = gimple_switch_index (as_a <gswitch *> (stmt)); |
| 0 | 456 else if (code == GIMPLE_GOTO) |
| 457 cond = gimple_goto_dest (stmt); | |
| 458 else | |
| 459 gcc_unreachable (); | |
| 460 | |
| 461 /* We can have conditionals which just test the state of a variable | |
| 462 rather than use a relational operator. These are simpler to handle. */ | |
| 463 if (TREE_CODE (cond) == SSA_NAME) | |
| 464 { | |
| 111 | 465 tree original_lhs = cond; |
| 0 | 466 cached_lhs = cond; |
| 467 | |
| 468 /* Get the variable's current value from the equivalence chains. | |
| 469 | |
| 470 It is possible to get loops in the SSA_NAME_VALUE chains | |
| 471 (consider threading the backedge of a loop where we have | |
| 111 | 472 a loop invariant SSA_NAME used in the condition). */ |
| 473 if (cached_lhs) | |
| 474 { | |
| 475 for (int i = 0; i < 2; i++) | |
| 476 { | |
| 477 if (TREE_CODE (cached_lhs) == SSA_NAME | |
| 478 && SSA_NAME_VALUE (cached_lhs)) | |
| 479 cached_lhs = SSA_NAME_VALUE (cached_lhs); | |
| 480 else | |
| 481 break; | |
| 482 } | |
| 483 } | |
| 0 | 484 |
| 485 /* If we haven't simplified to an invariant yet, then use the | |
| 486 pass specific callback to try and simplify it further. */ | |
| 487 if (cached_lhs && ! is_gimple_min_invariant (cached_lhs)) | |
| 111 | 488 { |
| 489 if (code == GIMPLE_SWITCH) | |
| 490 { | |
| 491 /* Replace the index operand of the GIMPLE_SWITCH with any LHS | |
| 492 we found before handing off to VRP. If simplification is | |
| 493 possible, the simplified value will be a CASE_LABEL_EXPR of | |
| 494 the label that is proven to be taken. */ | |
| 495 gswitch *dummy_switch = as_a<gswitch *> (gimple_copy (stmt)); | |
| 496 gimple_switch_set_index (dummy_switch, cached_lhs); | |
| 497 cached_lhs = (*simplify) (dummy_switch, stmt, | |
| 498 avail_exprs_stack, e->src); | |
| 499 ggc_free (dummy_switch); | |
| 500 } | |
| 501 else | |
| 502 cached_lhs = (*simplify) (stmt, stmt, avail_exprs_stack, e->src); | |
| 503 } | |
| 504 | |
| 505 /* We couldn't find an invariant. But, callers of this | |
| 506 function may be able to do something useful with the | |
| 507 unmodified destination. */ | |
| 508 if (!cached_lhs) | |
| 509 cached_lhs = original_lhs; | |
| 0 | 510 } |
| 511 else | |
| 512 cached_lhs = NULL; | |
| 513 | |
| 514 return cached_lhs; | |
| 515 } | |
| 516 | |
| 111 | 517 /* Recursive helper for simplify_control_stmt_condition. */ |
| 518 | |
| 519 static tree | |
| 520 simplify_control_stmt_condition_1 (edge e, | |
| 521 gimple *stmt, | |
| 522 class avail_exprs_stack *avail_exprs_stack, | |
| 523 tree op0, | |
| 524 enum tree_code cond_code, | |
| 525 tree op1, | |
| 526 gcond *dummy_cond, | |
| 527 pfn_simplify simplify, | |
| 528 unsigned limit) | |
| 529 { | |
| 530 if (limit == 0) | |
| 531 return NULL_TREE; | |
| 532 | |
| 533 /* We may need to canonicalize the comparison. For | |
| 534 example, op0 might be a constant while op1 is an | |
| 535 SSA_NAME. Failure to canonicalize will cause us to | |
| 536 miss threading opportunities. */ | |
| 537 if (tree_swap_operands_p (op0, op1)) | |
| 538 { | |
| 539 cond_code = swap_tree_comparison (cond_code); | |
| 540 std::swap (op0, op1); | |
| 541 } | |
| 542 | |
| 543 /* If the condition has the form (A & B) CMP 0 or (A | B) CMP 0 then | |
| 544 recurse into the LHS to see if there is a dominating ASSERT_EXPR | |
| 545 of A or of B that makes this condition always true or always false | |
| 546 along the edge E. */ | |
| 547 if ((cond_code == EQ_EXPR || cond_code == NE_EXPR) | |
| 548 && TREE_CODE (op0) == SSA_NAME | |
| 549 && integer_zerop (op1)) | |
| 550 { | |
| 551 gimple *def_stmt = SSA_NAME_DEF_STMT (op0); | |
| 552 if (gimple_code (def_stmt) != GIMPLE_ASSIGN) | |
| 553 ; | |
| 554 else if (gimple_assign_rhs_code (def_stmt) == BIT_AND_EXPR | |
| 555 || gimple_assign_rhs_code (def_stmt) == BIT_IOR_EXPR) | |
| 556 { | |
| 557 enum tree_code rhs_code = gimple_assign_rhs_code (def_stmt); | |
| 558 const tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
| 559 const tree rhs2 = gimple_assign_rhs2 (def_stmt); | |
| 560 | |
| 561 /* Is A != 0 ? */ | |
| 562 const tree res1 | |
