1 /* Loop invariant motion.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
27 #include "hard-reg-set.h"
28 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
37 #include "tree-pass.h"
41 #include "tree-affine.h"
42 #include "pointer-set.h"
44 /* TODO: Support for predicated code motion. I.e.
55 Where COND and INV are is invariants, but evaluating INV may trap or be
56 invalid from some other reason if !COND. This may be transformed to
66 /* A type for the list of statements that have to be moved in order to be able
67 to hoist an invariant computation. */
75 /* The auxiliary data kept for each statement. */
79 struct loop *max_loop; /* The outermost loop in that the statement
82 struct loop *tgt_loop; /* The loop out of that we want to move the
85 struct loop *always_executed_in;
86 /* The outermost loop for that we are sure
87 the statement is executed if the loop
90 unsigned cost; /* Cost of the computation performed by the
93 struct depend *depends; /* List of statements that must be also hoisted
94 out of the loop when this statement is
95 hoisted; i.e. those that define the operands
96 of the statement and are inside of the
100 #define LIM_DATA(STMT) (TREE_CODE (STMT) == PHI_NODE \
102 : (struct lim_aux_data *) (stmt_ann (STMT)->common.aux))
104 /* Description of a memory reference location. */
106 typedef struct mem_ref_loc
108 tree *ref; /* The reference itself. */
109 tree stmt; /* The statement in that it occurs. */
112 DEF_VEC_P(mem_ref_loc_p);
113 DEF_VEC_ALLOC_P(mem_ref_loc_p, heap);
115 /* The list of memory reference locations in a loop. */
117 typedef struct mem_ref_locs
119 VEC (mem_ref_loc_p, heap) *locs;
122 DEF_VEC_P(mem_ref_locs_p);
123 DEF_VEC_ALLOC_P(mem_ref_locs_p, heap);
125 /* Description of a memory reference. */
127 typedef struct mem_ref
129 tree mem; /* The memory itself. */
130 unsigned id; /* ID assigned to the memory reference
131 (its index in memory_accesses.refs_list) */
132 hashval_t hash; /* Its hash value. */
133 bitmap stored; /* The set of loops in that this memory locatio
135 VEC (mem_ref_locs_p, heap) *accesses_in_loop;
136 /* The locations of the accesses. Vector
137 indexed by the loop number. */
138 bitmap vops; /* Vops corresponding to this memory
141 /* The following sets are computed on demand. We keep both set and
142 its complement, so that we know whether the information was
143 already computed or not. */
144 bitmap indep_loop; /* The set of loops in that the memory
145 reference is independent, meaning:
146 If it is stored in the loop, this store
147 is independent on all other loads and
149 If it is only loaded, then it is independent
150 on all stores in the loop. */
151 bitmap dep_loop; /* The complement of INDEP_LOOP. */
153 bitmap indep_ref; /* The set of memory references on that
154 this reference is independent. */
155 bitmap dep_ref; /* The complement of DEP_REF. */
158 DEF_VEC_P(mem_ref_p);
159 DEF_VEC_ALLOC_P(mem_ref_p, heap);
162 DEF_VEC_ALLOC_P(bitmap, heap);
165 DEF_VEC_ALLOC_P(htab_t, heap);
167 /* Description of memory accesses in loops. */
171 /* The hash table of memory references accessed in loops. */
174 /* The list of memory references. */
175 VEC (mem_ref_p, heap) *refs_list;
177 /* The set of memory references accessed in each loop. */
178 VEC (bitmap, heap) *refs_in_loop;
180 /* The set of memory references accessed in each loop, including
182 VEC (bitmap, heap) *all_refs_in_loop;
184 /* The set of virtual operands clobbered in a given loop. */
185 VEC (bitmap, heap) *clobbered_vops;
187 /* Map from the pair (loop, virtual operand) to the set of refs that
188 touch the virtual operand in the loop. */
189 VEC (htab_t, heap) *vop_ref_map;
191 /* Cache for expanding memory addresses. */
192 struct pointer_map_t *ttae_cache;
195 static bool ref_indep_loop_p (struct loop *, mem_ref_p);
197 /* Minimum cost of an expensive expression. */
198 #define LIM_EXPENSIVE ((unsigned) PARAM_VALUE (PARAM_LIM_EXPENSIVE))
200 /* The outermost loop for that execution of the header guarantees that the
201 block will be executed. */
202 #define ALWAYS_EXECUTED_IN(BB) ((struct loop *) (BB)->aux)
204 /* Calls CBCK for each index in memory reference ADDR_P. There are two
205 kinds situations handled; in each of these cases, the memory reference
206 and DATA are passed to the callback:
208 Access to an array: ARRAY_{RANGE_}REF (base, index). In this case we also
209 pass the pointer to the index to the callback.
211 Pointer dereference: INDIRECT_REF (addr). In this case we also pass the
212 pointer to addr to the callback.
214 If the callback returns false, the whole search stops and false is returned.
215 Otherwise the function returns true after traversing through the whole
216 reference *ADDR_P. */
219 for_each_index (tree *addr_p, bool (*cbck) (tree, tree *, void *), void *data)
223 for (; ; addr_p = nxt)
225 switch (TREE_CODE (*addr_p))
228 return cbck (*addr_p, addr_p, data);
230 case MISALIGNED_INDIRECT_REF:
231 case ALIGN_INDIRECT_REF:
233 nxt = &TREE_OPERAND (*addr_p, 0);
234 return cbck (*addr_p, nxt, data);
237 case VIEW_CONVERT_EXPR:
240 nxt = &TREE_OPERAND (*addr_p, 0);
244 /* If the component has varying offset, it behaves like index
246 idx = &TREE_OPERAND (*addr_p, 2);
248 && !cbck (*addr_p, idx, data))
251 nxt = &TREE_OPERAND (*addr_p, 0);
255 case ARRAY_RANGE_REF:
256 nxt = &TREE_OPERAND (*addr_p, 0);
257 if (!cbck (*addr_p, &TREE_OPERAND (*addr_p, 1), data))
274 gcc_assert (is_gimple_min_invariant (*addr_p));
278 idx = &TMR_BASE (*addr_p);
280 && !cbck (*addr_p, idx, data))
282 idx = &TMR_INDEX (*addr_p);
284 && !cbck (*addr_p, idx, data))
294 /* If it is possible to hoist the statement STMT unconditionally,
295 returns MOVE_POSSIBLE.
296 If it is possible to hoist the statement STMT, but we must avoid making
297 it executed if it would not be executed in the original program (e.g.
298 because it may trap), return MOVE_PRESERVE_EXECUTION.
