1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2013 Free Software Foundation, Inc.
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* Interprocedural constant propagation (IPA-CP).
25 The goal of this transformation is to
27 1) discover functions which are always invoked with some arguments with the
28 same known constant values and modify the functions so that the
29 subsequent optimizations can take advantage of the knowledge, and
31 2) partial specialization - create specialized versions of functions
32 transformed in this way if some parameters are known constants only in
33 certain contexts but the estimated tradeoff between speedup and cost size
36 The algorithm also propagates types and attempts to perform type based
37 devirtualization. Types are propagated much like constants.
39 The algorithm basically consists of three stages. In the first, functions
40 are analyzed one at a time and jump functions are constructed for all known
41 call-sites. In the second phase, the pass propagates information from the
42 jump functions across the call to reveal what values are available at what
43 call sites, performs estimations of effects of known values on functions and
44 their callees, and finally decides what specialized extra versions should be
45 created. In the third, the special versions materialize and appropriate
48 The algorithm used is to a certain extent based on "Interprocedural Constant
49 Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
50 Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
51 Cooper, Mary W. Hall, and Ken Kennedy.
54 First stage - intraprocedural analysis
55 =======================================
57 This phase computes jump_function and modification flags.
59 A jump function for a call-site represents the values passed as an actual
60 arguments of a given call-site. In principle, there are three types of
63 Pass through - the caller's formal parameter is passed as an actual
64 argument, plus an operation on it can be performed.
65 Constant - a constant is passed as an actual argument.
66 Unknown - neither of the above.
68 All jump function types are described in detail in ipa-prop.h, together with
69 the data structures that represent them and methods of accessing them.
71 ipcp_generate_summary() is the main function of the first stage.
73 Second stage - interprocedural analysis
74 ========================================
76 This stage is itself divided into two phases. In the first, we propagate
77 known values over the call graph, in the second, we make cloning decisions.
78 It uses a different algorithm than the original Callahan's paper.
80 First, we traverse the functions topologically from callers to callees and,
81 for each strongly connected component (SCC), we propagate constants
82 according to previously computed jump functions. We also record what known
83 values depend on other known values and estimate local effects. Finally, we
84 propagate cumulative information about these effects from dependent values
85 to those on which they depend.
87 Second, we again traverse the call graph in the same topological order and
88 make clones for functions which we know are called with the same values in
89 all contexts and decide about extra specialized clones of functions just for
90 some contexts - these decisions are based on both local estimates and
91 cumulative estimates propagated from callees.
93 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
96 Third phase - materialization of clones, call statement updates.
97 ============================================
99 This stage is currently performed by call graph code (mainly in cgraphunit.c
100 and tree-inline.c) according to instructions inserted to the call graph by
105 #include "coretypes.h"
110 #include "ipa-prop.h"
111 #include "tree-flow.h"
112 #include "tree-pass.h"
114 #include "diagnostic.h"
115 #include "tree-pretty-print.h"
116 #include "tree-inline.h"
118 #include "ipa-inline.h"
119 #include "ipa-utils.h"
123 /* Describes a particular source for an IPA-CP value. */
125 struct ipcp_value_source
127 /* Aggregate offset of the source, negative if the source is scalar value of
128 the argument itself. */
129 HOST_WIDE_INT offset
;
130 /* The incoming edge that brought the value. */
131 struct cgraph_edge
*cs
;
132 /* If the jump function that resulted into his value was a pass-through or an
133 ancestor, this is the ipcp_value of the caller from which the described
134 value has been derived. Otherwise it is NULL. */
135 struct ipcp_value
*val
;
136 /* Next pointer in a linked list of sources of a value. */
137 struct ipcp_value_source
*next
;
138 /* If the jump function that resulted into his value was a pass-through or an
139 ancestor, this is the index of the parameter of the caller the jump
140 function references. */
144 /* Describes one particular value stored in struct ipcp_lattice. */
148 /* The actual value for the given parameter. This is either an IPA invariant
149 or a TREE_BINFO describing a type that can be used for
152 /* The list of sources from which this value originates. */
153 struct ipcp_value_source
*sources
;
154 /* Next pointers in a linked list of all values in a lattice. */
155 struct ipcp_value
*next
;
156 /* Next pointers in a linked list of values in a strongly connected component
158 struct ipcp_value
*scc_next
;
159 /* Next pointers in a linked list of SCCs of values sorted topologically
160 according their sources. */
161 struct ipcp_value
*topo_next
;
162 /* A specialized node created for this value, NULL if none has been (so far)
164 struct cgraph_node
*spec_node
;
165 /* Depth first search number and low link for topological sorting of
168 /* Time benefit and size cost that specializing the function for this value
169 would bring about in this function alone. */
170 int local_time_benefit
, local_size_cost
;
171 /* Time benefit and size cost that specializing the function for this value
172 can bring about in it's callees (transitively). */
173 int prop_time_benefit
, prop_size_cost
;
174 /* True if this valye is currently on the topo-sort stack. */
178 /* Lattice describing potential values of a formal parameter of a function, or
179 a part of an aggreagate. TOP is represented by a lattice with zero values
180 and with contains_variable and bottom flags cleared. BOTTOM is represented
181 by a lattice with the bottom flag set. In that case, values and
182 contains_variable flag should be disregarded. */
186 /* The list of known values and types in this lattice. Note that values are
187 not deallocated if a lattice is set to bottom because there may be value
188 sources referencing them. */
189 struct ipcp_value
*values
;
190 /* Number of known values and types in this lattice. */
192 /* The lattice contains a variable component (in addition to values). */
193 bool contains_variable
;
194 /* The value of the lattice is bottom (i.e. variable and unusable for any
199 /* Lattice with an offset to describe a part of an aggregate. */
201 struct ipcp_agg_lattice
: public ipcp_lattice
203 /* Offset that is being described by this lattice. */
204 HOST_WIDE_INT offset
;
205 /* Size so that we don't have to re-compute it every time we traverse the
206 list. Must correspond to TYPE_SIZE of all lat values. */
208 /* Next element of the linked list. */
209 struct ipcp_agg_lattice
*next
;
212 /* Structure containing lattices for a parameter itself and for pieces of
213 aggregates that are passed in the parameter or by a reference in a parameter
214 plus some other useful flags. */
216 struct ipcp_param_lattices
218 /* Lattice describing the value of the parameter itself. */
219 struct ipcp_lattice itself
;
220 /* Lattices describing aggregate parts. */
221 struct ipcp_agg_lattice
*aggs
;
222 /* Number of aggregate lattices */
224 /* True if aggregate data were passed by reference (as opposed to by
227 /* All aggregate lattices contain a variable component (in addition to
229 bool aggs_contain_variable
;
230 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
231 for any propagation). */
234 /* There is a virtual call based on this parameter. */
238 /* Allocation pools for values and their sources in ipa-cp. */
