1 /* High-level loop manipulation functions.
2 Copyright (C) 2004, 2005 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
))
89 if (!tree_expr_nonnegative_p (step
)
90 && may_negate_without_overflow_p (step
))
93 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
98 /* Gimplify the step if necessary. We put the computations in front of the
99 loop (i.e. the step should be loop invariant). */
100 step
= force_gimple_operand (step
, &stmts
, true, var
);
102 bsi_insert_on_edge_immediate (pe
, stmts
);
104 stmt
= build2_gimple (GIMPLE_MODIFY_STMT
, va
,
105 build2 (incr_op
, TREE_TYPE (base
), vb
, step
));
106 SSA_NAME_DEF_STMT (va
) = stmt
;
108 bsi_insert_after (incr_pos
, stmt
, BSI_NEW_STMT
);
110 bsi_insert_before (incr_pos
, stmt
, BSI_NEW_STMT
);
112 initial
= force_gimple_operand (base
, &stmts
, true, var
);
114 bsi_insert_on_edge_immediate (pe
, stmts
);
116 stmt
= create_phi_node (vb
, loop
->header
);
117 SSA_NAME_DEF_STMT (vb
) = stmt
;
118 add_phi_arg (stmt
, initial
, loop_preheader_edge (loop
));
119 add_phi_arg (stmt
, va
, loop_latch_edge (loop
));
122 /* Add exit phis for the USE on EXIT. */
125 add_exit_phis_edge (basic_block exit
, tree use
)
127 tree phi
, def_stmt
= SSA_NAME_DEF_STMT (use
);
128 basic_block def_bb
= bb_for_stmt (def_stmt
);
129 struct loop
*def_loop
;
133 /* Check that some of the edges entering the EXIT block exits a loop in
134 that USE is defined. */
135 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
137 def_loop
= find_common_loop (def_bb
->loop_father
, e
->src
->loop_father
);
138 if (!flow_bb_inside_loop_p (def_loop
, e
->dest
))
145 phi
= create_phi_node (use
, exit
);
146 create_new_def_for (PHI_RESULT (phi
), phi
, PHI_RESULT_PTR (phi
));
147 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
148 add_phi_arg (phi
, use
, e
);
151 /* Add exit phis for VAR that is used in LIVEIN.
152 Exits of the loops are stored in EXITS. */
155 add_exit_phis_var (tree var
, bitmap livein
, bitmap exits
)
159 basic_block def_bb
= bb_for_stmt (SSA_NAME_DEF_STMT (var
));
162 if (is_gimple_reg (var
))
163 bitmap_clear_bit (livein
, def_bb
->index
);
165 bitmap_set_bit (livein
, def_bb
->index
);
167 def
= BITMAP_ALLOC (NULL
);
168 bitmap_set_bit (def
, def_bb
->index
);
169 compute_global_livein (livein
, def
);
172 EXECUTE_IF_AND_IN_BITMAP (exits
, livein
, 0, index
, bi
)
174 add_exit_phis_edge (BASIC_BLOCK (index
), var
);
178 /* Add exit phis for the names marked in NAMES_TO_RENAME.