299 Otherwise return MOVE_IMPOSSIBLE. */
302 movement_possibility (tree stmt)
306 if (flag_unswitch_loops
307 && TREE_CODE (stmt) == COND_EXPR)
309 /* If we perform unswitching, force the operands of the invariant
310 condition to be moved out of the loop. */
311 return MOVE_POSSIBLE;
314 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
315 return MOVE_IMPOSSIBLE;
317 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_DEFS))
318 return MOVE_IMPOSSIBLE;
320 if (stmt_ends_bb_p (stmt))
321 return MOVE_IMPOSSIBLE;
323 if (stmt_ann (stmt)->has_volatile_ops)
324 return MOVE_IMPOSSIBLE;
326 lhs = GIMPLE_STMT_OPERAND (stmt, 0);
327 if (TREE_CODE (lhs) == SSA_NAME
328 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
329 return MOVE_IMPOSSIBLE;
331 rhs = GIMPLE_STMT_OPERAND (stmt, 1);
333 if (TREE_SIDE_EFFECTS (rhs)
334 || tree_could_throw_p (rhs))
335 return MOVE_IMPOSSIBLE;
337 if (TREE_CODE (lhs) != SSA_NAME
338 || tree_could_trap_p (rhs))
339 return MOVE_PRESERVE_EXECUTION;
341 if (get_call_expr_in (stmt))
343 /* While pure or const call is guaranteed to have no side effects, we
344 cannot move it arbitrarily. Consider code like
346 char *s = something ();
356 Here the strlen call cannot be moved out of the loop, even though
357 s is invariant. In addition to possibly creating a call with
358 invalid arguments, moving out a function call that is not executed
359 may cause performance regressions in case the call is costly and
360 not executed at all. */
361 return MOVE_PRESERVE_EXECUTION;
363 return MOVE_POSSIBLE;
366 /* Suppose that operand DEF is used inside the LOOP. Returns the outermost
367 loop to that we could move the expression using DEF if it did not have
368 other operands, i.e. the outermost loop enclosing LOOP in that the value
369 of DEF is invariant. */
372 outermost_invariant_loop (tree def, struct loop *loop)
376 struct loop *max_loop;
378 if (TREE_CODE (def) != SSA_NAME)
379 return superloop_at_depth (loop, 1);
381 def_stmt = SSA_NAME_DEF_STMT (def);
382 def_bb = bb_for_stmt (def_stmt);
384 return superloop_at_depth (loop, 1);
386 max_loop = find_common_loop (loop, def_bb->loop_father);
388 if (LIM_DATA (def_stmt) && LIM_DATA (def_stmt)->max_loop)
389 max_loop = find_common_loop (max_loop,
390 loop_outer (LIM_DATA (def_stmt)->max_loop));
391 if (max_loop == loop)
393 max_loop = superloop_at_depth (loop, loop_depth (max_loop) + 1);
398 /* Returns the outermost superloop of LOOP in that the expression EXPR is
402 outermost_invariant_loop_expr (tree expr, struct loop *loop)
404 enum tree_code_class codeclass = TREE_CODE_CLASS (TREE_CODE (expr));
406 struct loop *max_loop = superloop_at_depth (loop, 1), *aloop;
408 if (TREE_CODE (expr) == SSA_NAME
409 || TREE_CODE (expr) == INTEGER_CST
410 || is_gimple_min_invariant (expr))
411 return outermost_invariant_loop (expr, loop);
413 if (codeclass != tcc_unary
414 && codeclass != tcc_binary
415 && codeclass != tcc_expression
416 && codeclass != tcc_vl_exp
417 && codeclass != tcc_comparison)
420 nops = TREE_OPERAND_LENGTH (expr);
421 for (i = 0; i < nops; i++)
423 aloop = outermost_invariant_loop_expr (TREE_OPERAND (expr, i), loop);
427 if (flow_loop_nested_p (max_loop, aloop))
434 /* DATA is a structure containing information associated with a statement
435 inside LOOP. DEF is one of the operands of this statement.
437 Find the outermost loop enclosing LOOP in that value of DEF is invariant
438 and record this in DATA->max_loop field. If DEF itself is defined inside
439 this loop as well (i.e. we need to hoist it out of the loop if we want
440 to hoist the statement represented by DATA), record the statement in that
441 DEF is defined to the DATA->depends list. Additionally if ADD_COST is true,
442 add the cost of the computation of DEF to the DATA->cost.
444 If DEF is not invariant in LOOP, return false. Otherwise return TRUE. */
447 add_dependency (tree def, struct lim_aux_data *data, struct loop *loop,
450 tree def_stmt = SSA_NAME_DEF_STMT (def);
451 basic_block def_bb = bb_for_stmt (def_stmt);
452 struct loop *max_loop;
458 max_loop = outermost_invariant_loop (def, loop);
462 if (flow_loop_nested_p (data->max_loop, max_loop))
463 data->max_loop = max_loop;
465 if (!LIM_DATA (def_stmt))
469 /* Only add the cost if the statement defining DEF is inside LOOP,
470 i.e. if it is likely that by moving the invariants dependent
471 on it, we will be able to avoid creating a new register for
472 it (since it will be only used in these dependent invariants). */
473 && def_bb->loop_father == loop)
474 data->cost += LIM_DATA (def_stmt)->cost;
476 dep = XNEW (struct depend);
477 dep->stmt = def_stmt;
478 dep->next = data->depends;
484 /* Returns an estimate for a cost of statement STMT. TODO -- the values here
485 are just ad-hoc constants. The estimates should be based on target-specific
489 stmt_cost (tree stmt)
494 /* Always try to create possibilities for unswitching. */
495 if (TREE_CODE (stmt) == COND_EXPR)
496 return LIM_EXPENSIVE;
498 rhs = GENERIC_TREE_OPERAND (stmt, 1);
500 /* Hoisting memory references out should almost surely be a win. */
501 if (stmt_references_memory_p (stmt))
504 switch (TREE_CODE (rhs))
507 /* We should be hoisting calls if possible. */
509 /* Unless the call is a builtin_constant_p; this always folds to a
510 constant, so moving it is useless. */
511 rhs = get_callee_fndecl (rhs);
512 if (DECL_BUILT_IN_CLASS (rhs) == BUILT_IN_NORMAL
513 && DECL_FUNCTION_CODE (rhs) == BUILT_IN_CONSTANT_P)
530 /* Division and multiplication are usually expensive. */
546 /* Finds the outermost loop between OUTER and LOOP in that the memory reference
547 REF is independent. If REF is not independent in LOOP, NULL is returned
551 outermost_indep_loop (struct loop *outer, struct loop *loop, mem_ref_p ref)
555 if (bitmap_bit_p (ref->stored, loop->num))
560 aloop = superloop_at_depth (loop, loop_depth (aloop) + 1))
561 if (!bitmap_bit_p (ref->stored, aloop->num)
562 && ref_indep_loop_p (aloop, ref))
565 if (ref_indep_loop_p (loop, ref))
571 /* If there is a simple load or store to a memory reference in STMT, returns
572 the location of the memory reference, and sets IS_STORE accoring to whether
573 it is a store or load. Otherwise, returns NULL. */
576 simple_mem_ref_in_stmt (tree stmt, bool *is_store)
580 /* Recognize MEM = (SSA_NAME | invariant) and SSA_NAME = MEM patterns. */
581 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
584 lhs = &GIMPLE_STMT_OPERAND (stmt, 0);
585 rhs = &GIMPLE_STMT_OPERAND (stmt, 1);
587 if (TREE_CODE (*lhs) == SSA_NAME)
589 if (!is_gimple_addressable (*rhs))
595 else if (TREE_CODE (*rhs) == SSA_NAME
596 || is_gimple_min_invariant (*rhs))
605 /* Returns the memory reference contained in STMT. */
608 mem_ref_in_stmt (tree stmt)
611 tree *mem = simple_mem_ref_in_stmt (stmt, &store);
619 hash = iterative_hash_expr (*mem, 0);
620 ref = htab_find_with_hash (memory_accesses.refs, *mem, hash);
622 gcc_assert (ref != NULL);
626 /* Determine the outermost loop to that it is possible to hoist a statement
627 STMT and store it to LIM_DATA (STMT)->max_loop. To do this we determine
628 the outermost loop in that the value computed by STMT is invariant.
629 If MUST_PRESERVE_EXEC is true, additionally choose such a loop that
630 we preserve the fact whether STMT is executed. It also fills other related
631 information to LIM_DATA (STMT).
633 The function returns false if STMT cannot be hoisted outside of the loop it
634 is defined in, and true otherwise. */
637 determine_max_movement (tree stmt, bool must_preserve_exec)
639 basic_block bb = bb_for_stmt (stmt);
640 struct loop *loop = bb->loop_father;
642 struct lim_aux_data *lim_data = LIM_DATA (stmt);
646 if (must_preserve_exec)
647 level = ALWAYS_EXECUTED_IN (bb);
649 level = superloop_at_depth (loop, 1);
650 lim_data->max_loop = level;
652 FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_USE)
653 if (!add_dependency (val, lim_data, loop, true))
656 if (!ZERO_SSA_OPERANDS (stmt, SSA_OP_VIRTUAL_USES))
658 mem_ref_p ref = mem_ref_in_stmt (stmt);
663 = outermost_indep_loop (lim_data->max_loop, loop, ref);
664 if (!lim_data->max_loop)
669 FOR_EACH_SSA_TREE_OPERAND (val, stmt, iter, SSA_OP_VIRTUAL_USES)
671 if (!add_dependency (val, lim_data, loop, false))
677 lim_data->cost += stmt_cost (stmt);
682 /* Suppose that some statement in ORIG_LOOP is hoisted to the loop LEVEL,
683 and that one of the operands of this statement is computed by STMT.
