1 /* Lower GIMPLE_SWITCH expressions to something more efficient than
3 Copyright (C) 2006-2013 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 3, or (at your option) any
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
22 /* This file handles the lowering of GIMPLE_SWITCH to an indexed
23 load, or a series of bit-test-and-branch expressions. */
27 #include "coretypes.h"
33 #include "basic-block.h"
35 #include "gimple-ssa.h"
38 #include "tree-phinodes.h"
39 #include "tree-ssanames.h"
40 #include "tree-pass.h"
41 #include "gimple-pretty-print.h"
44 /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode
45 type in the GIMPLE type system that is language-independent? */
46 #include "langhooks.h"
48 /* Need to include expr.h and optabs.h for lshift_cheap_p. */
52 /* Maximum number of case bit tests.
53 FIXME: This should be derived from PARAM_CASE_VALUES_THRESHOLD and
54 targetm.case_values_threshold(), or be its own param. */
55 #define MAX_CASE_BIT_TESTS 3
57 /* Split the basic block at the statement pointed to by GSIP, and insert
58 a branch to the target basic block of E_TRUE conditional on tree
61 It is assumed that there is already an edge from the to-be-split
62 basic block to E_TRUE->dest block. This edge is removed, and the
63 profile information on the edge is re-used for the new conditional
66 The CFG is updated. The dominator tree will not be valid after
67 this transformation, but the immediate dominators are updated if
68 UPDATE_DOMINATORS is true.
70 Returns the newly created basic block. */
73 hoist_edge_and_branch_if_true (gimple_stmt_iterator
*gsip
,
74 tree cond
, edge e_true
,
75 bool update_dominators
)
80 basic_block new_bb
, split_bb
= gsi_bb (*gsip
);
81 bool dominated_e_true
= false;
83 gcc_assert (e_true
->src
== split_bb
);
86 && get_immediate_dominator (CDI_DOMINATORS
, e_true
->dest
) == split_bb
)
87 dominated_e_true
= true;
89 tmp
= force_gimple_operand_gsi (gsip
, cond
, /*simple=*/true, NULL
,
90 /*before=*/true, GSI_SAME_STMT
);
91 cond_stmt
= gimple_build_cond_from_tree (tmp
, NULL_TREE
, NULL_TREE
);
92 gsi_insert_before (gsip
, cond_stmt
, GSI_SAME_STMT
);
94 e_false
= split_block (split_bb
, cond_stmt
);
95 new_bb
= e_false
->dest
;
96 redirect_edge_pred (e_true
, split_bb
);
98 e_true
->flags
&= ~EDGE_FALLTHRU
;
99 e_true
->flags
|= EDGE_TRUE_VALUE
;
101 e_false
->flags
&= ~EDGE_FALLTHRU
;
102 e_false
->flags
|= EDGE_FALSE_VALUE
;
103 e_false
->probability
= REG_BR_PROB_BASE
- e_true
->probability
;
104 e_false
->count
= split_bb
->count
- e_true
->count
;
105 new_bb
->count
= e_false
->count
;
107 if (update_dominators
)
109 if (dominated_e_true
)
110 set_immediate_dominator (CDI_DOMINATORS
, e_true
->dest
, split_bb
);
111 set_immediate_dominator (CDI_DOMINATORS
, e_false
->dest
, split_bb
);
118 /* Determine whether "1 << x" is relatively cheap in word_mode. */
119 /* FIXME: This is the function that we need rtl.h and optabs.h for.
120 This function (and similar RTL-related cost code in e.g. IVOPTS) should
121 be moved to some kind of interface file for GIMPLE/RTL interactions. */
123 lshift_cheap_p (void)
125 /* FIXME: This should be made target dependent via this "this_target"
126 mechanism, similar to e.g. can_copy_init_p in gcse.c. */
127 static bool init
[2] = {false, false};
128 static bool cheap
[2] = {true, true};
131 /* If the targer has no lshift in word_mode, the operation will most
132 probably not be cheap. ??? Does GCC even work for such targets? */
133 if (optab_handler (ashl_optab
, word_mode
) == CODE_FOR_nothing
)
136 speed_p
= optimize_insn_for_speed_p ();
140 rtx reg
= gen_raw_REG (word_mode
, 10000);
141 int cost
= set_src_cost (gen_rtx_ASHIFT (word_mode
, const1_rtx
, reg
),
143 cheap
[speed_p
] = cost
< COSTS_N_INSNS (MAX_CASE_BIT_TESTS
);
144 init
[speed_p
] = true;
147 return cheap
[speed_p
];
150 /* Return true if a switch should be expanded as a bit test.
151 RANGE is the difference between highest and lowest case.
152 UNIQ is number of unique case node targets, not counting the default case.
153 COUNT is the number of comparisons needed, not counting the default case. */
156 expand_switch_using_bit_tests_p (tree range
,
160 return (((uniq
== 1 && count
>= 3)
161 || (uniq
== 2 && count
>= 5)
162 || (uniq
== 3 && count
>= 6))
164 && compare_tree_int (range
, GET_MODE_BITSIZE (word_mode
)) < 0
165 && compare_tree_int (range
, 0) > 0);
168 /* Implement switch statements with bit tests
170 A GIMPLE switch statement can be expanded to a short sequence of bit-wise
171 comparisons. "switch(x)" is converted into "if ((1 << (x-MINVAL)) & CST)"
172 where CST and MINVAL are integer constants. This is better than a series
173 of compare-and-banch insns in some cases, e.g. we can implement:
175 if ((x==4) || (x==6) || (x==9) || (x==11))
177 as a single bit test:
179 if ((1<<x) & ((1<<4)|(1<<6)|(1<<9)|(1<<11)))
181 This transformation is only applied if the number of case targets is small,
182 if CST constains at least 3 bits, and "1 << x" is cheap. The bit tests are
183 performed in "word_mode".
185 The following example shows the code the transformation generates:
191 case '0': case '1': case '2': case '3': case '4':
192 case '5': case '6': case '7': case '8': case '9':
193 case 'A': case 'B': case 'C': case 'D': case 'E':
205 if (tmp1 > (70 - 48)) goto L2;
207 tmp3 = 0b11111100000001111111111;
208 if ((tmp2 & tmp3) != 0) goto L1 ; else goto L2;
215 TODO: There are still some improvements to this transformation that could
218 * A narrower mode than word_mode could be used if that is cheaper, e.g.