179 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
180 names are used are stored in USE_BLOCKS. */
183 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap loop_exits
)
188 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
190 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
194 /* Returns a bitmap of all loop exit edge targets. */
197 get_loops_exits (void)
199 bitmap exits
= BITMAP_ALLOC (NULL
);
206 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
207 if (e
->src
!= ENTRY_BLOCK_PTR
208 && !flow_bb_inside_loop_p (e
->src
->loop_father
, bb
))
210 bitmap_set_bit (exits
, bb
->index
);
218 /* For USE in BB, if it is used outside of the loop it is defined in,
219 mark it for rewrite. Record basic block BB where it is used
220 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
223 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
228 struct loop
*def_loop
;
230 if (TREE_CODE (use
) != SSA_NAME
)
233 /* We don't need to keep virtual operands in loop-closed form. */
234 if (!is_gimple_reg (use
))
237 ver
= SSA_NAME_VERSION (use
);
238 def_bb
= bb_for_stmt (SSA_NAME_DEF_STMT (use
));
241 def_loop
= def_bb
->loop_father
;
243 /* If the definition is not inside loop, it is not interesting. */
244 if (!def_loop
->outer
)
247 if (!use_blocks
[ver
])
248 use_blocks
[ver
] = BITMAP_ALLOC (NULL
);
249 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
251 bitmap_set_bit (need_phis
, ver
);
254 /* For uses in STMT, mark names that are used outside of the loop they are
255 defined to rewrite. Record the set of blocks in that the ssa
256 names are defined to USE_BLOCKS and the ssa names themselves to
260 find_uses_to_rename_stmt (tree stmt
, bitmap
*use_blocks
, bitmap need_phis
)
264 basic_block bb
= bb_for_stmt (stmt
);
266 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_ALL_USES
)
267 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
270 /* Marks names that are used in BB and outside of the loop they are
271 defined in for rewrite. Records the set of blocks in that the ssa
272 names are defined to USE_BLOCKS. Record the SSA names that will
273 need exit PHIs in NEED_PHIS. */
276 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
278 block_stmt_iterator bsi
;
283 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
284 for (phi
= phi_nodes (e
->dest
); phi
; phi
= PHI_CHAIN (phi
))
285 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
286 use_blocks
, need_phis
);
288 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
289 find_uses_to_rename_stmt (bsi_stmt (bsi
), use_blocks
, need_phis
);
292 /* Marks names that are used outside of the loop they are defined in
293 for rewrite. Records the set of blocks in that the ssa
294 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
295 scan only blocks in this set. */
298 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
304 if (changed_bbs
&& !bitmap_empty_p (changed_bbs
))
306 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
308 find_uses_to_rename_bb (BASIC_BLOCK (index
), use_blocks
, need_phis
);
315 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
320 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
321 phi nodes to ensure that no variable is used outside the loop it is
324 This strengthening of the basic ssa form has several advantages:
326 1) Updating it during unrolling/peeling/versioning is trivial, since
327 we do not need to care about the uses outside of the loop.
328 2) The behavior of all uses of an induction variable is the same.
329 Without this, you need to distinguish the case when the variable
330 is used outside of the loop it is defined in, for example
332 for (i = 0; i < 100; i++)
334 for (j = 0; j < 100; j++)
342 Looking from the outer loop with the normal SSA form, the first use of k
343 is not well-behaved, while the second one is an induction variable with
346 If CHANGED_BBS is not NULL, we look for uses outside loops only in
347 the basic blocks in this set.
349 UPDATE_FLAG is used in the call to update_ssa. See
350 TODO_update_ssa* for documentation. */
353 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
355 bitmap loop_exits
= get_loops_exits ();
357 unsigned i
, old_num_ssa_names
;
358 bitmap names_to_rename
= BITMAP_ALLOC (NULL
);
360 /* If the pass has caused the SSA form to be out-of-date, update it
362 update_ssa (update_flag
);
364 old_num_ssa_names
= num_ssa_names
;
365 use_blocks
= XCNEWVEC (bitmap
, old_num_ssa_names
);
367 /* Find the uses outside loops. */
368 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
370 /* Add the PHI nodes on exits of the loops for the names we need to
372 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
374 for (i
= 0; i
< old_num_ssa_names
; i
++)
375 BITMAP_FREE (use_blocks
[i
]);
377 BITMAP_FREE (loop_exits
);
378 BITMAP_FREE (names_to_rename
);
380 /* Fix up all the names found to be used outside their original
382 update_ssa (TODO_update_ssa
);