684 Ensure that STMT (together with all the statements that define its
685 operands) is hoisted at least out of the loop LEVEL. */
688 set_level (tree stmt, struct loop *orig_loop, struct loop *level)
690 struct loop *stmt_loop = bb_for_stmt (stmt)->loop_father;
693 stmt_loop = find_common_loop (orig_loop, stmt_loop);
694 if (LIM_DATA (stmt) && LIM_DATA (stmt)->tgt_loop)
695 stmt_loop = find_common_loop (stmt_loop,
696 loop_outer (LIM_DATA (stmt)->tgt_loop));
697 if (flow_loop_nested_p (stmt_loop, level))
700 gcc_assert (LIM_DATA (stmt));
701 gcc_assert (level == LIM_DATA (stmt)->max_loop
702 || flow_loop_nested_p (LIM_DATA (stmt)->max_loop, level));
704 LIM_DATA (stmt)->tgt_loop = level;
705 for (dep = LIM_DATA (stmt)->depends; dep; dep = dep->next)
706 set_level (dep->stmt, orig_loop, level);
709 /* Determines an outermost loop from that we want to hoist the statement STMT.
710 For now we chose the outermost possible loop. TODO -- use profiling
711 information to set it more sanely. */
714 set_profitable_level (tree stmt)
716 set_level (stmt, bb_for_stmt (stmt)->loop_father, LIM_DATA (stmt)->max_loop);
719 /* Returns true if STMT is not a pure call. */
722 nonpure_call_p (tree stmt)
724 tree call = get_call_expr_in (stmt);
729 return TREE_SIDE_EFFECTS (call) != 0;
732 /* Releases the memory occupied by DATA. */
735 free_lim_aux_data (struct lim_aux_data *data)
737 struct depend *dep, *next;
739 for (dep = data->depends; dep; dep = next)
747 /* Rewrite a/b to a*(1/b). Return the invariant stmt to process. */
750 rewrite_reciprocal (block_stmt_iterator *bsi)
752 tree stmt, lhs, rhs, stmt1, stmt2, var, name, tmp;
754 stmt = bsi_stmt (*bsi);
755 lhs = GENERIC_TREE_OPERAND (stmt, 0);
756 rhs = GENERIC_TREE_OPERAND (stmt, 1);
758 /* stmt must be GIMPLE_MODIFY_STMT. */
759 var = create_tmp_var (TREE_TYPE (rhs), "reciptmp");
760 add_referenced_var (var);
762 tmp = build2 (RDIV_EXPR, TREE_TYPE (rhs),
763 build_real (TREE_TYPE (rhs), dconst1),
764 TREE_OPERAND (rhs, 1));
765 stmt1 = build_gimple_modify_stmt (var, tmp);
766 name = make_ssa_name (var, stmt1);
767 GIMPLE_STMT_OPERAND (stmt1, 0) = name;
768 tmp = build2 (MULT_EXPR, TREE_TYPE (rhs),
769 name, TREE_OPERAND (rhs, 0));
770 stmt2 = build_gimple_modify_stmt (lhs, tmp);
772 /* Replace division stmt with reciprocal and multiply stmts.
773 The multiply stmt is not invariant, so update iterator
774 and avoid rescanning. */
775 bsi_replace (bsi, stmt1, true);
776 bsi_insert_after (bsi, stmt2, BSI_NEW_STMT);
777 SSA_NAME_DEF_STMT (lhs) = stmt2;
779 /* Continue processing with invariant reciprocal statement. */
783 /* Check if the pattern at *BSI is a bittest of the form
784 (A >> B) & 1 != 0 and in this case rewrite it to A & (1 << B) != 0. */
787 rewrite_bittest (block_stmt_iterator *bsi)
789 tree stmt, lhs, rhs, var, name, use_stmt, stmt1, stmt2, t;
792 stmt = bsi_stmt (*bsi);
793 lhs = GENERIC_TREE_OPERAND (stmt, 0);
794 rhs = GENERIC_TREE_OPERAND (stmt, 1);
796 /* Verify that the single use of lhs is a comparison against zero. */
797 if (TREE_CODE (lhs) != SSA_NAME
798 || !single_imm_use (lhs, &use, &use_stmt)
799 || TREE_CODE (use_stmt) != COND_EXPR)
801 t = COND_EXPR_COND (use_stmt);
802 if (TREE_OPERAND (t, 0) != lhs
803 || (TREE_CODE (t) != NE_EXPR
804 && TREE_CODE (t) != EQ_EXPR)
805 || !integer_zerop (TREE_OPERAND (t, 1)))
808 /* Get at the operands of the shift. The rhs is TMP1 & 1. */
809 stmt1 = SSA_NAME_DEF_STMT (TREE_OPERAND (rhs, 0));
810 if (TREE_CODE (stmt1) != GIMPLE_MODIFY_STMT)
813 /* There is a conversion in between possibly inserted by fold. */
814 t = GIMPLE_STMT_OPERAND (stmt1, 1);
815 if (CONVERT_EXPR_P (t))
817 t = TREE_OPERAND (t, 0);
818 if (TREE_CODE (t) != SSA_NAME
819 || !has_single_use (t))
821 stmt1 = SSA_NAME_DEF_STMT (t);
822 if (TREE_CODE (stmt1) != GIMPLE_MODIFY_STMT)
824 t = GIMPLE_STMT_OPERAND (stmt1, 1);
827 /* Verify that B is loop invariant but A is not. Verify that with
828 all the stmt walking we are still in the same loop. */
829 if (TREE_CODE (t) == RSHIFT_EXPR
830 && loop_containing_stmt (stmt1) == loop_containing_stmt (stmt)
831 && outermost_invariant_loop_expr (TREE_OPERAND (t, 1),
832 loop_containing_stmt (stmt1)) != NULL
833 && outermost_invariant_loop_expr (TREE_OPERAND (t, 0),
834 loop_containing_stmt (stmt1)) == NULL)
836 tree a = TREE_OPERAND (t, 0);
837 tree b = TREE_OPERAND (t, 1);
840 var = create_tmp_var (TREE_TYPE (a), "shifttmp");
841 add_referenced_var (var);
842 t = fold_build2 (LSHIFT_EXPR, TREE_TYPE (a),
843 build_int_cst (TREE_TYPE (a), 1), b);
844 stmt1 = build_gimple_modify_stmt (var, t);
845 name = make_ssa_name (var, stmt1);
846 GIMPLE_STMT_OPERAND (stmt1, 0) = name;
849 t = fold_build2 (BIT_AND_EXPR, TREE_TYPE (a), a, name);
850 stmt2 = build_gimple_modify_stmt (var, t);
851 name = make_ssa_name (var, stmt2);
852 GIMPLE_STMT_OPERAND (stmt2, 0) = name;
854 /* Replace the SSA_NAME we compare against zero. Adjust
855 the type of zero accordingly. */
857 TREE_OPERAND (COND_EXPR_COND (use_stmt), 1)
858 = build_int_cst_type (TREE_TYPE (name), 0);
860 bsi_insert_before (bsi, stmt1, BSI_SAME_STMT);
861 bsi_replace (bsi, stmt2, true);
870 /* Determine the outermost loops in that statements in basic block BB are
871 invariant, and record them to the LIM_DATA associated with the statements.
