1 /* High-level loop manipulation functions.
2 Copyright (C) 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 2, 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 COPYING. If not, write to the Free
18 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
23 #include "coretypes.h"
28 #include "hard-reg-set.h"
29 #include "basic-block.h"
31 #include "diagnostic.h"
32 #include "tree-flow.h"
33 #include "tree-dump.h"
36 #include "tree-pass.h"
37 #include "cfglayout.h"
38 #include "tree-scalar-evolution.h"
40 #include "tree-inline.h"
42 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
43 It is expected that neither BASE nor STEP are shared with other expressions
44 (unless the sharing rules allow this). Use VAR as a base var_decl for it
45 (if NULL, a new temporary will be created). The increment will occur at
46 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
47 AFTER can be computed using standard_iv_increment_position. The ssa versions
48 of the variable before and after increment will be stored in VAR_BEFORE and
49 VAR_AFTER (unless they are NULL). */
52 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
53 block_stmt_iterator
*incr_pos
, bool after
,
54 tree
*var_before
, tree
*var_after
)
56 tree stmt
, initial
, step1
, stmts
;
58 enum tree_code incr_op
= PLUS_EXPR
;
59 edge pe
= loop_preheader_edge (loop
);
63 var
= create_tmp_var (TREE_TYPE (base
), "ivtmp");
64 add_referenced_var (var
);
67 vb
= make_ssa_name (var
, NULL_TREE
);
70 va
= make_ssa_name (var
, NULL_TREE
);
74 /* For easier readability of the created code, produce MINUS_EXPRs
76 if (TREE_CODE (step
) == INTEGER_CST
)
78 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
80 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
81 if (tree_int_cst_lt (step1
, step
))
91 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
92 && may_negate_without_overflow_p (step
))
95 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
100 /* Gimplify the step if necessary. We put the computations in front of the
101 loop (i.e. the step should be loop invariant). */
102 step
= force_gimple_operand (step
, &stmts
, true, var
);
104 bsi_insert_on_edge_immediate (pe
, stmts
);
106 stmt
= build_gimple_modify_stmt (va
,
107 build2 (incr_op
, TREE_TYPE (base
),
109 SSA_NAME_DEF_STMT (va
) = stmt
;
111 bsi_insert_after (incr_pos
, stmt
, BSI_NEW_STMT
);
113 bsi_insert_before (incr_pos
, stmt
, BSI_NEW_STMT
);
115 initial
= force_gimple_operand (base
, &stmts
, true, var
);
117 bsi_insert_on_edge_immediate (pe
, stmts
);
119 stmt
= create_phi_node (vb
, loop
->header
);
120 SSA_NAME_DEF_STMT (vb
) = stmt
;
121 add_phi_arg (stmt
, initial
, loop_preheader_edge (loop
));
122 add_phi_arg (stmt
, va
, loop_latch_edge (loop
));
125 /* Add exit phis for the USE on EXIT. */
128 add_exit_phis_edge (basic_block exit
, tree use
)
130 tree phi
, def_stmt
= SSA_NAME_DEF_STMT (use
);
131 basic_block def_bb
= bb_for_stmt (def_stmt
);
132 struct loop
*def_loop
;
136 /* Check that some of the edges entering the EXIT block exits a loop in
137 that USE is defined. */
138 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
140 def_loop
= find_common_loop (def_bb
->loop_father
, e
->src
->loop_father
);
141 if (!flow_bb_inside_loop_p (def_loop
, e
->dest
))
148 phi
= create_phi_node (use
, exit
);
149 create_new_def_for (PHI_RESULT (phi
), phi
, PHI_RESULT_PTR (phi
));
150 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
151 add_phi_arg (phi
, use
, e
);
154 /* Add exit phis for VAR that is used in LIVEIN.
155 Exits of the loops are stored in EXITS. */
158 add_exit_phis_var (tree var
, bitmap livein
, bitmap exits
)
162 basic_block def_bb
= bb_for_stmt (SSA_NAME_DEF_STMT (var
));
165 if (is_gimple_reg (var
))
166 bitmap_clear_bit (livein
, def_bb
->index
);
168 bitmap_set_bit (livein
, def_bb
->index
);
170 def
= BITMAP_ALLOC (NULL
);
171 bitmap_set_bit (def
, def_bb
->index
);
172 compute_global_livein (livein
, def
);
175 EXECUTE_IF_AND_IN_BITMAP (exits
, livein
, 0, index
, bi
)
177 add_exit_phis_edge (BASIC_BLOCK (index
), var
);
181 /* Add exit phis for the names marked in NAMES_TO_RENAME.