| 563 = simplify_control_stmt_condition_1 (e, def_stmt, avail_exprs_stack, | |
| 564 rhs1, NE_EXPR, op1, | |
| 565 dummy_cond, simplify, | |
| 566 limit - 1); | |
| 567 if (res1 == NULL_TREE) | |
| 568 ; | |
| 569 else if (rhs_code == BIT_AND_EXPR && integer_zerop (res1)) | |
| 570 { | |
| 571 /* If A == 0 then (A & B) != 0 is always false. */ | |
| 572 if (cond_code == NE_EXPR) | |
| 573 return boolean_false_node; | |
| 574 /* If A == 0 then (A & B) == 0 is always true. */ | |
| 575 if (cond_code == EQ_EXPR) | |
| 576 return boolean_true_node; | |
| 577 } | |
| 578 else if (rhs_code == BIT_IOR_EXPR && integer_nonzerop (res1)) | |
| 579 { | |
| 580 /* If A != 0 then (A | B) != 0 is always true. */ | |
| 581 if (cond_code == NE_EXPR) | |
| 582 return boolean_true_node; | |
| 583 /* If A != 0 then (A | B) == 0 is always false. */ | |
| 584 if (cond_code == EQ_EXPR) | |
| 585 return boolean_false_node; | |
| 586 } | |
| 587 | |
| 588 /* Is B != 0 ? */ | |
| 589 const tree res2 | |
| 590 = simplify_control_stmt_condition_1 (e, def_stmt, avail_exprs_stack, | |
| 591 rhs2, NE_EXPR, op1, | |
| 592 dummy_cond, simplify, | |
| 593 limit - 1); | |
| 594 if (res2 == NULL_TREE) | |
| 595 ; | |
| 596 else if (rhs_code == BIT_AND_EXPR && integer_zerop (res2)) | |
| 597 { | |
| 598 /* If B == 0 then (A & B) != 0 is always false. */ | |
| 599 if (cond_code == NE_EXPR) | |
| 600 return boolean_false_node; | |
| 601 /* If B == 0 then (A & B) == 0 is always true. */ | |
| 602 if (cond_code == EQ_EXPR) | |
| 603 return boolean_true_node; | |
| 604 } | |
| 605 else if (rhs_code == BIT_IOR_EXPR && integer_nonzerop (res2)) | |
| 606 { | |
| 607 /* If B != 0 then (A | B) != 0 is always true. */ | |
| 608 if (cond_code == NE_EXPR) | |
| 609 return boolean_true_node; | |
| 610 /* If B != 0 then (A | B) == 0 is always false. */ | |
| 611 if (cond_code == EQ_EXPR) | |
| 612 return boolean_false_node; | |
| 613 } | |
| 614 | |
| 615 if (res1 != NULL_TREE && res2 != NULL_TREE) | |
| 616 { | |
| 617 if (rhs_code == BIT_AND_EXPR | |
| 618 && TYPE_PRECISION (TREE_TYPE (op0)) == 1 | |
| 619 && integer_nonzerop (res1) | |
| 620 && integer_nonzerop (res2)) | |
| 621 { | |
| 622 /* If A != 0 and B != 0 then (bool)(A & B) != 0 is true. */ | |
| 623 if (cond_code == NE_EXPR) | |
| 624 return boolean_true_node; | |
| 625 /* If A != 0 and B != 0 then (bool)(A & B) == 0 is false. */ | |
| 626 if (cond_code == EQ_EXPR) | |
| 627 return boolean_false_node; | |
| 628 } | |
| 629 | |
| 630 if (rhs_code == BIT_IOR_EXPR | |
| 631 && integer_zerop (res1) | |
| 632 && integer_zerop (res2)) | |
| 633 { | |
| 634 /* If A == 0 and B == 0 then (A | B) != 0 is false. */ | |
| 635 if (cond_code == NE_EXPR) | |
| 636 return boolean_false_node; | |
| 637 /* If A == 0 and B == 0 then (A | B) == 0 is true. */ | |
| 638 if (cond_code == EQ_EXPR) | |
| 639 return boolean_true_node; | |
| 640 } | |
| 641 } | |
| 642 } | |
| 643 /* Handle (A CMP B) CMP 0. */ | |
| 644 else if (TREE_CODE_CLASS (gimple_assign_rhs_code (def_stmt)) | |
| 645 == tcc_comparison) | |
| 646 { | |
| 647 tree rhs1 = gimple_assign_rhs1 (def_stmt); | |
| 648 tree rhs2 = gimple_assign_rhs2 (def_stmt); | |
| 649 | |
| 650 tree_code new_cond = gimple_assign_rhs_code (def_stmt); | |
| 651 if (cond_code == EQ_EXPR) | |
| 652 new_cond = invert_tree_comparison (new_cond, false); | |
| 653 | |
| 654 tree res | |
| 655 = simplify_control_stmt_condition_1 (e, def_stmt, avail_exprs_stack, | |
| 656 rhs1, new_cond, rhs2, | |
| 657 dummy_cond, simplify, | |
| 658 limit - 1); | |
| 659 if (res != NULL_TREE && is_gimple_min_invariant (res)) | |
| 660 return res; | |
| 661 } | |
| 662 } | |
| 663 | |
| 664 gimple_cond_set_code (dummy_cond, cond_code); | |
| 665 gimple_cond_set_lhs (dummy_cond, op0); | |
| 666 gimple_cond_set_rhs (dummy_cond, op1); | |
| 667 | |
| 668 /* We absolutely do not care about any type conversions | |
| 669 we only care about a zero/nonzero value. */ | |
| 670 fold_defer_overflow_warnings (); | |
| 671 | |
| 672 tree res = fold_binary (cond_code, boolean_type_node, op0, op1); | |
| 673 if (res) | |
| 674 while (CONVERT_EXPR_P (res)) | |