872 Callback for walk_dominator_tree. */
875 determine_invariantness_stmt (struct dom_walk_data *dw_data ATTRIBUTE_UNUSED,
879 block_stmt_iterator bsi;
881 bool maybe_never = ALWAYS_EXECUTED_IN (bb) == NULL;
882 struct loop *outermost = ALWAYS_EXECUTED_IN (bb);
884 if (!loop_outer (bb->loop_father))
887 if (dump_file && (dump_flags & TDF_DETAILS))
888 fprintf (dump_file, "Basic block %d (loop %d -- depth %d):\n\n",
889 bb->index, bb->loop_father->num, loop_depth (bb->loop_father));
891 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
893 stmt = bsi_stmt (bsi);
895 pos = movement_possibility (stmt);
896 if (pos == MOVE_IMPOSSIBLE)
898 if (nonpure_call_p (stmt))
906 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
908 rhs = GIMPLE_STMT_OPERAND (stmt, 1);
910 /* If divisor is invariant, convert a/b to a*(1/b), allowing reciprocal
911 to be hoisted out of loop, saving expensive divide. */
912 if (pos == MOVE_POSSIBLE
913 && TREE_CODE (rhs) == RDIV_EXPR
914 && flag_unsafe_math_optimizations
915 && !flag_trapping_math
916 && outermost_invariant_loop_expr (TREE_OPERAND (rhs, 1),
917 loop_containing_stmt (stmt)) != NULL
918 && outermost_invariant_loop_expr (rhs,
919 loop_containing_stmt (stmt)) == NULL)
920 stmt = rewrite_reciprocal (&bsi);
922 /* If the shift count is invariant, convert (A >> B) & 1 to
923 A & (1 << B) allowing the bit mask to be hoisted out of the loop
924 saving an expensive shift. */
925 if (pos == MOVE_POSSIBLE
926 && TREE_CODE (rhs) == BIT_AND_EXPR
927 && integer_onep (TREE_OPERAND (rhs, 1))
928 && TREE_CODE (TREE_OPERAND (rhs, 0)) == SSA_NAME
929 && has_single_use (TREE_OPERAND (rhs, 0)))
930 stmt = rewrite_bittest (&bsi);
933 stmt_ann (stmt)->common.aux = xcalloc (1, sizeof (struct lim_aux_data));
934 LIM_DATA (stmt)->always_executed_in = outermost;
936 if (maybe_never && pos == MOVE_PRESERVE_EXECUTION)
939 if (!determine_max_movement (stmt, pos == MOVE_PRESERVE_EXECUTION))
941 LIM_DATA (stmt)->max_loop = NULL;
945 if (dump_file && (dump_flags & TDF_DETAILS))
947 print_generic_stmt_indented (dump_file, stmt, 0, 2);
948 fprintf (dump_file, " invariant up to level %d, cost %d.\n\n",
949 loop_depth (LIM_DATA (stmt)->max_loop),
950 LIM_DATA (stmt)->cost);
953 if (LIM_DATA (stmt)->cost >= LIM_EXPENSIVE)
954 set_profitable_level (stmt);
958 /* For each statement determines the outermost loop in that it is invariant,
959 statements on whose motion it depends and the cost of the computation.
960 This information is stored to the LIM_DATA structure associated with
964 determine_invariantness (void)
966 struct dom_walk_data walk_data;
968 memset (&walk_data, 0, sizeof (struct dom_walk_data));
969 walk_data.dom_direction = CDI_DOMINATORS;
970 walk_data.before_dom_children_before_stmts = determine_invariantness_stmt;
972 init_walk_dominator_tree (&walk_data);
973 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
974 fini_walk_dominator_tree (&walk_data);
977 /* Hoist the statements in basic block BB out of the loops prescribed by
978 data stored in LIM_DATA structures associated with each statement. Callback
979 for walk_dominator_tree. */
982 move_computations_stmt (struct dom_walk_data *dw_data ATTRIBUTE_UNUSED,
986 block_stmt_iterator bsi;
990 if (!loop_outer (bb->loop_father))
993 for (bsi = bsi_start (bb); !bsi_end_p (bsi); )
995 stmt = bsi_stmt (bsi);
997 if (!LIM_DATA (stmt))
1003 cost = LIM_DATA (stmt)->cost;
1004 level = LIM_DATA (stmt)->tgt_loop;
1005 free_lim_aux_data (LIM_DATA (stmt));
1006 stmt_ann (stmt)->common.aux = NULL;
1014 /* We do not really want to move conditionals out of the loop; we just
1015 placed it here to force its operands to be moved if necessary. */
1016 if (TREE_CODE (stmt) == COND_EXPR)
1019 if (dump_file && (dump_flags & TDF_DETAILS))
1021 fprintf (dump_file, "Moving statement\n");
1022 print_generic_stmt (dump_file, stmt, 0);
1023 fprintf (dump_file, "(cost %u) out of loop %d.\n\n",
1027 mark_virtual_ops_for_renaming (stmt);
1028 bsi_insert_on_edge (loop_preheader_edge (level), stmt);
1029 bsi_remove (&bsi, false);
1033 /* Hoist the statements out of the loops prescribed by data stored in
1034 LIM_DATA structures associated with each statement.*/
1037 move_computations (void)
1039 struct dom_walk_data walk_data;
1041 memset (&walk_data, 0, sizeof (struct dom_walk_data));
1042 walk_data.dom_direction = CDI_DOMINATORS;
1043 walk_data.before_dom_children_before_stmts = move_computations_stmt;
1045 init_walk_dominator_tree (&walk_data);
1046 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
1047 fini_walk_dominator_tree (&walk_data);
1049 bsi_commit_edge_inserts ();
1050 if (need_ssa_update_p ())
1051 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1054 /* Checks whether the statement defining variable *INDEX can be hoisted
1055 out of the loop passed in DATA. Callback for for_each_index. */
1058 may_move_till (tree ref, tree *index, void *data)
1060 struct loop *loop = (struct loop*) data, *max_loop;
1062 /* If REF is an array reference, check also that the step and the lower
1063 bound is invariant in LOOP. */
1064 if (TREE_CODE (ref) == ARRAY_REF)
1066 tree step = array_ref_element_size (ref);
1067 tree lbound = array_ref_low_bound (ref);
1069 max_loop = outermost_invariant_loop_expr (step, loop);
1073 max_loop = outermost_invariant_loop_expr (lbound, loop);
1078 max_loop = outermost_invariant_loop (*index, loop);
1085 /* Forces statements defining (invariant) SSA names in expression EXPR to be
1086 moved out of the LOOP. ORIG_LOOP is the loop in that EXPR is used. */
1089 force_move_till_expr (tree expr, struct loop *orig_loop, struct loop *loop)
1091 enum tree_code_class codeclass = TREE_CODE_CLASS (TREE_CODE (expr));
1094 if (TREE_CODE (expr) == SSA_NAME)
1096 tree stmt = SSA_NAME_DEF_STMT (expr);
1097 if (IS_EMPTY_STMT (stmt))
1100 set_level (stmt, orig_loop, loop);
1104 if (codeclass != tcc_unary
1105 && codeclass != tcc_binary
1106 && codeclass != tcc_expression
1107 && codeclass != tcc_vl_exp
1108 && codeclass != tcc_comparison)
1111 nops = TREE_OPERAND_LENGTH (expr);
1112 for (i = 0; i < nops; i++)
1113 force_move_till_expr (TREE_OPERAND (expr, i), orig_loop, loop);
1116 /* Forces statement defining invariants in REF (and *INDEX) to be moved out of
1117 the LOOP. The reference REF is used in the loop ORIG_LOOP. Callback for
1123 struct loop *orig_loop;
1127 force_move_till (tree ref, tree *index, void *data)
1130 struct fmt_data *fmt_data = (struct fmt_data *) data;
1132 if (TREE_CODE (ref) == ARRAY_REF)
1134 tree step = array_ref_element_size (ref);
1135 tree lbound = array_ref_low_bound (ref);
1137 force_move_till_expr (step, fmt_data->orig_loop, fmt_data->loop);
1138 force_move_till_expr (lbound, fmt_data->orig_loop, fmt_data->loop);
1141 if (TREE_CODE (*index) != SSA_NAME)
1144 stmt = SSA_NAME_DEF_STMT (*index);
1145 if (IS_EMPTY_STMT (stmt))
1148 set_level (stmt, fmt_data->orig_loop, fmt_data->loop);
1153 /* A hash function for struct mem_ref object OBJ. */
1156 memref_hash (const void *obj)
1158 const struct mem_ref *mem = obj;
1163 /* An equality function for struct mem_ref object OBJ1 with
1164 memory reference OBJ2. */
1167 memref_eq (const void *obj1, const void *obj2)