182 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
183 names are used are stored in USE_BLOCKS. */
186 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap loop_exits
)
191 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
193 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
197 /* Returns a bitmap of all loop exit edge targets. */
200 get_loops_exits (void)
202 bitmap exits
= BITMAP_ALLOC (NULL
);
209 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
210 if (e
->src
!= ENTRY_BLOCK_PTR
211 && !flow_bb_inside_loop_p (e
->src
->loop_father
, bb
))
213 bitmap_set_bit (exits
, bb
->index
);
221 /* For USE in BB, if it is used outside of the loop it is defined in,
222 mark it for rewrite. Record basic block BB where it is used
223 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
226 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
231 struct loop
*def_loop
;
233 if (TREE_CODE (use
) != SSA_NAME
)
236 /* We don't need to keep virtual operands in loop-closed form. */
237 if (!is_gimple_reg (use
))
240 ver
= SSA_NAME_VERSION (use
);
241 def_bb
= bb_for_stmt (SSA_NAME_DEF_STMT (use
));
244 def_loop
= def_bb
->loop_father
;
246 /* If the definition is not inside loop, it is not interesting. */
247 if (!def_loop
->outer
)
250 if (!use_blocks
[ver
])
251 use_blocks
[ver
] = BITMAP_ALLOC (NULL
);
252 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
254 bitmap_set_bit (need_phis
, ver
);
257 /* For uses in STMT, mark names that are used outside of the loop they are
258 defined to rewrite. Record the set of blocks in that the ssa
259 names are defined to USE_BLOCKS and the ssa names themselves to
263 find_uses_to_rename_stmt (tree stmt
, bitmap
*use_blocks
, bitmap need_phis
)
267 basic_block bb
= bb_for_stmt (stmt
);
269 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_ALL_USES
)
270 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
273 /* Marks names that are used in BB and outside of the loop they are
274 defined in for rewrite. Records the set of blocks in that the ssa
275 names are defined to USE_BLOCKS. Record the SSA names that will
276 need exit PHIs in NEED_PHIS. */
279 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
281 block_stmt_iterator bsi
;
286 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
287 for (phi
= phi_nodes (e
->dest
); phi
; phi
= PHI_CHAIN (phi
))
288 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
289 use_blocks
, need_phis
);
291 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
292 find_uses_to_rename_stmt (bsi_stmt (bsi
), use_blocks
, need_phis
);
295 /* Marks names that are used outside of the loop they are defined in
296 for rewrite. Records the set of blocks in that the ssa
297 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
298 scan only blocks in this set. */
301 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
307 if (changed_bbs
&& !bitmap_empty_p (changed_bbs
))
309 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
311 find_uses_to_rename_bb (BASIC_BLOCK (index
), use_blocks
, need_phis
);
318 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
323 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
324 phi nodes to ensure that no variable is used outside the loop it is
327 This strengthening of the basic ssa form has several advantages:
329 1) Updating it during unrolling/peeling/versioning is trivial, since
330 we do not need to care about the uses outside of the loop.
331 2) The behavior of all uses of an induction variable is the same.
332 Without this, you need to distinguish the case when the variable
333 is used outside of the loop it is defined in, for example
335 for (i = 0; i < 100; i++)
337 for (j = 0; j < 100; j++)
345 Looking from the outer loop with the normal SSA form, the first use of k
346 is not well-behaved, while the second one is an induction variable with
349 If CHANGED_BBS is not NULL, we look for uses outside loops only in
350 the basic blocks in this set.
