1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2023 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 "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
29 #include "gimple-pretty-print.h"
30 #include "fold-const.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
36 #include "tree-ssa-loop-ivopts.h"
37 #include "tree-ssa-loop-manip.h"
38 #include "tree-ssa-loop-niter.h"
39 #include "tree-ssa-loop.h"
40 #include "tree-into-ssa.h"
43 #include "tree-scalar-evolution.h"
44 #include "tree-inline.h"
46 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
47 so that we can free them all at once. */
48 static bitmap_obstack loop_renamer_obstack
;
50 /* Creates an induction variable with value BASE (+/-) STEP * iteration in LOOP.
51 If INCR_OP is PLUS_EXPR, the induction variable is BASE + STEP * iteration.
52 If INCR_OP is MINUS_EXPR, the induction variable is BASE - STEP * iteration.
53 It is expected that neither BASE nor STEP are shared with other expressions
54 (unless the sharing rules allow this). Use VAR as a base var_decl for it
55 (if NULL, a new temporary will be created). The increment will occur at
56 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
57 AFTER can be computed using standard_iv_increment_position. The ssa versions
58 of the variable before and after increment will be stored in VAR_BEFORE and
59 VAR_AFTER (unless they are NULL). */
62 create_iv (tree base
, tree_code incr_op
, tree step
, tree var
, class loop
*loop
,
63 gimple_stmt_iterator
*incr_pos
, bool after
, tree
*var_before
,
71 gcc_assert (incr_op
== PLUS_EXPR
|| incr_op
== MINUS_EXPR
);
72 edge pe
= loop_preheader_edge (loop
);
76 vb
= make_ssa_name (var
);
77 va
= make_ssa_name (var
);
81 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
82 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
89 /* For easier readability of the created code, produce MINUS_EXPRs
91 if (TREE_CODE (step
) == INTEGER_CST
)
93 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
95 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
96 if (tree_int_cst_lt (step1
, step
))
98 incr_op
= (incr_op
== PLUS_EXPR
? MINUS_EXPR
: PLUS_EXPR
);
106 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
107 && may_negate_without_overflow_p (step
))
109 incr_op
= (incr_op
== PLUS_EXPR
? MINUS_EXPR
: PLUS_EXPR
);
110 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
114 if (POINTER_TYPE_P (TREE_TYPE (base
)))
116 if (TREE_CODE (base
) == ADDR_EXPR
)
117 mark_addressable (TREE_OPERAND (base
, 0));
118 step
= convert_to_ptrofftype (step
);
119 if (incr_op
== MINUS_EXPR
)
120 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
121 incr_op
= POINTER_PLUS_EXPR
;
123 /* Gimplify the step if necessary. We put the computations in front of the
124 loop (i.e. the step should be loop invariant). */
125 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
127 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
129 stmt
= gimple_build_assign (va
, incr_op
, vb
, step
);
130 /* Prevent the increment from inheriting a bogus location if it is not put
131 immediately after a statement whose location is known. */
134 gimple_stmt_iterator gsi
= *incr_pos
;
135 if (!gsi_end_p (gsi
))
136 gsi_next_nondebug (&gsi
);
139 edge e
= single_succ_edge (gsi_bb (*incr_pos
));
140 gimple_set_location (stmt
, e
->goto_locus
);
142 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
146 gimple_stmt_iterator gsi
= *incr_pos
;
147 if (!gsi_end_p (gsi
) && is_gimple_debug (gsi_stmt (gsi
)))
148 gsi_next_nondebug (&gsi
);
149 if (!gsi_end_p (gsi
))
150 gimple_set_location (stmt
, gimple_location (gsi_stmt (gsi
)));
151 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
154 initial
= force_gimple_operand (base
, &stmts
, true, var
);
156 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
158 phi
= create_phi_node (vb
, loop
->header
);
159 add_phi_arg (phi
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
160 add_phi_arg (phi
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
163 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
164 both DEF_LOOP and USE_LOOP. */
166 static inline class loop
*
167 find_sibling_superloop (class loop
*use_loop
, class loop
*def_loop
)
169 unsigned ud
= loop_depth (use_loop
);
170 unsigned dd
= loop_depth (def_loop
);
171 gcc_assert (ud
> 0 && dd
> 0);
173 use_loop
= superloop_at_depth (use_loop
, dd
);
175 def_loop
= superloop_at_depth (def_loop
, ud
);
176 while (loop_outer (use_loop
) != loop_outer (def_loop
))
178 use_loop
= loop_outer (use_loop
);
179 def_loop
= loop_outer (def_loop
);
180 gcc_assert (use_loop
&& def_loop
);
185 /* DEF_BB is a basic block containing a DEF that needs rewriting into
186 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
187 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
188 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_BB).
189 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
190 DEF_LOOP_EXITS is a bitmap of loop exit blocks that exit the loop
191 containing DEF_BB or its outer loops.
193 Compute the subset of loop exit destinations that exit the loop
194 containing DEF_BB or one of its loop fathers, in which DEF is live.
