1 /* High-level loop manipulation functions.
2 Copyright (C) 2004-2013 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"
26 #include "basic-block.h"
27 #include "tree-flow.h"
29 #include "gimple-pretty-print.h"
31 #include "tree-pass.h" /* ??? for TODO_update_ssa but this isn't a pass. */
32 #include "tree-scalar-evolution.h"
34 #include "tree-inline.h"
35 #include "langhooks.h"
37 /* All bitmaps for rewriting into loop-closed SSA go on this obstack,
38 so that we can free them all at once. */
39 static bitmap_obstack loop_renamer_obstack
;
41 /* Creates an induction variable with value BASE + STEP * iteration in LOOP.
42 It is expected that neither BASE nor STEP are shared with other expressions
43 (unless the sharing rules allow this). Use VAR as a base var_decl for it
44 (if NULL, a new temporary will be created). The increment will occur at
45 INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and
46 AFTER can be computed using standard_iv_increment_position. The ssa versions
47 of the variable before and after increment will be stored in VAR_BEFORE and
48 VAR_AFTER (unless they are NULL). */
51 create_iv (tree base
, tree step
, tree var
, struct loop
*loop
,
52 gimple_stmt_iterator
*incr_pos
, bool after
,
53 tree
*var_before
, tree
*var_after
)
59 enum tree_code incr_op
= PLUS_EXPR
;
60 edge pe
= loop_preheader_edge (loop
);
64 vb
= make_ssa_name (var
, NULL
);
65 va
= make_ssa_name (var
, NULL
);
69 vb
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
70 va
= make_temp_ssa_name (TREE_TYPE (base
), NULL
, "ivtmp");
77 /* For easier readability of the created code, produce MINUS_EXPRs
79 if (TREE_CODE (step
) == INTEGER_CST
)
81 if (TYPE_UNSIGNED (TREE_TYPE (step
)))
83 step1
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
84 if (tree_int_cst_lt (step1
, step
))
94 if (!tree_expr_nonnegative_warnv_p (step
, &ovf
)
95 && may_negate_without_overflow_p (step
))
98 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
102 if (POINTER_TYPE_P (TREE_TYPE (base
)))
104 if (TREE_CODE (base
) == ADDR_EXPR
)
105 mark_addressable (TREE_OPERAND (base
, 0));
106 step
= convert_to_ptrofftype (step
);
107 if (incr_op
== MINUS_EXPR
)
108 step
= fold_build1 (NEGATE_EXPR
, TREE_TYPE (step
), step
);
109 incr_op
= POINTER_PLUS_EXPR
;
111 /* Gimplify the step if necessary. We put the computations in front of the
112 loop (i.e. the step should be loop invariant). */
113 step
= force_gimple_operand (step
, &stmts
, true, NULL_TREE
);
115 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
117 stmt
= gimple_build_assign_with_ops (incr_op
, va
, vb
, step
);
119 gsi_insert_after (incr_pos
, stmt
, GSI_NEW_STMT
);
121 gsi_insert_before (incr_pos
, stmt
, GSI_NEW_STMT
);
123 initial
= force_gimple_operand (base
, &stmts
, true, var
);
125 gsi_insert_seq_on_edge_immediate (pe
, stmts
);
127 stmt
= create_phi_node (vb
, loop
->header
);
128 add_phi_arg (stmt
, initial
, loop_preheader_edge (loop
), UNKNOWN_LOCATION
);
129 add_phi_arg (stmt
, va
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
132 /* Return the innermost superloop LOOP of USE_LOOP that is a superloop of
133 both DEF_LOOP and USE_LOOP. */
135 static inline struct loop
*
136 find_sibling_superloop (struct loop
*use_loop
, struct loop
*def_loop
)
138 unsigned ud
= loop_depth (use_loop
);
139 unsigned dd
= loop_depth (def_loop
);
140 gcc_assert (ud
> 0 && dd
> 0);
142 use_loop
= superloop_at_depth (use_loop
, dd
);
144 def_loop
= superloop_at_depth (def_loop
, ud
);
145 while (loop_outer (use_loop
) != loop_outer (def_loop
))
147 use_loop
= loop_outer (use_loop
);
148 def_loop
= loop_outer (def_loop
);
149 gcc_assert (use_loop
&& def_loop
);
154 /* DEF_BB is a basic block containing a DEF that needs rewriting into
155 loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing
156 uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in
157 USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B).
158 ALL_EXITS[I] is the set of all basic blocks that exit loop I.
160 Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB
161 or one of its loop fathers, in which DEF is live. This set is returned
162 in the bitmap LIVE_EXITS.