| 675 res = TREE_OPERAND (res, 0); | |
| 676 | |
| 677 fold_undefer_overflow_warnings ((res && is_gimple_min_invariant (res)), | |
| 678 stmt, WARN_STRICT_OVERFLOW_CONDITIONAL); | |
| 679 | |
| 680 /* If we have not simplified the condition down to an invariant, | |
| 681 then use the pass specific callback to simplify the condition. */ | |
| 682 if (!res | |
| 683 || !is_gimple_min_invariant (res)) | |
| 684 res = (*simplify) (dummy_cond, stmt, avail_exprs_stack, e->src); | |
| 685 | |
| 686 return res; | |
| 687 } | |
| 688 | |
| 689 /* Copy debug stmts from DEST's chain of single predecessors up to | |
| 690 SRC, so that we don't lose the bindings as PHI nodes are introduced | |
| 691 when DEST gains new predecessors. */ | |
| 692 void | |
| 693 propagate_threaded_block_debug_into (basic_block dest, basic_block src) | |
| 694 { | |
| 695 if (!MAY_HAVE_DEBUG_STMTS) | |
| 696 return; | |
| 697 | |
| 698 if (!single_pred_p (dest)) | |
| 699 return; | |
| 700 | |
| 701 gcc_checking_assert (dest != src); | |
| 702 | |
| 703 gimple_stmt_iterator gsi = gsi_after_labels (dest); | |
| 704 int i = 0; | |
| 705 const int alloc_count = 16; // ?? Should this be a PARAM? | |
| 706 | |
| 707 /* Estimate the number of debug vars overridden in the beginning of | |
| 708 DEST, to tell how many we're going to need to begin with. */ | |
| 709 for (gimple_stmt_iterator si = gsi; | |
| 710 i * 4 <= alloc_count * 3 && !gsi_end_p (si); gsi_next (&si)) | |
| 711 { | |
| 712 gimple *stmt = gsi_stmt (si); | |
| 713 if (!is_gimple_debug (stmt)) | |
| 714 break; | |
| 715 i++; | |
| 716 } | |
| 717 | |
| 718 auto_vec<tree, alloc_count> fewvars; | |
| 719 hash_set<tree> *vars = NULL; | |
| 720 | |
| 721 /* If we're already starting with 3/4 of alloc_count, go for a | |
| 722 hash_set, otherwise start with an unordered stack-allocated | |
| 723 VEC. */ | |
| 724 if (i * 4 > alloc_count * 3) | |
| 725 vars = new hash_set<tree>; | |
| 726 | |
| 727 /* Now go through the initial debug stmts in DEST again, this time | |
| 728 actually inserting in VARS or FEWVARS. Don't bother checking for | |
| 729 duplicates in FEWVARS. */ | |
| 730 for (gimple_stmt_iterator si = gsi; !gsi_end_p (si); gsi_next (&si)) | |
| 731 { | |
| 732 gimple *stmt = gsi_stmt (si); | |
| 733 if (!is_gimple_debug (stmt)) | |
| 734 break; | |
| 735 | |
| 736 tree var; | |
| 737 | |
| 738 if (gimple_debug_bind_p (stmt)) | |
| 739 var = gimple_debug_bind_get_var (stmt); | |
| 740 else if (gimple_debug_source_bind_p (stmt)) | |
| 741 var = gimple_debug_source_bind_get_var (stmt); | |
| 742 else | |
| 743 gcc_unreachable (); | |
| 744 | |
| 745 if (vars) | |
| 746 vars->add (var); | |
| 747 else | |
| 748 fewvars.quick_push (var); | |
| 749 } | |
| 750 | |
| 751 basic_block bb = dest; | |
| 752 | |
| 753 do | |
| 754 { | |
| 755 bb = single_pred (bb); | |
| 756 for (gimple_stmt_iterator si = gsi_last_bb (bb); | |
| 757 !gsi_end_p (si); gsi_prev (&si)) | |
| 758 { | |
| 759 gimple *stmt = gsi_stmt (si); | |
| 760 if (!is_gimple_debug (stmt)) | |
| 761 continue; | |
| 762 | |
| 763 tree var; | |
| 764 | |
| 765 if (gimple_debug_bind_p (stmt)) | |
| 766 var = gimple_debug_bind_get_var (stmt); | |
| 767 else if (gimple_debug_source_bind_p (stmt)) | |
| 768 var = gimple_debug_source_bind_get_var (stmt); | |
| 769 else | |
| 770 gcc_unreachable (); | |
| 771 | |
| 772 /* Discard debug bind overlaps. ??? Unlike stmts from src, | |
| 773 copied into a new block that will precede BB, debug bind | |
| 774 stmts in bypassed BBs may actually be discarded if | |
| 775 they're overwritten by subsequent debug bind stmts, which | |
| 776 might be a problem once we introduce stmt frontier notes | |
| 777 or somesuch. Adding `&& bb == src' to the condition | |
| 778 below will preserve all potentially relevant debug | |
| 779 notes. */ | |
| 780 if (vars && vars->add (var)) | |
| 781 continue; | |
| 782 else if (!vars) | |
| 783 { | |
| 784 int i = fewvars.length (); | |
| 785 while (i--) | |
| 786 if (fewvars[i] == var) | |
| 787 break; | |
| 788 if (i >= 0) | |
| 789 continue; | |
| 790 | |
| 791 if (fewvars.length () < (unsigned) alloc_count) | |
| 792 fewvars.quick_push (var); | |