1169 const struct mem_ref *mem1 = obj1;
1171 return operand_equal_p (mem1->mem, (tree) obj2, 0);
1174 /* Releases list of memory reference locations ACCS. */
1177 free_mem_ref_locs (mem_ref_locs_p accs)
1185 for (i = 0; VEC_iterate (mem_ref_loc_p, accs->locs, i, loc); i++)
1187 VEC_free (mem_ref_loc_p, heap, accs->locs);
1191 /* A function to free the mem_ref object OBJ. */
1194 memref_free (void *obj)
1196 struct mem_ref *mem = obj;
1198 mem_ref_locs_p accs;
1200 BITMAP_FREE (mem->stored);
1201 BITMAP_FREE (mem->indep_loop);
1202 BITMAP_FREE (mem->dep_loop);
1203 BITMAP_FREE (mem->indep_ref);
1204 BITMAP_FREE (mem->dep_ref);
1206 for (i = 0; VEC_iterate (mem_ref_locs_p, mem->accesses_in_loop, i, accs); i++)
1207 free_mem_ref_locs (accs);
1208 VEC_free (mem_ref_locs_p, heap, mem->accesses_in_loop);
1210 BITMAP_FREE (mem->vops);
1214 /* Allocates and returns a memory reference description for MEM whose hash
1215 value is HASH and id is ID. */
1218 mem_ref_alloc (tree mem, unsigned hash, unsigned id)
1220 mem_ref_p ref = XNEW (struct mem_ref);
1224 ref->stored = BITMAP_ALLOC (NULL);
1225 ref->indep_loop = BITMAP_ALLOC (NULL);
1226 ref->dep_loop = BITMAP_ALLOC (NULL);
1227 ref->indep_ref = BITMAP_ALLOC (NULL);
1228 ref->dep_ref = BITMAP_ALLOC (NULL);
1229 ref->accesses_in_loop = NULL;
1230 ref->vops = BITMAP_ALLOC (NULL);
1235 /* Allocates and returns the new list of locations. */
1237 static mem_ref_locs_p
1238 mem_ref_locs_alloc (void)
1240 mem_ref_locs_p accs = XNEW (struct mem_ref_locs);
1245 /* Records memory reference location *LOC in LOOP to the memory reference
1246 description REF. The reference occurs in statement STMT. */
1249 record_mem_ref_loc (mem_ref_p ref, struct loop *loop, tree stmt, tree *loc)
1251 mem_ref_loc_p aref = XNEW (struct mem_ref_loc);
1252 mem_ref_locs_p accs;
1253 bitmap ril = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
1255 if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop)
1256 <= (unsigned) loop->num)
1257 VEC_safe_grow_cleared (mem_ref_locs_p, heap, ref->accesses_in_loop,
1259 accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num);
1262 accs = mem_ref_locs_alloc ();
1263 VEC_replace (mem_ref_locs_p, ref->accesses_in_loop, loop->num, accs);
1269 VEC_safe_push (mem_ref_loc_p, heap, accs->locs, aref);
1270 bitmap_set_bit (ril, ref->id);
1273 /* Marks reference REF as stored in LOOP. */
1276 mark_ref_stored (mem_ref_p ref, struct loop *loop)
1279 loop != current_loops->tree_root
1280 && !bitmap_bit_p (ref->stored, loop->num);
1281 loop = loop_outer (loop))
1282 bitmap_set_bit (ref->stored, loop->num);
1285 /* Gathers memory references in statement STMT in LOOP, storing the
1286 information about them in the memory_accesses structure. Marks
1287 the vops accessed through unrecognized statements there as
1291 gather_mem_refs_stmt (struct loop *loop, tree stmt)
1303 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
1306 mem = simple_mem_ref_in_stmt (stmt, &is_stored);
1310 hash = iterative_hash_expr (*mem, 0);
1311 slot = htab_find_slot_with_hash (memory_accesses.refs, *mem, hash, INSERT);
1320 id = VEC_length (mem_ref_p, memory_accesses.refs_list);
1321 ref = mem_ref_alloc (*mem, hash, id);
1322 VEC_safe_push (mem_ref_p, heap, memory_accesses.refs_list, ref);
1325 if (dump_file && (dump_flags & TDF_DETAILS))
1327 fprintf (dump_file, "Memory reference %u: ", id);
1328 print_generic_expr (dump_file, ref->mem, TDF_SLIM);
1329 fprintf (dump_file, "\n");
1333 mark_ref_stored (ref, loop);
1335 FOR_EACH_SSA_TREE_OPERAND (vname, stmt, oi, SSA_OP_VIRTUAL_USES)
1336 bitmap_set_bit (ref->vops, DECL_UID (SSA_NAME_VAR (vname)));
1337 record_mem_ref_loc (ref, loop, stmt, mem);
1341 clvops = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
1342 FOR_EACH_SSA_TREE_OPERAND (vname, stmt, oi, SSA_OP_VIRTUAL_USES)
1343 bitmap_set_bit (clvops, DECL_UID (SSA_NAME_VAR (vname)));
1346 /* Gathers memory references in loops. */
1349 gather_mem_refs_in_loops (void)
1351 block_stmt_iterator bsi;
1356 bitmap lrefs, alrefs, alrefso;
1360 loop = bb->loop_father;
1361 if (loop == current_loops->tree_root)
1364 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
1365 gather_mem_refs_stmt (loop, bsi_stmt (bsi));
1368 /* Propagate the information about clobbered vops and accessed memory
1369 references up the loop hierarchy. */
1370 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1372 lrefs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
1373 alrefs = VEC_index (bitmap, memory_accesses.all_refs_in_loop, loop->num);
1374 bitmap_ior_into (alrefs, lrefs);
1376 if (loop_outer (loop) == current_loops->tree_root)
1379 clvi = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
1380 clvo = VEC_index (bitmap, memory_accesses.clobbered_vops,
1381 loop_outer (loop)->num);
1382 bitmap_ior_into (clvo, clvi);
1384 alrefso = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
1385 loop_outer (loop)->num);
1386 bitmap_ior_into (alrefso, alrefs);
1390 /* Element of the hash table that maps vops to memory references. */
1392 struct vop_to_refs_elt
1394 /* DECL_UID of the vop. */
1397 /* List of the all references. */
1400 /* List of stored references. */
1404 /* A hash function for struct vop_to_refs_elt object OBJ. */
1407 vtoe_hash (const void *obj)
1409 const struct vop_to_refs_elt *vtoe = obj;
1414 /* An equality function for struct vop_to_refs_elt object OBJ1 with
1415 uid of a vop OBJ2. */
1418 vtoe_eq (const void *obj1, const void *obj2)
1420 const struct vop_to_refs_elt *vtoe = obj1;
1421 const unsigned *uid = obj2;
1423 return vtoe->uid == *uid;
1426 /* A function to free the struct vop_to_refs_elt object. */
1429 vtoe_free (void *obj)
1431 struct vop_to_refs_elt *vtoe = obj;
1433 BITMAP_FREE (vtoe->refs_all);
1434 BITMAP_FREE (vtoe->refs_stored);
1438 /* Records REF to hashtable VOP_TO_REFS for the index VOP. STORED is true
1439 if the reference REF is stored. */
1442 record_vop_access (htab_t vop_to_refs, unsigned vop, unsigned ref, bool stored)
1444 void **slot = htab_find_slot_with_hash (vop_to_refs, &vop, vop, INSERT);
1445 struct vop_to_refs_elt *vtoe;
1449 vtoe = XNEW (struct vop_to_refs_elt);
1451 vtoe->refs_all = BITMAP_ALLOC (NULL);
1452 vtoe->refs_stored = BITMAP_ALLOC (NULL);
1458 bitmap_set_bit (vtoe->refs_all, ref);
1460 bitmap_set_bit (vtoe->refs_stored, ref);
1463 /* Returns the set of references that access VOP according to the table
1467 get_vop_accesses (htab_t vop_to_refs, unsigned vop)
1469 struct vop_to_refs_elt *vtoe = htab_find_with_hash (vop_to_refs, &vop, vop);
1470 return vtoe->refs_all;
1473 /* Returns the set of stores that access VOP according to the table
1477 get_vop_stores (htab_t vop_to_refs, unsigned vop)
1479 struct vop_to_refs_elt *vtoe = htab_find_with_hash (vop_to_refs, &vop, vop);
1480 return vtoe->refs_stored;
1483 /* Adds REF to mapping from virtual operands to references in LOOP. */
1486 add_vop_ref_mapping (struct loop *loop, mem_ref_p ref)
1488 htab_t map = VEC_index (htab_t, memory_accesses.vop_ref_map, loop->num);
1489 bool stored = bitmap_bit_p (ref->stored, loop->num);
1490 bitmap clobbers = VEC_index (bitmap, memory_accesses.clobbered_vops,
1495 EXECUTE_IF_AND_COMPL_IN_BITMAP (ref->vops, clobbers, 0, vop, bi)