164 Instead of computing the complete livein set of the def, we use the loop
165 nesting tree as a form of poor man's structure analysis. This greatly
166 speeds up the analysis, which is important because this function may be
167 called on all SSA names that need rewriting, one at a time. */
170 compute_live_loop_exits (bitmap live_exits
, bitmap use_blocks
,
171 bitmap
*loop_exits
, basic_block def_bb
)
175 vec
<basic_block
> worklist
;
176 struct loop
*def_loop
= def_bb
->loop_father
;
177 unsigned def_loop_depth
= loop_depth (def_loop
);
178 bitmap def_loop_exits
;
180 /* Normally the work list size is bounded by the number of basic
181 blocks in the largest loop. We don't know this number, but we
182 can be fairly sure that it will be relatively small. */
183 worklist
.create (MAX (8, n_basic_blocks
/ 128));
185 EXECUTE_IF_SET_IN_BITMAP (use_blocks
, 0, i
, bi
)
187 basic_block use_bb
= BASIC_BLOCK (i
);
188 struct loop
*use_loop
= use_bb
->loop_father
;
189 gcc_checking_assert (def_loop
!= use_loop
190 && ! flow_loop_nested_p (def_loop
, use_loop
));
191 if (! flow_loop_nested_p (use_loop
, def_loop
))
192 use_bb
= find_sibling_superloop (use_loop
, def_loop
)->header
;
193 if (bitmap_set_bit (live_exits
, use_bb
->index
))
194 worklist
.safe_push (use_bb
);
197 /* Iterate until the worklist is empty. */
198 while (! worklist
.is_empty ())
203 /* Pull a block off the worklist. */
204 basic_block bb
= worklist
.pop ();
206 /* Make sure we have at least enough room in the work list
207 for all predecessors of this block. */
208 worklist
.reserve (EDGE_COUNT (bb
->preds
));
210 /* For each predecessor block. */
211 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
213 basic_block pred
= e
->src
;
214 struct loop
*pred_loop
= pred
->loop_father
;
215 unsigned pred_loop_depth
= loop_depth (pred_loop
);
218 /* We should have met DEF_BB along the way. */
219 gcc_assert (pred
!= ENTRY_BLOCK_PTR
);
221 if (pred_loop_depth
>= def_loop_depth
)
223 if (pred_loop_depth
> def_loop_depth
)
224 pred_loop
= superloop_at_depth (pred_loop
, def_loop_depth
);
225 /* If we've reached DEF_LOOP, our train ends here. */
226 if (pred_loop
== def_loop
)
229 else if (! flow_loop_nested_p (pred_loop
, def_loop
))
230 pred
= find_sibling_superloop (pred_loop
, def_loop
)->header
;
232 /* Add PRED to the LIVEIN set. PRED_VISITED is true if
233 we had already added PRED to LIVEIN before. */
234 pred_visited
= !bitmap_set_bit (live_exits
, pred
->index
);
236 /* If we have visited PRED before, don't add it to the worklist.
237 If BB dominates PRED, then we're probably looking at a loop.
238 We're only interested in looking up in the dominance tree
239 because DEF_BB dominates all the uses. */
240 if (pred_visited
|| dominated_by_p (CDI_DOMINATORS
, pred
, bb
))
243 worklist
.quick_push (pred
);
248 def_loop_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
249 for (struct loop
*loop
= def_loop
;
250 loop
!= current_loops
->tree_root
;
251 loop
= loop_outer (loop
))
252 bitmap_ior_into (def_loop_exits
, loop_exits
[loop
->num
]);
253 bitmap_and_into (live_exits
, def_loop_exits
);
254 BITMAP_FREE (def_loop_exits
);
257 /* Add a loop-closing PHI for VAR in basic block EXIT. */
260 add_exit_phi (basic_block exit
, tree var
)
266 #ifdef ENABLE_CHECKING
267 /* Check that at least one of the edges entering the EXIT block exits
268 the loop, or a superloop of that loop, that VAR is defined in. */
269 gimple def_stmt
= SSA_NAME_DEF_STMT (var
);
270 basic_block def_bb
= gimple_bb (def_stmt
);
271 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
273 struct loop
*aloop
= find_common_loop (def_bb
->loop_father
,
274 e
->src
->loop_father
);
275 if (!flow_bb_inside_loop_p (aloop
, e
->dest
))
279 gcc_checking_assert (e
);
282 phi
= create_phi_node (NULL_TREE
, exit
);
283 create_new_def_for (var
, phi
, gimple_phi_result_ptr (phi
));
284 FOR_EACH_EDGE (e
, ei
, exit
->preds
)
285 add_phi_arg (phi
, var
, e
, UNKNOWN_LOCATION
);
287 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
289 fprintf (dump_file
, ";; Created LCSSA PHI: ");
290 print_gimple_stmt (dump_file
, phi
, 0, dump_flags
);
294 /* Add exit phis for VAR that is used in LIVEIN.
295 Exits of the loops are stored in LOOP_EXITS. */
298 add_exit_phis_var (tree var
, bitmap use_blocks
, bitmap
*loop_exits
)
302 basic_block def_bb
= gimple_bb (SSA_NAME_DEF_STMT (var
));
303 bitmap live_exits
= BITMAP_ALLOC (&loop_renamer_obstack
);
305 gcc_checking_assert (! bitmap_bit_p (use_blocks
, def_bb
->index
));
307 compute_live_loop_exits (live_exits
, use_blocks
, loop_exits
, def_bb
);
309 EXECUTE_IF_SET_IN_BITMAP (live_exits
, 0, index
, bi
)
311 add_exit_phi (BASIC_BLOCK (index
), var
);
314 BITMAP_FREE (live_exits
);
317 /* Add exit phis for the names marked in NAMES_TO_RENAME.