195 This set is returned in the bitmap LIVE_EXITS.
197 Instead of computing the complete livein set of the def, we use the loop
198 nesting tree as a form of poor man's structure analysis. This greatly
199 speeds up the analysis, which is important because this function may be
200 called on all SSA names that need rewriting, one at a time. */
203 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
204 basic_block def_bb
, bitmap def_loop_exits
)
208 class loop
*def_loop
= def_bb
->loop_father
;
209 unsigned def_loop_depth
= loop_depth (def_loop
);
211 /* Normally the work list size is bounded by the number of basic
212 blocks in the largest loop. We don't know this number, but we
213 can be fairly sure that it will be relatively small. */
214 auto_vec
<basic_block
, 8> worklist (MAX (8, n_basic_blocks_for_fn (cfun
) / 128));
216 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
218 basic_block use_bb
= BASIC_BLOCK_FOR_FN (cfun
, i
);
219 class loop
*use_loop
= use_bb
->loop_father
;
220 gcc_checking_assert (def_loop
!= use_loop
221 && ! flow_loop_nested_p (def_loop
, use_loop
));
222 if (! flow_loop_nested_p (use_loop
, def_loop
))
223 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
224 if (bitmap_set_bit (live_exits
, use_bb
->index
))
225 worklist
.safe_push (use_bb
);
228 /* Iterate until the worklist is empty. */
229 while (! worklist
.is_empty ())
234 /* Pull a block off the worklist. */
235 basic_block bb
= worklist
.pop ();
237 /* Make sure we have at least enough room in the work list
238 for all predecessors of this block. */
239 worklist
.reserve (EDGE_COUNT (bb
->preds
));
241 /* For each predecessor block. */
242 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
244 basic_block pred
= e
->src
;
245 class loop
*pred_loop
= pred
->loop_father
;
246 unsigned pred_loop_depth
= loop_depth (pred_loop
);
249 /* We should have met DEF_BB along the way. */
250 gcc_assert (pred
!= ENTRY_BLOCK_PTR_FOR_FN (cfun
));
252 if (pred_loop_depth
>= def_loop_depth
)
254 if (pred_loop_depth
> def_loop_depth
)
255 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
256 /* If we've reached DEF_LOOP, our train ends here. */
257 if (pred_loop
== def_loop
)
260 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
261 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
263 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
264 we had already added PRED to LIVEIN before. */
265 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
267 /* If we have visited PRED before, don't add it to the worklist.
268 If BB dominates PRED, then we're probably looking at a loop.
269 We're only interested in looking up in the dominance tree
270 because DEF_BB dominates all the uses. */
271 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
274 worklist
.quick_push (pred
);
278 bitmap_and_into (live_exits
, def_loop_exits
);
281 /* Add a loop-closing PHI for VAR in basic block EXIT. */
284 add_exit_phi (basic_block exit
, tree var
)
290 /* Check that at least one of the edges entering the EXIT block exits
291 the loop, or a superloop of that loop, that VAR is defined in. */
294 gimple
*def_stmt
= SSA_NAME_DEF_STMT (var
);
295 basic_block def_bb
= gimple_bb (def_stmt
);
296 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
298 class loop
*aloop
= find_common_loop (def_bb
->loop_father
,
299 e
->src
->loop_father
);
300 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
306 phi
= create_phi_node (NULL_TREE
, exit
);
307 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
308 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
309 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
311 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
313 fprintf (dump_file
, ";; Created LCSSA PHI: ");
314 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
318 /* Add exit phis for VAR that is used in LIVEIN.
319 Exits of the loops are stored in LOOP_EXITS. Returns the number
320 of PHIs added for VAR. */
323 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap def_loop_exits
)
327 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
329 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
331 auto_bitmap
live_exits (&loop_renamer_obstack
);
332 compute_live_loop_exits (live_exits
, use_blocks
, def_bb
, def_loop_exits
);
335 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
337 add_exit_phi (BASIC_BLOCK_FOR_FN (cfun
, index
), var
);
344 loop_name_cmp (const void *p1
, const void *p2
)
346 auto l1
= (const std::pair
<int, int> *)p1
;
347 auto l2
= (const std::pair
<int, int> *)p2
;
348 if (l1
->first
< l2
->first
)
350 else if (l1
->first
> l2
->first
)
355 /* Add exit phis for the names marked in NAMES_TO_RENAME.
356 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
357 names are used are stored in USE_BLOCKS. Returns whether any name
358 required multiple LC PHI nodes. */
361 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
)
365 bool multiple_p
= false;
367 /* Sort names_to_rename after definition loop so we can avoid re-computing
369 auto_vec
<std::pair
<int, int> > names (bitmap_count_bits (names_to_rename
));
370 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
372 tree name
= ssa_name (i
);
373 loop_p def_loop
= gimple_bb (SSA_NAME_DEF_STMT (name
))->loop_father
;
374 names
.quick_push (std::make_pair (def_loop
->num
, i
));
376 names
.qsort (loop_name_cmp
);
378 auto_bitmap
def_loop_exits (&loop_renamer_obstack
);
379 loop_p last_def_loop
= NULL
;
382 loop_p def_loop
= get_loop (cfun
, p
.first
);
383 if (def_loop
!= last_def_loop
)
385 bitmap_clear (def_loop_exits
);
386 last_def_loop
= def_loop
;
387 for (class loop
*loop
= def_loop
; loop
!= current_loops
->tree_root
;
388 loop
= loop_outer (loop
))
389 for (auto exit
= loop
->exits
->next
; exit
->e
; exit
= exit
->next
)
390 bitmap_set_bit (def_loop_exits
, exit
->e
->dest
->index
);
392 if (add_exit_phis_var (ssa_name (p
.second
), use_blocks
[p
.second
],
400 /* For USE in BB, if it is used outside of the loop it is defined in,
401 mark it for rewrite. Record basic block BB where it is used
402 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap.