1497 record_vop_access (map, vop, ref->id, stored);
1501 /* Create a mapping from virtual operands to references that touch them
1505 create_vop_ref_mapping_loop (struct loop *loop)
1507 bitmap refs = VEC_index (bitmap, memory_accesses.refs_in_loop, loop->num);
1513 EXECUTE_IF_SET_IN_BITMAP (refs, 0, i, bi)
1515 ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
1516 for (sloop = loop; sloop != current_loops->tree_root; sloop = loop_outer (sloop))
1517 add_vop_ref_mapping (sloop, ref);
1521 /* For each non-clobbered virtual operand and each loop, record the memory
1522 references in this loop that touch the operand. */
1525 create_vop_ref_mapping (void)
1530 FOR_EACH_LOOP (li, loop, 0)
1532 create_vop_ref_mapping_loop (loop);
1536 /* Gathers information about memory accesses in the loops. */
1539 analyze_memory_references (void)
1545 memory_accesses.refs
1546 = htab_create (100, memref_hash, memref_eq, memref_free);
1547 memory_accesses.refs_list = NULL;
1548 memory_accesses.refs_in_loop = VEC_alloc (bitmap, heap,
1549 number_of_loops ());
1550 memory_accesses.all_refs_in_loop = VEC_alloc (bitmap, heap,
1551 number_of_loops ());
1552 memory_accesses.clobbered_vops = VEC_alloc (bitmap, heap,
1553 number_of_loops ());
1554 memory_accesses.vop_ref_map = VEC_alloc (htab_t, heap,
1555 number_of_loops ());
1557 for (i = 0; i < number_of_loops (); i++)
1559 empty = BITMAP_ALLOC (NULL);
1560 VEC_quick_push (bitmap, memory_accesses.refs_in_loop, empty);
1561 empty = BITMAP_ALLOC (NULL);
1562 VEC_quick_push (bitmap, memory_accesses.all_refs_in_loop, empty);
1563 empty = BITMAP_ALLOC (NULL);
1564 VEC_quick_push (bitmap, memory_accesses.clobbered_vops, empty);
1565 hempty = htab_create (10, vtoe_hash, vtoe_eq, vtoe_free);
1566 VEC_quick_push (htab_t, memory_accesses.vop_ref_map, hempty);
1569 memory_accesses.ttae_cache = NULL;
1571 gather_mem_refs_in_loops ();
1572 create_vop_ref_mapping ();
1575 /* Returns true if a region of size SIZE1 at position 0 and a region of
1576 size SIZE2 at position DIFF cannot overlap. */
1579 cannot_overlap_p (aff_tree *diff, double_int size1, double_int size2)
1581 double_int d, bound;
1583 /* Unless the difference is a constant, we fail. */
1588 if (double_int_negative_p (d))
1590 /* The second object is before the first one, we succeed if the last
1591 element of the second object is before the start of the first one. */
1592 bound = double_int_add (d, double_int_add (size2, double_int_minus_one));
1593 return double_int_negative_p (bound);
1597 /* We succeed if the second object starts after the first one ends. */
1598 return double_int_scmp (size1, d) <= 0;
1602 /* Returns true if MEM1 and MEM2 may alias. TTAE_CACHE is used as a cache in
1603 tree_to_aff_combination_expand. */
1606 mem_refs_may_alias_p (tree mem1, tree mem2, struct pointer_map_t **ttae_cache)
1608 /* Perform BASE + OFFSET analysis -- if MEM1 and MEM2 are based on the same
1609 object and their offset differ in such a way that the locations cannot
1610 overlap, then they cannot alias. */
1611 aff_tree off1, off2;
1612 double_int size1, size2;
1615 /* If MEM1 and MEM2 are based on different variables, they cannot alias. */
1616 base1 = get_base_address (mem1);
1617 base2 = get_base_address (mem2);
1620 && !INDIRECT_REF_P (base1)
1622 && !INDIRECT_REF_P (base2)
1623 && !operand_equal_p (base1, base2, 0))
1626 /* With strict aliasing, it is impossible to access a scalar variable through
1627 anything but a pointer dereference or through a union (gcc extension). */
1628 if (flag_strict_aliasing)
1630 if (!INDIRECT_REF_P (mem1)
1632 && TREE_CODE (TREE_TYPE (base1)) != UNION_TYPE
1634 && !AGGREGATE_TYPE_P (TREE_TYPE (mem2)))
1636 if (!INDIRECT_REF_P (mem2)
1638 && TREE_CODE (TREE_TYPE (base2)) != UNION_TYPE
1640 && !AGGREGATE_TYPE_P (TREE_TYPE (mem1)))
1642 if (!alias_sets_conflict_p (get_alias_set (mem1), get_alias_set (mem2)))
1646 /* The expansion of addresses may be a bit expensive, thus we only do
1647 the check at -O2 and higher optimization levels. */
1651 get_inner_reference_aff (mem1, &off1, &size1);
1652 get_inner_reference_aff (mem2, &off2, &size2);
1653 aff_combination_expand (&off1, ttae_cache);
1654 aff_combination_expand (&off2, ttae_cache);
1655 aff_combination_scale (&off1, double_int_minus_one);
1656 aff_combination_add (&off2, &off1);
1658 if (cannot_overlap_p (&off2, size1, size2))
1664 /* Rewrites location LOC by TMP_VAR. */
1667 rewrite_mem_ref_loc (mem_ref_loc_p loc, tree tmp_var)
1669 mark_virtual_ops_for_renaming (loc->stmt);
1670 *loc->ref = tmp_var;
1671 update_stmt (loc->stmt);
1674 /* Adds all locations of REF in LOOP and its subloops to LOCS. */
1677 get_all_locs_in_loop (struct loop *loop, mem_ref_p ref,
1678 VEC (mem_ref_loc_p, heap) **locs)
1680 mem_ref_locs_p accs;
1683 bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
1685 struct loop *subloop;
1687 if (!bitmap_bit_p (refs, ref->id))
1690 if (VEC_length (mem_ref_locs_p, ref->accesses_in_loop)
1691 > (unsigned) loop->num)
1693 accs = VEC_index (mem_ref_locs_p, ref->accesses_in_loop, loop->num);
1696 for (i = 0; VEC_iterate (mem_ref_loc_p, accs->locs, i, loc); i++)
1697 VEC_safe_push (mem_ref_loc_p, heap, *locs, loc);
1701 for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
1702 get_all_locs_in_loop (subloop, ref, locs);
1705 /* Rewrites all references to REF in LOOP by variable TMP_VAR. */
1708 rewrite_mem_refs (struct loop *loop, mem_ref_p ref, tree tmp_var)
1712 VEC (mem_ref_loc_p, heap) *locs = NULL;
1714 get_all_locs_in_loop (loop, ref, &locs);
1715 for (i = 0; VEC_iterate (mem_ref_loc_p, locs, i, loc); i++)
1716 rewrite_mem_ref_loc (loc, tmp_var);
1717 VEC_free (mem_ref_loc_p, heap, locs);
1720 /* The name and the length of the currently generated variable
1722 #define MAX_LSM_NAME_LENGTH 40
1723 static char lsm_tmp_name[MAX_LSM_NAME_LENGTH + 1];
1724 static int lsm_tmp_name_length;
1726 /* Adds S to lsm_tmp_name. */
1729 lsm_tmp_name_add (const char *s)
1731 int l = strlen (s) + lsm_tmp_name_length;
1732 if (l > MAX_LSM_NAME_LENGTH)
1735 strcpy (lsm_tmp_name + lsm_tmp_name_length, s);
1736 lsm_tmp_name_length = l;
1739 /* Stores the name for temporary variable that replaces REF to
1743 gen_lsm_tmp_name (tree ref)
1747 switch (TREE_CODE (ref))
1749 case MISALIGNED_INDIRECT_REF:
1750 case ALIGN_INDIRECT_REF:
1752 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1753 lsm_tmp_name_add ("_");
1757 case VIEW_CONVERT_EXPR:
1758 case ARRAY_RANGE_REF:
1759 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1763 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1764 lsm_tmp_name_add ("_RE");
1768 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1769 lsm_tmp_name_add ("_IM");
1773 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1774 lsm_tmp_name_add ("_");
1775 name = get_name (TREE_OPERAND (ref, 1));
1778 lsm_tmp_name_add ("_");
1779 lsm_tmp_name_add (name);
1782 gen_lsm_tmp_name (TREE_OPERAND (ref, 0));
1783 lsm_tmp_name_add ("_I");
1787 ref = SSA_NAME_VAR (ref);
1792 name = get_name (ref);
1795 lsm_tmp_name_add (name);
1799 lsm_tmp_name_add ("S");
1803 lsm_tmp_name_add ("R");
1811 /* Determines name for temporary variable that replaces REF.