318 Exits of the loops are stored in EXITS. Sets of blocks where the ssa
319 names are used are stored in USE_BLOCKS. */
322 add_exit_phis (bitmap names_to_rename
, bitmap
*use_blocks
, bitmap
*loop_exits
)
327 EXECUTE_IF_SET_IN_BITMAP (names_to_rename
, 0, i
, bi
)
329 add_exit_phis_var (ssa_name (i
), use_blocks
[i
], loop_exits
);
333 /* Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets. */
336 get_loops_exits (bitmap
*loop_exits
)
343 FOR_EACH_LOOP (li
, loop
, 0)
345 vec
<edge
> exit_edges
= get_loop_exit_edges (loop
);
346 loop_exits
[loop
->num
] = BITMAP_ALLOC (&loop_renamer_obstack
);
347 FOR_EACH_VEC_ELT (exit_edges
, j
, e
)
348 bitmap_set_bit (loop_exits
[loop
->num
], e
->dest
->index
);
349 exit_edges
.release ();
353 /* For USE in BB, if it is used outside of the loop it is defined in,
354 mark it for rewrite. Record basic block BB where it is used
355 to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap. */
358 find_uses_to_rename_use (basic_block bb
, tree use
, bitmap
*use_blocks
,
363 struct loop
*def_loop
;
365 if (TREE_CODE (use
) != SSA_NAME
)
368 ver
= SSA_NAME_VERSION (use
);
369 def_bb
= gimple_bb (SSA_NAME_DEF_STMT (use
));
372 def_loop
= def_bb
->loop_father
;
374 /* If the definition is not inside a loop, it is not interesting. */
375 if (!loop_outer (def_loop
))
378 /* If the use is not outside of the loop it is defined in, it is not
380 if (flow_bb_inside_loop_p (def_loop
, bb
))
383 /* If we're seeing VER for the first time, we still have to allocate
384 a bitmap for its uses. */
385 if (bitmap_set_bit (need_phis
, ver
))
386 use_blocks
[ver
] = BITMAP_ALLOC (&loop_renamer_obstack
);
387 bitmap_set_bit (use_blocks
[ver
], bb
->index
);
390 /* For uses in STMT, mark names that are used outside of the loop they are
391 defined to rewrite. Record the set of blocks in that the ssa
392 names are defined to USE_BLOCKS and the ssa names themselves to
396 find_uses_to_rename_stmt (gimple stmt
, bitmap
*use_blocks
, bitmap need_phis
)
400 basic_block bb
= gimple_bb (stmt
);
402 if (is_gimple_debug (stmt
))
405 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
406 find_uses_to_rename_use (bb
, var
, use_blocks
, need_phis
);
409 /* Marks names that are used in BB and outside of the loop they are
410 defined in for rewrite. Records the set of blocks in that the ssa
411 names are defined to USE_BLOCKS. Record the SSA names that will
412 need exit PHIs in NEED_PHIS. */
415 find_uses_to_rename_bb (basic_block bb
, bitmap
*use_blocks
, bitmap need_phis
)
417 gimple_stmt_iterator bsi
;
421 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
422 for (bsi
= gsi_start_phis (e
->dest
); !gsi_end_p (bsi
); gsi_next (&bsi
))
424 gimple phi
= gsi_stmt (bsi
);
425 if (! virtual_operand_p (gimple_phi_result (phi
)))
426 find_uses_to_rename_use (bb
, PHI_ARG_DEF_FROM_EDGE (phi
, e
),
427 use_blocks
, need_phis
);
430 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
431 find_uses_to_rename_stmt (gsi_stmt (bsi
), use_blocks
, need_phis
);
434 /* Marks names that are used outside of the loop they are defined in
435 for rewrite. Records the set of blocks in that the ssa
436 names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL,
437 scan only blocks in this set. */
440 find_uses_to_rename (bitmap changed_bbs
, bitmap
*use_blocks
, bitmap need_phis
)
447 EXECUTE_IF_SET_IN_BITMAP (changed_bbs
, 0, index
, bi
)
448 find_uses_to_rename_bb (BASIC_BLOCK (index
), use_blocks
, need_phis
);
451 find_uses_to_rename_bb (bb
, use_blocks
, need_phis
);
454 /* Rewrites the program into a loop closed ssa form -- i.e. inserts extra
455 phi nodes to ensure that no variable is used outside the loop it is
458 This strengthening of the basic ssa form has several advantages:
460 1) Updating it during unrolling/peeling/versioning is trivial, since
461 we do not need to care about the uses outside of the loop.
462 The same applies to virtual operands which are also rewritten into
463 loop closed SSA form. Note that virtual operands are always live
465 2) The behavior of all uses of an induction variable is the same.
466 Without this, you need to distinguish the case when the variable
467 is used outside of the loop it is defined in, for example
469 for (i = 0; i < 100; i++)
471 for (j = 0; j < 100; j++)
479 Looking from the outer loop with the normal SSA form, the first use of k
480 is not well-behaved, while the second one is an induction variable with
483 If CHANGED_BBS is not NULL, we look for uses outside loops only in
484 the basic blocks in this set.
486 UPDATE_FLAG is used in the call to update_ssa. See
487 TODO_update_ssa* for documentation. */
490 rewrite_into_loop_closed_ssa (bitmap changed_bbs
, unsigned update_flag
)
494 bitmap names_to_rename
;
496 loops_state_set (LOOP_CLOSED_SSA
);
497 if (number_of_loops () <= 1)
500 /* If the pass has caused the SSA form to be out-of-date, update it
502 update_ssa (update_flag
);
504 bitmap_obstack_initialize (&loop_renamer_obstack
);
506 names_to_rename
= BITMAP_ALLOC (&loop_renamer_obstack
);
508 /* An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
509 that are the destination of an edge exiting loop number I. */
510 loop_exits
= XNEWVEC (bitmap
, number_of_loops ());
511 get_loops_exits (loop_exits
);
513 /* Uses of names to rename. We don't have to initialize this array,
514 because we know that we will only have entries for the SSA names
515 in NAMES_TO_RENAME. */
516 use_blocks
= XNEWVEC (bitmap
, num_ssa_names
);
518 /* Find the uses outside loops. */
519 find_uses_to_rename (changed_bbs
, use_blocks
, names_to_rename
);
521 /* Add the PHI nodes on exits of the loops for the names we need to
523 add_exit_phis (names_to_rename
, use_blocks
, loop_exits
);
525 bitmap_obstack_release (&loop_renamer_obstack
);
529 /* Fix up all the names found to be used outside their original
531 update_ssa (TODO_update_ssa
);
534 /* Check invariants of the loop closed ssa form for the USE in BB. */
537 check_loop_closed_ssa_use (basic_block bb
, tree use
)
542 if (TREE_CODE (use
) != SSA_NAME
|| virtual_operand_p (use
))
545 def
= SSA_NAME_DEF_STMT (use
);
546 def_bb
= gimple_bb (def
);
548 || flow_bb_inside_loop_p (def_bb
->loop_father
, bb
));
551 /* Checks invariants of loop closed ssa form in statement STMT in BB. */
554 check_loop_closed_ssa_stmt (basic_block bb
, gimple stmt
)
559 if (is_gimple_debug (stmt
))
562 FOR_EACH_SSA_TREE_OPERAND (var
, stmt
, iter
, SSA_OP_USE
)
563 check_loop_closed_ssa_use (bb
, var
);
566 /* Checks that invariants of the loop closed ssa form are preserved.
