1 /* Lower GIMPLE_SWITCH expressions to something more efficient than
3 Copyright (C) 2006, 2008, 2009, 2010, 2011, 2012
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it
9 under the terms of the GNU General Public License as published by the
10 Free Software Foundation; either version 3, or (at your option) any
13 GCC is distributed in the hope that it will be useful, but WITHOUT
14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not, write to the Free
20 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
23 /* This file handles the lowering of GIMPLE_SWITCH to an indexed
24 load, or a series of bit-test-and-branch expressions. */
28 #include "coretypes.h"
34 #include "basic-block.h"
35 #include "tree-flow.h"
36 #include "tree-flow-inline.h"
37 #include "tree-ssa-operands.h"
38 #include "tree-pass.h"
39 #include "gimple-pretty-print.h"
41 /* ??? For lang_hooks.types.type_for_mode, but is there a word_mode
42 type in the GIMPLE type system that is language-independent? */
43 #include "langhooks.h"
45 /* Need to include expr.h and optabs.h for lshift_cheap_p. */
49 /* Maximum number of case bit tests.
50 FIXME: This should be derived from PARAM_CASE_VALUES_THRESHOLD and
51 targetm.case_values_threshold(), or be its own param. */
52 #define MAX_CASE_BIT_TESTS 3
54 /* Split the basic block at the statement pointed to by GSIP, and insert
55 a branch to the target basic block of E_TRUE conditional on tree
58 It is assumed that there is already an edge from the to-be-split
59 basic block to E_TRUE->dest block. This edge is removed, and the
60 profile information on the edge is re-used for the new conditional
63 The CFG is updated. The dominator tree will not be valid after
64 this transformation, but the immediate dominators are updated if
65 UPDATE_DOMINATORS is true.
67 Returns the newly created basic block. */
70 hoist_edge_and_branch_if_true (gimple_stmt_iterator *gsip,
71 tree cond, edge e_true,
72 bool update_dominators)
77 basic_block new_bb, split_bb = gsi_bb (*gsip);
78 bool dominated_e_true = false;
80 gcc_assert (e_true->src == split_bb);
83 && get_immediate_dominator (CDI_DOMINATORS, e_true->dest) == split_bb)
84 dominated_e_true = true;
86 tmp = force_gimple_operand_gsi (gsip, cond, /*simple=*/true, NULL,
87 /*before=*/true, GSI_SAME_STMT);
88 cond_stmt = gimple_build_cond_from_tree (tmp, NULL_TREE, NULL_TREE);
89 gsi_insert_before (gsip, cond_stmt, GSI_SAME_STMT);
91 e_false = split_block (split_bb, cond_stmt);
92 new_bb = e_false->dest;
93 redirect_edge_pred (e_true, split_bb);
95 e_true->flags &= ~EDGE_FALLTHRU;
96 e_true->flags |= EDGE_TRUE_VALUE;
98 e_false->flags &= ~EDGE_FALLTHRU;
99 e_false->flags |= EDGE_FALSE_VALUE;
100 e_false->probability = REG_BR_PROB_BASE - e_true->probability;
101 e_false->count = split_bb->count - e_true->count;
102 new_bb->count = e_false->count;
104 if (update_dominators)
106 if (dominated_e_true)
107 set_immediate_dominator (CDI_DOMINATORS, e_true->dest, split_bb);
108 set_immediate_dominator (CDI_DOMINATORS, e_false->dest, split_bb);
115 /* Determine whether "1 << x" is relatively cheap in word_mode. */
116 /* FIXME: This is the function that we need rtl.h and optabs.h for.
117 This function (and similar RTL-related cost code in e.g. IVOPTS) should
118 be moved to some kind of interface file for GIMPLE/RTL interactions. */
120 lshift_cheap_p (void)
122 /* FIXME: This should be made target dependent via this "this_target"
123 mechanism, similar to e.g. can_copy_init_p in gcse.c. */
124 static bool init[2] = {false, false};
125 static bool cheap[2] = {true, true};
128 /* If the targer has no lshift in word_mode, the operation will most
129 probably not be cheap. ??? Does GCC even work for such targets? */
130 if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing)
133 speed_p = optimize_insn_for_speed_p ();
137 rtx reg = gen_raw_REG (word_mode, 10000);
138 int cost = set_src_cost (gen_rtx_ASHIFT (word_mode, const1_rtx, reg),
140 cheap[speed_p] = cost < COSTS_N_INSNS (MAX_CASE_BIT_TESTS);
141 init[speed_p] = true;
144 return cheap[speed_p];
147 /* Return true if a switch should be expanded as a bit test.
148 RANGE is the difference between highest and lowest case.
149 UNIQ is number of unique case node targets, not counting the default case.
150 COUNT is the number of comparisons needed, not counting the default case. */
153 expand_switch_using_bit_tests_p (tree range,
157 return (((uniq == 1 && count >= 3)
158 || (uniq == 2 && count >= 5)
159 || (uniq == 3 && count >= 6))
161 && compare_tree_int (range, GET_MODE_BITSIZE (word_mode)) < 0
162 && compare_tree_int (range, 0) > 0);
165 /* Implement switch statements with bit tests
167 A GIMPLE switch statement can be expanded to a short sequence of bit-wise
168 comparisons. "switch(x)" is converted into "if ((1 << (x-MINVAL)) & CST)"
169 where CST and MINVAL are integer constants. This is better than a series
170 of compare-and-banch insns in some cases, e.g. we can implement:
172 if ((x==4) || (x==6) || (x==9) || (x==11))
174 as a single bit test:
176 if ((1<<x) & ((1<<4)|(1<<6)|(1<<9)|(1<<11)))
178 This transformation is only applied if the number of case targets is small,
179 if CST constains at least 3 bits, and "1 << x" is cheap. The bit tests are
180 performed in "word_mode".