219 for x86_64 where a narrower-mode shift may result in smaller code.
221 * The compounded constant could be shifted rather than the one. The
222 test would be either on the sign bit or on the least significant bit,
223 depending on the direction of the shift. On some machines, the test
224 for the branch would be free if the bit to test is already set by the
227 This transformation was contributed by Roger Sayle, see this e-mail:
228 http://gcc.gnu.org/ml/gcc-patches/2003-01/msg01950.html
231 /* A case_bit_test represents a set of case nodes that may be
232 selected from using a bit-wise comparison. HI and LO hold
233 the integer to be tested against, TARGET_EDGE contains the
234 edge to the basic block to jump to upon success and BITS
235 counts the number of case nodes handled by this test,
236 typically the number of bits set in HI:LO. The LABEL field
237 is used to quickly identify all cases in this set without
238 looking at label_to_block for every case label. */
249 /* Comparison function for qsort to order bit tests by decreasing
250 probability of execution. Our best guess comes from a measured
251 profile. If the profile counts are equal, break even on the
252 number of case nodes, i.e. the node with the most cases gets
255 TODO: Actually this currently runs before a profile is available.
256 Therefore the case-as-bit-tests transformation should be done
257 later in the pass pipeline, or something along the lines of
258 "Efficient and effective branch reordering using profile data"
259 (Yang et. al., 2002) should be implemented (although, how good
260 is a paper is called "Efficient and effective ..." when the
261 latter is implied by the former, but oh well...). */
264 case_bit_test_cmp (const void *p1
, const void *p2
)
266 const struct case_bit_test
*const d1
= (const struct case_bit_test
*) p1
;
267 const struct case_bit_test
*const d2
= (const struct case_bit_test
*) p2
;
269 if (d2
->target_edge
->count
!= d1
->target_edge
->count
)
270 return d2
->target_edge
->count
- d1
->target_edge
->count
;
271 if (d2
->bits
!= d1
->bits
)
272 return d2
->bits
- d1
->bits
;
274 /* Stabilize the sort. */
275 return LABEL_DECL_UID (d2
->label
) - LABEL_DECL_UID (d1
->label
);
278 /* Expand a switch statement by a short sequence of bit-wise
279 comparisons. "switch(x)" is effectively converted into
280 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are
283 INDEX_EXPR is the value being switched on.
285 MINVAL is the lowest case value of in the case nodes,
286 and RANGE is highest value minus MINVAL. MINVAL and RANGE
287 are not guaranteed to be of the same type as INDEX_EXPR
288 (the gimplifier doesn't change the type of case label values,
289 and MINVAL and RANGE are derived from those values).
291 There *MUST* be MAX_CASE_BIT_TESTS or less unique case
295 emit_case_bit_tests (gimple swtch
, tree index_expr
,
296 tree minval
, tree range
)
298 struct case_bit_test test
[MAX_CASE_BIT_TESTS
];
299 unsigned int i
, j
, k
;
302 basic_block switch_bb
= gimple_bb (swtch
);
303 basic_block default_bb
, new_default_bb
, new_bb
;
305 bool update_dom
= dom_info_available_p (CDI_DOMINATORS
);
307 vec
<basic_block
> bbs_to_fix_dom
= vNULL
;
309 tree index_type
= TREE_TYPE (index_expr
);
310 tree unsigned_index_type
= unsigned_type_for (index_type
);
311 unsigned int branch_num
= gimple_switch_num_labels (swtch
);
313 gimple_stmt_iterator gsi
;
317 tree word_type_node
= lang_hooks
.types
.type_for_mode (word_mode
, 1);
318 tree word_mode_zero
= fold_convert (word_type_node
, integer_zero_node
);
319 tree word_mode_one
= fold_convert (word_type_node
, integer_one_node
);
321 memset (&test
, 0, sizeof (test
));
323 /* Get the edge for the default case. */
324 tmp
= gimple_switch_default_label (swtch
);
325 default_bb
= label_to_block (CASE_LABEL (tmp
));
326 default_edge
= find_edge (switch_bb
, default_bb
);
328 /* Go through all case labels, and collect the case labels, profile
329 counts, and other information we need to build the branch tests. */
331 for (i
= 1; i
< branch_num
; i
++)
334 tree cs
= gimple_switch_label (swtch
, i
);
335 tree label
= CASE_LABEL (cs
);
336 edge e
= find_edge (switch_bb
, label_to_block (label
));
337 for (k
= 0; k
< count
; k
++)
338 if (e
== test
[k
].target_edge
)
343 gcc_checking_assert (count
< MAX_CASE_BIT_TESTS
);
346 test
[k
].target_edge
= e
;
347 test
[k
].label
= label
;
354 lo
= tree_low_cst (int_const_binop (MINUS_EXPR
,
355 CASE_LOW (cs
), minval
),
357 if (CASE_HIGH (cs
) == NULL_TREE
)
360 hi
= tree_low_cst (int_const_binop (MINUS_EXPR
,
361 CASE_HIGH (cs
), minval
),
364 for (j
= lo
; j
<= hi
; j
++)
365 if (j
>= HOST_BITS_PER_WIDE_INT
)
366 test
[k
].hi
|= (HOST_WIDE_INT
) 1 << (j
- HOST_BITS_PER_INT
);
368 test
[k
].lo
|= (HOST_WIDE_INT
) 1 << j
;
371 qsort (test
, count
, sizeof (*test
), case_bit_test_cmp
);
373 /* We generate two jumps to the default case label.