567 Call verify_ssa when VERIFY_SSA_P is true. */
570 verify_loop_closed_ssa (bool verify_ssa_p
)
573 gimple_stmt_iterator bsi
;
578 if (number_of_loops () <= 1)
584 timevar_push (TV_VERIFY_LOOP_CLOSED
);
588 for (bsi
= gsi_start_phis (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
590 phi
= gsi_stmt (bsi
);
591 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
592 check_loop_closed_ssa_use (e
->src
,
593 PHI_ARG_DEF_FROM_EDGE (phi
, e
));
596 for (bsi
= gsi_start_bb (bb
); !gsi_end_p (bsi
); gsi_next (&bsi
))
597 check_loop_closed_ssa_stmt (bb
, gsi_stmt (bsi
));
600 timevar_pop (TV_VERIFY_LOOP_CLOSED
);
603 /* Split loop exit edge EXIT. The things are a bit complicated by a need to
604 preserve the loop closed ssa form. The newly created block is returned. */
607 split_loop_exit_edge (edge exit
)
609 basic_block dest
= exit
->dest
;
610 basic_block bb
= split_edge (exit
);
614 gimple_stmt_iterator psi
;
615 source_location locus
;
617 for (psi
= gsi_start_phis (dest
); !gsi_end_p (psi
); gsi_next (&psi
))
619 phi
= gsi_stmt (psi
);
620 op_p
= PHI_ARG_DEF_PTR_FROM_EDGE (phi
, single_succ_edge (bb
));
621 locus
= gimple_phi_arg_location_from_edge (phi
, single_succ_edge (bb
));
623 name
= USE_FROM_PTR (op_p
);
625 /* If the argument of the PHI node is a constant, we do not need
626 to keep it inside loop. */
627 if (TREE_CODE (name
) != SSA_NAME
)
630 /* Otherwise create an auxiliary phi node that will copy the value
631 of the SSA name out of the loop. */
632 new_name
= duplicate_ssa_name (name
, NULL
);
633 new_phi
= create_phi_node (new_name
, bb
);
634 add_phi_arg (new_phi
, name
, exit
, locus
);
635 SET_USE (op_p
, new_name
);
641 /* Returns the basic block in that statements should be emitted for induction
642 variables incremented at the end of the LOOP. */
645 ip_end_pos (struct loop
*loop
)
650 /* Returns the basic block in that statements should be emitted for induction
651 variables incremented just before exit condition of a LOOP. */
654 ip_normal_pos (struct loop
*loop
)
660 if (!single_pred_p (loop
->latch
))
663 bb
= single_pred (loop
->latch
);
664 last
= last_stmt (bb
);
666 || gimple_code (last
) != GIMPLE_COND
)
669 exit
= EDGE_SUCC (bb
, 0);
670 if (exit
->dest
== loop
->latch
)
671 exit
= EDGE_SUCC (bb
, 1);
673 if (flow_bb_inside_loop_p (loop
, exit
->dest
))
679 /* Stores the standard position for induction variable increment in LOOP
680 (just before the exit condition if it is available and latch block is empty,
681 end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if
682 the increment should be inserted after *BSI. */
685 standard_iv_increment_position (struct loop
*loop
, gimple_stmt_iterator
*bsi
,
688 basic_block bb
= ip_normal_pos (loop
), latch
= ip_end_pos (loop
);
689 gimple last
= last_stmt (latch
);
692 || (last
&& gimple_code (last
) != GIMPLE_LABEL
))
694 *bsi
= gsi_last_bb (latch
);
695 *insert_after
= true;
699 *bsi
= gsi_last_bb (bb
);
700 *insert_after
= false;
704 /* Copies phi node arguments for duplicated blocks. The index of the first
705 duplicated block is FIRST_NEW_BLOCK. */
708 copy_phi_node_args (unsigned first_new_block
)
712 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
713 BASIC_BLOCK (i
)->flags
|= BB_DUPLICATED
;
715 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
716 add_phi_args_after_copy_bb (BASIC_BLOCK (i
));
718 for (i
= first_new_block
; i
< (unsigned) last_basic_block
; i
++)
719 BASIC_BLOCK (i
)->flags
&= ~BB_DUPLICATED
;
723 /* The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also
724 updates the PHI nodes at start of the copied region. In order to
725 achieve this, only loops whose exits all lead to the same location
728 Notice that we do not completely update the SSA web after
729 duplication. The caller is responsible for calling update_ssa
730 after the loop has been duplicated. */
733 gimple_duplicate_loop_to_header_edge (struct loop
*loop
, edge e
,
734 unsigned int ndupl
, sbitmap wont_exit
,
735 edge orig
, vec
<edge
> *to_remove
,
738 unsigned first_new_block
;
740 if (!loops_state_satisfies_p (LOOPS_HAVE_SIMPLE_LATCHES
))
742 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
745 #ifdef ENABLE_CHECKING
746 /* ??? This forces needless update_ssa calls after processing each
747 loop instead of just once after processing all loops. We should
748 instead verify that loop-closed SSA form is up-to-date for LOOP
749 only (and possibly SSA form). For now just skip verifying if
750 there are to-be renamed variables. */
751 if (!need_ssa_update_p (cfun
)
752 && loops_state_satisfies_p (LOOP_CLOSED_SSA
))
753 verify_loop_closed_ssa (true);
756 first_new_block
= last_basic_block
;
757 if (!duplicate_loop_to_header_edge (loop
, e
, ndupl
, wont_exit
,
758 orig
, to_remove
, flags
))
761 /* Readd the removed phi args for e. */
762 flush_pending_stmts (e
);
764 /* Copy the phi node arguments. */
765 copy_phi_node_args (first_new_block
);
772 /* Returns true if we can unroll LOOP FACTOR times. Number
773 of iterations of the loop is returned in NITER. */
776 can_unroll_loop_p (struct loop
*loop
, unsigned factor
,
777 struct tree_niter_desc
*niter
)