352 UPDATE_FLAG is used in the call to update_ssa. See
353 TODO_update_ssa* for documentation. */
356 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
360 unsigned i
, old_num_ssa_names
;
361 bitmap names_to_rename
;
366 loop_exits
= get_loops_exits ();
367 names_to_rename
= BITMAP_ALLOC (NULL
);
369 /* If the pass has caused the SSA form to be out-of-date, update it
371 update_ssa (update_flag
);
373 old_num_ssa_names
= num_ssa_names
;
374 use_blocks
= XCNEWVEC (bitmap
, old_num_ssa_names
);
376 /* Find the uses outside loops. */
377 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
379 /* Add the PHI nodes on exits of the loops for the names we need to
381 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
383 for (i
= 0; i
< old_num_ssa_names
; i
++)
384 BITMAP_FREE (use_blocks
[i
]);
386 BITMAP_FREE (loop_exits
);
387 BITMAP_FREE (names_to_rename
);
389 /* Fix up all the names found to be used outside their original
391 update_ssa (TODO_update_ssa
);
393 current_loops
->state
|= LOOP_CLOSED_SSA
;
396 /* Check invariants of the loop closed ssa form for the USE in BB. */
399 check_loop_closed_ssa_use (basic_block bb
, tree use
)
404 if (TREE_CODE (use
) != SSA_NAME
|| !is_gimple_reg (use
))
407 def
= SSA_NAME_DEF_STMT (use
);
408 def_bb
= bb_for_stmt (def
);
410 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
413 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
416 check_loop_closed_ssa_stmt (basic_block bb
, tree stmt
)
421 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_ALL_USES
)
422 check_loop_closed_ssa_use (bb
, var
);
425 /* Checks that invariants of the loop closed ssa form are preserved. */
428 verify_loop_closed_ssa (void)
431 block_stmt_iterator bsi
;
435 if (current_loops
== NULL
)
442 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
443 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (phi
); i
++)
444 check_loop_closed_ssa_use (PHI_ARG_EDGE (phi
, i
)->src
,
445 PHI_ARG_DEF (phi
, i
));
447 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
448 check_loop_closed_ssa_stmt (bb
, bsi_stmt (bsi
));
452 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
453 preserve the loop closed ssa form. */
456 split_loop_exit_edge (edge exit
)
458 basic_block dest
= exit
->dest
;
459 basic_block bb
= split_edge (exit
);
460 tree phi
, new_phi
, new_name
, name
;
463 for (phi
= phi_nodes (dest
); phi
; phi
= PHI_CHAIN (phi
))
465 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
467 name
= USE_FROM_PTR (op_p
);
469 /* If the argument of the PHI node is a constant, we do not need
470 to keep it inside loop. */
471 if (TREE_CODE (name
) != SSA_NAME
)
474 /* Otherwise create an auxiliary phi node that will copy the value
475 of the SSA name out of the loop. */
476 new_name
= duplicate_ssa_name (name
, NULL
);
477 new_phi
= create_phi_node (new_name
, bb
);
478 SSA_NAME_DEF_STMT (new_name
) = new_phi
;
479 add_phi_arg (new_phi
, name
, exit
);
480 SET_USE (op_p
, new_name
);
484 /* Returns the basic block in that statements should be emitted for induction
485 variables incremented at the end of the LOOP. */
488 ip_end_pos (struct loop
*loop
)
493 /* Returns the basic block in that statements should be emitted for induction
494 variables incremented just before exit condition of a LOOP. */
497 ip_normal_pos (struct loop
*loop
)
503 if (!single_pred_p (loop
->latch
))
506 bb
= single_pred (loop
->latch
);
507 last
= last_stmt (bb
);
509 || TREE_CODE (last
) != COND_EXPR
)
512 exit
= EDGE_SUCC (bb
, 0);
513 if (exit
->dest
== loop
->latch
)
514 exit
= EDGE_SUCC (bb
, 1);
516 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
522 /* Stores the standard position for induction variable increment in LOOP
523 (just before the exit condition if it is available and latch block is empty,
524 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
525 the increment should be inserted after *BSI. */
528 standard_iv_increment_position (struct loop
*loop
, block_stmt_iterator
*bsi
,
531 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
532 tree last
= last_stmt (latch
);
535 || (last
&& TREE_CODE (last
) != LABEL_EXPR
))
537 *bsi
= bsi_last (latch
);
538 *insert_after
= true;
542 *bsi
= bsi_last (bb
);
543 *insert_after
= false;
547 /* Copies phi node arguments for duplicated blocks. The index of the first
548 duplicated block is FIRST_NEW_BLOCK. */
551 copy_phi_node_args (unsigned first_new_block