182 The following example shows the code the transformation generates:
188 case '0': case '1': case '2': case '3': case '4':
189 case '5': case '6': case '7': case '8': case '9':
190 case 'A': case 'B': case 'C': case 'D': case 'E':
202 if (tmp1 > (70 - 48)) goto L2;
204 tmp3 = 0b11111100000001111111111;
205 if ((tmp2 & tmp3) != 0) goto L1 ; else goto L2;
212 TODO: There are still some improvements to this transformation that could
215 * A narrower mode than word_mode could be used if that is cheaper, e.g.
216 for x86_64 where a narrower-mode shift may result in smaller code.
218 * The compounded constant could be shifted rather than the one. The
219 test would be either on the sign bit or on the least significant bit,
220 depending on the direction of the shift. On some machines, the test
221 for the branch would be free if the bit to test is already set by the
224 This transformation was contributed by Roger Sayle, see this e-mail:
225 http://gcc.gnu.org/ml/gcc-patches/2003-01/msg01950.html
228 /* A case_bit_test represents a set of case nodes that may be
229 selected from using a bit-wise comparison. HI and LO hold
230 the integer to be tested against, TARGET_EDGE contains the
231 edge to the basic block to jump to upon success and BITS
232 counts the number of case nodes handled by this test,
233 typically the number of bits set in HI:LO. The LABEL field
234 is used to quickly identify all cases in this set without
235 looking at label_to_block for every case label. */
246 /* Comparison function for qsort to order bit tests by decreasing
247 probability of execution. Our best guess comes from a measured
248 profile. If the profile counts are equal, break even on the
249 number of case nodes, i.e. the node with the most cases gets
252 TODO: Actually this currently runs before a profile is available.
253 Therefore the case-as-bit-tests transformation should be done
254 later in the pass pipeline, or something along the lines of
255 "Efficient and effective branch reordering using profile data"
256 (Yang et. al., 2002) should be implemented (although, how good
257 is a paper is called "Efficient and effective ..." when the
258 latter is implied by the former, but oh well...). */
261 case_bit_test_cmp (const void *p1, const void *p2)
263 const struct case_bit_test *const d1 = (const struct case_bit_test *) p1;
264 const struct case_bit_test *const d2 = (const struct case_bit_test *) p2;
266 if (d2->target_edge->count != d1->target_edge->count)
267 return d2->target_edge->count - d1->target_edge->count;
268 if (d2->bits != d1->bits)
269 return d2->bits - d1->bits;
271 /* Stabilize the sort. */
272 return LABEL_DECL_UID (d2->label) - LABEL_DECL_UID (d1->label);
275 /* Expand a switch statement by a short sequence of bit-wise
276 comparisons. "switch(x)" is effectively converted into
277 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are
280 INDEX_EXPR is the value being switched on.
282 MINVAL is the lowest case value of in the case nodes,
283 and RANGE is highest value minus MINVAL. MINVAL and RANGE
284 are not guaranteed to be of the same type as INDEX_EXPR
285 (the gimplifier doesn't change the type of case label values,
286 and MINVAL and RANGE are derived from those values).
288 There *MUST* be MAX_CASE_BIT_TESTS or less unique case
292 emit_case_bit_tests (gimple swtch, tree index_expr,
293 tree minval, tree range)
295 struct case_bit_test test[MAX_CASE_BIT_TESTS];
296 unsigned int i, j, k;
299 basic_block switch_bb = gimple_bb (swtch);
300 basic_block default_bb, new_default_bb, new_bb;
302 bool update_dom = dom_info_available_p (CDI_DOMINATORS);
304 VEC (basic_block, heap) *bbs_to_fix_dom = NULL;
306 tree index_type = TREE_TYPE (index_expr);
307 tree unsigned_index_type = unsigned_type_for (index_type);
308 unsigned int branch_num = gimple_switch_num_labels (swtch);
310 gimple_stmt_iterator gsi;
314 tree word_type_node = lang_hooks.types.type_for_mode (word_mode, 1);
315 tree word_mode_zero = fold_convert (word_type_node, integer_zero_node);
316 tree word_mode_one = fold_convert (word_type_node, integer_one_node);
318 memset (&test, 0, sizeof (test));
320 /* Get the edge for the default case. */
321 tmp = gimple_switch_label (swtch, 0);
322 default_bb = label_to_block (CASE_LABEL (tmp));
323 default_edge = find_edge (switch_bb, default_bb);
325 /* Go through all case labels, and collect the case labels, profile
326 counts, and other information we need to build the branch tests. */
328 for (i = 1; i < branch_num; i++)
331 tree cs = gimple_switch_label (swtch, i);
332 tree label = CASE_LABEL (cs);
333 edge e = find_edge (switch_bb, label_to_block (label));
334 for (k = 0; k < count; k++)
335 if (e == test[k].target_edge)
340 gcc_checking_assert (count < MAX_CASE_BIT_TESTS);
343 test[k].target_edge = e;
344 test[k].label = label;
351 lo = tree_low_cst (int_const_binop (MINUS_EXPR,
352 CASE_LOW (cs), minval),
354 if (CASE_HIGH (cs) == NULL_TREE)
357 hi = tree_low_cst (int_const_binop (MINUS_EXPR,
358 CASE_HIGH (cs), minval),
361 for (j = lo; j <= hi; j++)
362 if (j >= HOST_BITS_PER_WIDE_INT)
363 test[k].hi |= (HOST_WIDE_INT) 1 << (j - HOST_BITS_PER_INT);
365 test[k].lo |= (HOST_WIDE_INT) 1 << j;
368 qsort (test, count, sizeof(*test), case_bit_test_cmp);
370 /* We generate two jumps to the default case label.