781 /* Check whether unrolling is possible. We only want to unroll loops
782 for that we are able to determine number of iterations. We also
783 want to split the extra iterations of the loop from its end,
784 therefore we require that the loop has precisely one
787 exit
= single_dom_exit (loop
);
791 if (!number_of_iterations_exit (loop
, exit
, niter
, false)
792 || niter
->cmp
== ERROR_MARK
793 /* Scalar evolutions analysis might have copy propagated
794 the abnormal ssa names into these expressions, hence
795 emitting the computations based on them during loop
796 unrolling might create overlapping life ranges for
797 them, and failures in out-of-ssa. */
798 || contains_abnormal_ssa_name_p (niter
->may_be_zero
)
799 || contains_abnormal_ssa_name_p (niter
->control
.base
)
800 || contains_abnormal_ssa_name_p (niter
->control
.step
)
801 || contains_abnormal_ssa_name_p (niter
->bound
))
804 /* And of course, we must be able to duplicate the loop. */
805 if (!can_duplicate_loop_p (loop
))
808 /* The final loop should be small enough. */
809 if (tree_num_loop_insns (loop
, &eni_size_weights
) * factor
810 > (unsigned) PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
))
816 /* Determines the conditions that control execution of LOOP unrolled FACTOR
817 times. DESC is number of iterations of LOOP. ENTER_COND is set to
818 condition that must be true if the main loop can be entered.
819 EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing
820 how the exit from the unrolled loop should be controlled. */
823 determine_exit_conditions (struct loop
*loop
, struct tree_niter_desc
*desc
,
824 unsigned factor
, tree
*enter_cond
,
825 tree
*exit_base
, tree
*exit_step
,
826 enum tree_code
*exit_cmp
, tree
*exit_bound
)
829 tree base
= desc
->control
.base
;
830 tree step
= desc
->control
.step
;
831 tree bound
= desc
->bound
;
832 tree type
= TREE_TYPE (step
);
834 tree min
= lower_bound_in_type (type
, type
);
835 tree max
= upper_bound_in_type (type
, type
);
836 enum tree_code cmp
= desc
->cmp
;
837 tree cond
= boolean_true_node
, assum
;
839 /* For pointers, do the arithmetics in the type of step. */
840 base
= fold_convert (type
, base
);
841 bound
= fold_convert (type
, bound
);
843 *enter_cond
= boolean_false_node
;
844 *exit_base
= NULL_TREE
;
845 *exit_step
= NULL_TREE
;
846 *exit_cmp
= ERROR_MARK
;
847 *exit_bound
= NULL_TREE
;
848 gcc_assert (cmp
!= ERROR_MARK
);
850 /* We only need to be correct when we answer question
851 "Do at least FACTOR more iterations remain?" in the unrolled loop.
852 Thus, transforming BASE + STEP * i <> BOUND to
853 BASE + STEP * i < BOUND is ok. */
856 if (tree_int_cst_sign_bit (step
))
861 else if (cmp
== LT_EXPR
)
863 gcc_assert (!tree_int_cst_sign_bit (step
));
865 else if (cmp
== GT_EXPR
)
867 gcc_assert (tree_int_cst_sign_bit (step
));
872 /* The main body of the loop may be entered iff:
874 1) desc->may_be_zero is false.
875 2) it is possible to check that there are at least FACTOR iterations
876 of the loop, i.e., BOUND - step * FACTOR does not overflow.
877 3) # of iterations is at least FACTOR */
879 if (!integer_zerop (desc
->may_be_zero
))
880 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
881 invert_truthvalue (desc
->may_be_zero
),
884 bigstep
= fold_build2 (MULT_EXPR
, type
, step
,
885 build_int_cst_type (type
, factor
));
886 delta
= fold_build2 (MINUS_EXPR
, type
, bigstep
, step
);
888 assum
= fold_build2 (GE_EXPR
, boolean_type_node
,
890 fold_build2 (PLUS_EXPR
, type
, min
, delta
));
892 assum
= fold_build2 (LE_EXPR
, boolean_type_node
,
894 fold_build2 (PLUS_EXPR
, type
, max
, delta
));
895 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
897 bound
= fold_build2 (MINUS_EXPR
, type
, bound
, delta
);
898 assum
= fold_build2 (cmp
, boolean_type_node
, base
, bound
);
899 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
, assum
, cond
);
901 cond
= force_gimple_operand (unshare_expr (cond
), &stmts
, false, NULL_TREE
);
903 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
904 /* cond now may be a gimple comparison, which would be OK, but also any
905 other gimple rhs (say a && b). In this case we need to force it to
907 if (!is_gimple_condexpr (cond
))
909 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
911 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
915 base
= force_gimple_operand (unshare_expr (base
), &stmts
, true, NULL_TREE
);
917 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
918 bound
= force_gimple_operand (unshare_expr (bound
), &stmts
, true, NULL_TREE
);
920 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
923 *exit_step
= bigstep
;
928 /* Scales the frequencies of all basic blocks in LOOP that are strictly
929 dominated by BB by NUM/DEN. */
932 scale_dominated_blocks_in_loop (struct loop
*loop
, basic_block bb
,
940 for (son
= first_dom_son (CDI_DOMINATORS
, bb
);
942 son
= next_dom_son (CDI_DOMINATORS
, son
))
944 if (!flow_bb_inside_loop_p (loop
, son
))
946 scale_bbs_frequencies_int (&son
, 1, num
, den
);
947 scale_dominated_blocks_in_loop (loop
, son
, num
, den
);
951 /* Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP.