374 Split the default edge, so that we don't have to do any PHI node
376 new_default_bb
= split_edge (default_edge
);
380 bbs_to_fix_dom
.create (10);
381 bbs_to_fix_dom
.quick_push (switch_bb
);
382 bbs_to_fix_dom
.quick_push (default_bb
);
383 bbs_to_fix_dom
.quick_push (new_default_bb
);
386 /* Now build the test-and-branch code. */
388 gsi
= gsi_last_bb (switch_bb
);
390 /* idx = (unsigned)x - minval. */
391 idx
= fold_convert (unsigned_index_type
, index_expr
);
392 idx
= fold_build2 (MINUS_EXPR
, unsigned_index_type
, idx
,
393 fold_convert (unsigned_index_type
, minval
));
394 idx
= force_gimple_operand_gsi (&gsi
, idx
,
395 /*simple=*/true, NULL_TREE
,
396 /*before=*/true, GSI_SAME_STMT
);
398 /* if (idx > range) goto default */
399 range
= force_gimple_operand_gsi (&gsi
,
400 fold_convert (unsigned_index_type
, range
),
401 /*simple=*/true, NULL_TREE
,
402 /*before=*/true, GSI_SAME_STMT
);
403 tmp
= fold_build2 (GT_EXPR
, boolean_type_node
, idx
, range
);
404 new_bb
= hoist_edge_and_branch_if_true (&gsi
, tmp
, default_edge
, update_dom
);
406 bbs_to_fix_dom
.quick_push (new_bb
);
407 gcc_assert (gimple_bb (swtch
) == new_bb
);
408 gsi
= gsi_last_bb (new_bb
);
410 /* Any blocks dominated by the GIMPLE_SWITCH, but that are not successors
411 of NEW_BB, are still immediately dominated by SWITCH_BB. Make it so. */
414 vec
<basic_block
> dom_bbs
;
417 dom_bbs
= get_dominated_by (CDI_DOMINATORS
, new_bb
);
418 FOR_EACH_VEC_ELT (dom_bbs
, i
, dom_son
)
420 edge e
= find_edge (new_bb
, dom_son
);
421 if (e
&& single_pred_p (e
->dest
))
423 set_immediate_dominator (CDI_DOMINATORS
, dom_son
, switch_bb
);
424 bbs_to_fix_dom
.safe_push (dom_son
);
429 /* csui = (1 << (word_mode) idx) */
430 csui
= make_ssa_name (word_type_node
, NULL
);
431 tmp
= fold_build2 (LSHIFT_EXPR
, word_type_node
, word_mode_one
,
432 fold_convert (word_type_node
, idx
));
433 tmp
= force_gimple_operand_gsi (&gsi
, tmp
,
434 /*simple=*/false, NULL_TREE
,
435 /*before=*/true, GSI_SAME_STMT
);
436 shift_stmt
= gimple_build_assign (csui
, tmp
);
437 gsi_insert_before (&gsi
, shift_stmt
, GSI_SAME_STMT
);
438 update_stmt (shift_stmt
);
440 /* for each unique set of cases:
441 if (const & csui) goto target */
442 for (k
= 0; k
< count
; k
++)
444 tmp
= build_int_cst_wide (word_type_node
, test
[k
].lo
, test
[k
].hi
);
445 tmp
= fold_build2 (BIT_AND_EXPR
, word_type_node
, csui
, tmp
);
446 tmp
= force_gimple_operand_gsi (&gsi
, tmp
,
447 /*simple=*/true, NULL_TREE
,
448 /*before=*/true, GSI_SAME_STMT
);
449 tmp
= fold_build2 (NE_EXPR
, boolean_type_node
, tmp
, word_mode_zero
);
450 new_bb
= hoist_edge_and_branch_if_true (&gsi
, tmp
, test
[k
].target_edge
,
453 bbs_to_fix_dom
.safe_push (new_bb
);
454 gcc_assert (gimple_bb (swtch
) == new_bb
);
455 gsi
= gsi_last_bb (new_bb
);
458 /* We should have removed all edges now. */
459 gcc_assert (EDGE_COUNT (gsi_bb (gsi
)->succs
) == 0);
461 /* If nothing matched, go to the default label. */
462 make_edge (gsi_bb (gsi
), new_default_bb
, EDGE_FALLTHRU
);
464 /* The GIMPLE_SWITCH is now redundant. */
465 gsi_remove (&gsi
, true);
469 /* Fix up the dominator tree. */
470 iterate_fix_dominators (CDI_DOMINATORS
, bbs_to_fix_dom
, true);
471 bbs_to_fix_dom
.release ();
476 Switch initialization conversion
478 The following pass changes simple initializations of scalars in a switch
479 statement into initializations from a static array. Obviously, the values
480 must be constant and known at compile time and a default branch must be
481 provided. For example, the following code:
505 a_5 = PHI <a_1, a_2, a_3, a_4>
506 b_5 = PHI <b_1, b_2, b_3, b_4>
511 static const int = CSWTCH01[] = {6, 6, 5, 1, 1, 1, 1, 1, 1, 1, 1, 4};
512 static const int = CSWTCH02[] = {8, 8, 9, 16, 16, 16, 16, 16, 16, 16,
515 if (((unsigned) argc) - 1 < 11)
517 a_6 = CSWTCH02[argc - 1];
518 b_6 = CSWTCH01[argc - 1];
528 There are further constraints. Specifically, the range of values across all
529 case labels must not be bigger than SWITCH_CONVERSION_BRANCH_RATIO (default
530 eight) times the number of the actual switch branches.
532 This transformation was contributed by Martin Jambor, see this e-mail:
533 http://gcc.gnu.org/ml/gcc-patches/2008-07/msg00011.html */
535 /* The main structure of the pass. */
536 struct switch_conv_info
538 /* The expression used to decide the switch branch. */
541 /* The following integer constants store the minimum and maximum value
542 covered by the case labels. */
546 /* The difference between the above two numbers. Stored here because it
547 is used in all the conversion heuristics, as well as for some of the
548 transformation, and it is expensive to re-compute it all the time. */
551 /* Basic block that contains the actual GIMPLE_SWITCH. */
552 basic_block switch_bb
;
554 /* Basic block that is the target of the default case. */
555 basic_block default_bb
;
557 /* The single successor block of all branches out of the GIMPLE_SWITCH,
558 if such a block exists. Otherwise NULL. */
559 basic_block final_bb
;
561 /* The probability of the default edge in the replaced switch. */
564 /* The count of the default edge in the replaced switch. */
565 gcov_type default_count
;
567 /* Combined count of all other (non-default) edges in the replaced switch. */
568 gcov_type other_count
;
570 /* Number of phi nodes in the final bb (that we'll be replacing). */
573 /* Array of default values, in the same order as phi nodes. */
574 tree
*default_values
;
576 /* Constructors of new static arrays. */
577 vec
<constructor_elt
, va_gc
> **constructors
;
579 /* Array of ssa names that are initialized with a value from a new static
581 tree
*target_inbound_names
;
583 /* Array of ssa names that are initialized with the default value if the
584 switch expression is out of range. */
585 tree
*target_outbound_names
;
587 /* The first load statement that loads a temporary from a new static array.