| 793 else | |
| 794 { | |
| 795 vars = new hash_set<tree>; | |
| 796 for (i = 0; i < alloc_count; i++) | |
| 797 vars->add (fewvars[i]); | |
| 798 fewvars.release (); | |
| 799 vars->add (var); | |
| 800 } | |
| 801 } | |
| 802 | |
| 803 stmt = gimple_copy (stmt); | |
| 804 /* ??? Should we drop the location of the copy to denote | |
| 805 they're artificial bindings? */ | |
| 806 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
| 807 } | |
| 808 } | |
| 809 while (bb != src && single_pred_p (bb)); | |
| 810 | |
| 811 if (vars) | |
| 812 delete vars; | |
| 813 else if (fewvars.exists ()) | |
| 814 fewvars.release (); | |
| 815 } | |
| 816 | |
| 817 /* See if TAKEN_EDGE->dest is a threadable block with no side effecs (ie, it | |
| 818 need not be duplicated as part of the CFG/SSA updating process). | |
| 819 | |
| 820 If it is threadable, add it to PATH and VISITED and recurse, ultimately | |
| 821 returning TRUE from the toplevel call. Otherwise do nothing and | |
| 822 return false. | |
| 823 | |
| 824 DUMMY_COND, SIMPLIFY are used to try and simplify the condition at the | |
| 825 end of TAKEN_EDGE->dest. | |
| 826 | |
| 827 The available expression table is referenced via AVAIL_EXPRS_STACK. */ | |
| 828 | |
| 829 static bool | |
| 830 thread_around_empty_blocks (edge taken_edge, | |
| 831 gcond *dummy_cond, | |
| 832 class avail_exprs_stack *avail_exprs_stack, | |
| 833 pfn_simplify simplify, | |
| 834 bitmap visited, | |
| 835 vec<jump_thread_edge *> *path) | |
| 836 { | |
| 837 basic_block bb = taken_edge->dest; | |
| 838 gimple_stmt_iterator gsi; | |
| 839 gimple *stmt; | |
| 840 tree cond; | |
| 841 | |
| 842 /* The key property of these blocks is that they need not be duplicated | |
| 843 when threading. Thus they can not have visible side effects such | |
| 844 as PHI nodes. */ | |
| 845 if (!gsi_end_p (gsi_start_phis (bb))) | |
| 846 return false; | |
| 847 | |
| 848 /* Skip over DEBUG statements at the start of the block. */ | |
| 849 gsi = gsi_start_nondebug_bb (bb); | |
| 850 | |
| 851 /* If the block has no statements, but does have a single successor, then | |
| 852 it's just a forwarding block and we can thread through it trivially. | |
| 853 | |
| 854 However, note that just threading through empty blocks with single | |
| 855 successors is not inherently profitable. For the jump thread to | |
| 856 be profitable, we must avoid a runtime conditional. | |
| 857 | |
| 858 By taking the return value from the recursive call, we get the | |
| 859 desired effect of returning TRUE when we found a profitable jump | |
| 860 threading opportunity and FALSE otherwise. | |
| 861 | |
| 862 This is particularly important when this routine is called after | |
| 863 processing a joiner block. Returning TRUE too aggressively in | |
| 864 that case results in pointless duplication of the joiner block. */ | |
| 865 if (gsi_end_p (gsi)) | |
| 866 { | |
| 867 if (single_succ_p (bb)) | |
| 868 { | |
| 869 taken_edge = single_succ_edge (bb); | |
| 870 | |
| 871 if ((taken_edge->flags & EDGE_DFS_BACK) != 0) | |
| 872 return false; | |
| 873 | |
| 874 if (!bitmap_bit_p (visited, taken_edge->dest->index)) | |
| 875 { | |
| 876 jump_thread_edge *x | |
| 877 = new jump_thread_edge (taken_edge, EDGE_NO_COPY_SRC_BLOCK); | |
| 878 path->safe_push (x); | |
| 879 bitmap_set_bit (visited, taken_edge->dest->index); | |
| 880 return thread_around_empty_blocks (taken_edge, | |
| 881 dummy_cond, | |
| 882 avail_exprs_stack, | |
| 883 simplify, | |
| 884 visited, | |
| 885 path); | |
| 886 } | |
| 887 } | |
| 888 | |
| 889 /* We have a block with no statements, but multiple successors? */ | |
| 890 return false; | |
| 891 } | |
| 892 | |
| 893 /* The only real statements this block can have are a control | |
| 894 flow altering statement. Anything else stops the thread. */ | |
| 895 stmt = gsi_stmt (gsi); | |
| 896 if (gimple_code (stmt) != GIMPLE_COND | |
| 897 && gimple_code (stmt) != GIMPLE_GOTO | |
| 898 && gimple_code (stmt) != GIMPLE_SWITCH) | |
| 899 return false; | |
| 900 | |
| 901 /* Extract and simplify the condition. */ | |
| 902 cond = simplify_control_stmt_condition (taken_edge, stmt, | |