952 EXIT is the exit of the loop to that DESC corresponds.
954 If N is number of iterations of the loop and MAY_BE_ZERO is the condition
955 under that loop exits in the first iteration even if N != 0,
959 x = phi (init, next);
967 becomes (with possibly the exit conditions formulated a bit differently,
968 avoiding the need to create a new iv):
970 if (MAY_BE_ZERO || N < FACTOR)
975 x = phi (init, next);
986 } while (N >= FACTOR);
989 init' = phi (init, x);
993 x = phi (init', next);
1001 Before the loop is unrolled, TRANSFORM is called for it (only for the
1002 unrolled loop, but not for its versioned copy). DATA is passed to
1005 /* Probability in % that the unrolled loop is entered. Just a guess. */
1006 #define PROB_UNROLLED_LOOP_ENTERED 90
1009 tree_transform_and_unroll_loop (struct loop
*loop
, unsigned factor
,
1010 edge exit
, struct tree_niter_desc
*desc
,
1011 transform_callback transform
,
1015 tree ctr_before
, ctr_after
;
1016 tree enter_main_cond
, exit_base
, exit_step
, exit_bound
;
1017 enum tree_code exit_cmp
;
1018 gimple phi_old_loop
, phi_new_loop
, phi_rest
;
1019 gimple_stmt_iterator psi_old_loop
, psi_new_loop
;
1020 tree init
, next
, new_init
;
1021 struct loop
*new_loop
;
1022 basic_block rest
, exit_bb
;
1023 edge old_entry
, new_entry
, old_latch
, precond_edge
, new_exit
;
1024 edge new_nonexit
, e
;
1025 gimple_stmt_iterator bsi
;
1028 unsigned est_niter
, prob_entry
, scale_unrolled
, scale_rest
, freq_e
, freq_h
;
1029 unsigned new_est_niter
, i
, prob
;
1030 unsigned irr
= loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
;
1032 vec
<edge
> to_remove
= vNULL
;
1034 est_niter
= expected_loop_iterations (loop
);
1035 determine_exit_conditions (loop
, desc
, factor
,
1036 &enter_main_cond
, &exit_base
, &exit_step
,
1037 &exit_cmp
, &exit_bound
);
1039 /* Let us assume that the unrolled loop is quite likely to be entered. */
1040 if (integer_nonzerop (enter_main_cond
))
1041 prob_entry
= REG_BR_PROB_BASE
;
1043 prob_entry
= PROB_UNROLLED_LOOP_ENTERED
* REG_BR_PROB_BASE
/ 100;
1045 /* The values for scales should keep profile consistent, and somewhat close
1048 TODO: The current value of SCALE_REST makes it appear that the loop that
1049 is created by splitting the remaining iterations of the unrolled loop is
1050 executed the same number of times as the original loop, and with the same
1051 frequencies, which is obviously wrong. This does not appear to cause
1052 problems, so we do not bother with fixing it for now. To make the profile
1053 correct, we would need to change the probability of the exit edge of the
1054 loop, and recompute the distribution of frequencies in its body because
1055 of this change (scale the frequencies of blocks before and after the exit
1056 by appropriate factors). */
1057 scale_unrolled
= prob_entry
;
1058 scale_rest
= REG_BR_PROB_BASE
;
1060 new_loop
= loop_version (loop
, enter_main_cond
, NULL
,
1061 prob_entry
, scale_unrolled
, scale_rest
, true);
1062 gcc_assert (new_loop
!= NULL
);
1063 update_ssa (TODO_update_ssa
);
1065 /* Determine the probability of the exit edge of the unrolled loop. */
1066 new_est_niter
= est_niter
/ factor
;
1068 /* Without profile feedback, loops for that we do not know a better estimate
1069 are assumed to roll 10 times. When we unroll such loop, it appears to
1070 roll too little, and it may even seem to be cold. To avoid this, we
1071 ensure that the created loop appears to roll at least 5 times (but at
1072 most as many times as before unrolling). */
1073 if (new_est_niter
< 5)
1076 new_est_niter
= est_niter
;
1081 /* Prepare the cfg and update the phi nodes. Move the loop exit to the
1082 loop latch (and make its condition dummy, for the moment). */
1083 rest
= loop_preheader_edge (new_loop
)->src
;
1084 precond_edge
= single_pred_edge (rest
);
1085 split_edge (loop_latch_edge (loop
));
1086 exit_bb
= single_pred (loop
->latch
);
1088 /* Since the exit edge will be removed, the frequency of all the blocks
1089 in the loop that are dominated by it must be scaled by
1090 1 / (1 - exit->probability). */
1091 scale_dominated_blocks_in_loop (loop
, exit
->src
,
1093 REG_BR_PROB_BASE
- exit
->probability
);
1095 bsi
= gsi_last_bb (exit_bb
);
1096 exit_if
= gimple_build_cond (EQ_EXPR
, integer_zero_node
,
1098 NULL_TREE
, NULL_TREE
);
1100 gsi_insert_after (&bsi
, exit_if
, GSI_NEW_STMT
);
1101 new_exit
= make_edge (exit_bb
, rest
, EDGE_FALSE_VALUE
| irr
);
1102 rescan_loop_exit (new_exit
, true, false);
1104 /* Set the probability of new exit to the same of the old one. Fix
1105 the frequency of the latch block, by scaling it back by
1106 1 - exit->probability. */
1107 new_exit
->count
= exit
->count
;
1108 new_exit
->probability
= exit
->probability
;
1109 new_nonexit
= single_pred_edge (loop
->latch
);
1110 new_nonexit
->probability
= REG_BR_PROB_BASE
- exit
->probability
;
1111 new_nonexit
->flags
= EDGE_TRUE_VALUE
;
1112 new_nonexit
->count
-= exit
->count
;
1113 if (new_nonexit
->count
< 0)
1114 new_nonexit
->count
= 0;
1115 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1118 old_entry
= loop_preheader_edge (loop
);
1119 new_entry
= loop_preheader_edge (new_loop
);
1120 old_latch
= loop_latch_edge (loop
);
1121 for (psi_old_loop
= gsi_start_phis (loop
->header
),
1122 psi_new_loop
= gsi_start_phis (new_loop
->header
);
1123 !gsi_end_p (psi_old_loop
);
1124 gsi_next (&psi_old_loop
), gsi_next (&psi_new_loop
))
1126 phi_old_loop
= gsi_stmt (psi_old_loop
);
1127 phi_new_loop
= gsi_stmt (psi_new_loop
);
1129 init
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_entry
);
1130 op
= PHI_ARG_DEF_PTR_FROM_EDGE (phi_new_loop
, new_entry
);
1131 gcc_assert (operand_equal_for_phi_arg_p (init
, USE_FROM_PTR (op
)));
1132 next
= PHI_ARG_DEF_FROM_EDGE (phi_old_loop
, old_latch
);
1134 /* Prefer using original variable as a base for the new ssa name.