589 gimple arr_ref_first
;
591 /* The last load statement that loads a temporary from a new static array. */
594 /* String reason why the case wasn't a good candidate that is written to the
595 dump file, if there is one. */
598 /* Parameters for expand_switch_using_bit_tests. Should be computed
599 the same way as in expand_case. */
604 /* Collect information about GIMPLE_SWITCH statement SWTCH into INFO. */
607 collect_switch_conv_info (gimple swtch
, struct switch_conv_info
*info
)
609 unsigned int branch_num
= gimple_switch_num_labels (swtch
);
610 tree min_case
, max_case
;
611 unsigned int count
, i
;
615 memset (info
, 0, sizeof (*info
));
617 /* The gimplifier has already sorted the cases by CASE_LOW and ensured there
618 is a default label which is the first in the vector.
619 Collect the bits we can deduce from the CFG. */
620 info
->index_expr
= gimple_switch_index (swtch
);
621 info
->switch_bb
= gimple_bb (swtch
);
623 label_to_block (CASE_LABEL (gimple_switch_default_label (swtch
)));
624 e_default
= find_edge (info
->switch_bb
, info
->default_bb
);
625 info
->default_prob
= e_default
->probability
;
626 info
->default_count
= e_default
->count
;
627 FOR_EACH_EDGE (e
, ei
, info
->switch_bb
->succs
)
629 info
->other_count
+= e
->count
;
631 /* See if there is one common successor block for all branch
632 targets. If it exists, record it in FINAL_BB. */
633 FOR_EACH_EDGE (e
, ei
, info
->switch_bb
->succs
)
635 if (! single_pred_p (e
->dest
))
637 info
->final_bb
= e
->dest
;
642 FOR_EACH_EDGE (e
, ei
, info
->switch_bb
->succs
)
644 if (e
->dest
== info
->final_bb
)
647 if (single_pred_p (e
->dest
)
648 && single_succ_p (e
->dest
)
649 && single_succ (e
->dest
) == info
->final_bb
)
652 info
->final_bb
= NULL
;
656 /* Get upper and lower bounds of case values, and the covered range. */
657 min_case
= gimple_switch_label (swtch
, 1);
658 max_case
= gimple_switch_label (swtch
, branch_num
- 1);
660 info
->range_min
= CASE_LOW (min_case
);
661 if (CASE_HIGH (max_case
) != NULL_TREE
)
662 info
->range_max
= CASE_HIGH (max_case
);
664 info
->range_max
= CASE_LOW (max_case
);
667 int_const_binop (MINUS_EXPR
, info
->range_max
, info
->range_min
);
669 /* Get a count of the number of case labels. Single-valued case labels
670 simply count as one, but a case range counts double, since it may
671 require two compares if it gets lowered as a branching tree. */
673 for (i
= 1; i
< branch_num
; i
++)
675 tree elt
= gimple_switch_label (swtch
, i
);
678 && ! tree_int_cst_equal (CASE_LOW (elt
), CASE_HIGH (elt
)))
683 /* Get the number of unique non-default targets out of the GIMPLE_SWITCH
684 block. Assume a CFG cleanup would have already removed degenerate
685 switch statements, this allows us to just use EDGE_COUNT. */
686 info
->uniq
= EDGE_COUNT (gimple_bb (swtch
)->succs
) - 1;
689 /* Checks whether the range given by individual case statements of the SWTCH
690 switch statement isn't too big and whether the number of branches actually
691 satisfies the size of the new array. */
694 check_range (struct switch_conv_info
*info
)
696 gcc_assert (info
->range_size
);
697 if (!host_integerp (info
->range_size
, 1))
699 info
->reason
= "index range way too large or otherwise unusable";
703 if ((unsigned HOST_WIDE_INT
) tree_low_cst (info
->range_size
, 1)
704 > ((unsigned) info
->count
* SWITCH_CONVERSION_BRANCH_RATIO
))
706 info
->reason
= "the maximum range-branch ratio exceeded";
713 /* Checks whether all but the FINAL_BB basic blocks are empty. */
716 check_all_empty_except_final (struct switch_conv_info
*info
)
721 FOR_EACH_EDGE (e
, ei
, info
->switch_bb
->succs
)
723 if (e
->dest
== info
->final_bb
)
726 if (!empty_block_p (e
->dest
))
728 info
->reason
= "bad case - a non-final BB not empty";
736 /* This function checks whether all required values in phi nodes in final_bb
737 are constants. Required values are those that correspond to a basic block
738 which is a part of the examined switch statement. It returns true if the
739 phi nodes are OK, otherwise false. */
742 check_final_bb (struct switch_conv_info
*info
)
744 gimple_stmt_iterator gsi
;
747 for (gsi
= gsi_start_phis (info
->final_bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
749 gimple phi
= gsi_stmt (gsi
);
754 for (i
= 0; i
< gimple_phi_num_args (phi
); i
++)
756 basic_block bb
= gimple_phi_arg_edge (phi
, i
)->src
;
758 if (bb
== info
->switch_bb
759 || (single_pred_p (bb
) && single_pred (bb
) == info
->switch_bb
))
763 val
= gimple_phi_arg_def (phi
, i
);
764 if (!is_gimple_ip_invariant (val
))
766 info
->reason
= "non-invariant value from a case";
767 return false; /* Non-invariant argument. */
769 reloc
= initializer_constant_valid_p (val
, TREE_TYPE (val
));
770 if ((flag_pic
&& reloc
!= null_pointer_node
)
771 || (!flag_pic
&& reloc
== NULL_TREE
))
775 = "value from a case would need runtime relocations";
778 = "value from a case is not a valid initializer";
788 /* The following function allocates default_values, target_{in,out}_names and
789 constructors arrays. The last one is also populated with pointers to
790 vectors that will become constructors of new arrays. */
793 create_temp_arrays (struct switch_conv_info
*info
)
797 info
->default_values
= XCNEWVEC (tree
, info
->phi_count
* 3);
798 /* ??? Macros do not support multi argument templates in their
799 argument list. We create a typedef to work around that problem. */
800 typedef vec
<constructor_elt
, va_gc
> *vec_constructor_elt_gc
;
801 info
->constructors
= XCNEWVEC (vec_constructor_elt_gc
, info
->phi_count
);
802 info
->target_inbound_names
= info
->default_values
+ info
->phi_count
;
803 info
->target_outbound_names
= info
->target_inbound_names
+ info
->phi_count
;
804 for (i
= 0; i
< info
->phi_count
; i
++)
805 vec_alloc (info
->constructors
[i
], tree_low_cst (info
->range_size
, 1) + 1);
808 /* Free the arrays created by create_temp_arrays(). The vectors that are
809 created by that function are not freed here, however, because they have
810 already become constructors and must be preserved. */
813 free_temp_arrays (struct switch_conv_info
*info
)
815 XDELETEVEC (info
->constructors
);
816 XDELETEVEC (info
->default_values
);
819 /* Populate the array of default values in the order of phi nodes.