240 alloc_pool ipcp_values_pool
;
241 alloc_pool ipcp_sources_pool
;
242 alloc_pool ipcp_agg_lattice_pool
;
244 /* Maximal count found in program. */
246 static gcov_type max_count
;
248 /* Original overall size of the program. */
250 static long overall_size
, max_new_size
;
252 /* Head of the linked list of topologically sorted values. */
254 static struct ipcp_value
*values_topo
;
256 /* Return the param lattices structure corresponding to the Ith formal
257 parameter of the function described by INFO. */
258 static inline struct ipcp_param_lattices
*
259 ipa_get_parm_lattices (struct ipa_node_params
*info
, int i
)
261 gcc_assert (i
>= 0 && i
< ipa_get_param_count (info
));
262 gcc_checking_assert (!info
->ipcp_orig_node
);
263 gcc_checking_assert (info
->lattices
);
264 return &(info
->lattices
[i
]);
267 /* Return the lattice corresponding to the scalar value of the Ith formal
268 parameter of the function described by INFO. */
269 static inline struct ipcp_lattice
*
270 ipa_get_scalar_lat (struct ipa_node_params
*info
, int i
)
272 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
273 return &plats
->itself
;
276 /* Return whether LAT is a lattice with a single constant and without an
280 ipa_lat_is_single_const (struct ipcp_lattice
*lat
)
283 || lat
->contains_variable
284 || lat
->values_count
!= 1)
290 /* Return true iff the CS is an edge within a strongly connected component as
291 computed by ipa_reduced_postorder. */
294 edge_within_scc (struct cgraph_edge
*cs
)
296 struct ipa_dfs_info
*caller_dfs
= (struct ipa_dfs_info
*) cs
->caller
->symbol
.aux
;
297 struct ipa_dfs_info
*callee_dfs
;
298 struct cgraph_node
*callee
= cgraph_function_node (cs
->callee
, NULL
);
300 callee_dfs
= (struct ipa_dfs_info
*) callee
->symbol
.aux
;
303 && caller_dfs
->scc_no
== callee_dfs
->scc_no
);
306 /* Print V which is extracted from a value in a lattice to F. */
309 print_ipcp_constant_value (FILE * f
, tree v
)
311 if (TREE_CODE (v
) == TREE_BINFO
)
313 fprintf (f
, "BINFO ");
314 print_generic_expr (f
, BINFO_TYPE (v
), 0);
316 else if (TREE_CODE (v
) == ADDR_EXPR
317 && TREE_CODE (TREE_OPERAND (v
, 0)) == CONST_DECL
)
320 print_generic_expr (f
, DECL_INITIAL (TREE_OPERAND (v
, 0)), 0);
323 print_generic_expr (f
, v
, 0);
326 /* Print a lattice LAT to F. */
329 print_lattice (FILE * f
, struct ipcp_lattice
*lat
,
330 bool dump_sources
, bool dump_benefits
)
332 struct ipcp_value
*val
;
337 fprintf (f
, "BOTTOM\n");
341 if (!lat
->values_count
&& !lat
->contains_variable
)
343 fprintf (f
, "TOP\n");
347 if (lat
->contains_variable
)
349 fprintf (f
, "VARIABLE");
355 for (val
= lat
->values
; val
; val
= val
->next
)
357 if (dump_benefits
&& prev
)
359 else if (!dump_benefits
&& prev
)
364 print_ipcp_constant_value (f
, val
->value
);
368 struct ipcp_value_source
*s
;
370 fprintf (f
, " [from:");
371 for (s
= val
->sources
; s
; s
= s
->next
)
372 fprintf (f
, " %i(%i)", s
->cs
->caller
->symbol
.order
,
378 fprintf (f
, " [loc_time: %i, loc_size: %i, "
379 "prop_time: %i, prop_size: %i]\n",
380 val
->local_time_benefit
, val
->local_size_cost
,
381 val
->prop_time_benefit
, val
->prop_size_cost
);
387 /* Print all ipcp_lattices of all functions to F. */
390 print_all_lattices (FILE * f
, bool dump_sources
, bool dump_benefits
)
392 struct cgraph_node
*node
;
395 fprintf (f
, "\nLattices:\n");
396 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
398 struct ipa_node_params
*info
;
400 info
= IPA_NODE_REF (node
);
401 fprintf (f
, " Node: %s/%i:\n", cgraph_node_name (node
),
403 count
= ipa_get_param_count (info
);
404 for (i
= 0; i
< count
; i
++)
406 struct ipcp_agg_lattice
*aglat
;
407 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
408 fprintf (f
, " param [%d]: ", i
);
409 print_lattice (f
, &plats
->itself
, dump_sources
, dump_benefits
);
411 if (plats
->virt_call
)
412 fprintf (f
, " virt_call flag set\n");
414 if (plats
->aggs_bottom
)
416 fprintf (f
, " AGGS BOTTOM\n");
419 if (plats
->aggs_contain_variable
)
420 fprintf (f
, " AGGS VARIABLE\n");
421 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
423 fprintf (f
, " %soffset " HOST_WIDE_INT_PRINT_DEC
": ",
424 plats
->aggs_by_ref
? "ref " : "", aglat
->offset
);
425 print_lattice (f
, aglat
, dump_sources
, dump_benefits
);
431 /* Determine whether it is at all technically possible to create clones of NODE
432 and store this information in the ipa_node_params structure associated
436 determine_versionability (struct cgraph_node
*node
)
438 const char *reason
= NULL
;
440 /* There are a number of generic reasons functions cannot be versioned. We
441 also cannot remove parameters if there are type attributes such as fnspec
443 if (node
->symbol
.alias
|| node
->thunk
.thunk_p
)
444 reason
= "alias or thunk";
445 else if (!node
->local
.versionable
)
446 reason
= "not a tree_versionable_function";
447 else if (cgraph_function_body_availability (node
) <= AVAIL_OVERWRITABLE
)
448 reason
= "insufficient body availability";
450 if (reason
&& dump_file
&& !node
->symbol
.alias
&& !node
->thunk
.thunk_p
)
451 fprintf (dump_file
, "Function %s/%i is not versionable, reason: %s.\n",
452 cgraph_node_name (node
), node
->symbol
.order
, reason
);
454 node
->local
.versionable
= (reason
== NULL
);
457 /* Return true if it is at all technically possible to create clones of a
461 ipcp_versionable_function_p (struct cgraph_node
*node
)
463 return node
->local
.versionable
;
466 /* Structure holding accumulated information about callers of a node. */
468 struct caller_statistics
471 int n_calls
, n_hot_calls
, freq_sum
;
474 /* Initialize fields of STAT to zeroes. */
477 init_caller_stats (struct caller_statistics
*stats
)
479 stats
->count_sum
= 0;
481 stats
->n_hot_calls
= 0;
485 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
486 non-thunk incoming edges to NODE. */
489 gather_caller_stats (struct cgraph_node
*node
, void *data
)
491 struct caller_statistics
*stats
= (struct caller_statistics
*) data
;
492 struct cgraph_edge
*cs
;
494 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
495 if (cs
->caller
->thunk
.thunk_p
)
496 cgraph_for_node_and_aliases (cs
->caller
, gather_caller_stats
,
500 stats
->count_sum
+= cs
->count
;
501 stats
->freq_sum
+= cs
->frequency
;
503 if (cgraph_maybe_hot_edge_p (cs
))
504 stats
->n_hot_calls
++;
510 /* Return true if this NODE is viable candidate for cloning. */
513 ipcp_cloning_candidate_p (struct cgraph_node
*node
)
515 struct caller_statistics stats
;
517 gcc_checking_assert (cgraph_function_with_gimple_body_p (node
));
519 if (!flag_ipa_cp_clone
)
522 fprintf (dump_file
, "Not considering %s for cloning; "
523 "-fipa-cp-clone disabled.\n",
524 cgraph_node_name (node
));
528 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
)))
531 fprintf (dump_file
, "Not considering %s for cloning; "
532 "optimizing it for size.\n",
533 cgraph_node_name (node
));
537 init_caller_stats (&stats
);
538 cgraph_for_node_and_aliases (node
, gather_caller_stats
, &stats
, false);
540 if (inline_summary (node
)->self_size
< stats
.n_calls
)
543 fprintf (dump_file
, "Considering %s for cloning; code might shrink.\n",
544 cgraph_node_name (node
));
548 /* When profile is available and function is hot, propagate into it even if
549 calls seems cold; constant propagation can improve function's speed
553 if (stats
.count_sum
> node
->count
* 90 / 100)
556 fprintf (dump_file
, "Considering %s for cloning; "
557 "usually called directly.\n",
558 cgraph_node_name (node
));
562 if (!stats
.n_hot_calls
)
565 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
566 cgraph_node_name (node
));
570 fprintf (dump_file
, "Considering %s for cloning.\n",
571 cgraph_node_name (node
));
575 /* Arrays representing a topological ordering of call graph nodes and a stack
576 of noes used during constant propagation. */
580 struct cgraph_node
**order
;
581 struct cgraph_node
**stack
;
582 int nnodes
, stack_top
;
585 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
588 build_toporder_info (struct topo_info
*topo
)
590 topo
->order
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
591 topo
->stack