403 Note that for USEs in phis, BB should be the src of the edge corresponding to
404 the use, rather than the bb containing the phi. */
407 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
412 class loop
*def_loop
;
414 if (TREE_CODE (use
) != SSA_NAME
)
417 ver
= SSA_NAME_VERSION (use
);
418 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
421 def_loop
= def_bb
->loop_father
;
423 /* If the definition is not inside a loop, it is not interesting. */
424 if (!loop_outer (def_loop
))
427 /* If the use is not outside of the loop it is defined in, it is not
429 if (flow_bb_inside_loop_p (def_loop
, bb
))
432 /* If we're seeing VER for the first time, we still have to allocate
433 a bitmap for its uses. */
434 if (bitmap_set_bit (need_phis
, ver
))
435 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
436 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
439 /* For uses matching USE_FLAGS in STMT, mark names that are used outside of the
440 loop they are defined to rewrite. Record the set of blocks in which the ssa
441 names are used to USE_BLOCKS, and the ssa names themselves to NEED_PHIS. */
444 find_uses_to_rename_stmt (gimple
*stmt
, bitmap
*use_blocks
, bitmap need_phis
,
449 basic_block bb
= gimple_bb (stmt
);
451 if (is_gimple_debug (stmt
))
454 /* FOR_EACH_SSA_TREE_OPERAND iterator does not allows SSA_OP_VIRTUAL_USES
456 if (use_flags
== SSA_OP_VIRTUAL_USES
)
458 tree vuse
= gimple_vuse (stmt
);
459 if (vuse
!= NULL_TREE
)
460 find_uses_to_rename_use (bb
, gimple_vuse (stmt
), use_blocks
, need_phis
);
463 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, use_flags
)
464 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
467 /* Marks names matching USE_FLAGS that are used in BB and outside of the loop
468 they are defined in for rewrite. Records the set of blocks in which the ssa
469 names are used to USE_BLOCKS. Record the SSA names that will
470 need exit PHIs in NEED_PHIS. */
473 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
,
478 bool do_virtuals
= (use_flags
& SSA_OP_VIRTUAL_USES
) != 0;
479 bool do_nonvirtuals
= (use_flags
& SSA_OP_USE
) != 0;
481 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
482 for (gphi_iterator bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
);
485 gphi
*phi
= bsi
.phi ();
486 bool virtual_p
= virtual_operand_p (gimple_phi_result (phi
));
487 if ((virtual_p
&& do_virtuals
)
488 || (!virtual_p
&& do_nonvirtuals
))
489 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
490 use_blocks
, need_phis
);
493 for (gimple_stmt_iterator bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
);
495 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
,
499 /* Marks names matching USE_FLAGS that are used outside of the loop they are
500 defined in for rewrite. Records the set of blocks in which the ssa names are
501 used to USE_BLOCKS. Record the SSA names that will need exit PHIs in
502 NEED_PHIS. If CHANGED_BBS is not NULL, scan only blocks in this set. */
505 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
,
513 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
515 bb
= BASIC_BLOCK_FOR_FN (cfun
, index
);
517 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
, use_flags
);
520 FOR_EACH_BB_FN (bb
, cfun
)
521 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
, use_flags
);
524 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
525 phi nodes to ensure that no variable is used outside the loop it is
528 This strengthening of the basic ssa form has several advantages:
530 1) Updating it during unrolling/peeling/versioning is trivial, since
531 we do not need to care about the uses outside of the loop.
532 The same applies to virtual operands which are also rewritten into
533 loop closed SSA form. Note that virtual operands are always live
535 2) The behavior of all uses of an induction variable is the same.
536 Without this, you need to distinguish the case when the variable
537 is used outside of the loop it is defined in, for example
539 for (i = 0; i < 100; i++)
541 for (j = 0; j < 100; j++)
549 Looking from the outer loop with the normal SSA form, the first use of k
550 is not well-behaved, while the second one is an induction variable with
553 If CHANGED_BBS is not NULL, we look for uses outside loops only in the
554 basic blocks in this set.
556 USE_FLAGS allows us to specify whether we want virtual, non-virtual or
557 both variables rewritten.
559 UPDATE_FLAG is used in the call to update_ssa. See
560 TODO_update_ssa* for documentation. */
563 rewrite_into_loop_closed_ssa_1 (bitmap changed_bbs
, unsigned update_flag
,
567 bitmap names_to_rename
;
569 loops_state_set (LOOP_CLOSED_SSA
);
570 if (number_of_loops (cfun
) <= 1)
573 /* If the pass has caused the SSA form to be out-of-date, update it
575 if (update_flag
!= 0)
576 update_ssa (update_flag
);
577 else if (flag_checking
)
578 verify_ssa (true, true);
580 bitmap_obstack_initialize (&loop_renamer_obstack
);
582 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
584 /* Uses of names to rename. We don't have to initialize this array,
585 because we know that we will only have entries for the SSA names
586 in NAMES_TO_RENAME. */
587 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
588 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
, use_flags
);
590 if (!bitmap_empty_p (names_to_rename
))
592 bool release_recorded_exits_p
= false;
593 if (!loops_state_satisfies_p (LOOPS_HAVE_RECORDED_EXITS
))
595 /* Doing one scan over the whole function is cheaper than
596 traversing the loop tree and gathering BBs of each loop. */
597 record_loop_exits ();
598 release_recorded_exits_p
= true;
601 /* Add the PHI nodes on exits of the loops for the names we need to
602 rewrite. When no variable required multiple LC PHI nodes to be
603 inserted then we know that all uses outside of the loop are
604 dominated by the single LC SSA definition and no further PHI
605 node insertions are required. */
606 bool need_phis_p
= add_exit_phis (names_to_rename
, use_blocks
);
608 if (release_recorded_exits_p
)
609 release_recorded_exits (cfun
);
611 /* Fix up all the names found to be used outside their original
613 update_ssa (need_phis_p
? TODO_update_ssa
: TODO_update_ssa_no_phi
);
616 bitmap_obstack_release (&loop_renamer_obstack
);
620 /* Rewrites the defs and uses into a loop closed ssa form.