820 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch. */
823 gather_default_values (tree default_case
, struct switch_conv_info
*info
)
825 gimple_stmt_iterator gsi
;
826 basic_block bb
= label_to_block (CASE_LABEL (default_case
));
830 gcc_assert (CASE_LOW (default_case
) == NULL_TREE
);
832 if (bb
== info
->final_bb
)
833 e
= find_edge (info
->switch_bb
, bb
);
835 e
= single_succ_edge (bb
);
837 for (gsi
= gsi_start_phis (info
->final_bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
839 gimple phi
= gsi_stmt (gsi
);
840 tree val
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
842 info
->default_values
[i
++] = val
;
846 /* The following function populates the vectors in the constructors array with
847 future contents of the static arrays. The vectors are populated in the
848 order of phi nodes. SWTCH is the switch statement being converted. */
851 build_constructors (gimple swtch
, struct switch_conv_info
*info
)
853 unsigned i
, branch_num
= gimple_switch_num_labels (swtch
);
854 tree pos
= info
->range_min
;
856 for (i
= 1; i
< branch_num
; i
++)
858 tree cs
= gimple_switch_label (swtch
, i
);
859 basic_block bb
= label_to_block (CASE_LABEL (cs
));
862 gimple_stmt_iterator gsi
;
865 if (bb
== info
->final_bb
)
866 e
= find_edge (info
->switch_bb
, bb
);
868 e
= single_succ_edge (bb
);
871 while (tree_int_cst_lt (pos
, CASE_LOW (cs
)))
874 for (k
= 0; k
< info
->phi_count
; k
++)
878 elt
.index
= int_const_binop (MINUS_EXPR
, pos
, info
->range_min
);
880 = unshare_expr_without_location (info
->default_values
[k
]);
881 info
->constructors
[k
]->quick_push (elt
);
884 pos
= int_const_binop (PLUS_EXPR
, pos
, integer_one_node
);
886 gcc_assert (tree_int_cst_equal (pos
, CASE_LOW (cs
)));
890 high
= CASE_HIGH (cs
);
892 high
= CASE_LOW (cs
);
893 for (gsi
= gsi_start_phis (info
->final_bb
);
894 !gsi_end_p (gsi
); gsi_next (&gsi
))
896 gimple phi
= gsi_stmt (gsi
);
897 tree val
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
898 tree low
= CASE_LOW (cs
);
905 elt
.index
= int_const_binop (MINUS_EXPR
, pos
, info
->range_min
);
906 elt
.value
= unshare_expr_without_location (val
);
907 info
->constructors
[j
]->quick_push (elt
);
909 pos
= int_const_binop (PLUS_EXPR
, pos
, integer_one_node
);
910 } while (!tree_int_cst_lt (high
, pos
)
911 && tree_int_cst_lt (low
, pos
));
917 /* If all values in the constructor vector are the same, return the value.