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
593 topo
->nnodes
= ipa_reduced_postorder (topo
->order
, true, true, NULL
);
596 /* Free information about strongly connected components and the arrays in
600 free_toporder_info (struct topo_info
*topo
)
602 ipa_free_postorder_info ();
607 /* Add NODE to the stack in TOPO, unless it is already there. */
610 push_node_to_stack (struct topo_info
*topo
, struct cgraph_node
*node
)
612 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
613 if (info
->node_enqueued
)
615 info
->node_enqueued
= 1;
616 topo
->stack
[topo
->stack_top
++] = node
;
619 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
622 static struct cgraph_node
*
623 pop_node_from_stack (struct topo_info
*topo
)
627 struct cgraph_node
*node
;
629 node
= topo
->stack
[topo
->stack_top
];
630 IPA_NODE_REF (node
)->node_enqueued
= 0;
637 /* Set lattice LAT to bottom and return true if it previously was not set as
641 set_lattice_to_bottom (struct ipcp_lattice
*lat
)
643 bool ret
= !lat
->bottom
;
648 /* Mark lattice as containing an unknown value and return true if it previously
649 was not marked as such. */
652 set_lattice_contains_variable (struct ipcp_lattice
*lat
)
654 bool ret
= !lat
->contains_variable
;
655 lat
->contains_variable
= true;
659 /* Set all aggegate lattices in PLATS to bottom and return true if they were
660 not previously set as such. */
663 set_agg_lats_to_bottom (struct ipcp_param_lattices
*plats
)
665 bool ret
= !plats
->aggs_bottom
;
666 plats
->aggs_bottom
= true;
670 /* Mark all aggegate lattices in PLATS as containing an unknown value and
671 return true if they were not previously marked as such. */
674 set_agg_lats_contain_variable (struct ipcp_param_lattices
*plats
)
676 bool ret
= !plats
->aggs_contain_variable
;
677 plats
->aggs_contain_variable
= true;
681 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
682 return true is any of them has not been marked as such so far. */
685 set_all_contains_variable (struct ipcp_param_lattices
*plats
)
687 bool ret
= !plats
->itself
.contains_variable
|| !plats
->aggs_contain_variable
;
688 plats
->itself
.contains_variable
= true;
689 plats
->aggs_contain_variable
= true;
693 /* Initialize ipcp_lattices. */
696 initialize_node_lattices (struct cgraph_node
*node
)
698 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
699 struct cgraph_edge
*ie
;
700 bool disable
= false, variable
= false;
703 gcc_checking_assert (cgraph_function_with_gimple_body_p (node
));
704 if (!node
->local
.local
)
706 /* When cloning is allowed, we can assume that externally visible
707 functions are not called. We will compensate this by cloning
709 if (ipcp_versionable_function_p (node
)
710 && ipcp_cloning_candidate_p (node
))
716 if (disable
|| variable
)
718 for (i
= 0; i
< ipa_get_param_count (info
) ; i
++)
720 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
723 set_lattice_to_bottom (&plats
->itself
);
724 set_agg_lats_to_bottom (plats
);
727 set_all_contains_variable (plats
);
729 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
730 && !node
->symbol
.alias
&& !node
->thunk
.thunk_p
)
731 fprintf (dump_file
, "Marking all lattices of %s/%i as %s\n",
732 cgraph_node_name (node
), node
->symbol
.order
,
733 disable
? "BOTTOM" : "VARIABLE");
736 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
737 if (ie
->indirect_info
->polymorphic
738 && ie
->indirect_info
->param_index
>= 0)
740 gcc_checking_assert (ie
->indirect_info
->param_index
>= 0);
741 ipa_get_parm_lattices (info
,
742 ie
->indirect_info
->param_index
)->virt_call
= 1;
746 /* Return the result of a (possibly arithmetic) pass through jump function
747 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
748 determined or itself is considered an interprocedural invariant. */
751 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
)
755 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
757 else if (TREE_CODE (input
) == TREE_BINFO
)
760 gcc_checking_assert (is_gimple_ip_invariant (input
));
761 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc
))
763 restype
= boolean_type_node
;
765 restype
= TREE_TYPE (input
);
766 res
= fold_binary (ipa_get_jf_pass_through_operation (jfunc
), restype
,
767 input
, ipa_get_jf_pass_through_operand (jfunc
));
769 if (res
&& !is_gimple_ip_invariant (res
))
775 /* Return the result of an ancestor jump function JFUNC on the constant value
776 INPUT. Return NULL_TREE if that cannot be determined. */
779 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
781 if (TREE_CODE (input
) == TREE_BINFO
)
782 return get_binfo_at_offset (input
,
783 ipa_get_jf_ancestor_offset (jfunc
),
784 ipa_get_jf_ancestor_type (jfunc
));
785 else if (TREE_CODE (input
) == ADDR_EXPR
)
787 tree t
= TREE_OPERAND (input
, 0);
788 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
789 ipa_get_jf_ancestor_offset (jfunc
),
790 ipa_get_jf_ancestor_type (jfunc
), NULL
, false);
791 return build_fold_addr_expr (t
);
797 /* Determine whether JFUNC evaluates to a known value (that is either a
798 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
799 describes the caller node so that pass-through jump functions can be
803 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
)
805 if (jfunc
->type
== IPA_JF_CONST
)
806 return ipa_get_jf_constant (jfunc
);
807 else if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
808 return ipa_binfo_from_known_type_jfunc (jfunc
);
809 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
810 || jfunc
->type
== IPA_JF_ANCESTOR
)
815 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
816 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
818 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
820 if (info
->ipcp_orig_node
)
821 input
= info
->known_vals
[idx
];
824 struct ipcp_lattice
*lat
;
828 gcc_checking_assert (!flag_ipa_cp
);
831 lat
= ipa_get_scalar_lat (info
, idx
);
832 if (!ipa_lat_is_single_const (lat
))
834 input
= lat
->values
->value
;
840 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
841 return ipa_get_jf_pass_through_result (jfunc
, input
);
843 return ipa_get_jf_ancestor_result (jfunc
, input
);
850 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
851 bottom, not containing a variable component and without any known value at
855 ipcp_verify_propagated_values (void)
857 struct cgraph_node
*node
;
859 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
861 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
862 int i
, count
= ipa_get_param_count (info
);
864 for (i
= 0; i
< count
; i
++)
866 struct ipcp_lattice
*lat
= ipa_get_scalar_lat (info
, i
);
869 && !lat
->contains_variable
870 && lat
->values_count
== 0)
874 fprintf (dump_file
, "\nIPA lattices after constant "
876 print_all_lattices (dump_file
, true, false);
885 /* Return true iff X and Y should be considered equal values by IPA-CP. */
888 values_equal_for_ipcp_p (tree x
, tree y
)
890 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
895 if (TREE_CODE (x
) == TREE_BINFO
|| TREE_CODE (y
) == TREE_BINFO
)
898 if (TREE_CODE (x
) == ADDR_EXPR
899 && TREE_CODE (y
) == ADDR_EXPR
900 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
901 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
902 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
903 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
905 return operand_equal_p (x
, y
, 0);
908 /* Add a new value source to VAL, marking that a value comes from edge CS and
909 (if the underlying jump function is a pass-through or an ancestor one) from
910 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
911 is negative if the source was the scalar value of the parameter itself or
912 the offset within an aggregate. */
915 add_value_source (struct ipcp_value
*val
, struct cgraph_edge
*cs
,
916 struct ipcp_value
*src_val
, int src_idx
, HOST_WIDE_INT offset
)
918 struct ipcp_value_source
*src
;
920 src
= (struct ipcp_value_source
*) pool_alloc (ipcp_sources_pool
);
921 src
->offset
= offset
;
924 src
->index
= src_idx
;
926 src
->next
= val
->sources
;
930 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
931 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