621 If CHANGED_BBS is not NULL, we look for uses outside loops only in the basic
622 blocks in this set. UPDATE_FLAG is used in the call to update_ssa. See
623 TODO_update_ssa* for documentation. */
626 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
628 rewrite_into_loop_closed_ssa_1 (changed_bbs
, update_flag
, SSA_OP_ALL_USES
);
631 /* Check invariants of the loop closed ssa form for the def in DEF_BB. */
634 check_loop_closed_ssa_def (basic_block def_bb
, tree def
)
637 imm_use_iterator iterator
;
638 FOR_EACH_IMM_USE_FAST (use_p
, iterator
, def
)
640 if (is_gimple_debug (USE_STMT (use_p
)))
643 basic_block use_bb
= gimple_bb (USE_STMT (use_p
));
644 if (is_a
<gphi
*> (USE_STMT (use_p
)))
645 use_bb
= EDGE_PRED (use_bb
, PHI_ARG_INDEX_FROM_USE (use_p
))->src
;
647 gcc_assert (flow_bb_inside_loop_p (def_bb
->loop_father
, use_bb
));
651 /* Checks invariants of loop closed ssa form in BB. */
654 check_loop_closed_ssa_bb (basic_block bb
)
656 for (gphi_iterator bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
);
659 gphi
*phi
= bsi
.phi ();
661 check_loop_closed_ssa_def (bb
, PHI_RESULT (phi
));
664 for (gimple_stmt_iterator bsi
= gsi_start_nondebug_bb (bb
); !gsi_end_p (bsi
);
665 gsi_next_nondebug (&bsi
))
669 gimple
*stmt
= gsi_stmt (bsi
);
671 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_ALL_DEFS
)
672 check_loop_closed_ssa_def (bb
, var
);
676 /* Checks that invariants of the loop closed ssa form are preserved.
677 Call verify_ssa when VERIFY_SSA_P is true. Note all loops are checked
678 if LOOP is NULL, otherwise, only LOOP is checked. */
681 verify_loop_closed_ssa (bool verify_ssa_p
, class loop
*loop
)
683 if (number_of_loops (cfun
) <= 1)
686 timevar_push (TV_VERIFY_LOOP_CLOSED
);
693 verify_ssa (false, true);
695 FOR_EACH_BB_FN (bb
, cfun
)
696 if (bb
->loop_father
&& bb
->loop_father
->num
> 0)
697 check_loop_closed_ssa_bb (bb
);
701 basic_block
*bbs
= get_loop_body (loop
);
703 /* We do not have loop-local SSA verification so just
704 check there's no update queued. */
706 gcc_assert (!need_ssa_update_p (cfun
));
708 for (unsigned i
= 0; i
< loop
->num_nodes
; ++i
)
709 check_loop_closed_ssa_bb (bbs
[i
]);
714 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
717 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
718 preserve the loop closed ssa form. If COPY_CONSTANTS_P is true then
719 forwarder PHIs are also created for constant arguments.
720 The newly created block is returned. */
723 split_loop_exit_edge (edge exit
, bool copy_constants_p
)
725 basic_block dest
= exit
->dest
;
726 basic_block bb
= split_edge (exit
);
733 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
736 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
737 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
739 name
= USE_FROM_PTR (op_p
);
741 /* If the argument of the PHI node is a constant, we do not need
742 to keep it inside loop. */
743 if (TREE_CODE (name
) != SSA_NAME
744 && !copy_constants_p
)
747 /* Otherwise create an auxiliary phi node that will copy the value
748 of the SSA name out of the loop. */
749 new_name
= duplicate_ssa_name (PHI_RESULT (phi
), NULL
);
750 new_phi
= create_phi_node (new_name
, bb
);
751 add_phi_arg (new_phi
, name
, exit
, locus
);
752 SET_USE (op_p
, new_name
);
758 /* Returns the basic block in that statements should be emitted for induction
759 variables incremented at the end of the LOOP. */
762 ip_end_pos (class loop
*loop
)
767 /* Returns the basic block in that statements should be emitted for induction
768 variables incremented just before exit condition of a LOOP. */
771 ip_normal_pos (class loop
*loop
)
776 if (!single_pred_p (loop
->latch
))
779 bb
= single_pred (loop
->latch
);
780 if (!safe_is_a
<gcond
*> (*gsi_last_bb (bb
)))
783 exit
= EDGE_SUCC (bb
, 0);
784 if (exit
->dest
== loop
->latch
)
785 exit
= EDGE_SUCC (bb
, 1);
787 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
793 /* Stores the standard position for induction variable increment in LOOP
794 (just before the exit condition if it is available and latch block is empty,
795 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
796 the increment should be inserted after *BSI. */
799 standard_iv_increment_position (class loop
*loop
, gimple_stmt_iterator
*bsi
,
802 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
803 gimple
*last
= last_nondebug_stmt (latch
);
806 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
808 *bsi
= gsi_last_bb (latch
);
809 *insert_after
= true;
813 *bsi
= gsi_last_bb (bb
);
814 *insert_after
= false;
818 /* Copies phi node arguments for duplicated blocks. The index of the first
819 duplicated block is FIRST_NEW_BLOCK. */
822 copy_phi_node_args (unsigned first_new_block
)
826 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
827 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
|= BB_DUPLICATED
;
829 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
830 add_phi_args_after_copy_bb (BASIC_BLOCK_FOR_FN (cfun
, i
));
832 for (i
= first_new_block
; i
< (unsigned) last_basic_block_for_fn (cfun
); i
++)
833 BASIC_BLOCK_FOR_FN (cfun
, i
)->flags
&= ~BB_DUPLICATED
;
837 /* The same as cfgloopmanip.cc:duplicate_loop_body_to_header_edge, but also
838 updates the PHI nodes at start of the copied region. In order to
839 achieve this, only loops whose exits all lead to the same location
842 Notice that we do not completely update the SSA web after
843 duplication. The caller is responsible for calling update_ssa
844 after the loop has been duplicated. */
847 gimple_duplicate_loop_body_to_header_edge (class loop
*loop
, edge e
,
849 sbitmap wont_exit
, edge orig
,
850 vec
<edge
> *to_remove
, int flags
)
852 unsigned first_new_block
;
854 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
856 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
859 first_new_block
= last_basic_block_for_fn (cfun
);
860 if (!duplicate_loop_body_to_header_edge (loop
, e
, ndupl
, wont_exit
, orig
,
864 /* Readd the removed phi args for e. */
865 flush_pending_stmts (e
);
867 /* Copy the phi node arguments. */
868 copy_phi_node_args (first_new_block
);