918 Otherwise return NULL_TREE. Not supposed to be called for empty
922 constructor_contains_same_values_p (vec
<constructor_elt
, va_gc
> *vec
)
925 tree prev
= NULL_TREE
;
926 constructor_elt
*elt
;
928 FOR_EACH_VEC_SAFE_ELT (vec
, i
, elt
)
932 else if (!operand_equal_p (elt
->value
, prev
, OEP_ONLY_CONST
))
938 /* Return type which should be used for array elements, either TYPE,
939 or for integral type some smaller integral type that can still hold
940 all the constants. */
943 array_value_type (gimple swtch
, tree type
, int num
,
944 struct switch_conv_info
*info
)
946 unsigned int i
, len
= vec_safe_length (info
->constructors
[num
]);
947 constructor_elt
*elt
;
948 enum machine_mode mode
;
952 if (!INTEGRAL_TYPE_P (type
))
955 mode
= GET_CLASS_NARROWEST_MODE (GET_MODE_CLASS (TYPE_MODE (type
)));
956 if (GET_MODE_SIZE (TYPE_MODE (type
)) <= GET_MODE_SIZE (mode
))
959 if (len
< (optimize_bb_for_size_p (gimple_bb (swtch
)) ? 2 : 32))
962 FOR_EACH_VEC_SAFE_ELT (info
->constructors
[num
], i
, elt
)
966 if (TREE_CODE (elt
->value
) != INTEGER_CST
)
969 cst
= TREE_INT_CST (elt
->value
);
972 unsigned int prec
= GET_MODE_BITSIZE (mode
);
973 if (prec
> HOST_BITS_PER_WIDE_INT
)
976 if (sign
>= 0 && cst
== cst
.zext (prec
))
978 if (sign
== 0 && cst
== cst
.sext (prec
))
983 if (sign
<= 0 && cst
== cst
.sext (prec
))
992 mode
= GET_MODE_WIDER_MODE (mode
);
994 || GET_MODE_SIZE (mode
) >= GET_MODE_SIZE (TYPE_MODE (type
)))
1000 sign
= TYPE_UNSIGNED (type
) ? 1 : -1;
1001 smaller_type
= lang_hooks
.types
.type_for_mode (mode
, sign
>= 0);
1002 if (GET_MODE_SIZE (TYPE_MODE (type
))
1003 <= GET_MODE_SIZE (TYPE_MODE (smaller_type
)))
1006 return smaller_type
;
1009 /* Create an appropriate array type and declaration and assemble a static array
1010 variable. Also create a load statement that initializes the variable in
1011 question with a value from the static array. SWTCH is the switch statement
1012 being converted, NUM is the index to arrays of constructors, default values
1013 and target SSA names for this particular array. ARR_INDEX_TYPE is the type
1014 of the index of the new array, PHI is the phi node of the final BB that
1015 corresponds to the value that will be loaded from the created array. TIDX
1016 is an ssa name of a temporary variable holding the index for loads from the
1020 build_one_array (gimple swtch
, int num
, tree arr_index_type
, gimple phi
,
1021 tree tidx
, struct switch_conv_info
*info
)
1025 gimple_stmt_iterator gsi
= gsi_for_stmt (swtch
);
1026 location_t loc
= gimple_location (swtch
);
1028 gcc_assert (info
->default_values
[num
]);
1030 name
= copy_ssa_name (PHI_RESULT (phi
), NULL
);
1031 info
->target_inbound_names
[num
] = name
;
1033 cst
= constructor_contains_same_values_p (info
->constructors
[num
]);
1035 load
= gimple_build_assign (name
, cst
);
1038 tree array_type
, ctor
, decl
, value_type
, fetch
, default_type
;
1040 default_type
= TREE_TYPE (info
->default_values
[num
]);
1041 value_type
= array_value_type (swtch
, default_type
, num
, info
);
1042 array_type
= build_array_type (value_type
, arr_index_type
);
1043 if (default_type
!= value_type
)
1046 constructor_elt
*elt
;
1048 FOR_EACH_VEC_SAFE_ELT (info
->constructors
[num
], i
, elt
)
1049 elt
->value
= fold_convert (value_type
, elt
->value
);
1051 ctor
= build_constructor (array_type
, info
->constructors
[num
]);
1052 TREE_CONSTANT (ctor
) = true;
1053 TREE_STATIC (ctor
) = true;
1055 decl
= build_decl (loc
, VAR_DECL
, NULL_TREE
, array_type
);
1056 TREE_STATIC (decl
) = 1;
1057 DECL_INITIAL (decl
) = ctor
;
1059 DECL_NAME (decl
) = create_tmp_var_name ("CSWTCH");
1060 DECL_ARTIFICIAL (decl
) = 1;
1061 TREE_CONSTANT (decl
) = 1;
1062 TREE_READONLY (decl
) = 1;
1063 varpool_finalize_decl (decl
);
1065 fetch
= build4 (ARRAY_REF
, value_type
, decl
, tidx
, NULL_TREE
,
1067 if (default_type
!= value_type
)
1069 fetch
= fold_convert (default_type
, fetch
);
1070 fetch
= force_gimple_operand_gsi (&gsi
, fetch
, true, NULL_TREE
,
1071 true, GSI_SAME_STMT
);
1073 load
= gimple_build_assign (name
, fetch
);
1076 gsi_insert_before (&gsi
, load
, GSI_SAME_STMT
);
1078 info
->arr_ref_last
= load
;
1081 /* Builds and initializes static arrays initialized with values gathered from
1082 the SWTCH switch statement. Also creates statements that load values from
1086 build_arrays (gimple swtch
, struct switch_conv_info
*info
)
1088 tree arr_index_type
;
1089 tree tidx
, sub
, utype
;
1091 gimple_stmt_iterator gsi
;
1093 location_t loc
= gimple_location (swtch
);
1095 gsi
= gsi_for_stmt (swtch
);
1097 /* Make sure we do not generate arithmetics in a subrange. */
1098 utype
= TREE_TYPE (info
->index_expr
);
1099 if (TREE_TYPE (utype
))
1100 utype
= lang_hooks
.types
.type_for_mode (TYPE_MODE (TREE_TYPE (utype
)), 1);
1102 utype
= lang_hooks
.types
.type_for_mode (TYPE_MODE (utype
), 1);
1104 arr_index_type
= build_index_type (info
->range_size
);
1105 tidx
= make_ssa_name (utype
, NULL
);
1106 sub
= fold_build2_loc (loc
, MINUS_EXPR
, utype
,
1107 fold_convert_loc (loc
, utype
, info
->index_expr
),
1108 fold_convert_loc (loc
, utype
, info
->range_min
));
1109 sub
= force_gimple_operand_gsi (&gsi
, sub
,
1110 false, NULL
, true, GSI_SAME_STMT
);
1111 stmt
= gimple_build_assign (tidx
, sub
);
1113 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1115 info
->arr_ref_first
= stmt
;
1117 for (gsi
= gsi_start_phis (info
->final_bb
), i
= 0;
1118 !gsi_end_p (gsi
); gsi_next (&gsi
), i
++)
1119 build_one_array (swtch
, i
, arr_index_type
, gsi_stmt (gsi
), tidx
, info
);
1122 /* Generates and appropriately inserts loads of default values at the position
1123 given by BSI. Returns the last inserted statement. */
1126 gen_def_assigns (gimple_stmt_iterator
*gsi
, struct switch_conv_info
*info
)
1129 gimple assign
= NULL
;
1131 for (i
= 0; i
< info
->phi_count
; i
++)
1133 tree name
= copy_ssa_name (info
->target_inbound_names
[i
], NULL
);
1134 info
->target_outbound_names
[i
] = name
;
1135 assign
= gimple_build_assign (name
, info
->default_values
[i
]);
1136 gsi_insert_before (gsi
, assign
, GSI_SAME_STMT
);
1137 update_stmt (assign
);
1142 /* Deletes the unused bbs and edges that now contain the switch statement and
1143 its empty branch bbs. BBD is the now dead BB containing the original switch
1144 statement, FINAL is the last BB of the converted switch statement (in terms
1148 prune_bbs (basic_block bbd
, basic_block final
)
1153 for (ei
= ei_start (bbd
->succs
); (e
= ei_safe_edge (ei
)); )
1159 delete_basic_block (bb
);
1161 delete_basic_block (bbd
);
1164 /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge
1165 from the basic block loading values from an array and E2F from the basic
1166 block loading default values. BBF is the last switch basic block (see the
1167 bbf description in the comment below). */
1170 fix_phi_nodes (edge e1f
, edge e2f
, basic_block bbf
,
1171 struct switch_conv_info
*info
)
1173 gimple_stmt_iterator gsi
;
1176 for (gsi
= gsi_start_phis (bbf
), i
= 0;
1177 !gsi_end_p (gsi
); gsi_next (&gsi
), i
++)
1179 gimple phi
= gsi_stmt (gsi
);
1180 add_phi_arg (phi
, info
->target_inbound_names
[i
], e1f
, UNKNOWN_LOCATION
);
1181 add_phi_arg (phi
, info
->target_outbound_names
[i
], e2f
, UNKNOWN_LOCATION
);
1185 /* Creates a check whether the switch expression value actually falls into the
1186 range given by all the cases. If it does not, the temporaries are loaded
1187 with default values instead. SWTCH is the switch statement being converted.