371 Split the default edge, so that we don't have to do any PHI node
373 new_default_bb = split_edge (default_edge);
377 bbs_to_fix_dom = VEC_alloc (basic_block, heap, 10);
378 VEC_quick_push (basic_block, bbs_to_fix_dom, switch_bb);
379 VEC_quick_push (basic_block, bbs_to_fix_dom, default_bb);
380 VEC_quick_push (basic_block, bbs_to_fix_dom, new_default_bb);
383 /* Now build the test-and-branch code. */
385 gsi = gsi_last_bb (switch_bb);
387 /* idx = (unsigned) (x - minval) */
388 idx = fold_build2 (MINUS_EXPR, index_type, index_expr,
389 fold_convert (index_type, minval));
390 idx = fold_convert (unsigned_index_type, idx);
391 idx = force_gimple_operand_gsi (&gsi, idx,
392 /*simple=*/true, NULL_TREE,
393 /*before=*/true, GSI_SAME_STMT);
395 /* if (idx > range) goto default */
396 range = force_gimple_operand_gsi (&gsi,
397 fold_convert (unsigned_index_type, range),
398 /*simple=*/true, NULL_TREE,
399 /*before=*/true, GSI_SAME_STMT);
400 tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range);
401 new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, default_edge, update_dom);
403 VEC_quick_push (basic_block, bbs_to_fix_dom, new_bb);
404 gcc_assert (gimple_bb (swtch) == new_bb);
405 gsi = gsi_last_bb (new_bb);
407 /* Any blocks dominated by the GIMPLE_SWITCH, but that are not successors
408 of NEW_BB, are still immediately dominated by SWITCH_BB. Make it so. */
411 VEC (basic_block, heap) *dom_bbs;
414 dom_bbs = get_dominated_by (CDI_DOMINATORS, new_bb);
415 FOR_EACH_VEC_ELT (basic_block, dom_bbs, i, dom_son)
417 edge e = find_edge (new_bb, dom_son);
418 if (e && single_pred_p (e->dest))
420 set_immediate_dominator (CDI_DOMINATORS, dom_son, switch_bb);
421 VEC_safe_push (basic_block, heap, bbs_to_fix_dom, dom_son);
423 VEC_free (basic_block, heap, dom_bbs);
426 /* csui = (1 << (word_mode) idx) */
427 tmp = create_tmp_var (word_type_node, "csui");
428 csui = make_ssa_name (tmp, NULL);
429 tmp = fold_build2 (LSHIFT_EXPR, word_type_node, word_mode_one,
430 fold_convert (word_type_node, idx));
431 tmp = force_gimple_operand_gsi (&gsi, tmp,
432 /*simple=*/false, NULL_TREE,
433 /*before=*/true, GSI_SAME_STMT);
434 shift_stmt = gimple_build_assign (csui, tmp);
435 SSA_NAME_DEF_STMT (csui) = shift_stmt;
436 gsi_insert_before (&gsi, shift_stmt, GSI_SAME_STMT);
437 update_stmt (shift_stmt);
439 /* for each unique set of cases:
440 if (const & csui) goto target */
441 for (k = 0; k < count; k++)
443 tmp = build_int_cst_wide (word_type_node, test[k].lo, test[k].hi);
444 tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp);
445 tmp = force_gimple_operand_gsi (&gsi, tmp,
446 /*simple=*/true, NULL_TREE,
447 /*before=*/true, GSI_SAME_STMT);
448 tmp = fold_build2 (NE_EXPR, boolean_type_node, tmp, word_mode_zero);
449 new_bb = hoist_edge_and_branch_if_true (&gsi, tmp, test[k].target_edge,
452 VEC_safe_push (basic_block, heap, bbs_to_fix_dom, new_bb);
453 gcc_assert (gimple_bb (swtch) == new_bb);
454 gsi = gsi_last_bb (new_bb);
457 /* We should have removed all edges now. */
458 gcc_assert (EDGE_COUNT (gsi_bb (gsi)->succs) == 0);
460 /* If nothing matched, go to the default label. */
461 make_edge (gsi_bb (gsi), new_default_bb, EDGE_FALLTHRU);
463 /* The GIMPLE_SWITCH is now redundant. */
464 gsi_remove (&gsi, true);
468 /* Fix up the dominator tree. */
469 iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true);
470 VEC_free (basic_block, heap, bbs_to_fix_dom);
475 Switch initialization conversion
477 The following pass changes simple initializations of scalars in a switch
478 statement into initializations from a static array. Obviously, the values
479 must be constant and known at compile time and a default branch must be
480 provided. For example, the following code:
504 a_5 = PHI <a_1, a_2, a_3, a_4>
505 b_5 = PHI <b_1, b_2, b_3, b_4>
510 static const int = CSWTCH01[] = {6, 6, 5, 1, 1, 1, 1, 1, 1, 1, 1, 4};
511 static const int = CSWTCH02[] = {8, 8, 9, 16, 16, 16, 16, 16, 16, 16,
514 if (((unsigned) argc) - 1 < 11)
516 a_6 = CSWTCH02[argc - 1];
517 b_6 = CSWTCH01[argc - 1];
527 There are further constraints. Specifically, the range of values across all
528 case labels must not be bigger than SWITCH_CONVERSION_BRANCH_RATIO (default
529 eight) times the number of the actual switch branches.