932 have the same meaning. */
935 add_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
936 struct cgraph_edge
*cs
, struct ipcp_value
*src_val
,
937 int src_idx
, HOST_WIDE_INT offset
)
939 struct ipcp_value
*val
;
944 for (val
= lat
->values
; val
; val
= val
->next
)
945 if (values_equal_for_ipcp_p (val
->value
, newval
))
947 if (edge_within_scc (cs
))
949 struct ipcp_value_source
*s
;
950 for (s
= val
->sources
; s
; s
= s
->next
)
957 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
961 if (lat
->values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
963 /* We can only free sources, not the values themselves, because sources
964 of other values in this this SCC might point to them. */
965 for (val
= lat
->values
; val
; val
= val
->next
)
969 struct ipcp_value_source
*src
= val
->sources
;
970 val
->sources
= src
->next
;
971 pool_free (ipcp_sources_pool
, src
);
976 return set_lattice_to_bottom (lat
);
980 val
= (struct ipcp_value
*) pool_alloc (ipcp_values_pool
);
981 memset (val
, 0, sizeof (*val
));
983 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
985 val
->next
= lat
->values
;
990 /* Like above but passes a special value of offset to distinguish that the
991 origin is the scalar value of the parameter rather than a part of an
995 add_scalar_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
996 struct cgraph_edge
*cs
,
997 struct ipcp_value
*src_val
, int src_idx
)
999 return add_value_to_lattice (lat
, newval
, cs
, src_val
, src_idx
, -1);
1002 /* Propagate values through a pass-through jump function JFUNC associated with
1003 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1004 is the index of the source parameter. */
1007 propagate_vals_accross_pass_through (struct cgraph_edge
*cs
,
1008 struct ipa_jump_func
*jfunc
,
1009 struct ipcp_lattice
*src_lat
,
1010 struct ipcp_lattice
*dest_lat
,
1013 struct ipcp_value
*src_val
;
1016 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1017 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1018 ret
|= add_scalar_value_to_lattice (dest_lat
, src_val
->value
, cs
,
1020 /* Do not create new values when propagating within an SCC because if there
1021 are arithmetic functions with circular dependencies, there is infinite
1022 number of them and we would just make lattices bottom. */
1023 else if (edge_within_scc (cs
))
1024 ret
= set_lattice_contains_variable (dest_lat
);
1026 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1028 tree cstval
= src_val
->value
;
1030 if (TREE_CODE (cstval
) == TREE_BINFO
)
1032 ret
|= set_lattice_contains_variable (dest_lat
);
1035 cstval
= ipa_get_jf_pass_through_result (jfunc
, cstval
);
1038 ret
|= add_scalar_value_to_lattice (dest_lat
, cstval
, cs
, src_val
,
1041 ret
|= set_lattice_contains_variable (dest_lat
);
1047 /* Propagate values through an ancestor jump function JFUNC associated with
1048 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1049 is the index of the source parameter. */
1052 propagate_vals_accross_ancestor (struct cgraph_edge
*cs
,
1053 struct ipa_jump_func
*jfunc
,
1054 struct ipcp_lattice
*src_lat
,
1055 struct ipcp_lattice
*dest_lat
,
1058 struct ipcp_value
*src_val
;
1061 if (edge_within_scc (cs
))
1062 return set_lattice_contains_variable (dest_lat
);
1064 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1066 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1069 ret
|= add_scalar_value_to_lattice (dest_lat
, t
, cs
, src_val
, src_idx
);
1071 ret
|= set_lattice_contains_variable (dest_lat
);
1077 /* Propagate scalar values across jump function JFUNC that is associated with
1078 edge CS and put the values into DEST_LAT. */
1081 propagate_scalar_accross_jump_function (struct cgraph_edge
*cs
,
1082 struct ipa_jump_func
*jfunc
,
1083 struct ipcp_lattice
*dest_lat
)
1085 if (dest_lat
->bottom
)
1088 if (jfunc
->type
== IPA_JF_CONST
1089 || jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1093 if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1095 val
= ipa_binfo_from_known_type_jfunc (jfunc
);
1097 return set_lattice_contains_variable (dest_lat
);
1100 val
= ipa_get_jf_constant (jfunc
);
1101 return add_scalar_value_to_lattice (dest_lat
, val
, cs
, NULL
, 0);
1103 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1104 || jfunc
->type
== IPA_JF_ANCESTOR
)
1106 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1107 struct ipcp_lattice
*src_lat
;
1111 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1112 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1114 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1116 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1117 if (src_lat
->bottom
)
1118 return set_lattice_contains_variable (dest_lat
);
1120 /* If we would need to clone the caller and cannot, do not propagate. */
1121 if (!ipcp_versionable_function_p (cs
->caller
)
1122 && (src_lat
->contains_variable
1123 || (src_lat
->values_count
> 1)))
1124 return set_lattice_contains_variable (dest_lat
);
1126 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1127 ret
= propagate_vals_accross_pass_through (cs
, jfunc
, src_lat
,
1130 ret
= propagate_vals_accross_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1133 if (src_lat
->contains_variable
)
1134 ret
|= set_lattice_contains_variable (dest_lat
);
1139 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1140 use it for indirect inlining), we should propagate them too. */
1141 return set_lattice_contains_variable (dest_lat
);
1144 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1145 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1146 other cases, return false). If there are no aggregate items, set
1147 aggs_by_ref to NEW_AGGS_BY_REF. */
1150 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1151 bool new_aggs_by_ref
)
1153 if (dest_plats
->aggs
)
1155 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1157 set_agg_lats_to_bottom (dest_plats
);
1162 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1166 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1167 already existing lattice for the given OFFSET and SIZE, marking all skipped
1168 lattices as containing variable and checking for overlaps. If there is no
1169 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1170 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1171 unless there are too many already. If there are two many, return false. If
1172 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1173 skipped lattices were newly marked as containing variable, set *CHANGE to
1177 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
1178 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
1179 struct ipcp_agg_lattice
***aglat
,
1180 bool pre_existing
, bool *change
)
1182 gcc_checking_assert (offset
>= 0);
1184 while (**aglat
&& (**aglat
)->offset
< offset
)
1186 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
1188 set_agg_lats_to_bottom (dest_plats
);
1191 *change
|= set_lattice_contains_variable (**aglat
);
1192 *aglat
= &(**aglat
)->next
;
1195 if (**aglat
&& (**aglat
)->offset
== offset
)
1197 if ((**aglat
)->size
!= val_size
1199 && (**aglat
)->next
->offset
< offset
+ val_size
))
1201 set_agg_lats_to_bottom (dest_plats
);
1204 gcc_checking_assert (!(**aglat
)->next
1205 || (**aglat
)->next
->offset
>= offset
+ val_size
);
1210 struct ipcp_agg_lattice
*new_al
;
1212 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
1214 set_agg_lats_to_bottom (dest_plats
);
1217 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
1219 dest_plats
->aggs_count
++;
1220 new_al
= (struct ipcp_agg_lattice
*) pool_alloc (ipcp_agg_lattice_pool
);
1221 memset (new_al
, 0, sizeof (*new_al
));
1223 new_al
->offset
= offset
;
1224 new_al
->size
= val_size
;
1225 new_al
->contains_variable
= pre_existing
;
1227 new_al
->next
= **aglat
;
1233 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1234 containing an unknown value. */
1237 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
1242 ret
|= set_lattice_contains_variable (aglat
);
1243 aglat
= aglat
->next
;
1248 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1249 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1250 parameter used for lattice value sources. Return true if DEST_PLATS changed
