1 /* Tree switch conversion for GNU compiler.
2 Copyright (C) 2017-2020 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 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/>. */
20 #ifndef TREE_SWITCH_CONVERSION_H
21 #define TREE_SWITCH_CONVERSION_H
23 namespace tree_switch_conversion
{
25 /* Type of cluster. */
34 #define PRINT_CASE(f,c) print_generic_expr (f, c)
36 /* Abstract base class for representing a cluster of cases.
38 Here is the inheritance hierarachy, and the enum_cluster_type
39 values for the concrete subclasses:
42 |-simple_cluster (SIMPLE_CASE)
44 |-jump_table_cluster (JUMP_TABLE)
45 `-bit_test_cluster (BIT_TEST). */
51 cluster (tree case_label_expr
, basic_block case_bb
, profile_probability prob
,
52 profile_probability subtree_prob
);
59 virtual cluster_type
get_type () = 0;
61 /* Get low value covered by a cluster. */
62 virtual tree
get_low () = 0;
64 /* Get high value covered by a cluster. */
65 virtual tree
get_high () = 0;
67 /* Debug content of a cluster. */
68 virtual void debug () = 0;
70 /* Dump content of a cluster. */
71 virtual void dump (FILE *f
, bool details
= false) = 0;
73 /* Emit GIMPLE code to handle the cluster. */
74 virtual void emit (tree
, tree
, tree
, basic_block
) = 0;
76 /* Return true if a cluster handles only a single case value and the
77 value is not a range. */
78 virtual bool is_single_value_p ()
83 /* Return range of a cluster. If value would overflow in type of LOW,
85 static unsigned HOST_WIDE_INT
get_range (tree low
, tree high
)
87 tree r
= fold_build2 (MINUS_EXPR
, TREE_TYPE (low
), high
, low
);
88 if (!tree_fits_uhwi_p (r
))
91 return tree_to_uhwi (r
) + 1;
95 tree m_case_label_expr
;
97 /* Basic block of the case. */
98 basic_block m_case_bb
;
100 /* Probability of taking this cluster. */
101 profile_probability m_prob
;
103 /* Probability of reaching subtree rooted at this node. */
104 profile_probability m_subtree_prob
;
107 /* Default constructor. */
111 cluster::cluster (tree case_label_expr
, basic_block case_bb
,
112 profile_probability prob
, profile_probability subtree_prob
):
113 m_case_label_expr (case_label_expr
), m_case_bb (case_bb
), m_prob (prob
),
114 m_subtree_prob (subtree_prob
)
118 /* Subclass of cluster representing a simple contiguous range
121 class simple_cluster
: public cluster
125 simple_cluster (tree low
, tree high
, tree case_label_expr
,
126 basic_block case_bb
, profile_probability prob
);
157 dump (FILE *f
, bool details ATTRIBUTE_UNUSED
= false)
159 PRINT_CASE (f
, get_low ());
160 if (get_low () != get_high ())
163 PRINT_CASE (f
, get_high ());
168 void emit (tree
, tree
, tree
, basic_block
)
173 bool is_single_value_p ()
175 return tree_int_cst_equal (get_low (), get_high ());
178 /* Low value of the case. */
181 /* High value of the case. */
184 /* True if case is a range. */
188 simple_cluster::simple_cluster (tree low
, tree high
, tree case_label_expr
,
189 basic_block case_bb
, profile_probability prob
):
190 cluster (case_label_expr
, case_bb
, prob
, prob
),
191 m_low (low
), m_high (high
)
193 m_range_p
= m_high
!= NULL
;
198 /* Abstract subclass of jump table and bit test cluster,
199 handling a collection of simple_cluster instances. */
201 class group_cluster
: public cluster
205 group_cluster (vec
<cluster
*> &clusters
, unsigned start
, unsigned end
);
213 return m_cases
[0]->get_low ();
219 return m_cases
[m_cases
.length () - 1]->get_high ();
228 void dump (FILE *f
, bool details
= false);
230 /* List of simple clusters handled by the group. */
231 vec
<simple_cluster
*> m_cases
;
234 /* Concrete subclass of group_cluster representing a collection
235 of cases to be implemented as a jump table.