531 This transformation was contributed by Martin Jambor, see this e-mail:
532 http://gcc.gnu.org/ml/gcc-patches/2008-07/msg00011.html */
534 /* The main structure of the pass. */
535 struct switch_conv_info
537 /* The expression used to decide the switch branch. */
540 /* The following integer constants store the minimum and maximum value
541 covered by the case labels. */
545 /* The difference between the above two numbers. Stored here because it
546 is used in all the conversion heuristics, as well as for some of the
547 transformation, and it is expensive to re-compute it all the time. */
550 /* Basic block that contains the actual GIMPLE_SWITCH. */
551 basic_block switch_bb;
553 /* Basic block that is the target of the default case. */
554 basic_block default_bb;
556 /* The single successor block of all branches out of the GIMPLE_SWITCH,
557 if such a block exists. Otherwise NULL. */
558 basic_block final_bb;
560 /* The probability of the default edge in the replaced switch. */
563 /* The count of the default edge in the replaced switch. */
564 gcov_type default_count;
566 /* Combined count of all other (non-default) edges in the replaced switch. */
567 gcov_type other_count;
569 /* Number of phi nodes in the final bb (that we'll be replacing). */
572 /* Array of default values, in the same order as phi nodes. */
573 tree *default_values;
575 /* Constructors of new static arrays. */
576 VEC (constructor_elt, gc) **constructors;
578 /* Array of ssa names that are initialized with a value from a new static
580 tree *target_inbound_names;
582 /* Array of ssa names that are initialized with the default value if the
583 switch expression is out of range. */
584 tree *target_outbound_names;
586 /* The first load statement that loads a temporary from a new static array.
588 gimple arr_ref_first;
590 /* The last load statement that loads a temporary from a new static array. */
593 /* String reason why the case wasn't a good candidate that is written to the
594 dump file, if there is one. */
597 /* Parameters for expand_switch_using_bit_tests. Should be computed
598 the same way as in expand_case. */
603 /* Collect information about GIMPLE_SWITCH statement SWTCH into INFO. */
606 collect_switch_conv_info (gimple swtch, struct switch_conv_info *info)
608 unsigned int branch_num = gimple_switch_num_labels (swtch);
609 tree min_case, max_case;
610 unsigned int count, i;
614 memset (info, 0, sizeof (*info));
616 /* The gimplifier has already sorted the cases by CASE_LOW and ensured there
617 is a default label which is the first in the vector. */
618 gcc_assert (CASE_LOW (gimple_switch_label (swtch, 0)) == NULL_TREE);
620 /* Collect the bits we can deduce from the CFG. */
621 info->index_expr = gimple_switch_index (swtch);
622 info->switch_bb = gimple_bb (swtch);
624 label_to_block (CASE_LABEL (gimple_switch_label (swtch, 0)));
625 e_default = find_edge (info->switch_bb, info->default_bb);
626 info->default_prob = e_default->probability;
627 info->default_count = e_default->count;
628 FOR_EACH_EDGE (e, ei, info->switch_bb->succs)
630 info->other_count += e->count;
632 /* See if there is one common successor block for all branch
633 targets. If it exists, record it in FINAL_BB. */
634 FOR_EACH_EDGE (e, ei, info->switch_bb->succs)
636 if (! single_pred_p (e->dest))
638 info->final_bb = e->dest;
643 FOR_EACH_EDGE (e, ei, info->switch_bb->succs)
645 if (e->dest == info->final_bb)
648 if (single_pred_p (e->dest)
649 && single_succ_p (e->dest)
650 && single_succ (e->dest) == info->final_bb)
653 info->final_bb = NULL;
657 /* Get upper and lower bounds of case values, and the covered range. */
658 min_case = gimple_switch_label (swtch, 1);
659 max_case = gimple_switch_label (swtch, branch_num - 1);
661 info->range_min = CASE_LOW (min_case);
662 if (CASE_HIGH (max_case) != NULL_TREE)
663 info->range_max = CASE_HIGH (max_case);
665 info->range_max = CASE_LOW (max_case);
668 int_const_binop (MINUS_EXPR, info->range_max, info->range_min);
670 /* Get a count of the number of case labels. Single-valued case labels
671 simply count as one, but a case range counts double, since it may
672 require two compares if it gets lowered as a branching tree. */
674 for (i = 1; i < branch_num; i++)
676 tree elt = gimple_switch_label (swtch, i);
679 && ! tree_int_cst_equal (CASE_LOW (elt), CASE_HIGH (elt)))
684 /* Get the number of unique non-default targets out of the GIMPLE_SWITCH
685 block. Assume a CFG cleanup would have already removed degenerate
686 switch statements, this allows us to just use EDGE_COUNT. */
687 info->uniq = EDGE_COUNT (gimple_bb (swtch)->succs) - 1;
690 /* Checks whether the range given by individual case statements of the SWTCH
691 switch statement isn't too big and whether the number of branches actually
692 satisfies the size of the new array. */
695 check_range (struct switch_conv_info *info)