1254 merge_aggregate_lattices (struct cgraph_edge
*cs
,
1255 struct ipcp_param_lattices
*dest_plats
,
1256 struct ipcp_param_lattices
*src_plats
,
1257 int src_idx
, HOST_WIDE_INT offset_delta
)
1259 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1260 struct ipcp_agg_lattice
**dst_aglat
;
1263 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
1265 if (src_plats
->aggs_bottom
)
1266 return set_agg_lats_contain_variable (dest_plats
);
1267 if (src_plats
->aggs_contain_variable
)
1268 ret
|= set_agg_lats_contain_variable (dest_plats
);
1269 dst_aglat
= &dest_plats
->aggs
;
1271 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
1273 src_aglat
= src_aglat
->next
)
1275 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
1279 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
1280 &dst_aglat
, pre_existing
, &ret
))
1282 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
1284 dst_aglat
= &(*dst_aglat
)->next
;
1285 if (src_aglat
->bottom
)
1287 ret
|= set_lattice_contains_variable (new_al
);
1290 if (src_aglat
->contains_variable
)
1291 ret
|= set_lattice_contains_variable (new_al
);
1292 for (struct ipcp_value
*val
= src_aglat
->values
;
1295 ret
|= add_value_to_lattice (new_al
, val
->value
, cs
, val
, src_idx
,
1298 else if (dest_plats
->aggs_bottom
)
1301 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
1305 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1306 pass-through JFUNC and if so, whether it has conform and conforms to the
1307 rules about propagating values passed by reference. */
1310 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
1311 struct ipa_jump_func
*jfunc
)
1313 return src_plats
->aggs
1314 && (!src_plats
->aggs_by_ref
1315 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
1318 /* Propagate scalar values across jump function JFUNC that is associated with
1319 edge CS and put the values into DEST_LAT. */
1322 propagate_aggs_accross_jump_function (struct cgraph_edge
*cs
,
1323 struct ipa_jump_func
*jfunc
,
1324 struct ipcp_param_lattices
*dest_plats
)
1328 if (dest_plats
->aggs_bottom
)
1331 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1332 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1334 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1335 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1336 struct ipcp_param_lattices
*src_plats
;
1338 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1339 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
1341 /* Currently we do not produce clobber aggregate jump
1342 functions, replace with merging when we do. */
1343 gcc_assert (!jfunc
->agg
.items
);
1344 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
1348 ret
|= set_agg_lats_contain_variable (dest_plats
);
1350 else if (jfunc
->type
== IPA_JF_ANCESTOR
1351 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
1353 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1354 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1355 struct ipcp_param_lattices
*src_plats
;
1357 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1358 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
1360 /* Currently we do not produce clobber aggregate jump
1361 functions, replace with merging when we do. */
1362 gcc_assert (!jfunc
->agg
.items
);
1363 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
1364 ipa_get_jf_ancestor_offset (jfunc
));
1366 else if (!src_plats
->aggs_by_ref
)
1367 ret
|= set_agg_lats_to_bottom (dest_plats
);
1369 ret
|= set_agg_lats_contain_variable (dest_plats
);
1371 else if (jfunc
->agg
.items
)
1373 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1374 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
1375 struct ipa_agg_jf_item
*item
;
1378 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
1381 FOR_EACH_VEC_ELT (*jfunc
->agg
.items
, i
, item
)
1383 HOST_WIDE_INT val_size
;
1385 if (item
->offset
< 0)
1387 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
1388 val_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (item
->value
)), 1);
1390 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
1391 &aglat
, pre_existing
, &ret
))
1393 ret
|= add_value_to_lattice (*aglat
, item
->value
, cs
, NULL
, 0, 0);
1394 aglat
= &(*aglat
)->next
;
1396 else if (dest_plats
->aggs_bottom
)
1400 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
1403 ret
|= set_agg_lats_contain_variable (dest_plats
);
1408 /* Propagate constants from the caller to the callee of CS. INFO describes the
1412 propagate_constants_accross_call (struct cgraph_edge
*cs
)
1414 struct ipa_node_params
*callee_info
;
1415 enum availability availability
;
1416 struct cgraph_node
*callee
, *alias_or_thunk
;
1417 struct ipa_edge_args
*args
;
1419 int i
, args_count
, parms_count
;
1421 callee
= cgraph_function_node (cs
->callee
, &availability
);
1422 if (!callee
->symbol
.definition
)
1424 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee
));
1425 callee_info
= IPA_NODE_REF (callee
);
1427 args
= IPA_EDGE_REF (cs
);
1428 args_count
= ipa_get_cs_argument_count (args
);
1429 parms_count
= ipa_get_param_count (callee_info
);
1431 /* If this call goes through a thunk we must not propagate to the first (0th)
1432 parameter. However, we might need to uncover a thunk from below a series
1433 of aliases first. */
1434 alias_or_thunk
= cs
->callee
;
1435 while (alias_or_thunk
->symbol
.alias
)
1436 alias_or_thunk
= cgraph_alias_target (alias_or_thunk
);
1437 if (alias_or_thunk
->thunk
.thunk_p
)
1439 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
1446 for (; (i
< args_count
) && (i
< parms_count
); i
++)
1448 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
1449 struct ipcp_param_lattices
*dest_plats
;
1451 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
1452 if (availability
== AVAIL_OVERWRITABLE
)
1453 ret
|= set_all_contains_variable (dest_plats
);
1456 ret
|= propagate_scalar_accross_jump_function (cs
, jump_func
,
1457 &dest_plats
->itself
);
1458 ret
|= propagate_aggs_accross_jump_function (cs
, jump_func
,
1462 for (; i
< parms_count
; i
++)
1463 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
1468 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1469 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1470 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1471 is not NULL, KNOWN_AGGS is ignored. */
1474 ipa_get_indirect_edge_target_1 (struct cgraph_edge
*ie
,
1475 vec
<tree
> known_vals
,
1476 vec
<tree
> known_binfos
,
1477 vec
<ipa_agg_jump_function_p
> known_aggs
,
1478 struct ipa_agg_replacement_value
*agg_reps
)
1480 int param_index
= ie
->indirect_info
->param_index
;
1481 HOST_WIDE_INT token
, anc_offset
;
1485 if (param_index
== -1
1486 || known_vals
.length () <= (unsigned int) param_index
)
1489 if (!ie
->indirect_info
->polymorphic
)
1493 if (ie
->indirect_info
->agg_contents
)
1500 if (agg_reps
->index
== param_index
1501 && agg_reps
->offset
== ie
->indirect_info
->offset
1502 && agg_reps
->by_ref
== ie
->indirect_info
->by_ref
)
1504 t
= agg_reps
->value
;
1507 agg_reps
= agg_reps
->next
;
1510 else if (known_aggs
.length () > (unsigned int) param_index
)
1512 struct ipa_agg_jump_function
*agg
;
1513 agg
= known_aggs
[param_index
];
1514 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1515 ie
->indirect_info
->by_ref
);
1521 t
= known_vals
[param_index
];
1524 TREE_CODE (t
) == ADDR_EXPR
1525 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
1526 return TREE_OPERAND (t
, 0);
1531 gcc_assert (!ie
->indirect_info
->agg_contents
);
1532 token
= ie
->indirect_info
->otr_token
;
1533 anc_offset
= ie
->indirect_info
->offset
;
1534 otr_type
= ie
->indirect_info
->otr_type
;
1536 t
= known_vals
[param_index
];
1537 if (!t
&& known_binfos
.length () > (unsigned int) param_index
)
1538 t
= known_binfos
[param_index
];
1542 if (TREE_CODE (t
) != TREE_BINFO
)
1545 binfo
= gimple_extract_devirt_binfo_from_cst
1546 (t
, ie
->indirect_info
->otr_type
);
1549 binfo
= get_binfo_at_offset (binfo
, anc_offset
, otr_type
);
1552 return gimple_get_virt_method_for_binfo (token
, binfo
);
1558 binfo
= get_binfo_at_offset (t
, anc_offset
, otr_type
);