| 903 avail_exprs_stack, dummy_cond, | |
| 904 simplify); | |
| 905 | |
| 906 /* If the condition can be statically computed and we have not already | |
| 907 visited the destination edge, then add the taken edge to our thread | |
| 908 path. */ | |
| 909 if (cond != NULL_TREE | |
| 910 && (is_gimple_min_invariant (cond) | |
| 911 || TREE_CODE (cond) == CASE_LABEL_EXPR)) | |
| 912 { | |
| 913 if (TREE_CODE (cond) == CASE_LABEL_EXPR) | |
| 914 taken_edge = find_edge (bb, label_to_block (CASE_LABEL (cond))); | |
| 915 else | |
| 916 taken_edge = find_taken_edge (bb, cond); | |
| 917 | |
| 918 if ((taken_edge->flags & EDGE_DFS_BACK) != 0) | |
| 919 return false; | |
| 920 | |
| 921 if (bitmap_bit_p (visited, taken_edge->dest->index)) | |
| 922 return false; | |
| 923 bitmap_set_bit (visited, taken_edge->dest->index); | |
| 924 | |
| 925 jump_thread_edge *x | |
| 926 = new jump_thread_edge (taken_edge, EDGE_NO_COPY_SRC_BLOCK); | |
| 927 path->safe_push (x); | |
| 928 | |
| 929 thread_around_empty_blocks (taken_edge, | |
| 930 dummy_cond, | |
| 931 avail_exprs_stack, | |
| 932 simplify, | |
| 933 visited, | |
| 934 path); | |
| 935 return true; | |
| 936 } | |
| 937 | |
| 938 return false; | |
| 939 } | |
| 940 | |
| 0 | 941 /* We are exiting E->src, see if E->dest ends with a conditional |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
942 jump which has a known value when reached via E. |
| 0 | 943 |
| 111 | 944 E->dest can have arbitrary side effects which, if threading is |
| 945 successful, will be maintained. | |
| 946 | |
| 0 | 947 Special care is necessary if E is a back edge in the CFG as we |
| 948 may have already recorded equivalences for E->dest into our | |
| 949 various tables, including the result of the conditional at | |
| 950 the end of E->dest. Threading opportunities are severely | |
| 951 limited in that case to avoid short-circuiting the loop | |
| 952 incorrectly. | |
| 953 | |
| 954 DUMMY_COND is a shared cond_expr used by condition simplification as scratch, | |
| 955 to avoid allocating memory. | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
956 |
| 0 | 957 STACK is used to undo temporary equivalences created during the walk of |
| 958 E->dest. | |
| 959 | |
| 111 | 960 SIMPLIFY is a pass-specific function used to simplify statements. |
| 961 | |
| 962 Our caller is responsible for restoring the state of the expression | |
| 963 and const_and_copies stacks. | |
| 0 | 964 |
| 111 | 965 Positive return value is success. Zero return value is failure, but |
| 966 the block can still be duplicated as a joiner in a jump thread path, | |
| 967 negative indicates the block should not be duplicated and thus is not | |
| 968 suitable for a joiner in a jump threading path. */ | |
| 0 | 969 |
| 111 | 970 static int |
| 971 thread_through_normal_block (edge e, | |
| 972 gcond *dummy_cond, | |
| 973 const_and_copies *const_and_copies, | |
| 974 avail_exprs_stack *avail_exprs_stack, | |
| 975 pfn_simplify simplify, | |
| 976 vec<jump_thread_edge *> *path, | |
| 977 bitmap visited) | |
| 978 { | |
| 979 /* We want to record any equivalences created by traversing E. */ | |
| 980 record_temporary_equivalences (e, const_and_copies, avail_exprs_stack); | |
| 0 | 981 |
| 111 | 982 /* PHIs create temporary equivalences. |
| 983 Note that if we found a PHI that made the block non-threadable, then | |
| 984 we need to bubble that up to our caller in the same manner we do | |
| 985 when we prematurely stop processing statements below. */ | |
| 986 if (!record_temporary_equivalences_from_phis (e, const_and_copies)) | |
| 987 return -1; | |
| 0 | 988 |
| 989 /* Now walk each statement recording any context sensitive | |
| 990 temporary equivalences we can detect. */ | |
| 111 | 991 gimple *stmt |
| 992 = record_temporary_equivalences_from_stmts_at_dest (e, const_and_copies, | |
| 993 avail_exprs_stack, | |
| 994 simplify); | |
| 995 | |
| 996 /* There's two reasons STMT might be null, and distinguishing | |
| 997 between them is important. | |
| 998 | |
| 999 First the block may not have had any statements. For example, it | |