697 gcc_assert (info->range_size);
698 if (!host_integerp (info->range_size, 1))
700 info->reason = "index range way too large or otherwise unusable";
704 if ((unsigned HOST_WIDE_INT) tree_low_cst (info->range_size, 1)
705 > ((unsigned) info->count * SWITCH_CONVERSION_BRANCH_RATIO))
707 info->reason = "the maximum range-branch ratio exceeded";
714 /* Checks whether all but the FINAL_BB basic blocks are empty. */
717 check_all_empty_except_final (struct switch_conv_info *info)
722 FOR_EACH_EDGE (e, ei, info->switch_bb->succs)
724 if (e->dest == info->final_bb)
727 if (!empty_block_p (e->dest))
729 info->reason = "bad case - a non-final BB not empty";
737 /* This function checks whether all required values in phi nodes in final_bb
738 are constants. Required values are those that correspond to a basic block
739 which is a part of the examined switch statement. It returns true if the
740 phi nodes are OK, otherwise false. */
743 check_final_bb (struct switch_conv_info *info)
745 gimple_stmt_iterator gsi;
748 for (gsi = gsi_start_phis (info->final_bb); !gsi_end_p (gsi); gsi_next (&gsi))
750 gimple phi = gsi_stmt (gsi);
755 for (i = 0; i < gimple_phi_num_args (phi); i++)
757 basic_block bb = gimple_phi_arg_edge (phi, i)->src;
759 if (bb == info->switch_bb
760 || (single_pred_p (bb) && single_pred (bb) == info->switch_bb))
764 val = gimple_phi_arg_def (phi, i);
765 if (!is_gimple_ip_invariant (val))
767 info->reason = "non-invariant value from a case";
768 return false; /* Non-invariant argument. */
770 reloc = initializer_constant_valid_p (val, TREE_TYPE (val));
771 if ((flag_pic && reloc != null_pointer_node)
772 || (!flag_pic && reloc == NULL_TREE))
776 = "value from a case would need runtime relocations";
779 = "value from a case is not a valid initializer";
789 /* The following function allocates default_values, target_{in,out}_names and
790 constructors arrays. The last one is also populated with pointers to
791 vectors that will become constructors of new arrays. */
794 create_temp_arrays (struct switch_conv_info *info)
798 info->default_values = XCNEWVEC (tree, info->phi_count * 3);
799 info->constructors = XCNEWVEC (VEC (constructor_elt, gc) *, info->phi_count);
800 info->target_inbound_names = info->default_values + info->phi_count;
801 info->target_outbound_names = info->target_inbound_names + info->phi_count;
802 for (i = 0; i < info->phi_count; i++)
803 info->constructors[i]
804 = VEC_alloc (constructor_elt, gc, tree_low_cst (info->range_size, 1) + 1);
807 /* Free the arrays created by create_temp_arrays(). The vectors that are
808 created by that function are not freed here, however, because they have
809 already become constructors and must be preserved. */
812 free_temp_arrays (struct switch_conv_info *info)
814 XDELETEVEC (info->constructors);
815 XDELETEVEC (info->default_values);
818 /* Populate the array of default values in the order of phi nodes.
819 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch. */
822 gather_default_values (tree default_case, struct switch_conv_info *info)
824 gimple_stmt_iterator gsi;
825 basic_block bb = label_to_block (CASE_LABEL (default_case));
829 gcc_assert (CASE_LOW (default_case) == NULL_TREE);
831 if (bb == info->final_bb)
832 e = find_edge (info->switch_bb, bb);
834 e = single_succ_edge (bb);
836 for (gsi = gsi_start_phis (info->final_bb); !gsi_end_p (gsi); gsi_next (&gsi))
838 gimple phi = gsi_stmt (gsi);
839 tree val = PHI_ARG_DEF_FROM_EDGE (phi, e);
841 info->default_values[i++] = val;
845 /* The following function populates the vectors in the constructors array with
846 future contents of the static arrays. The vectors are populated in the
847 order of phi nodes. SWTCH is the switch statement being converted. */
850 build_constructors (gimple swtch, struct switch_conv_info *info)
852 unsigned i, branch_num = gimple_switch_num_labels (swtch);
853 tree pos = info->range_min;
855 for (i = 1; i < branch_num; i++)
857 tree cs = gimple_switch_label (swtch, i);
858 basic_block bb = label_to_block (CASE_LABEL (cs));
861 gimple_stmt_iterator gsi;
864 if (bb == info->final_bb)
865 e = find_edge (info->switch_bb, bb);
867 e = single_succ_edge (bb);
870 while (tree_int_cst_lt (pos, CASE_LOW (cs)))
873 for (k = 0; k < info->phi_count; k++)
875 constructor_elt *elt;
877 elt = VEC_quick_push (constructor_elt,
878 info->constructors[k], NULL);
879 elt->index = int_const_binop (MINUS_EXPR, pos,
881 elt->value = info->default_values[k];
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);
903 constructor_elt *elt;
905 elt = VEC_quick_push (constructor_elt,
906 info->constructors[j], NULL);
907 elt->index = int_const_binop (MINUS_EXPR, pos, info->range_min);
910 pos = int_const_binop (PLUS_EXPR, pos, integer_one_node);
911 } while (!tree_int_cst_lt (high, pos)
912 && tree_int_cst_lt (low, pos));
918 /* If all values in the constructor vector are the same, return the value.