1812 The name is accumulated into the lsm_tmp_name variable.
1813 N is added to the name of the temporary. */
1816 get_lsm_tmp_name (tree ref, unsigned n)
1820 lsm_tmp_name_length = 0;
1821 gen_lsm_tmp_name (ref);
1822 lsm_tmp_name_add ("_lsm");
1827 lsm_tmp_name_add (ns);
1829 return lsm_tmp_name;
1832 /* Executes store motion of memory reference REF from LOOP.
1833 Exits from the LOOP are stored in EXITS. The initialization of the
1834 temporary variable is put to the preheader of the loop, and assignments
1835 to the reference from the temporary variable are emitted to exits. */
1838 execute_sm (struct loop *loop, VEC (edge, heap) *exits, mem_ref_p ref)
1843 struct fmt_data fmt_data;
1846 if (dump_file && (dump_flags & TDF_DETAILS))
1848 fprintf (dump_file, "Executing store motion of ");
1849 print_generic_expr (dump_file, ref->mem, 0);
1850 fprintf (dump_file, " from loop %d\n", loop->num);
1853 tmp_var = make_rename_temp (TREE_TYPE (ref->mem),
1854 get_lsm_tmp_name (ref->mem, ~0));
1856 fmt_data.loop = loop;
1857 fmt_data.orig_loop = loop;
1858 for_each_index (&ref->mem, force_move_till, &fmt_data);
1860 rewrite_mem_refs (loop, ref, tmp_var);
1862 /* Emit the load & stores. */
1863 load = build_gimple_modify_stmt (tmp_var, unshare_expr (ref->mem));
1864 get_stmt_ann (load)->common.aux = xcalloc (1, sizeof (struct lim_aux_data));
1865 LIM_DATA (load)->max_loop = loop;
1866 LIM_DATA (load)->tgt_loop = loop;
1868 /* Put this into the latch, so that we are sure it will be processed after
1869 all dependencies. */
1870 bsi_insert_on_edge (loop_latch_edge (loop), load);
1872 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
1874 store = build_gimple_modify_stmt (unshare_expr (ref->mem), tmp_var);
1875 bsi_insert_on_edge (ex, store);
1879 /* Hoists memory references MEM_REFS out of LOOP. EXITS is the list of exit
1880 edges of the LOOP. */
1883 hoist_memory_references (struct loop *loop, bitmap mem_refs,
1884 VEC (edge, heap) *exits)
1890 EXECUTE_IF_SET_IN_BITMAP (mem_refs, 0, i, bi)
1892 ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
1893 execute_sm (loop, exits, ref);
1897 /* Returns true if REF is always accessed in LOOP. */
1900 ref_always_accessed_p (struct loop *loop, mem_ref_p ref)
1902 VEC (mem_ref_loc_p, heap) *locs = NULL;
1906 struct loop *must_exec;
1908 get_all_locs_in_loop (loop, ref, &locs);
1909 for (i = 0; VEC_iterate (mem_ref_loc_p, locs, i, loc); i++)
1911 if (!LIM_DATA (loc->stmt))
1914 must_exec = LIM_DATA (loc->stmt)->always_executed_in;
1918 if (must_exec == loop
1919 || flow_loop_nested_p (must_exec, loop))
1925 VEC_free (mem_ref_loc_p, heap, locs);
1930 /* Returns true if REF1 and REF2 are independent. */
1933 refs_independent_p (mem_ref_p ref1, mem_ref_p ref2)
1936 || bitmap_bit_p (ref1->indep_ref, ref2->id))
1938 if (bitmap_bit_p (ref1->dep_ref, ref2->id))
1941 if (dump_file && (dump_flags & TDF_DETAILS))
1942 fprintf (dump_file, "Querying dependency of refs %u and %u: ",
1943 ref1->id, ref2->id);
1945 if (mem_refs_may_alias_p (ref1->mem, ref2->mem,
1946 &memory_accesses.ttae_cache))
1948 bitmap_set_bit (ref1->dep_ref, ref2->id);
1949 bitmap_set_bit (ref2->dep_ref, ref1->id);
1950 if (dump_file && (dump_flags & TDF_DETAILS))
1951 fprintf (dump_file, "dependent.\n");
1956 bitmap_set_bit (ref1->indep_ref, ref2->id);
1957 bitmap_set_bit (ref2->indep_ref, ref1->id);
1958 if (dump_file && (dump_flags & TDF_DETAILS))
1959 fprintf (dump_file, "independent.\n");
1964 /* Records the information whether REF is independent in LOOP (according
1968 record_indep_loop (struct loop *loop, mem_ref_p ref, bool indep)
1971 bitmap_set_bit (ref->indep_loop, loop->num);
1973 bitmap_set_bit (ref->dep_loop, loop->num);
1976 /* Returns true if REF is independent on all other memory references in
1980 ref_indep_loop_p_1 (struct loop *loop, mem_ref_p ref)
1982 bitmap clobbers, refs_to_check, refs;
1985 bool ret = true, stored = bitmap_bit_p (ref->stored, loop->num);
1989 /* If the reference is clobbered, it is not independent. */
1990 clobbers = VEC_index (bitmap, memory_accesses.clobbered_vops, loop->num);
1991 if (bitmap_intersect_p (ref->vops, clobbers))
1994 refs_to_check = BITMAP_ALLOC (NULL);
1996 map = VEC_index (htab_t, memory_accesses.vop_ref_map, loop->num);
1997 EXECUTE_IF_AND_COMPL_IN_BITMAP (ref->vops, clobbers, 0, i, bi)
2000 refs = get_vop_accesses (map, i);
2002 refs = get_vop_stores (map, i);
2004 bitmap_ior_into (refs_to_check, refs);
2007 EXECUTE_IF_SET_IN_BITMAP (refs_to_check, 0, i, bi)
2009 aref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
2010 if (!refs_independent_p (ref, aref))
2013 record_indep_loop (loop, aref, false);
2018 BITMAP_FREE (refs_to_check);
2022 /* Returns true if REF is independent on all other memory references in
2023 LOOP. Wrapper over ref_indep_loop_p_1, caching its results. */
2026 ref_indep_loop_p (struct loop *loop, mem_ref_p ref)
2030 if (bitmap_bit_p (ref->indep_loop, loop->num))
2032 if (bitmap_bit_p (ref->dep_loop, loop->num))
2035 ret = ref_indep_loop_p_1 (loop, ref);
2037 if (dump_file && (dump_flags & TDF_DETAILS))
2038 fprintf (dump_file, "Querying dependencies of ref %u in loop %d: %s\n",
2039 ref->id, loop->num, ret ? "independent" : "dependent");
2041 record_indep_loop (loop, ref, ret);
2046 /* Returns true if we can perform store motion of REF from LOOP. */
2049 can_sm_ref_p (struct loop *loop, mem_ref_p ref)
2051 /* Unless the reference is stored in the loop, there is nothing to do. */
2052 if (!bitmap_bit_p (ref->stored, loop->num))
2055 /* It should be movable. */
2056 if (!is_gimple_reg_type (TREE_TYPE (ref->mem))
2057 || TREE_THIS_VOLATILE (ref->mem)
2058 || !for_each_index (&ref->mem, may_move_till, loop))
2061 /* If it can trap, it must be always executed in LOOP. */