875 /* Returns true if we can unroll LOOP FACTOR times. Number
876 of iterations of the loop is returned in NITER. */
879 can_unroll_loop_p (class loop
*loop
, unsigned factor
,
880 class tree_niter_desc
*niter
)
884 /* Check whether unrolling is possible. We only want to unroll loops
885 for that we are able to determine number of iterations. We also
886 want to split the extra iterations of the loop from its end,
887 therefore we require that the loop has precisely one
890 exit
= single_dom_exit (loop
);
894 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
895 || niter
->cmp
== ERROR_MARK
896 /* Scalar evolutions analysis might have copy propagated
897 the abnormal ssa names into these expressions, hence
898 emitting the computations based on them during loop
899 unrolling might create overlapping life ranges for
900 them, and failures in out-of-ssa. */
901 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
902 || contains_abnormal_ssa_name_p (niter
->control
.base
)
903 || contains_abnormal_ssa_name_p (niter
->control
.step
)
904 || contains_abnormal_ssa_name_p (niter
->bound
))
907 /* And of course, we must be able to duplicate the loop. */
908 if (!can_duplicate_loop_p (loop
))
911 /* The final loop should be small enough. */
912 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
913 > (unsigned) param_max_unrolled_insns
)
919 /* Determines the conditions that control execution of LOOP unrolled FACTOR
920 times. DESC is number of iterations of LOOP. ENTER_COND is set to
921 condition that must be true if the main loop can be entered.
922 If the loop does not always iterate an exact multiple of FACTOR times,
923 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
924 how the exit from the unrolled loop should be controlled. Otherwise,
925 the trees are set to null and EXIT_CMP is set to ERROR_MARK. */
928 determine_exit_conditions (class loop
*loop
, class tree_niter_desc
*desc
,
929 unsigned factor
, tree
*enter_cond
,
930 tree
*exit_base
, tree
*exit_step
,
931 enum tree_code
*exit_cmp
, tree
*exit_bound
)
934 tree base
= desc
->control
.base
;
935 tree step
= desc
->control
.step
;
936 tree bound
= desc
->bound
;
937 tree type
= TREE_TYPE (step
);
939 tree min
= lower_bound_in_type (type
, type
);
940 tree max
= upper_bound_in_type (type
, type
);
941 enum tree_code cmp
= desc
->cmp
;
942 tree cond
= boolean_true_node
, assum
;
944 /* For pointers, do the arithmetics in the type of step. */
945 base
= fold_convert (type
, base
);
946 bound
= fold_convert (type
, bound
);
948 *enter_cond
= boolean_false_node
;
949 *exit_base
= NULL_TREE
;
950 *exit_step
= NULL_TREE
;
951 *exit_cmp
= ERROR_MARK
;
952 *exit_bound
= NULL_TREE
;
953 gcc_assert (cmp
!= ERROR_MARK
);
955 /* We only need to be correct when we answer question
956 "Do at least FACTOR more iterations remain?" in the unrolled loop.
957 Thus, transforming BASE + STEP * i <> BOUND to
958 BASE + STEP * i < BOUND is ok. */
961 if (tree_int_cst_sign_bit (step
))
966 else if (cmp
== LT_EXPR
)
968 gcc_assert (!tree_int_cst_sign_bit (step
));
970 else if (cmp
== GT_EXPR
)
972 gcc_assert (tree_int_cst_sign_bit (step
));
977 /* The main body of the loop may be entered iff:
979 1) desc->may_be_zero is false.
980 2) it is possible to check that there are at least FACTOR iterations
981 of the loop, i.e., BOUND - step * FACTOR does not overflow.
982 3) # of iterations is at least FACTOR */
984 if (!integer_zerop (desc
->may_be_zero
))
985 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
986 invert_truthvalue (desc
->may_be_zero
),
989 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
990 build_int_cst_type (type
, factor
));
991 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
993 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
995 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
997 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
999 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
1000 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
1002 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
1003 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
1004 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
1006 if (integer_nonzerop (cond
)
1007 && integer_zerop (desc
->may_be_zero
))
1009 /* Convert the latch count to an iteration count. */
1010 tree niter
= fold_build2 (PLUS_EXPR
, type
, desc
->niter
,
1011 build_one_cst (type
));
1012 if (multiple_of_p (type
, niter
, build_int_cst (type
, factor
)))
1016 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
1018 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1019 /* cond now may be a gimple comparison, which would be OK, but also any
1020 other gimple rhs (say a && b). In this case we need to force it to
1022 if (!is_gimple_condexpr_for_cond (cond
))
1024 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
1026 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1030 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
1032 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1033 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
1035 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1038 *exit_step
= bigstep
;
1040 *exit_bound
= bound
;
1043 /* Scales the frequencies of all basic blocks in LOOP that are strictly
1044 dominated by BB by NUM/DEN. */
1047 scale_dominated_blocks_in_loop (class loop
*loop
, basic_block bb
,
1048 profile_count num
, profile_count den
)
1052 if (!den
.nonzero_p () && !(num
== profile_count::zero ()))
1055 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
1057 son
= next_dom_son (CDI_DOMINATORS
, son
))
1059 if (!flow_bb_inside_loop_p (loop
, son
))
1061 scale_bbs_frequencies_profile_count (&son
, 1, num
, den
);
1062 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
1066 /* Return estimated niter for LOOP after unrolling by FACTOR times. */
1069 niter_for_unrolled_loop (class loop
*loop
, unsigned factor
)
1071 gcc_assert (factor
!= 0);
1072 bool profile_p
= false;
1073 gcov_type est_niter
= expected_loop_iterations_unbounded (loop
, &profile_p
);
1074 /* Note that this is really CEIL (est_niter + 1, factor) - 1, where the