1189 bb0 is the bb with the switch statement, however, we'll end it with a
1192 bb1 is the bb to be used when the range check went ok. It is derived from
1195 bb2 is the bb taken when the expression evaluated outside of the range
1196 covered by the created arrays. It is populated by loads of default
1199 bbF is a fall through for both bb1 and bb2 and contains exactly what
1200 originally followed the switch statement.
1202 bbD contains the switch statement (in the end). It is unreachable but we
1203 still need to strip off its edges.
1207 gen_inbound_check (gimple swtch
, struct switch_conv_info
*info
)
1209 tree label_decl1
= create_artificial_label (UNKNOWN_LOCATION
);
1210 tree label_decl2
= create_artificial_label (UNKNOWN_LOCATION
);
1211 tree label_decl3
= create_artificial_label (UNKNOWN_LOCATION
);
1212 gimple label1
, label2
, label3
;
1219 gimple_stmt_iterator gsi
;
1220 basic_block bb0
, bb1
, bb2
, bbf
, bbd
;
1221 edge e01
, e02
, e21
, e1d
, e1f
, e2f
;
1222 location_t loc
= gimple_location (swtch
);
1224 gcc_assert (info
->default_values
);
1226 bb0
= gimple_bb (swtch
);
1228 tidx
= gimple_assign_lhs (info
->arr_ref_first
);
1229 utype
= TREE_TYPE (tidx
);
1231 /* (end of) block 0 */
1232 gsi
= gsi_for_stmt (info
->arr_ref_first
);
1235 bound
= fold_convert_loc (loc
, utype
, info
->range_size
);
1236 cond_stmt
= gimple_build_cond (LE_EXPR
, tidx
, bound
, NULL_TREE
, NULL_TREE
);
1237 gsi_insert_before (&gsi
, cond_stmt
, GSI_SAME_STMT
);
1238 update_stmt (cond_stmt
);
1241 label2
= gimple_build_label (label_decl2
);
1242 gsi_insert_before (&gsi
, label2
, GSI_SAME_STMT
);
1243 last_assign
= gen_def_assigns (&gsi
, info
);
1246 label1
= gimple_build_label (label_decl1
);
1247 gsi_insert_before (&gsi
, label1
, GSI_SAME_STMT
);
1250 gsi
= gsi_start_bb (info
->final_bb
);
1251 label3
= gimple_build_label (label_decl3
);
1252 gsi_insert_before (&gsi
, label3
, GSI_SAME_STMT
);
1255 e02
= split_block (bb0
, cond_stmt
);
1258 e21
= split_block (bb2
, last_assign
);
1262 e1d
= split_block (bb1
, info
->arr_ref_last
);
1266 /* flags and profiles of the edge for in-range values */
1267 e01
= make_edge (bb0
, bb1
, EDGE_TRUE_VALUE
);
1268 e01
->probability
= REG_BR_PROB_BASE
- info
->default_prob
;
1269 e01
->count
= info
->other_count
;
1271 /* flags and profiles of the edge taking care of out-of-range values */
1272 e02
->flags
&= ~EDGE_FALLTHRU
;
1273 e02
->flags
|= EDGE_FALSE_VALUE
;
1274 e02
->probability
= info
->default_prob
;
1275 e02
->count
= info
->default_count
;
1277 bbf
= info
->final_bb
;
1279 e1f
= make_edge (bb1
, bbf
, EDGE_FALLTHRU
);
1280 e1f
->probability
= REG_BR_PROB_BASE
;
1281 e1f
->count
= info
->other_count
;
1283 e2f
= make_edge (bb2
, bbf
, EDGE_FALLTHRU
);
1284 e2f
->probability
= REG_BR_PROB_BASE
;
1285 e2f
->count
= info
->default_count
;
1287 /* frequencies of the new BBs */
1288 bb1
->frequency
= EDGE_FREQUENCY (e01
);
1289 bb2
->frequency
= EDGE_FREQUENCY (e02
);
1290 bbf
->frequency
= EDGE_FREQUENCY (e1f
) + EDGE_FREQUENCY (e2f
);
1292 /* Tidy blocks that have become unreachable. */
1293 prune_bbs (bbd
, info
->final_bb
);
1295 /* Fixup the PHI nodes in bbF. */
1296 fix_phi_nodes (e1f
, e2f
, bbf
, info
);
1298 /* Fix the dominator tree, if it is available. */
1299 if (dom_info_available_p (CDI_DOMINATORS
))
1301 vec
<basic_block
> bbs_to_fix_dom
;
1303 set_immediate_dominator (CDI_DOMINATORS
, bb1
, bb0
);
1304 set_immediate_dominator (CDI_DOMINATORS
, bb2
, bb0
);
1305 if (! get_immediate_dominator (CDI_DOMINATORS
, bbf
))
1306 /* If bbD was the immediate dominator ... */
1307 set_immediate_dominator (CDI_DOMINATORS
, bbf
, bb0
);
1309 bbs_to_fix_dom
.create (4);
1310 bbs_to_fix_dom
.quick_push (bb0
);
1311 bbs_to_fix_dom
.quick_push (bb1
);
1312 bbs_to_fix_dom
.quick_push (bb2
);
1313 bbs_to_fix_dom
.quick_push (bbf
);
1315 iterate_fix_dominators (CDI_DOMINATORS
, bbs_to_fix_dom
, true);
1316 bbs_to_fix_dom
.release ();
1320 /* The following function is invoked on every switch statement (the current one
1321 is given in SWTCH) and runs the individual phases of switch conversion on it
1322 one after another until one fails or the conversion is completed.