)
555 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
556 BASIC_BLOCK (i
)->flags
|= BB_DUPLICATED
;
558 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
559 add_phi_args_after_copy_bb (BASIC_BLOCK (i
));
561 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
562 BASIC_BLOCK (i
)->flags
&= ~BB_DUPLICATED
;
566 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
567 updates the PHI nodes at start of the copied region. In order to
568 achieve this, only loops whose exits all lead to the same location
571 Notice that we do not completely update the SSA web after
572 duplication. The caller is responsible for calling update_ssa
573 after the loop has been duplicated. */
576 tree_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
577 unsigned int ndupl
, sbitmap wont_exit
,
578 edge orig
, VEC (edge
, heap
) **to_remove
,
581 unsigned first_new_block
;
583 if (!(current_loops
->state
& LOOPS_HAVE_SIMPLE_LATCHES
))
585 if (!(current_loops
->state
& LOOPS_HAVE_PREHEADERS
))
588 #ifdef ENABLE_CHECKING
589 verify_loop_closed_ssa ();
592 first_new_block
= last_basic_block
;
593 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
594 orig
, to_remove
, flags
))
597 /* Readd the removed phi args for e. */
598 flush_pending_stmts (e
);
600 /* Copy the phi node arguments. */
601 copy_phi_node_args (first_new_block
);
608 /* Returns true if we can unroll LOOP FACTOR times. Number
609 of iterations of the loop is returned in NITER. */
612 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
613 struct tree_niter_desc
*niter
)
617 /* Check whether unrolling is possible. We only want to unroll loops
618 for that we are able to determine number of iterations. We also
619 want to split the extra iterations of the loop from its end,
620 therefore we require that the loop has precisely one
623 exit
= single_dom_exit (loop
);
627 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
628 || niter
->cmp
== ERROR_MARK
629 /* Scalar evolutions analysis might have copy propagated
630 the abnormal ssa names into these expressions, hence
631 emitting the computations based on them during loop
632 unrolling might create overlapping life ranges for
633 them, and failures in out-of-ssa. */
634 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
635 || contains_abnormal_ssa_name_p (niter
->control
.base
)
636 || contains_abnormal_ssa_name_p (niter
->control
.step
)
637 || contains_abnormal_ssa_name_p (niter
->bound
))
640 /* And of course, we must be able to duplicate the loop. */
641 if (!can_duplicate_loop_p (loop
))
644 /* The final loop should be small enough. */
645 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
646 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
652 /* Determines the conditions that control execution of LOOP unrolled FACTOR
653 times. DESC is number of iterations of LOOP. ENTER_COND is set to
654 condition that must be true if the main loop can be entered.
655 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
656 how the exit from the unrolled loop should be controlled. */
659 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
660 unsigned factor
, tree
*enter_cond
,
661 tree
*exit_base
, tree
*exit_step
,
662 enum tree_code
*exit_cmp
, tree
*exit_bound
)
665 tree base
= desc
->control
.base
;
666 tree step
= desc
->control
.step
;
667 tree bound
= desc
->bound
;
668 tree type
= TREE_TYPE (base
);
670 tree min
= lower_bound_in_type (type
, type
);
671 tree max
= upper_bound_in_type (type
, type
);
672 enum tree_code cmp
= desc
->cmp
;
673 tree cond
= boolean_true_node
, assum
;
675 *enter_cond
= boolean_false_node
;
676 *exit_base
= NULL_TREE
;
677 *exit_step
= NULL_TREE
;
678 *exit_cmp
= ERROR_MARK
;
679 *exit_bound
= NULL_TREE
;
680 gcc_assert (cmp
!= ERROR_MARK
);
682 /* We only need to be correct when we answer question
683 "Do at least FACTOR more iterations remain?" in the unrolled loop.
684 Thus, transforming BASE + STEP * i <> BOUND to
685 BASE + STEP * i < BOUND is ok. */
688 if (tree_int_cst_sign_bit (step
))
693 else if (cmp
== LT_EXPR
)
695 gcc_assert (!tree_int_cst_sign_bit (step
));
697 else if (cmp
== GT_EXPR
)
699 gcc_assert (tree_int_cst_sign_bit (step
));
704 /* The main body of the loop may be entered iff:
706 1) desc->may_be_zero is false.
707 2) it is possible to check that there are at least FACTOR iterations
708 of the loop, i.e., BOUND - step * FACTOR does not overflow.