919 Otherwise return NULL_TREE. Not supposed to be called for empty
923 constructor_contains_same_values_p (VEC (constructor_elt, gc) *vec)
926 tree prev = NULL_TREE;
927 constructor_elt *elt;
929 FOR_EACH_VEC_ELT (constructor_elt, vec, i, elt)
933 else if (!operand_equal_p (elt->value, prev, OEP_ONLY_CONST))
939 /* Return type which should be used for array elements, either TYPE,
940 or for integral type some smaller integral type that can still hold
941 all the constants. */
944 array_value_type (gimple swtch, tree type, int num,
945 struct switch_conv_info *info)
947 unsigned int i, len = VEC_length (constructor_elt, info->constructors[num]);
948 constructor_elt *elt;
949 enum machine_mode mode;
953 if (!INTEGRAL_TYPE_P (type))
956 mode = GET_CLASS_NARROWEST_MODE (GET_MODE_CLASS (TYPE_MODE (type)));
957 if (GET_MODE_SIZE (TYPE_MODE (type)) <= GET_MODE_SIZE (mode))
960 if (len < (optimize_bb_for_size_p (gimple_bb (swtch)) ? 2 : 32))
963 FOR_EACH_VEC_ELT (constructor_elt, info->constructors[num], i, elt)
967 if (TREE_CODE (elt->value) != INTEGER_CST)
970 cst = TREE_INT_CST (elt->value);
973 unsigned int prec = GET_MODE_BITSIZE (mode);
974 if (prec > HOST_BITS_PER_WIDE_INT)
978 && double_int_equal_p (cst, double_int_zext (cst, prec)))
981 && double_int_equal_p (cst, double_int_sext (cst, prec)))
987 && double_int_equal_p (cst, double_int_sext (cst, prec)))
996 mode = GET_MODE_WIDER_MODE (mode);
998 || GET_MODE_SIZE (mode) >= GET_MODE_SIZE (TYPE_MODE (type)))
1004 sign = TYPE_UNSIGNED (type) ? 1 : -1;
1005 smaller_type = lang_hooks.types.type_for_mode (mode, sign >= 0);
1006 if (GET_MODE_SIZE (TYPE_MODE (type))
1007 <= GET_MODE_SIZE (TYPE_MODE (smaller_type)))
1010 return smaller_type;
1013 /* Create an appropriate array type and declaration and assemble a static array
1014 variable. Also create a load statement that initializes the variable in
1015 question with a value from the static array. SWTCH is the switch statement
1016 being converted, NUM is the index to arrays of constructors, default values
1017 and target SSA names for this particular array. ARR_INDEX_TYPE is the type
1018 of the index of the new array, PHI is the phi node of the final BB that
1019 corresponds to the value that will be loaded from the created array. TIDX
1020 is an ssa name of a temporary variable holding the index for loads from the
1024 build_one_array (gimple swtch, int num, tree arr_index_type, gimple phi,
1025 tree tidx, struct switch_conv_info *info)
1029 gimple_stmt_iterator gsi = gsi_for_stmt (swtch);
1030 location_t loc = gimple_location (swtch);
1032 gcc_assert (info->default_values[num]);
1034 name = make_ssa_name (SSA_NAME_VAR (PHI_RESULT (phi)), NULL);
1035 info->target_inbound_names[num] = name;
1037 cst = constructor_contains_same_values_p (info->constructors[num]);
1039 load = gimple_build_assign (name, cst);
1042 tree array_type, ctor, decl, value_type, fetch, default_type;
1044 default_type = TREE_TYPE (info->default_values[num]);
1045 value_type = array_value_type (swtch, default_type, num, info);
1046 array_type = build_array_type (value_type, arr_index_type);
1047 if (default_type != value_type)
1050 constructor_elt *elt;
1052 FOR_EACH_VEC_ELT (constructor_elt, info->constructors[num], i, elt)
1053 elt->value = fold_convert (value_type, elt->value);
1055 ctor = build_constructor (array_type, info->constructors[num]);
1056 TREE_CONSTANT (ctor) = true;
1057 TREE_STATIC (ctor) = true;
1059 decl = build_decl (loc, VAR_DECL, NULL_TREE, array_type);
1060 TREE_STATIC (decl) = 1;
1061 DECL_INITIAL (decl) = ctor;
1063 DECL_NAME (decl) = create_tmp_var_name ("CSWTCH");
1064 DECL_ARTIFICIAL (decl) = 1;
1065 TREE_CONSTANT (decl) = 1;
1066 TREE_READONLY (decl) = 1;
1067 varpool_finalize_decl (decl);
1069 fetch = build4 (ARRAY_REF, value_type, decl, tidx, NULL_TREE,
1071 if (default_type != value_type)
1073 fetch = fold_convert (default_type, fetch);
1074 fetch = force_gimple_operand_gsi (&gsi, fetch, true, NULL_TREE,
1075 true, GSI_SAME_STMT);
1077 load = gimple_build_assign (name, fetch);
1080 SSA_NAME_DEF_STMT (name) = load;
1081 gsi_insert_before (&gsi, load, GSI_SAME_STMT);
1083 info->arr_ref_last = load;
1086 /* Builds and initializes static arrays initialized with values gathered from
1087 the SWTCH switch statement. Also creates statements that load values from
1091 build_arrays (gimple swtch, struct switch_conv_info *info)
1093 tree arr_index_type;
1094 tree tidx, sub, tmp, utype;
1096 gimple_stmt_iterator gsi;
1098 location_t loc = gimple_location (swtch);
1100 gsi = gsi_for_stmt (swtch);
1102 /* Make sure we do not generate arithmetics in a subrange. */
1103 utype = TREE_TYPE (info->index_expr);
1104 if (TREE_TYPE (utype))
1105 utype = lang_hooks.types.type_for_mode (TYPE_MODE (TREE_TYPE (utype)), 1);
1107 utype = lang_hooks.types.type_for_mode (TYPE_MODE (utype), 1);
1109 arr_index_type = build_index_type (info->range_size);
1110 tmp = create_tmp_var (utype, "csui");
1111 tidx = make_ssa_name (tmp, NULL);
1112 sub = fold_build2_loc (loc, MINUS_EXPR, utype,
1113 fold_convert_loc (loc, utype, info->index_expr),
1114 fold_convert_loc (loc, utype, info->range_min));
1115 sub = force_gimple_operand_gsi (&gsi, sub,
1116 false, NULL, true, GSI_SAME_STMT);
1117 stmt = gimple_build_assign (tidx, sub);
1118 SSA_NAME_DEF_STMT (tidx) = stmt;
1120 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
1122 info->arr_ref_first = stmt;
1124 for (gsi = gsi_start_phis (info->final_bb), i = 0;
1125 !gsi_end_p (gsi); gsi_next (&gsi), i++)
1126 build_one_array (swtch, i, arr_index_type, gsi_stmt (gsi), tidx, info);
1129 /* Generates and appropriately inserts loads of default values at the position
1130 given by BSI. Returns the last inserted statement. */
1133 gen_def_assigns (gimple_stmt_iterator *gsi, struct switch_conv_info *info)
1136 gimple assign = NULL;