| 1000 might have some PHIs and unconditionally transfer control elsewhere. | |
| 1001 Such blocks are suitable for jump threading, particularly as a | |
| 1002 joiner block. | |
| 1003 | |
| 1004 The second reason would be if we did not process all the statements | |
| 1005 in the block (because there were too many to make duplicating the | |
| 1006 block profitable. If we did not look at all the statements, then | |
| 1007 we may not have invalidated everything needing invalidation. Thus | |
| 1008 we must signal to our caller that this block is not suitable for | |
| 1009 use as a joiner in a threading path. */ | |
| 0 | 1010 if (!stmt) |
| 111 | 1011 { |
| 1012 /* First case. The statement simply doesn't have any instructions, but | |
| 1013 does have PHIs. */ | |
| 1014 if (gsi_end_p (gsi_start_nondebug_bb (e->dest)) | |
| 1015 && !gsi_end_p (gsi_start_phis (e->dest))) | |
| 1016 return 0; | |
| 1017 | |
| 1018 /* Second case. */ | |
| 1019 return -1; | |
| 1020 } | |
| 0 | 1021 |
| 1022 /* If we stopped at a COND_EXPR or SWITCH_EXPR, see if we know which arm | |
| 1023 will be taken. */ | |
| 1024 if (gimple_code (stmt) == GIMPLE_COND | |
| 1025 || gimple_code (stmt) == GIMPLE_GOTO | |
| 1026 || gimple_code (stmt) == GIMPLE_SWITCH) | |
| 1027 { | |
| 1028 tree cond; | |
| 1029 | |
| 1030 /* Extract and simplify the condition. */ | |
| 111 | 1031 cond = simplify_control_stmt_condition (e, stmt, avail_exprs_stack, |
| 1032 dummy_cond, simplify); | |
| 1033 | |
| 1034 if (!cond) | |
| 1035 return 0; | |
| 0 | 1036 |
| 111 | 1037 if (is_gimple_min_invariant (cond) |
| 1038 || TREE_CODE (cond) == CASE_LABEL_EXPR) | |
| 0 | 1039 { |
| 111 | 1040 edge taken_edge; |
| 1041 if (TREE_CODE (cond) == CASE_LABEL_EXPR) | |
| 1042 taken_edge = find_edge (e->dest, | |
| 1043 label_to_block (CASE_LABEL (cond))); | |
| 1044 else | |
| 1045 taken_edge = find_taken_edge (e->dest, cond); | |
| 1046 | |
| 0 | 1047 basic_block dest = (taken_edge ? taken_edge->dest : NULL); |
| 1048 | |
| 111 | 1049 /* DEST could be NULL for a computed jump to an absolute |
| 1050 address. */ | |
| 1051 if (dest == NULL | |
| 1052 || dest == e->dest | |
| 1053 || (taken_edge->flags & EDGE_DFS_BACK) != 0 | |
| 1054 || bitmap_bit_p (visited, dest->index)) | |
| 1055 return 0; | |
| 1056 | |
| 1057 /* Only push the EDGE_START_JUMP_THREAD marker if this is | |
| 1058 first edge on the path. */ | |
| 1059 if (path->length () == 0) | |
| 1060 { | |
| 1061 jump_thread_edge *x | |
| 1062 = new jump_thread_edge (e, EDGE_START_JUMP_THREAD); | |
| 1063 path->safe_push (x); | |
| 1064 } | |
| 1065 | |
| 1066 jump_thread_edge *x | |
| 1067 = new jump_thread_edge (taken_edge, EDGE_COPY_SRC_BLOCK); | |
| 1068 path->safe_push (x); | |
| 1069 | |
| 1070 /* See if we can thread through DEST as well, this helps capture | |
| 1071 secondary effects of threading without having to re-run DOM or | |
| 1072 VRP. | |
| 1073 | |
| 1074 We don't want to thread back to a block we have already | |
| 1075 visited. This may be overly conservative. */ | |
| 1076 bitmap_set_bit (visited, dest->index); | |
| 1077 bitmap_set_bit (visited, e->dest->index); | |
| 1078 thread_around_empty_blocks (taken_edge, | |
| 1079 dummy_cond, | |
| 1080 avail_exprs_stack, | |
| 1081 simplify, | |
| 1082 visited, | |
| 1083 path); | |
| 1084 return 1; | |
| 1085 } | |
| 1086 } | |
| 1087 return 0; | |
| 1088 } | |
| 1089 | |
| 1090 /* We are exiting E->src, see if E->dest ends with a conditional | |
| 1091 jump which has a known value when reached via E. | |
| 1092 | |
| 1093 DUMMY_COND is a shared cond_expr used by condition simplification as scratch, | |
| 1094 to avoid allocating memory. | |
| 1095 | |
| 1096 CONST_AND_COPIES is used to undo temporary equivalences created during the | |
| 1097 walk of E->dest. | |
| 1098 | |
| 1099 The available expression table is referenced vai AVAIL_EXPRS_STACK. | |
| 1100 | |
| 1101 SIMPLIFY is a pass-specific function used to simplify statements. */ | |
| 0 | 1102 |
| 111 | 1103 static void |
| 1104 thread_across_edge (gcond *dummy_cond, | |
| 1105 edge e, | |
| 1106 class const_and_copies *const_and_copies, | |