709 3) # of iterations is at least FACTOR */
711 if (!integer_zerop (desc
->may_be_zero
))
712 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
713 invert_truthvalue (desc
->may_be_zero
),
716 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
717 build_int_cst_type (type
, factor
));
718 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
720 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
722 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
724 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
726 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
727 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
729 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
730 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
731 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
733 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
735 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
736 /* cond now may be a gimple comparison, which would be OK, but also any
737 other gimple rhs (say a && b). In this case we need to force it to
739 if (!is_gimple_condexpr (cond
))
741 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
743 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
747 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
749 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
750 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
752 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
755 *exit_step
= bigstep
;
760 /* Scales the frequencies of all basic blocks in LOOP that are strictly
761 dominated by BB by NUM/DEN. */
764 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
772 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
774 son
= next_dom_son (CDI_DOMINATORS
, son
))
776 if (!flow_bb_inside_loop_p (loop
, son
))
778 scale_bbs_frequencies_int (&son
, 1, num
, den
);
779 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
783 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
784 EXIT is the exit of the loop to that DESC corresponds.
786 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
787 under that loop exits in the first iteration even if N != 0,
791 x = phi (init, next);
799 becomes (with possibly the exit conditions formulated a bit differently,
800 avoiding the need to create a new iv):
802 if (MAY_BE_ZERO || N < FACTOR)
807 x = phi (init, next);
818 } while (N >= FACTOR);
821 init' = phi (init, x);
825 x = phi (init', next);
833 Before the loop is unrolled, TRANSFORM is called for it (only for the
834 unrolled loop, but not for its versioned copy). DATA is passed to
837 /* Probability in % that the unrolled loop is entered. Just a guess. */
838 #define PROB_UNROLLED_LOOP_ENTERED 90
841 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
842 edge exit
, struct tree_niter_desc
*desc
,
843 transform_callback transform
,
846 tree exit_if
, ctr_before
, ctr_after
;
847 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
848 enum tree_code exit_cmp
;
849 tree phi_old_loop
, phi_new_loop
, phi_rest
, init
, next
, new_init
, var
;
850 struct loop
*new_loop
;
851 basic_block rest
, exit_bb
;
852 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
854 block_stmt_iterator bsi
;
857 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
858 unsigned new_est_niter
, i
, prob
;
859 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
861 VEC (edge
, heap
) *to_remove
= NULL
;
863 est_niter
= expected_loop_iterations (loop
);
864 determine_exit_conditions (loop
, desc
, factor
,
865 &enter_main_cond
, &exit_base
, &exit_step
,
866 &exit_cmp
, &exit_bound
);
868 /* Let us assume that the unrolled loop is quite likely to be entered. */
869 if (integer_nonzerop (enter_main_cond
))
870 prob_entry
= REG_BR_PROB_BASE
;
872 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
874 /* The values for scales should keep profile consistent, and somewhat close
877 TODO: The current value of SCALE_REST makes it appear that the loop that
878 is created by splitting the remaining iterations of the unrolled loop is
879 executed the same number of times as the original loop, and with the same
880 frequencies, which is obviously wrong. This does not appear to cause
881 problems, so we do not bother with fixing it for now. To make the profile
882 correct, we would need to change the probability of the exit edge of the
883 loop, and recompute the distribution of frequencies in its body because
884 of this change (scale the frequencies of blocks before and after the exit
885 by appropriate factors). */
886 scale_unrolled
= prob_entry
;
887 scale_rest
= REG_BR_PROB_BASE
;
889 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
890 prob_entry
, scale_unrolled
, scale_rest
, true);
891 gcc_assert (new_loop
!= NULL
);
892 update_ssa (TODO_update_ssa
);
894 /* Determine the probability of the exit edge of the unrolled loop. */
895 new_est_niter
= est_niter
/ factor
;
897 /* Without profile feedback, loops for that we do not know a better estimate
898 are assumed to roll 10 times. When we unroll such loop, it appears to
899 roll too little, and it may even seem to be cold. To avoid this, we
900 ensure that the created loop appears to roll at least 5 times (but at
901 most as many times as before unrolling). */
902 if (new_est_niter
< 5)
905 new_est_niter
= est_niter
;
910 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
911 loop latch (and make its condition dummy, for the moment). */
912 rest
= loop_preheader_edge (new_loop
)->src
;
913 precond_edge
= single_pred_edge (rest
);
914 split_edge (loop_latch_edge (loop
));
915 exit_bb
= single_pred (loop
->latch
);
917 /* Since the exit edge will be removed, the frequency of all the blocks
918 in the loop that are dominated by it must be scaled by
919 1 / (1 - exit->probability). */
920 scale_dominated_blocks_in_loop (loop
, exit
->src
,
922 REG_BR_PROB_BASE
- exit
->probability
);
924 bsi
= bsi_last (exit_bb
);
925 exit_if
= build3 (COND_EXPR
, void_type_node
, boolean_true_node
,
926 NULL_TREE
, NULL_TREE
);
928 bsi_insert_after (&bsi
, exit_if
, BSI_NEW_STMT
);
929 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
930 rescan_loop_exit (new_exit
, true, false);
932 /* Set the probability of new exit to the same of the old one. Fix
933 the frequency of the latch block, by scaling it back by
934 1 - exit->probability. */
935 new_exit
->count
= exit
->count
;
936 new_exit
->probability
= exit
->probability
;
937 new_nonexit
= single_pred_edge (loop
->latch
);
938 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
939 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
940 new_nonexit
->count
-= exit
->count
;
941 if (new_nonexit
->count
< 0)
942 new_nonexit
->count
= 0;
943 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
946 old_entry
= loop_preheader_edge (loop
);
947 new_entry
= loop_preheader_edge (new_loop
);
948 old_latch
= loop_latch_edge (loop
);
949 for (phi_old_loop
= phi_nodes (loop
->header
),
950 phi_new_loop
= phi_nodes (new_loop
->header
);
952 phi_old_loop
= PHI_CHAIN (phi_old_loop
),
953 phi_new_loop
= PHI_CHAIN (phi_new_loop
))
955 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
956 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
957 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
958 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
960 /* Prefer using original variable as a base for the new ssa name.