1135 This is necessary for virtual ops, and useful in order to avoid
1136 losing debug info for real ops. */
1137 if (TREE_CODE (next
) == SSA_NAME
1138 && useless_type_conversion_p (TREE_TYPE (next
),
1140 new_init
= copy_ssa_name (next
, NULL
);
1141 else if (TREE_CODE (init
) == SSA_NAME
1142 && useless_type_conversion_p (TREE_TYPE (init
),
1144 new_init
= copy_ssa_name (init
, NULL
);
1145 else if (useless_type_conversion_p (TREE_TYPE (next
), TREE_TYPE (init
)))
1146 new_init
= make_temp_ssa_name (TREE_TYPE (next
), NULL
, "unrinittmp");
1148 new_init
= make_temp_ssa_name (TREE_TYPE (init
), NULL
, "unrinittmp");
1150 phi_rest
= create_phi_node (new_init
, rest
);
1152 add_phi_arg (phi_rest
, init
, precond_edge
, UNKNOWN_LOCATION
);
1153 add_phi_arg (phi_rest
, next
, new_exit
, UNKNOWN_LOCATION
);
1154 SET_USE (op
, new_init
);
1159 /* Transform the loop. */
1161 (*transform
) (loop
, data
);
1163 /* Unroll the loop and remove the exits in all iterations except for the
1165 wont_exit
= sbitmap_alloc (factor
);
1166 bitmap_ones (wont_exit
);
1167 bitmap_clear_bit (wont_exit
, factor
- 1);
1169 ok
= gimple_duplicate_loop_to_header_edge
1170 (loop
, loop_latch_edge (loop
), factor
- 1,
1171 wont_exit
, new_exit
, &to_remove
, DLTHE_FLAG_UPDATE_FREQ
);
1175 FOR_EACH_VEC_ELT (to_remove
, i
, e
)
1177 ok
= remove_path (e
);
1180 to_remove
.release ();
1181 update_ssa (TODO_update_ssa
);
1183 /* Ensure that the frequencies in the loop match the new estimated
1184 number of iterations, and change the probability of the new
1186 freq_h
= loop
->header
->frequency
;
1187 freq_e
= EDGE_FREQUENCY (loop_preheader_edge (loop
));
1189 scale_loop_frequencies (loop
, freq_e
* (new_est_niter
+ 1), freq_h
);
1191 exit_bb
= single_pred (loop
->latch
);
1192 new_exit
= find_edge (exit_bb
, rest
);
1193 new_exit
->count
= loop_preheader_edge (loop
)->count
;
1194 new_exit
->probability
= REG_BR_PROB_BASE
/ (new_est_niter
+ 1);
1196 rest
->count
+= new_exit
->count
;
1197 rest
->frequency
+= EDGE_FREQUENCY (new_exit
);
1199 new_nonexit
= single_pred_edge (loop
->latch
);
1200 prob
= new_nonexit
->probability
;
1201 new_nonexit
->probability
= REG_BR_PROB_BASE
- new_exit
->probability
;
1202 new_nonexit
->count
= exit_bb
->count
- new_exit
->count
;
1203 if (new_nonexit
->count
< 0)
1204 new_nonexit
->count
= 0;
1206 scale_bbs_frequencies_int (&loop
->latch
, 1, new_nonexit
->probability
,
1209 /* Finally create the new counter for number of iterations and add the new
1210 exit instruction. */
1211 bsi
= gsi_last_nondebug_bb (exit_bb
);
1212 exit_if
= gsi_stmt (bsi
);
1213 create_iv (exit_base
, exit_step
, NULL_TREE
, loop
,
1214 &bsi
, false, &ctr_before
, &ctr_after
);
1215 gimple_cond_set_code (exit_if
, exit_cmp
);
1216 gimple_cond_set_lhs (exit_if
, ctr_after
);
1217 gimple_cond_set_rhs (exit_if
, exit_bound
);
1218 update_stmt (exit_if
);
1220 #ifdef ENABLE_CHECKING
1221 verify_flow_info ();
1222 verify_loop_structure ();
1223 verify_loop_closed_ssa (true);
1227 /* Wrapper over tree_transform_and_unroll_loop for case we do not
1228 want to transform the loop before unrolling. The meaning
1229 of the arguments is the same as for tree_transform_and_unroll_loop. */
1232 tree_unroll_loop (struct loop
*loop
, unsigned factor
,
1233 edge exit
, struct tree_niter_desc
*desc
)
1235 tree_transform_and_unroll_loop (loop
, factor
, exit
, desc
,
1239 /* Rewrite the phi node at position PSI in function of the main
1240 induction variable MAIN_IV and insert the generated code at GSI. */
1243 rewrite_phi_with_iv (loop_p loop
,
1244 gimple_stmt_iterator
*psi
,
1245 gimple_stmt_iterator
*gsi
,
1249 gimple stmt
, phi
= gsi_stmt (*psi
);
1250 tree atype
, mtype
, val
, res
= PHI_RESULT (phi
);