1075 "+ 1" converts latch iterations to loop iterations and the "- 1"
1077 gcov_type new_est_niter
= est_niter
/ factor
;
1079 if (est_niter
== -1)
1082 /* Without profile feedback, loops for which we do not know a better estimate
1083 are assumed to roll 10 times. When we unroll such loop, it appears to
1084 roll too little, and it may even seem to be cold. To avoid this, we
1085 ensure that the created loop appears to roll at least 5 times (but at
1086 most as many times as before unrolling). Don't do adjustment if profile
1087 feedback is present. */
1088 if (new_est_niter
< 5 && !profile_p
)
1091 new_est_niter
= est_niter
;
1096 if (loop
->any_upper_bound
)
1098 /* As above, this is really CEIL (upper_bound + 1, factor) - 1. */
1099 widest_int bound
= wi::udiv_floor (loop
->nb_iterations_upper_bound
,
1101 if (wi::ltu_p (bound
, new_est_niter
))
1102 new_est_niter
= bound
.to_uhwi ();
1105 return new_est_niter
;
1108 /* Unroll LOOP FACTOR times. LOOP is known to have a single exit edge
1109 whose source block dominates the latch. DESC describes the number of
1112 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
1113 under that loop exits in the first iteration even if N != 0,
1117 x = phi (init, next);
1125 becomes (with possibly the exit conditions formulated a bit differently,
1126 avoiding the need to create a new iv):
1128 if (MAY_BE_ZERO || N < FACTOR)
1133 x = phi (init, next);
1144 } while (N >= FACTOR);
1147 init' = phi (init, x);
1151 x = phi (init', next);
1159 Before the loop is unrolled, TRANSFORM is called for it (only for the
1160 unrolled loop, but not for its versioned copy). DATA is passed to
1163 /* Probability in % that the unrolled loop is entered. Just a guess. */
1164 #define PROB_UNROLLED_LOOP_ENTERED 90
1167 tree_transform_and_unroll_loop (class loop
*loop
, unsigned factor
,
1168 class tree_niter_desc
*desc
,
1169 transform_callback transform
,
1172 gcov_type new_est_niter
= niter_for_unrolled_loop (loop
, factor
);
1173 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1175 enum tree_code exit_cmp
;
1176 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1177 determine_exit_conditions (loop
, desc
, factor
,
1178 &enter_main_cond
, &exit_base
, &exit_step
,
1179 &exit_cmp
, &exit_bound
);
1180 bool single_loop_p
= !exit_base
;
1182 /* Let us assume that the unrolled loop is quite likely to be entered. */
1183 profile_probability prob_entry
;
1184 if (integer_nonzerop (enter_main_cond
))
1185 prob_entry
= profile_probability::always ();
1187 prob_entry
= profile_probability::guessed_always ()
1188 .apply_scale (PROB_UNROLLED_LOOP_ENTERED
, 100);
1190 gcond
*exit_if
= nullptr;
1191 class loop
*new_loop
= nullptr;
1195 edge exit
= single_dom_exit (loop
);
1197 /* The values for scales should keep profile consistent, and somewhat
1200 TODO: The current value of SCALE_REST makes it appear that the loop
1201 that is created by splitting the remaining iterations of the unrolled
1202 loop is executed the same number of times as the original loop, and
1203 with the same frequencies, which is obviously wrong. This does not
1204 appear to cause problems, so we do not bother with fixing it for now.
1205 To make the profile correct, we would need to change the probability
1206 of the exit edge of the loop, and recompute the distribution of
1207 frequencies in its body because of this change (scale the frequencies
1208 of blocks before and after the exit by appropriate factors). */
1209 profile_probability scale_unrolled
= prob_entry
;
1210 new_loop
= loop_version (loop
, enter_main_cond
, NULL
, prob_entry
,
1211 prob_entry
.invert (), scale_unrolled
,
1212 profile_probability::guessed_always (),
1214 gcc_assert (new_loop
!= NULL
);
1215 update_ssa (TODO_update_ssa_no_phi
);
1217 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1218 loop latch (and make its condition dummy, for the moment). */
1219 basic_block rest
= loop_preheader_edge (new_loop
)->src
;
1220 edge precond_edge
= single_pred_edge (rest
);
1221 split_edge (loop_latch_edge (loop
));
1222 basic_block exit_bb
= single_pred (loop
->latch
);
1224 /* Since the exit edge will be removed, the frequency of all the blocks
1225 in the loop that are dominated by it must be scaled by
1226 1 / (1 - exit->probability). */
1227 if (exit
->probability
.initialized_p ())
1228 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1229 /* We are scaling up here so
1230 probability does not fit. */
1231 loop
->header
->count
,
1233 - loop
->header
->count
.apply_probability
1234 (exit
->probability
));
1236 gimple_stmt_iterator bsi
= gsi_last_bb (exit_bb
);
1237 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1239 NULL_TREE
, NULL_TREE
);
1241 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1242 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1243 rescan_loop_exit (new_exit
, true, false);
1245 /* Set the probability of new exit to the same of the old one. Fix
1246 the frequency of the latch block, by scaling it back by
1247 1 - exit->probability. */
1248 new_exit
->probability
= exit
->probability
;
1249 edge new_nonexit
= single_pred_edge (loop
->latch
);
1250 new_nonexit
->probability
= exit
->probability
.invert ();
1251 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1252 if (new_nonexit
->probability
.initialized_p ())
1253 scale_bbs_frequencies (&loop
->latch
, 1, new_nonexit
->probability
);
1255 edge old_entry
= loop_preheader_edge (loop
);
1256 edge new_entry
= loop_preheader_edge (new_loop
);
1257 edge old_latch
= loop_latch_edge (loop
);
1258 for (gphi_iterator psi_old_loop
= gsi_start_phis (loop
->header
),
1259 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1260 !gsi_end_p (psi_old_loop
);
1261 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1263 gphi
*phi_old_loop
= psi_old_loop
.phi ();
1264 gphi
*phi_new_loop
= psi_new_loop
.phi ();
1266 tree init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1268 = PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1269 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1270 tree next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1272 /* Prefer using original variable as a base for the new ssa name.