236 The "emit" vfunc gernerates a nested switch statement which
237 is later lowered to a jump table. */
239 class jump_table_cluster
: public group_cluster
243 jump_table_cluster (vec
<cluster
*> &clusters
, unsigned start
, unsigned end
)
244 : group_cluster (clusters
, start
, end
)
253 void emit (tree index_expr
, tree index_type
,
254 tree default_label_expr
, basic_block default_bb
);
256 /* Find jump tables of given CLUSTERS, where all members of the vector
257 are of type simple_cluster. New clusters are returned. */
258 static vec
<cluster
*> find_jump_tables (vec
<cluster
*> &clusters
);
260 /* Return true when cluster starting at START and ending at END (inclusive)
261 can build a jump-table. */
262 static bool can_be_handled (const vec
<cluster
*> &clusters
, unsigned start
,
265 /* Return true if cluster starting at START and ending at END (inclusive)
266 is profitable transformation. */
267 static bool is_beneficial (const vec
<cluster
*> &clusters
, unsigned start
,
270 /* Return the smallest number of different values for which it is best
271 to use a jump-table instead of a tree of conditional branches. */
272 static inline unsigned int case_values_threshold (void);
274 /* Return whether jump table expansion is allowed. */
275 static bool is_enabled (void);
278 /* A GIMPLE switch statement can be expanded to a short sequence of bit-wise
279 comparisons. "switch(x)" is converted into "if ((1 << (x-MINVAL)) & CST)"
280 where CST and MINVAL are integer constants. This is better than a series
281 of compare-and-banch insns in some cases, e.g. we can implement:
283 if ((x==4) || (x==6) || (x==9) || (x==11))
285 as a single bit test:
287 if ((1<<x) & ((1<<4)|(1<<6)|(1<<9)|(1<<11)))
289 This transformation is only applied if the number of case targets is small,
290 if CST constains at least 3 bits, and "1 << x" is cheap. The bit tests are
291 performed in "word_mode".
293 The following example shows the code the transformation generates:
299 case '0': case '1': case '2': case '3': case '4':
300 case '5': case '6': case '7': case '8': case '9':
301 case 'A': case 'B': case 'C': case 'D': case 'E':
313 if (tmp1 > (70 - 48)) goto L2;
315 tmp3 = 0b11111100000001111111111;
316 if ((tmp2 & tmp3) != 0) goto L1 ; else goto L2;
323 TODO: There are still some improvements to this transformation that could
326 * A narrower mode than word_mode could be used if that is cheaper, e.g.
327 for x86_64 where a narrower-mode shift may result in smaller code.
329 * The compounded constant could be shifted rather than the one. The
330 test would be either on the sign bit or on the least significant bit,
331 depending on the direction of the shift. On some machines, the test
332 for the branch would be free if the bit to test is already set by the
335 This transformation was contributed by Roger Sayle, see this e-mail:
336 http://gcc.gnu.org/ml/gcc-patches/2003-01/msg01950.html
339 class bit_test_cluster
: public group_cluster
343 bit_test_cluster (vec
<cluster
*> &clusters
, unsigned start
, unsigned end
,
344 bool handles_entire_switch
)
345 :group_cluster (clusters
, start
, end
),
346 m_handles_entire_switch (handles_entire_switch
)
355 /* Expand a switch statement by a short sequence of bit-wise
356 comparisons. "switch(x)" is effectively converted into
357 "if ((1 << (x-MINVAL)) & CST)" where CST and MINVAL are
360 INDEX_EXPR is the value being switched on.
362 MINVAL is the lowest case value of in the case nodes,
363 and RANGE is highest value minus MINVAL. MINVAL and RANGE
364 are not guaranteed to be of the same type as INDEX_EXPR
365 (the gimplifier doesn't change the type of case label values,
366 and MINVAL and RANGE are derived from those values).