1323 Returns NULL on success, or a pointer to a string with the reason why the
1324 conversion failed. */
1327 process_switch (gimple swtch
)
1329 struct switch_conv_info info
;
1331 /* Group case labels so that we get the right results from the heuristics
1332 that decide on the code generation approach for this switch. */
1333 group_case_labels_stmt (swtch
);
1335 /* If this switch is now a degenerate case with only a default label,
1336 there is nothing left for us to do. */
1337 if (gimple_switch_num_labels (swtch
) < 2)
1338 return "switch is a degenerate case";
1340 collect_switch_conv_info (swtch
, &info
);
1342 /* No error markers should reach here (they should be filtered out
1343 during gimplification). */
1344 gcc_checking_assert (TREE_TYPE (info
.index_expr
) != error_mark_node
);
1346 /* A switch on a constant should have been optimized in tree-cfg-cleanup. */
1347 gcc_checking_assert (! TREE_CONSTANT (info
.index_expr
));
1349 if (info
.uniq
<= MAX_CASE_BIT_TESTS
)
1351 if (expand_switch_using_bit_tests_p (info
.range_size
,
1352 info
.uniq
, info
.count
))
1355 fputs (" expanding as bit test is preferable\n", dump_file
);
1356 emit_case_bit_tests (swtch
, info
.index_expr
,
1357 info
.range_min
, info
.range_size
);
1359 loops_state_set (LOOPS_NEED_FIXUP
);
1364 /* This will be expanded as a decision tree in stmt.c:expand_case. */
1365 return " expanding as jumps is preferable";
1368 /* If there is no common successor, we cannot do the transformation. */
1369 if (! info
.final_bb
)
1370 return "no common successor to all case label target blocks found";
1372 /* Check the case label values are within reasonable range: */
1373 if (!check_range (&info
))
1375 gcc_assert (info
.reason
);
1379 /* For all the cases, see whether they are empty, the assignments they
1380 represent constant and so on... */
1381 if (! check_all_empty_except_final (&info
))
1383 gcc_assert (info
.reason
);
1386 if (!check_final_bb (&info
))
1388 gcc_assert (info
.reason
);
1392 /* At this point all checks have passed and we can proceed with the
1395 create_temp_arrays (&info
);
1396 gather_default_values (gimple_switch_default_label (swtch
), &info
);
1397 build_constructors (swtch
, &info
);
1399 build_arrays (swtch
, &info
); /* Build the static arrays and assignments. */
1400 gen_inbound_check (swtch
, &info
); /* Build the bounds check. */
1403 free_temp_arrays (&info
);
1407 /* The main function of the pass scans statements for switches and invokes
1408 process_switch on them. */
1411 do_switchconv (void)
1417 const char *failure_reason
;
1418 gimple stmt
= last_stmt (bb
);
1419 if (stmt
&& gimple_code (stmt
) == GIMPLE_SWITCH
)
1423 expanded_location loc
= expand_location (gimple_location (stmt
));
1425 fprintf (dump_file
, "beginning to process the following "
1426 "SWITCH statement (%s:%d) : ------- \n",
1427 loc
.file
, loc
.line
);
1428 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
1429 putc ('\n', dump_file
);
1432 failure_reason
= process_switch (stmt
);
1433 if (! failure_reason
)
1437 fputs ("Switch converted\n", dump_file
);
1438 fputs ("--------------------------------\n", dump_file
);
1441 /* Make no effort to update the post-dominator tree. It is actually not
1442 that hard for the transformations we have performed, but it is not
1443 supported by iterate_fix_dominators. */
1444 free_dominance_info (CDI_POST_DOMINATORS
);
1450 fputs ("Bailing out - ", dump_file
);
1451 fputs (failure_reason
, dump_file
);
1452 fputs ("\n--------------------------------\n", dump_file
);
1461 /* The pass gate. */
1464 switchconv_gate (void)
1466 return flag_tree_switch_conversion
!= 0;
1471 const pass_data pass_data_convert_switch
=
1473 GIMPLE_PASS
, /* type */
1474 "switchconv", /* name */
1475 OPTGROUP_NONE
, /* optinfo_flags */
1476 true, /* has_gate */
1477 true, /* has_execute */
1478 TV_TREE_SWITCH_CONVERSION
, /* tv_id */
1479 ( PROP_cfg
| PROP_ssa
), /* properties_required */
1480 0, /* properties_provided */
1481 0, /* properties_destroyed */
1482 0, /* todo_flags_start */
1483 ( TODO_update_ssa
| TODO_verify_ssa
1485 | TODO_verify_flow
), /* todo_flags_finish */
1488 class pass_convert_switch
: public gimple_opt_pass
1491 pass_convert_switch (gcc::context
*ctxt
)
1492 : gimple_opt_pass (pass_data_convert_switch
, ctxt
)
1495 /* opt_pass methods: */
1496 bool gate () { return switchconv_gate (); }
1497 unsigned int execute () { return do_switchconv (); }
1499 }; // class pass_convert_switch
1504 make_pass_convert_switch (gcc::context
*ctxt
)
1506 return new pass_convert_switch (ctxt
);