1138 for (i = 0; i < info->phi_count; i++)
1141 = make_ssa_name (SSA_NAME_VAR (info->target_inbound_names[i]), NULL);
1143 info->target_outbound_names[i] = name;
1144 assign = gimple_build_assign (name, info->default_values[i]);
1145 SSA_NAME_DEF_STMT (name) = assign;
1146 gsi_insert_before (gsi, assign, GSI_SAME_STMT);
1147 update_stmt (assign);
1152 /* Deletes the unused bbs and edges that now contain the switch statement and
1153 its empty branch bbs. BBD is the now dead BB containing the original switch
1154 statement, FINAL is the last BB of the converted switch statement (in terms
1158 prune_bbs (basic_block bbd, basic_block final)
1163 for (ei = ei_start (bbd->succs); (e = ei_safe_edge (ei)); )
1169 delete_basic_block (bb);
1171 delete_basic_block (bbd);
1174 /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge
1175 from the basic block loading values from an array and E2F from the basic
1176 block loading default values. BBF is the last switch basic block (see the
1177 bbf description in the comment below). */
1180 fix_phi_nodes (edge e1f, edge e2f, basic_block bbf,
1181 struct switch_conv_info *info)
1183 gimple_stmt_iterator gsi;
1186 for (gsi = gsi_start_phis (bbf), i = 0;
1187 !gsi_end_p (gsi); gsi_next (&gsi), i++)
1189 gimple phi = gsi_stmt (gsi);
1190 add_phi_arg (phi, info->target_inbound_names[i], e1f, UNKNOWN_LOCATION);
1191 add_phi_arg (phi, info->target_outbound_names[i], e2f, UNKNOWN_LOCATION);
1195 /* Creates a check whether the switch expression value actually falls into the
1196 range given by all the cases. If it does not, the temporaries are loaded
1197 with default values instead. SWTCH is the switch statement being converted.
1199 bb0 is the bb with the switch statement, however, we'll end it with a
1202 bb1 is the bb to be used when the range check went ok. It is derived from
1205 bb2 is the bb taken when the expression evaluated outside of the range
1206 covered by the created arrays. It is populated by loads of default
1209 bbF is a fall through for both bb1 and bb2 and contains exactly what
1210 originally followed the switch statement.
1212 bbD contains the switch statement (in the end). It is unreachable but we
1213 still need to strip off its edges.
1217 gen_inbound_check (gimple swtch, struct switch_conv_info *info)
1219 tree label_decl1 = create_artificial_label (UNKNOWN_LOCATION);
1220 tree label_decl2 = create_artificial_label (UNKNOWN_LOCATION);
1221 tree label_decl3 = create_artificial_label (UNKNOWN_LOCATION);
1222 gimple label1, label2, label3;
1229 gimple_stmt_iterator gsi;
1230 basic_block bb0, bb1, bb2, bbf, bbd;
1231 edge e01, e02, e21, e1d, e1f, e2f;
1232 location_t loc = gimple_location (swtch);
1234 gcc_assert (info->default_values);
1236 bb0 = gimple_bb (swtch);
1238 tidx = gimple_assign_lhs (info->arr_ref_first);
1239 utype = TREE_TYPE (tidx);
1241 /* (end of) block 0 */
1242 gsi = gsi_for_stmt (info->arr_ref_first);
1245 bound = fold_convert_loc (loc, utype, info->range_size);
1246 cond_stmt = gimple_build_cond (LE_EXPR, tidx, bound, NULL_TREE, NULL_TREE);
1247 gsi_insert_before (&gsi, cond_stmt, GSI_SAME_STMT);
1248 update_stmt (cond_stmt);
1251 label2 = gimple_build_label (label_decl2);
1252 gsi_insert_before (&gsi, label2, GSI_SAME_STMT);
1253 last_assign = gen_def_assigns (&gsi, info);
1256 label1 = gimple_build_label (label_decl1);
1257 gsi_insert_before (&gsi, label1, GSI_SAME_STMT);
1260 gsi = gsi_start_bb (info->final_bb);
1261 label3 = gimple_build_label (label_decl3);
1262 gsi_insert_before (&gsi, label3, GSI_SAME_STMT);
1265 e02 = split_block (bb0, cond_stmt);
1268 e21 = split_block (bb2, last_assign);
1272 e1d = split_block (bb1, info->arr_ref_last);
1276 /* flags and profiles of the edge for in-range values */
1277 e01 = make_edge (bb0, bb1, EDGE_TRUE_VALUE);
1278 e01->probability = REG_BR_PROB_BASE - info->default_prob;
1279 e01->count = info->other_count;
1281 /* flags and profiles of the edge taking care of out-of-range values */
1282 e02->flags &= ~EDGE_FALLTHRU;
1283 e02->flags |= EDGE_FALSE_VALUE;
1284 e02->probability = info->default_prob;
1285 e02->count = info->default_count;
1287 bbf = info->final_bb;
1289 e1f = make_edge (bb1, bbf, EDGE_FALLTHRU);
1290 e1f->probability = REG_BR_PROB_BASE;
1291 e1f->count = info->other_count;
1293 e2f = make_edge (bb2, bbf, EDGE_FALLTHRU);
1294 e2f->probability = REG_BR_PROB_BASE;
1295 e2f->count = info->default_count;
1297 /* frequencies of the new BBs */
1298 bb1->frequency = EDGE_FREQUENCY (e01);
1299 bb2->frequency = EDGE_FREQUENCY (e02);
1300 bbf->frequency = EDGE_FREQUENCY (e1f) + EDGE_FREQUENCY (e2f);
1302 /* Tidy blocks that have become unreachable. */
1303 prune_bbs (bbd, info->final_bb);
1305 /* Fixup the PHI nodes in bbF. */
1306 fix_phi_nodes (e1f, e2f, bbf, info);
1308 /* Fix the dominator tree, if it is available. */
1309 if (dom_info_available_p (CDI_DOMINATORS))
1311 VEC (basic_block, heap) *bbs_to_fix_dom;
1313 set_immediate_dominator (CDI_DOMINATORS, bb1, bb0);
1314 set_immediate_dominator (CDI_DOMINATORS, bb2, bb0);
1315 if (! get_immediate_dominator (CDI_DOMINATORS, bbf))
1316 /* If bbD was the immediate dominator ... */
1317 set_immediate_dominator (CDI_DOMINATORS, bbf, bb0);
1319 bbs_to_fix_dom = VEC_alloc (basic_block, heap, 4);
1320 VEC_quick_push (basic_block, bbs_to_fix_dom, bb0);
1321 VEC_quick_push (basic_block, bbs_to_fix_dom, bb1);
1322 VEC_quick_push (basic_block, bbs_to_fix_dom, bb2);
1323 VEC_quick_push (basic_block, bbs_to_fix_dom, bbf);
1325 iterate_fix_dominators (CDI_DOMINATORS, bbs_to_fix_dom, true);
1326 VEC_free (basic_block, heap, bbs_to_fix_dom);
1330 /* The following function is invoked on every switch statement (the current one
1331 is given in SWTCH) and runs the individual phases of switch conversion on it
1332 one after another until one fails or the conversion is completed.