367 MAXVAL is MINVAL + RANGE.
369 There *MUST* be max_case_bit_tests or less unique case
371 void emit (tree index_expr
, tree index_type
,
372 tree default_label_expr
, basic_block default_bb
);
374 /* Find bit tests of given CLUSTERS, where all members of the vector
375 are of type simple_cluster. New clusters are returned. */
376 static vec
<cluster
*> find_bit_tests (vec
<cluster
*> &clusters
);
378 /* Return true when RANGE of case values with UNIQ labels
379 can build a bit test. */
380 static bool can_be_handled (unsigned HOST_WIDE_INT range
, unsigned uniq
);
382 /* Return true when cluster starting at START and ending at END (inclusive)
383 can build a bit test. */
384 static bool can_be_handled (const vec
<cluster
*> &clusters
, unsigned start
,
387 /* Return true when COUNT of cases of UNIQ labels is beneficial for bit test
389 static bool is_beneficial (unsigned count
, unsigned uniq
);
391 /* Return true if cluster starting at START and ending at END (inclusive)
392 is profitable transformation. */
393 static bool is_beneficial (const vec
<cluster
*> &clusters
, unsigned start
,
396 /* Split the basic block at the statement pointed to by GSIP, and insert
397 a branch to the target basic block of E_TRUE conditional on tree
400 It is assumed that there is already an edge from the to-be-split
401 basic block to E_TRUE->dest block. This edge is removed, and the
402 profile information on the edge is re-used for the new conditional
405 The CFG is updated. The dominator tree will not be valid after
406 this transformation, but the immediate dominators are updated if
407 UPDATE_DOMINATORS is true.
409 Returns the newly created basic block. */
410 static basic_block
hoist_edge_and_branch_if_true (gimple_stmt_iterator
*gsip
,
413 profile_probability prob
);
415 /* True when the jump table handles an entire switch statement. */
416 bool m_handles_entire_switch
;
418 /* Maximum number of different basic blocks that can be handled by
420 static const int m_max_case_bit_tests
= 3;
423 /* Helper struct to find minimal clusters. */
425 class min_cluster_item
429 min_cluster_item (unsigned count
, unsigned start
, unsigned non_jt_cases
):
430 m_count (count
), m_start (start
), m_non_jt_cases (non_jt_cases
)
433 /* Count of clusters. */
436 /* Index where is cluster boundary. */
439 /* Total number of cases that will not be in a jump table. */
440 unsigned m_non_jt_cases
;
443 /* Helper struct to represent switch decision tree. */
448 /* Empty Constructor. */
451 /* Return true when it has a child. */
454 return m_left
!= NULL
|| m_right
!= NULL
;
457 /* Left son in binary tree. */
458 case_tree_node
*m_left
;
460 /* Right son in binary tree; also node chain. */
461 case_tree_node
*m_right
;
463 /* Parent of node in binary tree. */
464 case_tree_node
*m_parent
;
466 /* Cluster represented by this tree node. */
471 case_tree_node::case_tree_node ():
472 m_left (NULL
), m_right (NULL
), m_parent (NULL
), m_c (NULL
)
477 jump_table_cluster::case_values_threshold (void)
479 unsigned int threshold
= param_case_values_threshold
;
482 threshold
= targetm
.case_values_threshold ();
487 /* Return whether jump table expansion is allowed. */
488 bool jump_table_cluster::is_enabled (void)
490 /* If neither casesi or tablejump is available, or flag_jump_tables
491 over-ruled us, we really have no choice. */
492 if (!targetm
.have_casesi () && !targetm
.have_tablejump ())
494 if (!flag_jump_tables
)
496 #ifndef ASM_OUTPUT_ADDR_DIFF_ELT
504 /* A case_bit_test represents a set of case nodes that may be
505 selected from using a bit-wise comparison. HI and LO hold
506 the integer to be tested against, TARGET_EDGE contains the
507 edge to the basic block to jump to upon success and BITS
508 counts the number of case nodes handled by this test,
509 typically the number of bits set in HI:LO. The LABEL field
510 is used to quickly identify all cases in this set without
511 looking at label_to_block for every case label. */
517 basic_block target_bb
;
521 /* Comparison function for qsort to order bit tests by decreasing
522 probability of execution. */
523 static int cmp (const void *p1
, const void *p2
);
526 class switch_decision_tree
530 switch_decision_tree (gswitch
*swtch
): m_switch (swtch
), m_phi_mapping (),
531 m_case_bbs (), m_case_node_pool ("struct case_node pool"),
536 /* Analyze switch statement and return true when the statement is expanded
538 bool analyze_switch_statement ();
540 /* Attempt to expand CLUSTERS as a decision tree. Return true when
542 bool try_switch_expansion (vec
<cluster
*> &clusters
);
543 /* Compute the number of case labels that correspond to each outgoing edge of
544 switch statement. Record this information in the aux field of the edge.