1561 return gimple_get_virt_method_for_binfo (token
, binfo
);
1566 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1567 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1568 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1571 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
1572 vec
<tree
> known_vals
,
1573 vec
<tree
> known_binfos
,
1574 vec
<ipa_agg_jump_function_p
> known_aggs
)
1576 return ipa_get_indirect_edge_target_1 (ie
, known_vals
, known_binfos
,
1580 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1581 and KNOWN_BINFOS. */
1584 devirtualization_time_bonus (struct cgraph_node
*node
,
1585 vec
<tree
> known_csts
,
1586 vec
<tree
> known_binfos
,
1587 vec
<ipa_agg_jump_function_p
> known_aggs
)
1589 struct cgraph_edge
*ie
;
1592 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1594 struct cgraph_node
*callee
;
1595 struct inline_summary
*isummary
;
1598 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_binfos
,
1603 /* Only bare minimum benefit for clearly un-inlineable targets. */
1605 callee
= cgraph_get_node (target
);
1606 if (!callee
|| !callee
->symbol
.definition
)
1608 isummary
= inline_summary (callee
);
1609 if (!isummary
->inlinable
)
1612 /* FIXME: The values below need re-considering and perhaps also
1613 integrating into the cost metrics, at lest in some very basic way. */
1614 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
1616 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
1618 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
1619 || DECL_DECLARED_INLINE_P (callee
->symbol
.decl
))
1626 /* Return time bonus incurred because of HINTS. */
1629 hint_time_bonus (inline_hints hints
)
1632 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
1633 result
+= PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
1634 if (hints
& INLINE_HINT_array_index
)
1635 result
+= PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS
);
1639 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1640 and SIZE_COST and with the sum of frequencies of incoming edges to the
1641 potential new clone in FREQUENCIES. */
1644 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
1645 int freq_sum
, gcov_type count_sum
, int size_cost
)
1647 if (time_benefit
== 0
1648 || !flag_ipa_cp_clone
1649 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
)))
1652 gcc_assert (size_cost
> 0);
1656 int factor
= (count_sum
* 1000) / max_count
;
1657 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* factor
)
1660 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1661 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1662 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1663 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1664 ", threshold: %i\n",
1665 time_benefit
, size_cost
, (HOST_WIDE_INT
) count_sum
,
1666 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1668 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1672 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* freq_sum
)
1675 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1676 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1677 "size: %i, freq_sum: %i) -> evaluation: "
1678 HOST_WIDEST_INT_PRINT_DEC
", threshold: %i\n",
1679 time_benefit
, size_cost
, freq_sum
, evaluation
,
1680 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1682 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1686 /* Return all context independent values from aggregate lattices in PLATS in a
1687 vector. Return NULL if there are none. */
1689 static vec
<ipa_agg_jf_item_t
, va_gc
> *
1690 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
1692 vec
<ipa_agg_jf_item_t
, va_gc
> *res
= NULL
;
1694 if (plats
->aggs_bottom
1695 || plats
->aggs_contain_variable
1696 || plats
->aggs_count
== 0)
1699 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
1701 aglat
= aglat
->next
)
1702 if (ipa_lat_is_single_const (aglat
))
1704 struct ipa_agg_jf_item item
;
1705 item
.offset
= aglat
->offset
;
1706 item
.value
= aglat
->values
->value
;
1707 vec_safe_push (res
, item
);
1712 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1713 them with values of parameters that are known independent of the context.
1714 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1715 movement cost of all removable parameters will be stored in it. */
1718 gather_context_independent_values (struct ipa_node_params
*info
,
1719 vec
<tree
> *known_csts
,
1720 vec
<tree
> *known_binfos
,
1721 vec
<ipa_agg_jump_function_t
> *known_aggs
,
1722 int *removable_params_cost
)
1724 int i
, count
= ipa_get_param_count (info
);
1727 known_csts
->create (0);
1728 known_binfos
->create (0);
1729 known_csts
->safe_grow_cleared (count
);
1730 known_binfos
->safe_grow_cleared (count
);
1733 known_aggs
->create (0);
1734 known_aggs
->safe_grow_cleared (count
);
1737 if (removable_params_cost
)
1738 *removable_params_cost
= 0;
1740 for (i
= 0; i
< count
; i
++)
1742 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1743 struct ipcp_lattice
*lat
= &plats
->itself
;
1745 if (ipa_lat_is_single_const (lat
))
1747 struct ipcp_value
*val
= lat
->values
;
1748 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1750 (*known_csts
)[i
] = val
->value
;
1751 if (removable_params_cost
)
1752 *removable_params_cost
1753 += estimate_move_cost (TREE_TYPE (val
->value
));
1756 else if (plats
->virt_call
)
1758 (*known_binfos
)[i
] = val
->value
;
1761 else if (removable_params_cost
1762 && !ipa_is_param_used (info
, i
))
1763 *removable_params_cost
+= ipa_get_param_move_cost (info
, i
);
1765 else if (removable_params_cost
1766 && !ipa_is_param_used (info
, i
))
1767 *removable_params_cost
1768 += ipa_get_param_move_cost (info
, i
);
1772 vec
<ipa_agg_jf_item_t
, va_gc
> *agg_items
;
1773 struct ipa_agg_jump_function
*ajf
;
1775 agg_items
= context_independent_aggregate_values (plats
);
1776 ajf
= &(*known_aggs
)[i
];
1777 ajf
->items
= agg_items
;
1778 ajf
->by_ref
= plats
->aggs_by_ref
;
1779 ret
|= agg_items
!= NULL
;
1786 /* The current interface in ipa-inline-analysis requires a pointer vector.
1789 FIXME: That interface should be re-worked, this is slightly silly. Still,
1790 I'd like to discuss how to change it first and this demonstrates the
1793 static vec
<ipa_agg_jump_function_p
>
1794 agg_jmp_p_vec_for_t_vec (vec
<ipa_agg_jump_function_t
> known_aggs
)
1796 vec
<ipa_agg_jump_function_p
> ret
;
1797 struct ipa_agg_jump_function
*ajf
;
1800 ret
.create (known_aggs
.length ());
1801 FOR_EACH_VEC_ELT (known_aggs
, i
, ajf
)
1802 ret
.quick_push (ajf
);
1806 /* Iterate over known values of parameters of NODE and estimate the local
1807 effects in terms of time and size they have. */
1810 estimate_local_effects (struct cgraph_node
*node
)
1812 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1813 int i
, count
= ipa_get_param_count (info
);
1814 vec
<tree
> known_csts
, known_binfos
;
1815 vec
<ipa_agg_jump_function_t
> known_aggs
;
1816 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
;
1818 int base_time
= inline_summary (node
)->time
;
1819 int removable_params_cost
;
1821 if (!count
|| !ipcp_versionable_function_p (node
))
1824 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1825 fprintf (dump_file
, "\nEstimating effects for %s/%i, base_time: %i.\n",
1826 cgraph_node_name (node
), node
->symbol
.order
, base_time
);
1828 always_const
= gather_context_independent_values (info
, &known_csts
,
1829 &known_binfos
, &known_aggs
,
1830 &removable_params_cost
);
1831 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
1834 struct caller_statistics stats
;
1838 init_caller_stats (&stats
);
1839 cgraph_for_node_and_aliases (node
, gather_caller_stats
, &stats
, false);
1840 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1841 known_aggs_ptrs
, &size
, &time
, &hints
);
1842 time
-= devirtualization_time_bonus (node
, known_csts
, known_binfos
,
1844 time
-= hint_time_bonus (hints
);
1845 time
-= removable_params_cost
;
1846 size
-= stats
.n_calls
* removable_params_cost
;
1849 fprintf (dump_file
, " - context independent values, size: %i, "
1850 "time_benefit: %i\n", size
, base_time
- time
);
1853 || cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