385 /* Check invariants of the loop closed ssa form for the USE in BB. */
388 check_loop_closed_ssa_use (basic_block bb
, tree use
)
393 if (TREE_CODE (use
) != SSA_NAME
|| !is_gimple_reg (use
))
396 def
= SSA_NAME_DEF_STMT (use
);
397 def_bb
= bb_for_stmt (def
);
399 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
402 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
405 check_loop_closed_ssa_stmt (basic_block bb
, tree stmt
)
410 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_ALL_USES
)
411 check_loop_closed_ssa_use (bb
, var
);
414 /* Checks that invariants of the loop closed ssa form are preserved. */
417 verify_loop_closed_ssa (void)
420 block_stmt_iterator bsi
;
424 if (current_loops
== NULL
)
431 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
432 for (i
= 0; i
< (unsigned) PHI_NUM_ARGS (phi
); i
++)
433 check_loop_closed_ssa_use (PHI_ARG_EDGE (phi
, i
)->src
,
434 PHI_ARG_DEF (phi
, i
));
436 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
437 check_loop_closed_ssa_stmt (bb
, bsi_stmt (bsi
));
441 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
442 preserve the loop closed ssa form. */
445 split_loop_exit_edge (edge exit
)
447 basic_block dest
= exit
->dest
;
448 basic_block bb
= split_edge (exit
);
449 tree phi
, new_phi
, new_name
, name
;
452 for (phi
= phi_nodes (dest
); phi
; phi
= PHI_CHAIN (phi
))
454 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
456 name
= USE_FROM_PTR (op_p
);
458 /* If the argument of the PHI node is a constant, we do not need
459 to keep it inside loop. */
460 if (TREE_CODE (name
) != SSA_NAME
)
463 /* Otherwise create an auxiliary phi node that will copy the value
464 of the SSA name out of the loop. */
465 new_name
= duplicate_ssa_name (name
, NULL
);
466 new_phi
= create_phi_node (new_name
, bb
);
467 SSA_NAME_DEF_STMT (new_name
) = new_phi
;
468 add_phi_arg (new_phi
, name
, exit
);
469 SET_USE (op_p
, new_name
);
473 /* Returns the basic block in that statements should be emitted for induction
474 variables incremented at the end of the LOOP. */
477 ip_end_pos (struct loop
*loop
)
482 /* Returns the basic block in that statements should be emitted for induction
483 variables incremented just before exit condition of a LOOP. */
486 ip_normal_pos (struct loop
*loop
)
492 if (!single_pred_p (loop
->latch
))
495 bb
= single_pred (loop
->latch
);
496 last
= last_stmt (bb
);
497 if (TREE_CODE (last
) != COND_EXPR
)
500 exit
= EDGE_SUCC (bb
, 0);
501 if (exit
->dest
== loop
->latch
)
502 exit
= EDGE_SUCC (bb
, 1);
504 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
510 /* Stores the standard position for induction variable increment in LOOP
511 (just before the exit condition if it is available and latch block is empty,
512 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
513 the increment should be inserted after *BSI. */
516 standard_iv_increment_position (struct loop
*loop
, block_stmt_iterator
*bsi
,
519 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
520 tree last
= last_stmt (latch
);
523 || (last
&& TREE_CODE (last
) != LABEL_EXPR
))
525 *bsi
= bsi_last (latch
);
526 *insert_after
= true;
530 *bsi
= bsi_last (bb
);
531 *insert_after
= false;
535 /* Copies phi node arguments for duplicated blocks. The index of the first
536 duplicated block is FIRST_NEW_BLOCK. */
539 copy_phi_node_args (unsigned first_new_block
)
543 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
544 BASIC_BLOCK (i
)->flags
|= BB_DUPLICATED
;
546 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
547 add_phi_args_after_copy_bb (BASIC_BLOCK (i
));
549 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
550 BASIC_BLOCK (i
)->flags
&= ~BB_DUPLICATED
;
554 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
555 updates the PHI nodes at start of the copied region. In order to
556 achieve this, only loops whose exits all lead to the same location
559 Notice that we do not completely update the SSA web after
560 duplication. The caller is responsible for calling update_ssa
561 after the loop has been duplicated. */
564 tree_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
565 unsigned int ndupl
, sbitmap wont_exit
,
566 edge orig
, VEC (edge
, heap
) **to_remove
,
569 unsigned first_new_block
;
571 if (!(current_loops
->state
& LOOPS_HAVE_SIMPLE_LATCHES
))
573 if (!(current_loops
->state
& LOOPS_HAVE_PREHEADERS
))
576 #ifdef ENABLE_CHECKING
577 verify_loop_closed_ssa ();
580 first_new_block
= last_basic_block
;
581 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
582 orig
, to_remove
, flags
))
585 /* Readd the removed phi args for e. */
586 flush_pending_stmts (e
);
588 /* Copy the phi node arguments. */
589 copy_phi_node_args (first_new_block
);
596 /* Build if (COND) goto THEN_LABEL; else goto ELSE_LABEL; */
599 build_if_stmt (tree cond
, tree then_label
, tree else_label
)
601 return build3 (COND_EXPR
, void_type_node
,
603 build1 (GOTO_EXPR
, void_type_node
, then_label
),
604 build1 (GOTO_EXPR
, void_type_node