| 1107 class avail_exprs_stack *avail_exprs_stack, | |
| 1108 pfn_simplify simplify) | |
| 1109 { | |
| 1110 bitmap visited = BITMAP_ALLOC (NULL); | |
| 1111 | |
| 1112 const_and_copies->push_marker (); | |
| 1113 avail_exprs_stack->push_marker (); | |
| 1114 | |
| 1115 stmt_count = 0; | |
| 1116 | |
| 1117 vec<jump_thread_edge *> *path = new vec<jump_thread_edge *> (); | |
| 1118 bitmap_clear (visited); | |
| 1119 bitmap_set_bit (visited, e->src->index); | |
| 1120 bitmap_set_bit (visited, e->dest->index); | |
| 1121 | |
| 1122 int threaded; | |
| 1123 if ((e->flags & EDGE_DFS_BACK) == 0) | |
| 1124 threaded = thread_through_normal_block (e, dummy_cond, | |
| 1125 const_and_copies, | |
| 1126 avail_exprs_stack, | |
| 1127 simplify, path, | |
| 1128 visited); | |
| 1129 else | |
| 1130 threaded = 0; | |
| 1131 | |
| 1132 if (threaded > 0) | |
| 1133 { | |
| 1134 propagate_threaded_block_debug_into (path->last ()->e->dest, | |
| 1135 e->dest); | |
| 1136 const_and_copies->pop_to_marker (); | |
| 1137 avail_exprs_stack->pop_to_marker (); | |
| 1138 BITMAP_FREE (visited); | |
| 1139 register_jump_thread (path); | |
| 1140 return; | |
| 1141 } | |
| 1142 else | |
| 1143 { | |
| 1144 /* Negative and zero return values indicate no threading was possible, | |
| 1145 thus there should be no edges on the thread path and no need to walk | |
| 1146 through the vector entries. */ | |
| 1147 gcc_assert (path->length () == 0); | |
| 1148 path->release (); | |
| 1149 delete path; | |
| 1150 | |
| 1151 /* A negative status indicates the target block was deemed too big to | |
| 1152 duplicate. Just quit now rather than trying to use the block as | |
| 1153 a joiner in a jump threading path. | |
| 1154 | |
| 1155 This prevents unnecessary code growth, but more importantly if we | |
| 1156 do not look at all the statements in the block, then we may have | |
| 1157 missed some invalidations if we had traversed a backedge! */ | |
| 1158 if (threaded < 0) | |
| 1159 { | |
| 1160 BITMAP_FREE (visited); | |
| 1161 const_and_copies->pop_to_marker (); | |
| 1162 avail_exprs_stack->pop_to_marker (); | |
| 1163 return; | |
| 0 | 1164 } |
| 1165 } | |
| 1166 | |
| 111 | 1167 /* We were unable to determine what out edge from E->dest is taken. However, |
| 1168 we might still be able to thread through successors of E->dest. This | |
| 1169 often occurs when E->dest is a joiner block which then fans back out | |
| 1170 based on redundant tests. | |
| 1171 | |
| 1172 If so, we'll copy E->dest and redirect the appropriate predecessor to | |
| 1173 the copy. Within the copy of E->dest, we'll thread one or more edges | |
| 1174 to points deeper in the CFG. | |
| 1175 | |
| 1176 This is a stopgap until we have a more structured approach to path | |
| 1177 isolation. */ | |
| 1178 { | |
| 1179 edge taken_edge; | |
| 1180 edge_iterator ei; | |
| 1181 bool found; | |
| 1182 | |
| 1183 /* If E->dest has abnormal outgoing edges, then there's no guarantee | |
| 1184 we can safely redirect any of the edges. Just punt those cases. */ | |
| 1185 FOR_EACH_EDGE (taken_edge, ei, e->dest->succs) | |
| 1186 if (taken_edge->flags & EDGE_ABNORMAL) | |
| 1187 { | |
| 1188 const_and_copies->pop_to_marker (); | |
| 1189 avail_exprs_stack->pop_to_marker (); | |
| 1190 BITMAP_FREE (visited); | |
| 1191 return; | |
| 1192 } | |
| 1193 | |
| 1194 /* Look at each successor of E->dest to see if we can thread through it. */ | |
| 1195 FOR_EACH_EDGE (taken_edge, ei, e->dest->succs) | |
| 1196 { | |
| 1197 if ((e->flags & EDGE_DFS_BACK) != 0 | |
| 1198 || (taken_edge->flags & EDGE_DFS_BACK) != 0) | |
| 1199 continue; | |
| 1200 | |
| 1201 /* Push a fresh marker so we can unwind the equivalences created | |
| 1202 for each of E->dest's successors. */ | |
| 1203 const_and_copies->push_marker (); | |
| 1204 avail_exprs_stack->push_marker (); | |
| 1205 | |
| 1206 /* Avoid threading to any block we have already visited. */ | |
| 1207 bitmap_clear (visited); | |
| 1208 bitmap_set_bit (visited, e->src->index); | |
| 1209 bitmap_set_bit (visited, e->dest->index); | |
| 1210 bitmap_set_bit (visited, taken_edge->dest->index); | |