1333 Returns NULL on success, or a pointer to a string with the reason why the
1334 conversion failed. */
1337 process_switch (gimple swtch)
1339 struct switch_conv_info info;
1341 /* Group case labels so that we get the right results from the heuristics
1342 that decide on the code generation approach for this switch. */
1343 group_case_labels_stmt (swtch);
1345 /* If this switch is now a degenerate case with only a default label,
1346 there is nothing left for us to do. */
1347 if (gimple_switch_num_labels (swtch) < 2)
1348 return "switch is a degenerate case";
1350 collect_switch_conv_info (swtch, &info);
1352 /* No error markers should reach here (they should be filtered out
1353 during gimplification). */
1354 gcc_checking_assert (TREE_TYPE (info.index_expr) != error_mark_node);
1356 /* A switch on a constant should have been optimized in tree-cfg-cleanup. */
1357 gcc_checking_assert (! TREE_CONSTANT (info.index_expr));
1359 if (info.uniq <= MAX_CASE_BIT_TESTS)
1361 if (expand_switch_using_bit_tests_p (info.range_size,
1362 info.uniq, info.count))
1365 fputs (" expanding as bit test is preferable\n", dump_file);
1366 emit_case_bit_tests (swtch, info.index_expr,
1367 info.range_min, info.range_size);
1372 /* This will be expanded as a decision tree in stmt.c:expand_case. */
1373 return " expanding as jumps is preferable";
1376 /* If there is no common successor, we cannot do the transformation. */
1377 if (! info.final_bb)
1378 return "no common successor to all case label target blocks found";
1380 /* Check the case label values are within reasonable range: */
1381 if (!check_range (&info))
1383 gcc_assert (info.reason);
1387 /* For all the cases, see whether they are empty, the assignments they
1388 represent constant and so on... */
1389 if (! check_all_empty_except_final (&info))
1391 gcc_assert (info.reason);
1394 if (!check_final_bb (&info))
1396 gcc_assert (info.reason);
1400 /* At this point all checks have passed and we can proceed with the
1403 create_temp_arrays (&info);
1404 gather_default_values (gimple_switch_label (swtch, 0), &info);
1405 build_constructors (swtch, &info);
1407 build_arrays (swtch, &info); /* Build the static arrays and assignments. */
1408 gen_inbound_check (swtch, &info); /* Build the bounds check. */
1411 free_temp_arrays (&info);
1415 /* The main function of the pass scans statements for switches and invokes
1416 process_switch on them. */
1419 do_switchconv (void)
1425 const char *failure_reason;
1426 gimple stmt = last_stmt (bb);
1427 if (stmt && gimple_code (stmt) == GIMPLE_SWITCH)
1431 expanded_location loc = expand_location (gimple_location (stmt));
1433 fprintf (dump_file, "beginning to process the following "
1434 "SWITCH statement (%s:%d) : ------- \n",
1435 loc.file, loc.line);
1436 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1437 putc ('\n', dump_file);
1440 failure_reason = process_switch (stmt);
1441 if (! failure_reason)
1445 fputs ("Switch converted\n", dump_file);
1446 fputs ("--------------------------------\n", dump_file);
1449 /* Make no effort to update the post-dominator tree. It is actually not
1450 that hard for the transformations we have performed, but it is not
1451 supported by iterate_fix_dominators. */
1452 free_dominance_info (CDI_POST_DOMINATORS);
1458 fputs ("Bailing out - ", dump_file);
1459 fputs (failure_reason, dump_file);
1460 fputs ("\n--------------------------------\n", dump_file);
1469 /* The pass gate. */
1472 switchconv_gate (void)
1474 return flag_tree_switch_conversion != 0;
1477 struct gimple_opt_pass pass_convert_switch =
1481 "switchconv", /* name */
1482 switchconv_gate, /* gate */
1483 do_switchconv, /* execute */
1486 0, /* static_pass_number */
1487 TV_TREE_SWITCH_CONVERSION, /* tv_id */
1488 PROP_cfg | PROP_ssa, /* properties_required */
1489 0, /* properties_provided */
1490 0, /* properties_destroyed */
1491 0, /* todo_flags_start */
1493 | TODO_ggc_collect | TODO_verify_ssa
1495 | TODO_verify_flow /* todo_flags_finish */