1273 This is necessary for virtual ops, and useful in order to avoid
1274 losing debug info for real ops. */
1276 if (TREE_CODE (next
) == SSA_NAME
1277 && useless_type_conversion_p (TREE_TYPE (next
),
1279 new_init
= copy_ssa_name (next
);
1280 else if (TREE_CODE (init
) == SSA_NAME
1281 && useless_type_conversion_p (TREE_TYPE (init
),
1283 new_init
= copy_ssa_name (init
);
1284 else if (useless_type_conversion_p (TREE_TYPE (next
),
1286 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
,
1289 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
,
1292 gphi
*phi_rest
= create_phi_node (new_init
, rest
);
1293 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1294 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1295 SET_USE (op
, new_init
);
1300 /* The epilog loop latch executes at most factor - 1 times.
1301 Since the epilog is entered unconditionally it will need to handle
1302 up to factor executions of its body. */
1303 new_loop
->any_upper_bound
= 1;
1304 new_loop
->nb_iterations_upper_bound
= factor
- 1;
1307 new_exit
= single_dom_exit (loop
);
1309 /* Transform the loop. */
1311 (*transform
) (loop
, data
);
1313 /* Unroll the loop and remove the exits in all iterations except for the
1315 auto_sbitmap
wont_exit (factor
);
1316 bitmap_ones (wont_exit
);
1317 bitmap_clear_bit (wont_exit
, factor
- 1);
1319 auto_vec
<edge
> to_remove
;
1321 = gimple_duplicate_loop_body_to_header_edge (loop
, loop_latch_edge (loop
),
1322 factor
- 1, wont_exit
,
1323 new_exit
, &to_remove
,
1324 DLTHE_FLAG_UPDATE_FREQ
);
1327 for (edge e
: to_remove
)
1329 ok
= remove_path (e
);
1332 update_ssa (TODO_update_ssa
);
1334 new_exit
= single_dom_exit (loop
);
1337 /* Ensure that the frequencies in the loop match the new estimated
1338 number of iterations, and change the probability of the new
1341 profile_count freq_h
= loop
->header
->count
;
1342 profile_count freq_e
= (loop_preheader_edge (loop
))->count ();
1343 if (freq_h
.nonzero_p ())
1345 /* Avoid dropping loop body profile counter to 0 because of zero
1346 count in loop's preheader. */
1347 if (freq_h
.nonzero_p () && !(freq_e
== profile_count::zero ()))
1348 freq_e
= freq_e
.force_nonzero ();
1349 scale_loop_frequencies (loop
, freq_e
.probability_in (freq_h
));
1352 basic_block rest
= new_exit
->dest
;
1353 new_exit
->probability
1354 = (profile_probability::always () / (new_est_niter
+ 1));
1356 rest
->count
+= new_exit
->count ();
1358 edge new_nonexit
= single_pred_edge (loop
->latch
);
1359 profile_probability prob
= new_nonexit
->probability
;
1360 new_nonexit
->probability
= new_exit
->probability
.invert ();
1361 prob
= new_nonexit
->probability
/ prob
;
1362 if (prob
.initialized_p ())
1363 scale_bbs_frequencies (&loop
->latch
, 1, prob
);
1365 /* Finally create the new counter for number of iterations and add
1366 the new exit instruction. */
1367 tree ctr_before
, ctr_after
;
1368 gimple_stmt_iterator bsi
= gsi_last_nondebug_bb (new_exit
->src
);
1369 exit_if
= as_a
<gcond
*> (gsi_stmt (bsi
));
1370 create_iv (exit_base
, PLUS_EXPR
, exit_step
, NULL_TREE
, loop
,
1371 &bsi
, false, &ctr_before
, &ctr_after
);
1372 gimple_cond_set_code (exit_if
, exit_cmp
);
1373 gimple_cond_set_lhs (exit_if
, ctr_after
);
1374 gimple_cond_set_rhs (exit_if
, exit_bound
);
1375 update_stmt (exit_if
);
1379 /* gimple_duplicate_loop_to_header_edge has adjusted the loop body's
1380 original profile counts in line with the unroll factor. However,
1381 the old counts might not have been consistent with the old
1384 Therefore, if the iteration count is known exactly, make sure that the
1385 profile counts of the loop header (and any other blocks that might be
1386 executed in the final iteration) are consistent with the combination
1387 of (a) the incoming profile count and (b) the new iteration count. */
1388 profile_count in_count
= loop_preheader_edge (loop
)->count ();
1389 profile_count old_header_count
= loop
->header
->count
;
1390 if (in_count
.nonzero_p ()
1391 && old_header_count
.nonzero_p ()
1392 && TREE_CODE (desc
->niter
) == INTEGER_CST
)
1394 /* The + 1 converts latch counts to iteration counts. */
1395 profile_count new_header_count
= in_count
* (new_est_niter
+ 1);
1396 basic_block
*body
= get_loop_body (loop
);
1397 scale_bbs_frequencies_profile_count (body
, loop
->num_nodes
,
1403 /* gimple_duplicate_loop_to_header_edge discarded FACTOR - 1
1404 exit edges and adjusted the loop body's profile counts for the
1405 new probabilities of the remaining non-exit edges. However,
1406 the remaining exit edge still has the same probability as it
1407 did before, even though it is now more likely.