1855 info
->do_clone_for_all_contexts
= true;
1859 fprintf (dump_file
, " Decided to specialize for all "
1860 "known contexts, code not going to grow.\n");
1862 else if (good_cloning_opportunity_p (node
, base_time
- time
,
1863 stats
.freq_sum
, stats
.count_sum
,
1866 if (size
+ overall_size
<= max_new_size
)
1868 info
->do_clone_for_all_contexts
= true;
1870 overall_size
+= size
;
1873 fprintf (dump_file
, " Decided to specialize for all "
1874 "known contexts, growth deemed beneficial.\n");
1876 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1877 fprintf (dump_file
, " Not cloning for all contexts because "
1878 "max_new_size would be reached with %li.\n",
1879 size
+ overall_size
);
1883 for (i
= 0; i
< count
; i
++)
1885 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1886 struct ipcp_lattice
*lat
= &plats
->itself
;
1887 struct ipcp_value
*val
;
1896 for (val
= lat
->values
; val
; val
= val
->next
)
1898 int time
, size
, time_benefit
;
1901 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1903 known_csts
[i
] = val
->value
;
1904 known_binfos
[i
] = NULL_TREE
;
1905 emc
= estimate_move_cost (TREE_TYPE (val
->value
));
1907 else if (plats
->virt_call
)
1909 known_csts
[i
] = NULL_TREE
;
1910 known_binfos
[i
] = val
->value
;
1916 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1917 known_aggs_ptrs
, &size
, &time
,
1919 time_benefit
= base_time
- time
1920 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
1922 + hint_time_bonus (hints
)
1923 + removable_params_cost
+ emc
;
1925 gcc_checking_assert (size
>=0);
1926 /* The inliner-heuristics based estimates may think that in certain
1927 contexts some functions do not have any size at all but we want
1928 all specializations to have at least a tiny cost, not least not to
1933 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1935 fprintf (dump_file
, " - estimates for value ");
1936 print_ipcp_constant_value (dump_file
, val
->value
);
1937 fprintf (dump_file
, " for ");
1938 ipa_dump_param (dump_file
, info
, i
);
1939 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
1940 time_benefit
, size
);
1943 val
->local_time_benefit
= time_benefit
;
1944 val
->local_size_cost
= size
;
1946 known_binfos
[i
] = NULL_TREE
;
1947 known_csts
[i
] = NULL_TREE
;
1950 for (i
= 0; i
< count
; i
++)
1952 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1953 struct ipa_agg_jump_function
*ajf
;
1954 struct ipcp_agg_lattice
*aglat
;
1956 if (plats
->aggs_bottom
|| !plats
->aggs
)
1959 ajf
= &known_aggs
[i
];
1960 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
1962 struct ipcp_value
*val
;
1963 if (aglat
->bottom
|| !aglat
->values
1964 /* If the following is true, the one value is in known_aggs. */
1965 || (!plats
->aggs_contain_variable
1966 && ipa_lat_is_single_const (aglat
)))
1969 for (val
= aglat
->values
; val
; val
= val
->next
)
1971 int time
, size
, time_benefit
;
1972 struct ipa_agg_jf_item item
;
1975 item
.offset
= aglat
->offset
;
1976 item
.value
= val
->value
;
1977 vec_safe_push (ajf
->items
, item
);
1979 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1980 known_aggs_ptrs
, &size
, &time
,
1982 time_benefit
= base_time
- time
1983 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
1985 + hint_time_bonus (hints
);
1986 gcc_checking_assert (size
>=0);
1990 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1992 fprintf (dump_file
, " - estimates for value ");
1993 print_ipcp_constant_value (dump_file
, val
->value
);
1994 fprintf (dump_file
, " for ");
1995 ipa_dump_param (dump_file
, info
, i
);
1996 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
1997 "]: time_benefit: %i, size: %i\n",
1998 plats
->aggs_by_ref
? "ref " : "",
1999 aglat
->offset
, time_benefit
, size
);
2002 val
->local_time_benefit
= time_benefit
;
2003 val
->local_size_cost
= size
;
2009 for (i
= 0; i
< count
; i
++)
2010 vec_free (known_aggs
[i
].items
);
2012 known_csts
.release ();
2013 known_binfos
.release ();
2014 known_aggs
.release ();
2015 known_aggs_ptrs
.release ();
2019 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2020 topological sort of values. */
2023 add_val_to_toposort (struct ipcp_value
*cur_val
)
2025 static int dfs_counter
= 0;
2026 static struct ipcp_value
*stack
;
2027 struct ipcp_value_source
*src
;
2033 cur_val
->dfs
= dfs_counter
;
2034 cur_val
->low_link
= dfs_counter
;
2036 cur_val
->topo_next
= stack
;
2038 cur_val
->on_stack
= true;
2040 for (src
= cur_val
->sources
; src
; src
= src
->next
)
2043 if (src
->val
->dfs
== 0)
2045 add_val_to_toposort (src
->val
);
2046 if (src
->val
->low_link
< cur_val
->low_link
)
2047 cur_val
->low_link
= src
->val
->low_link
;
2049 else if (src
->val
->on_stack
2050 && src
->val
->dfs
< cur_val
->low_link
)
2051 cur_val
->low_link
= src
->val
->dfs
;
2054 if (cur_val
->dfs
== cur_val
->low_link
)
2056 struct ipcp_value
*v
, *scc_list
= NULL
;
2061 stack
= v
->topo_next
;
2062 v
->on_stack
= false;
2064 v
->scc_next
= scc_list
;
2067 while (v
!= cur_val
);
2069 cur_val
->topo_next
= values_topo
;
2070 values_topo
= cur_val
;
2074 /* Add all values in lattices associated with NODE to the topological sort if
2075 they are not there yet. */
2078 add_all_node_vals_to_toposort (struct cgraph_node
*node
)
2080 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2081 int i
, count
= ipa_get_param_count (info
);
2083 for (i
= 0; i
< count
; i
++)
2085 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2086 struct ipcp_lattice
*lat
= &plats
->itself
;
2087 struct ipcp_agg_lattice
*aglat
;
2088 struct ipcp_value
*val
;
2091 for (val
= lat
->values
; val
; val
= val
->next
)
2092 add_val_to_toposort (val
);
2094 if (!plats
->aggs_bottom
)
2095 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2097 for (val
= aglat
->values
; val
; val
= val
->next
)
2098 add_val_to_toposort (val
);
2102 /* One pass of constants propagation along the call graph edges, from callers
2103 to callees (requires topological ordering in TOPO), iterate over strongly
2104 connected components. */
2107 propagate_constants_topo (struct topo_info
*topo
)
2111 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
2113 struct cgraph_node
*v
, *node
= topo
->order
[i
];
2114 struct ipa_dfs_info
*node_dfs_info
;
2116 if (!cgraph_function_with_gimple_body_p (node
))
2119 node_dfs_info
= (struct ipa_dfs_info
*) node
->symbol
.aux
;
2120 /* First, iteratively propagate within the strongly connected component
2121 until all lattices stabilize. */
2122 v
= node_dfs_info
->next_cycle
;
2125 push_node_to_stack (topo
, v
);
2126 v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
;
2132 struct cgraph_edge
*cs
;
2134 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2135 if (edge_within_scc (cs
)
2136 && propagate_constants_accross_call (cs
))
2137 push_node_to_stack (topo
, cs
->callee
);
2138 v
= pop_node_from_stack (topo
);
2141 /* Afterwards, propagate along edges leading out of the SCC, calculates
2142 the local effects of the discovered constants and all valid values to
2143 their topological sort. */
2147 struct cgraph_edge
*cs
;
2149 estimate_local_effects (v
);
2150 add_all_node_vals_to_toposort (v
);
2151 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2152 if (!edge_within_scc (cs
))
2153 propagate_constants_accross_call (cs
);
2155 v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
;
2161 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2162 the bigger one if otherwise. */
2165 safe_add (int a
, int b
)
2167 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
2168 return a
> b
? a
: b
;
2174 /* Propagate the estimated effects of individual values along the topological
2175 from the dependent values to those they depend on. */
2178 propagate_effects (void)
2180 struct ipcp_value
*base
;
2182 for (base
= values_topo
; base
; base
= base
->topo_next
)
2184 struct ipcp_value_source
*src
;
2185 struct ipcp_value
*val
;
2186 int time
= 0, size
= 0;
2188 for (val
= base
; val
; val
= val
->scc_next
)
2190 time
= safe_add (time
,
2191 val
->local_time_benefit
+ val
->prop_time_benefit
);
2192 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
2195 for (val
= base
; val
; val