546 void compute_cases_per_edge ();
548 /* Before switch transformation, record all SSA_NAMEs defined in switch BB
549 and used in a label basic block. */
550 void record_phi_operand_mapping ();
552 /* Append new operands to PHI statements that were introduced due to
553 addition of new edges to case labels. */
554 void fix_phi_operands_for_edges ();
556 /* Generate a decision tree, switching on INDEX_EXPR and jumping to
557 one of the labels in CASE_LIST or to the DEFAULT_LABEL.
559 We generate a binary decision tree to select the appropriate target
561 void emit (basic_block bb
, tree index_expr
,
562 profile_probability default_prob
, tree index_type
);
564 /* Emit step-by-step code to select a case for the value of INDEX.
565 The thus generated decision tree follows the form of the
566 case-node binary tree NODE, whose nodes represent test conditions.
567 DEFAULT_PROB is probability of cases leading to default BB.
568 INDEX_TYPE is the type of the index of the switch. */
569 basic_block
emit_case_nodes (basic_block bb
, tree index
,
570 case_tree_node
*node
,
571 profile_probability default_prob
,
572 tree index_type
, location_t
);
574 /* Take an ordered list of case nodes
575 and transform them into a near optimal binary tree,
576 on the assumption that any target code selection value is as
579 The transformation is performed by splitting the ordered
580 list into two equal sections plus a pivot. The parts are
581 then attached to the pivot as left and right branches. Each
582 branch is then transformed recursively. */
583 static void balance_case_nodes (case_tree_node
**head
,
584 case_tree_node
*parent
);
586 /* Dump ROOT, a list or tree of case nodes, to file F. */
587 static void dump_case_nodes (FILE *f
, case_tree_node
*root
, int indent_step
,
590 /* Add an unconditional jump to CASE_BB that happens in basic block BB. */
591 static void emit_jump (basic_block bb
, basic_block case_bb
);
593 /* Generate code to compare OP0 with OP1 so that the condition codes are
594 set and to jump to LABEL_BB if the condition is true.
595 COMPARISON is the GIMPLE comparison (EQ, NE, GT, etc.).
596 PROB is the probability of jumping to LABEL_BB. */
597 static basic_block
emit_cmp_and_jump_insns (basic_block bb
, tree op0
,
598 tree op1
, tree_code comparison
,
599 basic_block label_bb
,
600 profile_probability prob
,
603 /* Generate code to jump to LABEL if OP0 and OP1 are equal in mode MODE.