, else_label
));
607 /* Returns true if we can unroll LOOP FACTOR times. Number
608 of iterations of the loop is returned in NITER. */
611 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
612 struct tree_niter_desc
*niter
)
616 /* Check whether unrolling is possible. We only want to unroll loops
617 for that we are able to determine number of iterations. We also
618 want to split the extra iterations of the loop from its end,
619 therefore we require that the loop has precisely one
622 exit
= single_dom_exit (loop
);
626 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
627 || niter
->cmp
== ERROR_MARK
628 /* Scalar evolutions analysis might have copy propagated
629 the abnormal ssa names into these expressions, hence
630 emitting the computations based on them during loop
631 unrolling might create overlapping life ranges for
632 them, and failures in out-of-ssa. */
633 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
634 || contains_abnormal_ssa_name_p (niter
->control
.base
)
635 || contains_abnormal_ssa_name_p (niter
->control
.step
)
636 || contains_abnormal_ssa_name_p (niter
->bound
))
639 /* And of course, we must be able to duplicate the loop. */
640 if (!can_duplicate_loop_p (loop
))
643 /* The final loop should be small enough. */
644 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
645 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
651 /* Determines the conditions that control execution of LOOP unrolled FACTOR
652 times. DESC is number of iterations of LOOP. ENTER_COND is set to
653 condition that must be true if the main loop can be entered.
654 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
655 how the exit from the unrolled loop should be controlled. */
658 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
659 unsigned factor
, tree
*enter_cond
,
660 tree
*exit_base
, tree
*exit_step
,
661 enum tree_code
*exit_cmp
, tree
*exit_bound
)
664 tree base
= desc
->control
.base
;
665 tree step
= desc
->control
.step
;
666 tree bound
= desc
->bound
;
667 tree type
= TREE_TYPE (base
);
669 tree min
= lower_bound_in_type (type
, type
);
670 tree max
= upper_bound_in_type (type
, type
);
671 enum tree_code cmp
= desc
->cmp
;
672 tree cond
= boolean_true_node
, assum
;
674 *enter_cond
= boolean_false_node
;
675 *exit_base
= NULL_TREE
;
676 *exit_step
= NULL_TREE
;
677 *exit_cmp
= ERROR_MARK
;
678 *exit_bound
= NULL_TREE
;
679 gcc_assert (cmp
!= ERROR_MARK
);
681 /* We only need to be correct when we answer question
682 "Do at least FACTOR more iterations remain?" in the unrolled loop.
683 Thus, transforming BASE + STEP * i <> BOUND to
684 BASE + STEP * i < BOUND is ok. */
687 if (tree_int_cst_sign_bit (step
))
692 else if (cmp
== LT_EXPR
)
694 gcc_assert (!tree_int_cst_sign_bit (step
));
696 else if (cmp
== GT_EXPR
)
698 gcc_assert (tree_int_cst_sign_bit (step
));
703 /* The main body of the loop may be entered iff:
705 1) desc->may_be_zero is false.
706 2) it is possible to check that there are at least FACTOR iterations
707 of the loop, i.e., BOUND - step * FACTOR does not overflow.
708 3) # of iterations is at least FACTOR */
710 if (!integer_zerop (desc
->may_be_zero
))
711 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
712 invert_truthvalue (desc
->may_be_zero
),
715 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
716 build_int_cst_type (type
, factor
));
717 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
719 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
721 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
723 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
725 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
726 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
728 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
729 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
730 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
732 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
734 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
735 /* cond now may be a gimple comparison, which would be OK, but also any
736 other gimple rhs (say a && b). In this case we need to force it to
738 if (!is_gimple_condexpr (cond
))
740 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
742 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
746 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
748 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
749 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
751 bsi_insert_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
754 *exit_step
= bigstep
;
759 /* Scales the frequencies of all basic blocks in LOOP that are strictly
760 dominated by BB by NUM/DEN. */
763 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
771 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
773 son
= next_dom_son (CDI_DOMINATORS
, son
))
775 if (!flow_bb_inside_loop_p (loop
, son
))
777 scale_bbs_frequencies_int (&son
, 1, num
, den
);