| 1211 vec<jump_thread_edge *> *path = new vec<jump_thread_edge *> (); | |
| 1212 | |
| 1213 /* Record whether or not we were able to thread through a successor | |
| 1214 of E->dest. */ | |
| 1215 jump_thread_edge *x = new jump_thread_edge (e, EDGE_START_JUMP_THREAD); | |
| 1216 path->safe_push (x); | |
| 1217 | |
| 1218 x = new jump_thread_edge (taken_edge, EDGE_COPY_SRC_JOINER_BLOCK); | |
| 1219 path->safe_push (x); | |
| 1220 found = false; | |
| 1221 found = thread_around_empty_blocks (taken_edge, | |
| 1222 dummy_cond, | |
| 1223 avail_exprs_stack, | |
| 1224 simplify, | |
| 1225 visited, | |
| 1226 path); | |
| 1227 | |
| 1228 if (!found) | |
| 1229 found = thread_through_normal_block (path->last ()->e, dummy_cond, | |
| 1230 const_and_copies, | |
| 1231 avail_exprs_stack, | |
| 1232 simplify, path, | |
| 1233 visited) > 0; | |
| 1234 | |
| 1235 /* If we were able to thread through a successor of E->dest, then | |
| 1236 record the jump threading opportunity. */ | |
| 1237 if (found) | |
| 1238 { | |
| 1239 propagate_threaded_block_debug_into (path->last ()->e->dest, | |
| 1240 taken_edge->dest); | |
| 1241 register_jump_thread (path); | |
| 1242 } | |
| 1243 else | |
| 1244 delete_jump_thread_path (path); | |
| 1245 | |
| 1246 /* And unwind the equivalence table. */ | |
| 1247 avail_exprs_stack->pop_to_marker (); | |
| 1248 const_and_copies->pop_to_marker (); | |
| 1249 } | |
| 1250 BITMAP_FREE (visited); | |
| 1251 } | |
| 1252 | |
| 1253 const_and_copies->pop_to_marker (); | |
| 1254 avail_exprs_stack->pop_to_marker (); | |
| 0 | 1255 } |
| 111 | 1256 |
| 1257 /* Examine the outgoing edges from BB and conditionally | |
| 1258 try to thread them. | |
| 1259 | |
| 1260 DUMMY_COND is a shared cond_expr used by condition simplification as scratch, | |
| 1261 to avoid allocating memory. | |
| 1262 | |
| 1263 CONST_AND_COPIES is used to undo temporary equivalences created during the | |
| 1264 walk of E->dest. | |
| 1265 | |
| 1266 The available expression table is referenced vai AVAIL_EXPRS_STACK. | |
| 1267 | |
| 1268 SIMPLIFY is a pass-specific function used to simplify statements. */ | |
| 1269 | |
| 1270 void | |
| 1271 thread_outgoing_edges (basic_block bb, gcond *dummy_cond, | |
| 1272 class const_and_copies *const_and_copies, | |
| 1273 class avail_exprs_stack *avail_exprs_stack, | |
| 1274 tree (*simplify) (gimple *, gimple *, | |
| 1275 class avail_exprs_stack *, | |
| 1276 basic_block)) | |
| 1277 { | |
| 1278 int flags = (EDGE_IGNORE | EDGE_COMPLEX | EDGE_ABNORMAL); | |
| 1279 gimple *last; | |
| 1280 | |
| 1281 /* If we have an outgoing edge to a block with multiple incoming and | |
| 1282 outgoing edges, then we may be able to thread the edge, i.e., we | |
| 1283 may be able to statically determine which of the outgoing edges | |
| 1284 will be traversed when the incoming edge from BB is traversed. */ | |
| 1285 if (single_succ_p (bb) | |
| 1286 && (single_succ_edge (bb)->flags & flags) == 0 | |
| 1287 && potentially_threadable_block (single_succ (bb))) | |
| 1288 { | |
| 1289 thread_across_edge (dummy_cond, single_succ_edge (bb), | |
| 1290 const_and_copies, avail_exprs_stack, | |
| 1291 simplify); | |
| 1292 } | |
| 1293 else if ((last = last_stmt (bb)) | |
| 1294 && gimple_code (last) == GIMPLE_COND | |
| 1295 && EDGE_COUNT (bb->succs) == 2 | |
| 1296 && (EDGE_SUCC (bb, 0)->flags & flags) == 0 | |
| 1297 && (EDGE_SUCC (bb, 1)->flags & flags) == 0) | |
| 1298 { | |
| 1299 edge true_edge, false_edge; | |
| 1300 | |
| 1301 extract_true_false_edges_from_block (bb, &true_edge, &false_edge); | |
| 1302 | |
| 1303 /* Only try to thread the edge if it reaches a target block with | |
| 1304 more than one predecessor and more than one successor. */ | |
| 1305 if (potentially_threadable_block (true_edge->dest)) | |
| 1306 thread_across_edge (dummy_cond, true_edge, | |
| 1307 const_and_copies, avail_exprs_stack, simplify); | |
| 1308 | |
| 1309 /* Similarly for the ELSE arm. */ | |
| 1310 if (potentially_threadable_block (false_edge->dest)) | |
| 1311 thread_across_edge (dummy_cond, false_edge, | |
| 1312 const_and_copies, avail_exprs_stack, simplify); | |
| 1313 } | |
| 1314 } |