1252 if (virtual_operand_p (res
) || res
== main_iv
)
1258 if (!simple_iv (loop
, loop
, res
, &iv
, true))
1264 remove_phi_node (psi
, false);
1266 atype
= TREE_TYPE (res
);
1267 mtype
= POINTER_TYPE_P (atype
) ? sizetype
: atype
;
1268 val
= fold_build2 (MULT_EXPR
, mtype
, unshare_expr (iv
.step
),
1269 fold_convert (mtype
, main_iv
));
1270 val
= fold_build2 (POINTER_TYPE_P (atype
)
1271 ? POINTER_PLUS_EXPR
: PLUS_EXPR
,
1272 atype
, unshare_expr (iv
.base
), val
);
1273 val
= force_gimple_operand_gsi (gsi
, val
, false, NULL_TREE
, true,
1275 stmt
= gimple_build_assign (res
, val
);
1276 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1277 SSA_NAME_DEF_STMT (res
) = stmt
;
1280 /* Rewrite all the phi nodes of LOOP in function of the main induction
1281 variable MAIN_IV. */
1284 rewrite_all_phi_nodes_with_iv (loop_p loop
, tree main_iv
)
1287 basic_block
*bbs
= get_loop_body_in_dom_order (loop
);
1288 gimple_stmt_iterator psi
;
1290 for (i
= 0; i
< loop
->num_nodes
; i
++)
1292 basic_block bb
= bbs
[i
];
1293 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1295 if (bb
->loop_father
!= loop
)
1298 for (psi
= gsi_start_phis (bb
); !gsi_end_p (psi
); )
1299 rewrite_phi_with_iv (loop
, &psi
, &gsi
, main_iv
);
1305 /* Bases all the induction variables in LOOP on a single induction
1306 variable (unsigned with base 0 and step 1), whose final value is
1307 compared with *NIT. When the IV type precision has to be larger
1308 than *NIT type precision, *NIT is converted to the larger type, the
1309 conversion code is inserted before the loop, and *NIT is updated to
1310 the new definition. When BUMP_IN_LATCH is true, the induction
1311 variable is incremented in the loop latch, otherwise it is
1312 incremented in the loop header. Return the induction variable that
1316 canonicalize_loop_ivs (struct loop
*loop
, tree
*nit
, bool bump_in_latch
)
1318 unsigned precision
= TYPE_PRECISION (TREE_TYPE (*nit
));
1319 unsigned original_precision
= precision
;
1320 tree type
, var_before
;
1321 gimple_stmt_iterator gsi
, psi
;
1323 edge exit
= single_dom_exit (loop
);
1325 enum machine_mode mode
;
1326 bool unsigned_p
= false;
1328 for (psi
= gsi_start_phis (loop
->header
);
1329 !gsi_end_p (psi
); gsi_next (&psi
))
1331 gimple phi
= gsi_stmt (psi
);
1332 tree res
= PHI_RESULT (phi
);
1335 type
= TREE_TYPE (res
);
1336 if (virtual_operand_p (res
)
1337 || (!INTEGRAL_TYPE_P (type
)
1338 && !POINTER_TYPE_P (type
))
1339 || TYPE_PRECISION (type
) < precision
)
1342 uns
= POINTER_TYPE_P (type
) | TYPE_UNSIGNED (type
);
1344 if (TYPE_PRECISION (type
) > precision
)
1349 precision
= TYPE_PRECISION (type
);
1352 mode
= smallest_mode_for_size (precision
, MODE_INT
);
1353 precision
= GET_MODE_PRECISION (mode
);
1354 type
= build_nonstandard_integer_type (precision
, unsigned_p
);
1356 if (original_precision
!= precision
)
1358 *nit
= fold_convert (type
, *nit
);
1359 *nit
= force_gimple_operand (*nit
, &stmts
, true, NULL_TREE
);
1361 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop
), stmts
);
1365 gsi
= gsi_last_bb (loop
->latch
);
1367 gsi
= gsi_last_nondebug_bb (loop
->header
);
1368 create_iv (build_int_cst_type (type
, 0), build_int_cst (type
, 1), NULL_TREE
,
1369 loop
, &gsi
, bump_in_latch
, &var_before
, NULL
);
1371 rewrite_all_phi_nodes_with_iv (loop
, var_before
);
1373 stmt
= last_stmt (exit
->src
);
1374 /* Make the loop exit if the control condition is not satisfied. */
1375 if (exit
->flags
& EDGE_TRUE_VALUE
)
1379 extract_true_false_edges_from_block (exit
->src
, &te
, &fe
);
1380 te
->flags
= EDGE_FALSE_VALUE
;
1381 fe
->flags
= EDGE_TRUE_VALUE
;
1383 gimple_cond_set_code (stmt
, LT_EXPR
);
1384 gimple_cond_set_lhs (stmt
, var_before
);
1385 gimple_cond_set_rhs (stmt
, *nit
);