961 This is necessary for virtual ops, and useful in order to avoid
962 losing debug info for real ops. */
963 if (TREE_CODE (next
) == SSA_NAME
)
964 var
= SSA_NAME_VAR (next
);
965 else if (TREE_CODE (init
) == SSA_NAME
)
966 var
= SSA_NAME_VAR (init
);
969 var
= create_tmp_var (TREE_TYPE (init
), "unrinittmp");
970 add_referenced_var (var
);
973 new_init
= make_ssa_name (var
, NULL_TREE
);
974 phi_rest
= create_phi_node (new_init
, rest
);
975 SSA_NAME_DEF_STMT (new_init
) = phi_rest
;
977 add_phi_arg (phi_rest
, init
, precond_edge
);
978 add_phi_arg (phi_rest
, next
, new_exit
);
979 SET_USE (op
, new_init
);
984 /* Transform the loop. */
986 (*transform
) (loop
, data
);
988 /* Unroll the loop and remove the exits in all iterations except for the
990 wont_exit
= sbitmap_alloc (factor
);
991 sbitmap_ones (wont_exit
);
992 RESET_BIT (wont_exit
, factor
- 1);
994 ok
= tree_duplicate_loop_to_header_edge
995 (loop
, loop_latch_edge (loop
), factor
- 1,
996 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1000 for (i
= 0; VEC_iterate (edge
, to_remove
, i
, e
); i
++)
1002 ok
= remove_path (e
);
1005 VEC_free (edge
, heap
, to_remove
);
1006 update_ssa (TODO_update_ssa
);
1008 /* Ensure that the frequencies in the loop match the new estimated
1009 number of iterations, and change the probability of the new
1011 freq_h
= loop
->header
->frequency
;
1012 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1014 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1016 exit_bb
= single_pred (loop
->latch
);
1017 new_exit
= find_edge (exit_bb
, rest
);
1018 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1019 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1021 rest
->count
+= new_exit
->count
;
1022 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1024 new_nonexit
= single_pred_edge (loop
->latch
);
1025 prob
= new_nonexit
->probability
;
1026 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1027 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1028 if (new_nonexit
->count
< 0)
1029 new_nonexit
->count
= 0;
1030 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1033 /* Finally create the new counter for number of iterations and add the new
1034 exit instruction. */
1035 bsi
= bsi_last (exit_bb
);
1036 exit_if
= bsi_stmt (bsi
);
1037 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1038 &bsi
, false, &ctr_before
, &ctr_after
);
1039 COND_EXPR_COND (exit_if
) = build2 (exit_cmp
, boolean_type_node
, ctr_after
,
1041 update_stmt (exit_if
);
1043 #ifdef ENABLE_CHECKING
1044 verify_flow_info ();
1045 verify_dominators (CDI_DOMINATORS
);
1046 verify_loop_structure ();
1047 verify_loop_closed_ssa ();
1051 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1052 want to transform the loop before unrolling. The meaning
1053 of the arguments is the same as for tree_transform_and_unroll_loop. */
1056 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1057 edge exit
, struct tree_niter_desc
*desc
)
1059 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,