604 PROB is the probability of jumping to LABEL_BB. */
605 static basic_block
do_jump_if_equal (basic_block bb
, tree op0
, tree op1
,
606 basic_block label_bb
,
607 profile_probability prob
,
610 /* Reset the aux field of all outgoing edges of switch basic block. */
611 static inline void reset_out_edges_aux (gswitch
*swtch
);
613 /* Switch statement. */
616 /* Map of PHI nodes that have to be fixed after expansion. */
617 hash_map
<tree
, tree
> m_phi_mapping
;
619 /* List of basic blocks that belong to labels of the switch. */
620 auto_vec
<basic_block
> m_case_bbs
;
622 /* Basic block with default label. */
623 basic_block m_default_bb
;
625 /* A pool for case nodes. */
626 object_allocator
<case_tree_node
> m_case_node_pool
;
628 /* Balanced tree of case nodes. */
629 case_tree_node
*m_case_list
;
633 Switch initialization conversion
635 The following pass changes simple initializations of scalars in a switch
636 statement into initializations from a static array. Obviously, the values
637 must be constant and known at compile time and a default branch must be
638 provided. For example, the following code:
662 a_5 = PHI <a_1, a_2, a_3, a_4>
663 b_5 = PHI <b_1, b_2, b_3, b_4>
668 static const int = CSWTCH01[] = {6, 6, 5, 1, 1, 1, 1, 1, 1, 1, 1, 4};
669 static const int = CSWTCH02[] = {8, 8, 9, 16, 16, 16, 16, 16, 16, 16,
672 if (((unsigned) argc) - 1 < 11)
674 a_6 = CSWTCH02[argc - 1];
675 b_6 = CSWTCH01[argc - 1];
685 There are further constraints. Specifically, the range of values across all
686 case labels must not be bigger than param_switch_conversion_branch_ratio
687 (default eight) times the number of the actual switch branches.
689 This transformation was contributed by Martin Jambor, see this e-mail:
690 http://gcc.gnu.org/ml/gcc-patches/2008-07/msg00011.html */
692 /* The main structure of the pass. */
693 class switch_conversion
697 switch_conversion ();
700 ~switch_conversion ();
702 /* The following function is invoked on every switch statement (the current
703 one is given in SWTCH) and runs the individual phases of switch
704 conversion on it one after another until one fails or the conversion
705 is completed. On success, NULL is in m_reason, otherwise points
706 to a string with the reason why the conversion failed. */
707 void expand (gswitch
*swtch
);
709 /* Collection information about SWTCH statement. */
710 void collect (gswitch
*swtch
);
712 /* Checks whether the range given by individual case statements of the switch
713 switch statement isn't too big and whether the number of branches actually
714 satisfies the size of the new array. */
717 /* Checks whether all but the final BB basic blocks are empty. */
718 bool check_all_empty_except_final ();
720 /* This function checks whether all required values in phi nodes in final_bb
721 are constants. Required values are those that correspond to a basic block
722 which is a part of the examined switch statement. It returns true if the
723 phi nodes are OK, otherwise false. */
724 bool check_final_bb ();
726 /* The following function allocates default_values, target_{in,out}_names and
727 constructors arrays. The last one is also populated with pointers to
728 vectors that will become constructors of new arrays. */
729 void create_temp_arrays ();
731 /* Populate the array of default values in the order of phi nodes.