1409 Therefore, all blocks executed after a failed exit test now have
1410 a profile count that is too high, and the sum of the profile counts
1411 for the header's incoming edges is greater than the profile count
1412 of the header itself.
1414 Adjust the profile counts of all code in the loop body after
1415 the exit test so that the sum of the counts on entry to the
1417 profile_count old_latch_count
= loop_latch_edge (loop
)->count ();
1418 profile_count new_latch_count
= loop
->header
->count
- in_count
;
1419 if (old_latch_count
.nonzero_p () && new_latch_count
.nonzero_p ())
1420 scale_dominated_blocks_in_loop (loop
, new_exit
->src
, new_latch_count
,
1423 /* Set the probability of the exit edge based on NEW_EST_NITER
1424 (which estimates latch counts rather than iteration counts).
1425 Update the probabilities of other edges to match.
1427 If the profile counts are large enough to give the required
1428 precision, the updates above will have made
1430 e->dest->count / e->src->count ~= new e->probability
1432 for every outgoing edge e of NEW_EXIT->src. */
1433 profile_probability new_exit_prob
1434 = profile_probability::always () / (new_est_niter
+ 1);
1435 change_edge_frequency (new_exit
, new_exit_prob
);
1438 checking_verify_flow_info ();
1439 checking_verify_loop_structure ();
1440 checking_verify_loop_closed_ssa (true, loop
);
1442 checking_verify_loop_closed_ssa (true, new_loop
);
1445 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1446 want to transform the loop before unrolling. The meaning
1447 of the arguments is the same as for tree_transform_and_unroll_loop. */
1450 tree_unroll_loop (class loop
*loop
, unsigned factor
,
1451 class tree_niter_desc
*desc
)
1453 tree_transform_and_unroll_loop (loop
, factor
, desc
, NULL
, NULL
);
1456 /* Rewrite the phi node at position PSI in function of the main
1457 induction variable MAIN_IV and insert the generated code at GSI. */
1460 rewrite_phi_with_iv (loop_p loop
,
1462 gimple_stmt_iterator
*gsi
,
1467 gphi
*phi
= psi
->phi ();
1468 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1470 if (virtual_operand_p (res
) || res
== main_iv
)
1476 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1482 remove_phi_node (psi
, false);
1484 atype
= TREE_TYPE (res
);
1485 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1486 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1487 fold_convert (mtype
, main_iv
));
1488 val
= fold_build2 (POINTER_TYPE_P (atype
)
1489 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1490 atype
, unshare_expr (iv
.base
), val
);
1491 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1493 stmt
= gimple_build_assign (res
, val
);
1494 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1497 /* Rewrite all the phi nodes of LOOP in function of the main induction
1498 variable MAIN_IV. */
1501 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1504 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1507 for (i
= 0; i
< loop
->num_nodes
; i
++)
1509 basic_block bb
= bbs
[i
];
1510 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1512 if (bb
->loop_father
!= loop
)
1515 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1516 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1522 /* Bases all the induction variables in LOOP on a single induction variable
1523 (with base 0 and step 1), whose final value is compared with *NIT. When the
1524 IV type precision has to be larger than *NIT type precision, *NIT is
1525 converted to the larger type, the conversion code is inserted before the
1526 loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true,
1527 the induction variable is incremented in the loop latch, otherwise it is
1528 incremented in the loop header. Return the induction variable that was
1532 canonicalize_loop_ivs (class loop
*loop
, tree
*nit
, bool bump_in_latch
)
1534 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1535 unsigned original_precision
= precision
;
1536 tree type
, var_before
;
1537 gimple_stmt_iterator gsi
;
1540 edge exit
= single_dom_exit (loop
);
1542 bool unsigned_p
= false;
1544 for (psi
= gsi_start_phis (loop
->header
);
1545 !gsi_end_p (psi
); gsi_next (&psi
))
1547 gphi
*phi
= psi
.phi ();
1548 tree res
= PHI_RESULT (phi
);
1551 type
= TREE_TYPE (res
);
1552 if (virtual_operand_p (res
)
1553 || (!INTEGRAL_TYPE_P (type
)
1554 && !POINTER_TYPE_P (type
))
1555 || TYPE_PRECISION (type
) < precision
)
1558 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1560 if (TYPE_PRECISION (type
) > precision
)
1565 precision
= TYPE_PRECISION (type
);
1568 scalar_int_mode mode
= smallest_int_mode_for_size (precision
);
1569 precision
= GET_MODE_PRECISION (mode
);
1570 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1572 if (original_precision
!= precision
1573 || TYPE_UNSIGNED (TREE_TYPE (*nit
)) != unsigned_p
)
1575 *nit
= fold_convert (type
, *nit
);
1576 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1578 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1582 gsi
= gsi_last_bb (loop
->latch
);
1584 gsi
= gsi_last_nondebug_bb (loop
->header
);
1585 create_iv (build_int_cst_type (type
, 0), PLUS_EXPR
, build_int_cst (type
, 1),
1586 NULL_TREE
, loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1588 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1590 stmt
= as_a
<gcond
*> (*gsi_last_bb (exit
->src
));
1591 /* Make the loop exit if the control condition is not satisfied. */
1592 if (exit
->flags
& EDGE_TRUE_VALUE
)
1596 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1597 te
->flags
= EDGE_FALSE_VALUE
;
1598 fe
->flags
= EDGE_TRUE_VALUE
;
1600 gimple_cond_set_code (stmt
, LT_EXPR
);
1601 gimple_cond_set_lhs (stmt
, var_before
);
1602 gimple_cond_set_rhs (stmt
, *nit
);