2062 if (tree_could_trap_p (ref->mem)
2063 && !ref_always_accessed_p (loop, ref))
2066 /* And it must be independent on all other memory references
2068 if (!ref_indep_loop_p (loop, ref))
2074 /* Marks the references in LOOP for that store motion should be performed
2075 in REFS_TO_SM. SM_EXECUTED is the set of references for that store
2076 motion was performed in one of the outer loops. */
2079 find_refs_for_sm (struct loop *loop, bitmap sm_executed, bitmap refs_to_sm)
2081 bitmap refs = VEC_index (bitmap, memory_accesses.all_refs_in_loop,
2087 EXECUTE_IF_AND_COMPL_IN_BITMAP (refs, sm_executed, 0, i, bi)
2089 ref = VEC_index (mem_ref_p, memory_accesses.refs_list, i);
2090 if (can_sm_ref_p (loop, ref))
2091 bitmap_set_bit (refs_to_sm, i);
2095 /* Checks whether LOOP (with exits stored in EXITS array) is suitable
2096 for a store motion optimization (i.e. whether we can insert statement
2100 loop_suitable_for_sm (struct loop *loop ATTRIBUTE_UNUSED,
2101 VEC (edge, heap) *exits)
2106 for (i = 0; VEC_iterate (edge, exits, i, ex); i++)
2107 if (ex->flags & EDGE_ABNORMAL)
2113 /* Try to perform store motion for all memory references modified inside
2114 LOOP. SM_EXECUTED is the bitmap of the memory references for that
2115 store motion was executed in one of the outer loops. */
2118 store_motion_loop (struct loop *loop, bitmap sm_executed)
2120 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
2121 struct loop *subloop;
2122 bitmap sm_in_loop = BITMAP_ALLOC (NULL);
2124 if (loop_suitable_for_sm (loop, exits))
2126 find_refs_for_sm (loop, sm_executed, sm_in_loop);
2127 hoist_memory_references (loop, sm_in_loop, exits);
2129 VEC_free (edge, heap, exits);
2131 bitmap_ior_into (sm_executed, sm_in_loop);
2132 for (subloop = loop->inner; subloop != NULL; subloop = subloop->next)
2133 store_motion_loop (subloop, sm_executed);
2134 bitmap_and_compl_into (sm_executed, sm_in_loop);
2135 BITMAP_FREE (sm_in_loop);
2138 /* Try to perform store motion for all memory references modified inside
2145 bitmap sm_executed = BITMAP_ALLOC (NULL);
2147 for (loop = current_loops->tree_root->inner; loop != NULL; loop = loop->next)
2148 store_motion_loop (loop, sm_executed);
2150 BITMAP_FREE (sm_executed);
2151 bsi_commit_edge_inserts ();
2154 /* Fills ALWAYS_EXECUTED_IN information for basic blocks of LOOP, i.e.
2155 for each such basic block bb records the outermost loop for that execution
2156 of its header implies execution of bb. CONTAINS_CALL is the bitmap of
2157 blocks that contain a nonpure call. */
2160 fill_always_executed_in (struct loop *loop, sbitmap contains_call)
2162 basic_block bb = NULL, *bbs, last = NULL;
2165 struct loop *inn_loop = loop;
2167 if (!loop->header->aux)
2169 bbs = get_loop_body_in_dom_order (loop);
2171 for (i = 0; i < loop->num_nodes; i++)
2176 if (dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
2179 if (TEST_BIT (contains_call, bb->index))
2182 FOR_EACH_EDGE (e, ei, bb->succs)
2183 if (!flow_bb_inside_loop_p (loop, e->dest))
2188 /* A loop might be infinite (TODO use simple loop analysis
2189 to disprove this if possible). */
2190 if (bb->flags & BB_IRREDUCIBLE_LOOP)
2193 if (!flow_bb_inside_loop_p (inn_loop, bb))
2196 if (bb->loop_father->header == bb)
2198 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
2201 /* In a loop that is always entered we may proceed anyway.
2202 But record that we entered it and stop once we leave it. */
2203 inn_loop = bb->loop_father;
2210 if (last == loop->header)
2212 last = get_immediate_dominator (CDI_DOMINATORS, last);
2218 for (loop = loop->inner; loop; loop = loop->next)
2219 fill_always_executed_in (loop, contains_call);
2222 /* Compute the global information needed by the loop invariant motion pass. */
2225 tree_ssa_lim_initialize (void)
2227 sbitmap contains_call = sbitmap_alloc (last_basic_block);
2228 block_stmt_iterator bsi;
2232 sbitmap_zero (contains_call);
2235 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2237 if (nonpure_call_p (bsi_stmt (bsi)))
2241 if (!bsi_end_p (bsi))
2242 SET_BIT (contains_call, bb->index);
2245 for (loop = current_loops->tree_root->inner; loop; loop = loop->next)
2246 fill_always_executed_in (loop, contains_call);
2248 sbitmap_free (contains_call);
2251 /* Cleans up after the invariant motion pass. */
2254 tree_ssa_lim_finalize (void)
2266 VEC_free (mem_ref_p, heap, memory_accesses.refs_list);
2267 htab_delete (memory_accesses.refs);
2269 for (i = 0; VEC_iterate (bitmap, memory_accesses.refs_in_loop, i, b); i++)
2271 VEC_free (bitmap, heap, memory_accesses.refs_in_loop);
2273 for (i = 0; VEC_iterate (bitmap, memory_accesses.all_refs_in_loop, i, b); i++)
2275 VEC_free (bitmap, heap, memory_accesses.all_refs_in_loop);
2277 for (i = 0; VEC_iterate (bitmap, memory_accesses.clobbered_vops, i, b); i++)
2279 VEC_free (bitmap, heap, memory_accesses.clobbered_vops);
2281 for (i = 0; VEC_iterate (htab_t, memory_accesses.vop_ref_map, i, h); i++)
2283 VEC_free (htab_t, heap, memory_accesses.vop_ref_map);
2285 if (memory_accesses.ttae_cache)
2286 pointer_map_destroy (memory_accesses.ttae_cache);
2289 /* Moves invariants from loops. Only "expensive" invariants are moved out --
2290 i.e. those that are likely to be win regardless of the register pressure. */
2295 tree_ssa_lim_initialize ();
2297 /* Gathers information about memory accesses in the loops. */
2298 analyze_memory_references ();
2300 /* For each statement determine the outermost loop in that it is
2301 invariant and cost for computing the invariant. */
2302 determine_invariantness ();
2304 /* Execute store motion. Force the necessary invariants to be moved
2305 out of the loops as well. */
2308 /* Move the expressions that are expensive enough. */
2309 move_computations ();
2311 tree_ssa_lim_finalize ();