732 DEFAULT_CASE is the CASE_LABEL_EXPR for the default switch branch
733 if the range is non-contiguous or the default case has standard
734 structure, otherwise it is the first non-default case instead. */
735 void gather_default_values (tree default_case
);
737 /* The following function populates the vectors in the constructors array with
738 future contents of the static arrays. The vectors are populated in the
739 order of phi nodes. */
740 void build_constructors ();
742 /* If all values in the constructor vector are products of a linear function
743 a * x + b, then return true. When true, COEFF_A and COEFF_B and
744 coefficients of the linear function. Note that equal values are special
745 case of a linear function with a and b equal to zero. */
746 bool contains_linear_function_p (vec
<constructor_elt
, va_gc
> *vec
,
747 wide_int
*coeff_a
, wide_int
*coeff_b
);
749 /* Return type which should be used for array elements, either TYPE's
750 main variant or, for integral types, some smaller integral type
751 that can still hold all the constants. */
752 tree
array_value_type (tree type
, int num
);
754 /* Create an appropriate array type and declaration and assemble a static
755 array variable. Also create a load statement that initializes
756 the variable in question with a value from the static array. SWTCH is
757 the switch statement being converted, NUM is the index to
758 arrays of constructors, default values and target SSA names
759 for this particular array. ARR_INDEX_TYPE is the type of the index
760 of the new array, PHI is the phi node of the final BB that corresponds
761 to the value that will be loaded from the created array. TIDX
762 is an ssa name of a temporary variable holding the index for loads from the
764 void build_one_array (int num
, tree arr_index_type
,
765 gphi
*phi
, tree tidx
);
767 /* Builds and initializes static arrays initialized with values gathered from
768 the switch statement. Also creates statements that load values from
770 void build_arrays ();
772 /* Generates and appropriately inserts loads of default values at the position
773 given by GSI. Returns the last inserted statement. */
774 gassign
*gen_def_assigns (gimple_stmt_iterator
*gsi
);
776 /* Deletes the unused bbs and edges that now contain the switch statement and
777 its empty branch bbs. BBD is the now dead BB containing
778 the original switch statement, FINAL is the last BB of the converted
779 switch statement (in terms of succession). */
780 void prune_bbs (basic_block bbd
, basic_block final
, basic_block default_bb
);
782 /* Add values to phi nodes in final_bb for the two new edges. E1F is the edge
783 from the basic block loading values from an array and E2F from the basic
784 block loading default values. BBF is the last switch basic block (see the
785 bbf description in the comment below). */
786 void fix_phi_nodes (edge e1f
, edge e2f
, basic_block bbf
);
788 /* Creates a check whether the switch expression value actually falls into the
789 range given by all the cases. If it does not, the temporaries are loaded
790 with default values instead. */
791 void gen_inbound_check ();
793 /* Switch statement for which switch conversion takes place. */
796 /* The expression used to decide the switch branch. */
799 /* The following integer constants store the minimum and maximum value
800 covered by the case labels. */
804 /* The difference between the above two numbers. Stored here because it
805 is used in all the conversion heuristics, as well as for some of the
806 transformation, and it is expensive to re-compute it all the time. */
809 /* Basic block that contains the actual GIMPLE_SWITCH. */
810 basic_block m_switch_bb
;
812 /* Basic block that is the target of the default case. */
813 basic_block m_default_bb
;
815 /* The single successor block of all branches out of the GIMPLE_SWITCH,
816 if such a block exists. Otherwise NULL. */
817 basic_block m_final_bb
;
819 /* The probability of the default edge in the replaced switch. */
820 profile_probability m_default_prob
;
822 /* Number of phi nodes in the final bb (that we'll be replacing). */
825 /* Constructors of new static arrays. */
826 vec
<constructor_elt
, va_gc
> **m_constructors
;
828 /* Array of default values, in the same order as phi nodes. */
829 tree
*m_default_values
;
831 /* Array of ssa names that are initialized with a value from a new static
833 tree
*m_target_inbound_names
;
835 /* Array of ssa names that are initialized with the default value if the
836 switch expression is out of range. */
837 tree
*m_target_outbound_names
;
842 /* The first load statement that loads a temporary from a new static array.
844 gimple
*m_arr_ref_first
;
846 /* The last load statement that loads a temporary from a new static array. */
847 gimple
*m_arr_ref_last
;
849 /* String reason why the case wasn't a good candidate that is written to the
850 dump file, if there is one. */
851 const char *m_reason
;
853 /* True if default case is not used for any value between range_min and
854 range_max inclusive. */
855 bool m_contiguous_range
;
857 /* True if default case does not have the required shape for other case
859 bool m_default_case_nonstandard
;
861 /* Number of uniq labels for non-default edges. */
864 /* Count is number of non-default edges. */
865 unsigned int m_count
;
867 /* True if CFG has been changed. */
872 switch_decision_tree::reset_out_edges_aux (gswitch
*swtch
)
874 basic_block bb
= gimple_bb (swtch
);
877 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
881 } // tree_switch_conversion namespace
883 #endif // TREE_SWITCH_CONVERSION_H