778 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
782 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
783 EXIT is the exit of the loop to that DESC corresponds.
785 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
786 under that loop exits in the first iteration even if N != 0,
790 x = phi (init, next);
798 becomes (with possibly the exit conditions formulated a bit differently,
799 avoiding the need to create a new iv):
801 if (MAY_BE_ZERO || N < FACTOR)
806 x = phi (init, next);
817 } while (N >= FACTOR);
820 init' = phi (init, x);
824 x = phi (init', next);
832 Before the loop is unrolled, TRANSFORM is called for it (only for the
833 unrolled loop, but not for its versioned copy). DATA is passed to
836 /* Probability in % that the unrolled loop is entered. Just a guess. */
837 #define PROB_UNROLLED_LOOP_ENTERED 90
840 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
841 edge exit
, struct tree_niter_desc
*desc
,
842 transform_callback transform
,
845 tree exit_if
, ctr_before
, ctr_after
;
846 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
847 enum tree_code exit_cmp
;
848 tree phi_old_loop
, phi_new_loop
, phi_rest
, init
, next
, new_init
, var
;
849 struct loop
*new_loop
;
850 basic_block rest
, exit_bb
;
851 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
853 block_stmt_iterator bsi
;
856 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
857 unsigned new_est_niter
, i
, prob
;
858 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
860 VEC (edge
, heap
) *to_remove
= NULL
;
862 est_niter
= expected_loop_iterations (loop
);
863 determine_exit_conditions (loop
, desc
, factor
,
864 &enter_main_cond
, &exit_base
, &exit_step
,
865 &exit_cmp
, &exit_bound
);
867 /* Let us assume that the unrolled loop is quite likely to be entered. */
868 if (integer_nonzerop (enter_main_cond
))
869 prob_entry
= REG_BR_PROB_BASE
;
871 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
873 /* The values for scales should keep profile consistent, and somewhat close
876 TODO: The current value of SCALE_REST makes it appear that the loop that
877 is created by splitting the remaining iterations of the unrolled loop is
878 executed the same number of times as the original loop, and with the same
879 frequencies, which is obviously wrong. This does not appear to cause
880 problems, so we do not bother with fixing it for now. To make the profile
881 correct, we would need to change the probability of the exit edge of the
882 loop, and recompute the distribution of frequencies in its body because
883 of this change (scale the frequencies of blocks before and after the exit
884 by appropriate factors). */
885 scale_unrolled
= prob_entry
;
886 scale_rest
= REG_BR_PROB_BASE
;
888 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
889 prob_entry
, scale_unrolled
, scale_rest
, true);
890 gcc_assert (new_loop
!= NULL
);
891 update_ssa (TODO_update_ssa
);
893 /* Determine the probability of the exit edge of the unrolled loop. */
894 new_est_niter
= est_niter
/ factor
;
896 /* Without profile feedback, loops for that we do not know a better estimate
897 are assumed to roll 10 times. When we unroll such loop, it appears to
898 roll too little, and it may even seem to be cold. To avoid this, we
899 ensure that the created loop appears to roll at least 5 times (but at
900 most as many times as before unrolling). */
901 if (new_est_niter
< 5)
904 new_est_niter
= est_niter
;
909 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
910 loop latch (and make its condition dummy, for the moment). */
911 rest
= loop_preheader_edge (new_loop
)->src
;
912 precond_edge
= single_pred_edge (rest
);
913 split_edge (loop_latch_edge (loop
));
914 exit_bb
= single_pred (loop
->latch
);
916 /* Since the exit edge will be removed, the frequency of all the blocks
917 in the loop that are dominated by it must be scaled by
918 1 / (1 - exit->probability). */
919 scale_dominated_blocks_in_loop (loop
, exit
->src
,
921 REG_BR_PROB_BASE
- exit
->probability
);
923 bsi
= bsi_last (exit_bb
);
924 exit_if
= build_if_stmt (boolean_true_node
,
925 tree_block_label (loop
->latch
),
926 tree_block_label (rest
));
927 bsi_insert_after (&bsi
, exit_if
, BSI_NEW_STMT
);
928 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
929 rescan_loop_exit (new_exit
, true, false);
931 /* Set the probability of new exit to the same of the old one. Fix
932 the frequency of the latch block, by scaling it back by
933 1 - exit->probability. */
934 new_exit
->count
= exit
->count
;
935 new_exit
->probability
= exit
->probability
;
936 new_nonexit
= single_pred_edge (loop
->latch
);
937 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
938 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
939 new_nonexit
->count
-= exit
->count
;
940 if (new_nonexit
->count
< 0)
941 new_nonexit
->count
= 0;
942 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
945 old_entry
= loop_preheader_edge (loop
);
946 new_entry
= loop_preheader_edge (new_loop
);
947 old_latch
= loop_latch_edge (loop
);
948 for (phi_old_loop
= phi_nodes (loop
->header
),
949 phi_new_loop
= phi_nodes (new_loop
->header
);
951 phi_old_loop
= PHI_CHAIN (phi_old_loop
),
952 phi_new_loop
= PHI_CHAIN (phi_new_loop
))
954 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
955 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
956 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
957 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
959 /* Prefer using original variable as a base for the new ssa name.
960 This is necessary for virtual ops, and useful in order to avoid
961 losing debug info for real ops. */
962 if (TREE_CODE (next
) == SSA_NAME
)
963 var
= SSA_NAME_VAR (next
);
964 else if (TREE_CODE (init
) == SSA_NAME
)
965 var
= SSA_NAME_VAR (init
);
968 var
= create_tmp_var (TREE_TYPE (init
), "unrinittmp");
969 add_referenced_var (var
);
972 new_init
= make_ssa_name (var
, NULL_TREE
);
973 phi_rest
= create_phi_node (new_init
, rest
);
974 SSA_NAME_DEF_STMT (new_init
) = phi_rest
;
976 add_phi_arg (phi_rest
, init
, precond_edge
);
977 add_phi_arg (phi_rest
, next
, new_exit
);
978 SET_USE (op
, new_init
);
983 /* Transform the loop. */
985 (*transform
) (loop
, data
);
987 /* Unroll the loop and remove the exits in all iterations except for the
989 wont_exit
= sbitmap_alloc (factor
);
990 sbitmap_ones (wont_exit
);
991 RESET_BIT (wont_exit
, factor
- 1);
993 ok
= tree_duplicate_loop_to_header_edge
994 (loop
, loop_latch_edge (loop
), factor
- 1,
995 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
999 for (i
= 0; VEC_iterate (edge
, to_remove
, i
, e
); i
++)
1001 ok
= remove_path (e
);
1004 VEC_free (edge
, heap
, to_remove
);
1005 update_ssa (TODO_update_ssa
);
1007 /* Ensure that the frequencies in the loop match the new estimated
1008 number of iterations, and change the probability of the new
1010 freq_h
= loop
->header
->frequency
;
1011 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1013 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1015 exit_bb
= single_pred (loop
->latch
);
1016 new_exit
= find_edge (exit_bb
, rest
);
1017 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1018 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1020 rest
->count
+= new_exit
->count
;
1021 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1023 new_nonexit
= single_pred_edge (loop
->latch
);
1024 prob
= new_nonexit
->probability
;
1025 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1026 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1027 if (new_nonexit
->count
< 0)
1028 new_nonexit
->count
= 0;
1029 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1032 /* Finally create the new counter for number of iterations and add the new
1033 exit instruction. */
1034 bsi
= bsi_last (exit_bb
);
1035 exit_if
= bsi_stmt (bsi
);
1036 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1037 &bsi
, false, &ctr_before
, &ctr_after
);
1038 COND_EXPR_COND (exit_if
) = build2 (exit_cmp
, boolean_type_node
, ctr_after
,
1040 update_stmt (exit_if
);
1042 #ifdef ENABLE_CHECKING
1043 verify_flow_info ();
1044 verify_dominators (CDI_DOMINATORS
);
1045 verify_loop_structure ();
1046 verify_loop_closed_ssa ();
1050 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1051 want to transform the loop before unrolling. The meaning
1052 of the arguments is the same as for tree_transform_and_unroll_loop. */
1055 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1056 edge exit
, struct tree_niter_desc
*desc
)
1058 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,