= val
->scc_next
)
2196 for (src
= val
->sources
; src
; src
= src
->next
)
2198 && cgraph_maybe_hot_edge_p (src
->cs
))
2200 src
->val
->prop_time_benefit
= safe_add (time
,
2201 src
->val
->prop_time_benefit
);
2202 src
->val
->prop_size_cost
= safe_add (size
,
2203 src
->val
->prop_size_cost
);
2209 /* Propagate constants, binfos and their effects from the summaries
2210 interprocedurally. */
2213 ipcp_propagate_stage (struct topo_info
*topo
)
2215 struct cgraph_node
*node
;
2218 fprintf (dump_file
, "\n Propagating constants:\n\n");
2221 ipa_update_after_lto_read ();
2224 FOR_EACH_DEFINED_FUNCTION (node
)
2226 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2228 determine_versionability (node
);
2229 if (cgraph_function_with_gimple_body_p (node
))
2231 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
2232 ipa_get_param_count (info
));
2233 initialize_node_lattices (node
);
2235 if (node
->symbol
.definition
&& !node
->symbol
.alias
)
2236 overall_size
+= inline_summary (node
)->self_size
;
2237 if (node
->count
> max_count
)
2238 max_count
= node
->count
;
2241 max_new_size
= overall_size
;
2242 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
2243 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
2244 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
2247 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
2248 overall_size
, max_new_size
);
2250 propagate_constants_topo (topo
);
2251 #ifdef ENABLE_CHECKING
2252 ipcp_verify_propagated_values ();
2254 propagate_effects ();
2258 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
2259 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
2263 /* Discover newly direct outgoing edges from NODE which is a new clone with
2264 known KNOWN_VALS and make them direct. */
2267 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
2268 vec
<tree
> known_vals
,
2269 struct ipa_agg_replacement_value
*aggvals
)
2271 struct cgraph_edge
*ie
, *next_ie
;
2274 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
2278 next_ie
= ie
->next_callee
;
2279 target
= ipa_get_indirect_edge_target_1 (ie
, known_vals
, vNULL
, vNULL
,
2283 bool agg_contents
= ie
->indirect_info
->agg_contents
;
2284 bool polymorphic
= ie
->indirect_info
->polymorphic
;
2285 bool param_index
= ie
->indirect_info
->param_index
;
2286 struct cgraph_edge
*cs
= ipa_make_edge_direct_to_target (ie
, target
);
2289 if (cs
&& !agg_contents
&& !polymorphic
)
2291 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2292 int c
= ipa_get_controlled_uses (info
, param_index
);
2293 if (c
!= IPA_UNDESCRIBED_USE
)
2295 struct ipa_ref
*to_del
;
2298 ipa_set_controlled_uses (info
, param_index
, c
);
2299 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2300 fprintf (dump_file
, " controlled uses count of param "
2301 "%i bumped down to %i\n", param_index
, c
);
2303 && (to_del
= ipa_find_reference ((symtab_node
) node
,
2304 (symtab_node
) cs
->callee
,
2307 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2308 fprintf (dump_file
, " and even removing its "
2309 "cloning-created reference\n");
2310 ipa_remove_reference (to_del
);
2316 /* Turning calls to direct calls will improve overall summary. */
2318 inline_update_overall_summary (node
);
2321 /* Vector of pointers which for linked lists of clones of an original crgaph
2324 static vec
<cgraph_edge_p
> next_edge_clone
;
2327 grow_next_edge_clone_vector (void)
2329 if (next_edge_clone
.length ()
2330 <= (unsigned) cgraph_edge_max_uid
)
2331 next_edge_clone
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
2334 /* Edge duplication hook to grow the appropriate linked list in
2338 ipcp_edge_duplication_hook (struct cgraph_edge
*src
, struct cgraph_edge
*dst
,
2339 __attribute__((unused
)) void *data
)
2341 grow_next_edge_clone_vector ();
2342 next_edge_clone
[dst
->uid
] = next_edge_clone
[src
->uid
];
2343 next_edge_clone
[src
->uid
] = dst
;
2346 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2347 parameter with the given INDEX. */
2350 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDEST_INT offset
,
2353 struct ipa_agg_replacement_value
*aggval
;
2355 aggval
= ipa_get_agg_replacements_for_node (node
);
2358 if (aggval
->offset
== offset
2359 && aggval
->index
== index
)
2360 return aggval
->value
;
2361 aggval
= aggval
->next
;
2366 /* Return true if edge CS does bring about the value described by SRC. */
2369 cgraph_edge_brings_value_p (struct cgraph_edge
*cs
,
2370 struct ipcp_value_source
*src
)
2372 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2373 struct ipa_node_params
*dst_info
= IPA_NODE_REF (cs
->callee
);
2375 if ((dst_info
->ipcp_orig_node
&& !dst_info
->is_all_contexts_clone
)
2376 || caller_info
->node_dead
)
2381 if (caller_info
->ipcp_orig_node
)
2384 if (src
->offset
== -1)
2385 t
= caller_info
->known_vals
[src
->index
];
2387 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
2388 return (t
!= NULL_TREE
2389 && values_equal_for_ipcp_p (src
->val
->value
, t
));
2393 struct ipcp_agg_lattice
*aglat
;
2394 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
2396 if (src
->offset
== -1)
2397 return (ipa_lat_is_single_const (&plats
->itself
)
2398 && values_equal_for_ipcp_p (src
->val
->value
,
2399 plats
->itself
.values
->value
));
2402 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
2404 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2405 if (aglat
->offset
== src
->offset
)
2406 return (ipa_lat_is_single_const (aglat
)
2407 && values_equal_for_ipcp_p (src
->val
->value
,
2408 aglat
->values
->value
));
2414 /* Get the next clone in the linked list of clones of an edge. */
2416 static inline struct cgraph_edge
*
2417 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
2419 return next_edge_clone
[cs
->uid
];
2422 /* Given VAL, iterate over all its sources and if they still hold, add their
2423 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2427 get_info_about_necessary_edges (struct ipcp_value
*val
, int *freq_sum
,
2428 gcov_type
*count_sum
, int *caller_count
)
2430 struct ipcp_value_source
*src
;
2431 int freq
= 0, count
= 0;
2435 for (src
= val
->sources
; src
; src
= src
->next
)
2437 struct cgraph_edge
*cs
= src
->cs
;
2440 if (cgraph_edge_brings_value_p (cs
, src
))
2443 freq
+= cs
->frequency
;
2445 hot
|= cgraph_maybe_hot_edge_p (cs
);
2447 cs
= get_next_cgraph_edge_clone (cs
);
2453 *caller_count
= count
;
2457 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2458 their number is known and equal to CALLER_COUNT. */
2460 static vec
<cgraph_edge_p
>
2461 gather_edges_for_value (struct ipcp_value
*val
, int caller_count
)
2463 struct ipcp_value_source
*src
;
2464 vec
<cgraph_edge_p
> ret
;
2466 ret
.create (caller_count
);
2467 for (src
= val
->sources
; src
; src
= src
->next
)
2469 struct cgraph_edge
*cs
= src
->cs
;
2472 if (cgraph_edge_brings_value_p (cs
, src
))
2473 ret
.quick_push (cs
);
2474 cs
= get_next_cgraph_edge_clone (cs
);
2481 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2482 Return it or NULL if for some reason it cannot be created. */
2484 static struct ipa_replace_map
*
2485 get_replacement_map (struct ipa_node_params
*info
, tree value
, int parm_num
)
2487 struct ipa_replace_map
*replace_map
;
2490 replace_map
= ggc_alloc_ipa_replace_map ();
2493 fprintf (dump_file
, " replacing ");
2494 ipa_dump_param (dump_file
, info
, parm_num
);
2496 fprintf (dump_file
, " with const ");
2497 print_generic_expr (dump_file
, value
, 0);
2498 fprintf (dump_file
, "\n");
2500 replace_map
->old_tree
= NULL
;
2501 replace_map
->parm_num
= parm_num
;
2502 replace_map
->new_tree
= value
;
2503 replace_map
->replace_p
= true;
2504 replace_map
->ref_p
= false;
2509 /* Dump new profiling counts */
2512 dump_profile_updates (struct cgraph_node
*orig_node
,
2513 struct cgraph_node
*new_node
)
2515 struct cgraph_edge
*cs
;
2517 fprintf (dump_file
, " setting count of the specialized node to "
2518 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) new_node
->count
);
2519 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2520 fprintf (dump_file
, " edge to %s has count "
2521 HOST_WIDE_INT_PRINT_DEC
"\n",
2522 cgraph_node_name (cs
->callee
), (HOST_WIDE_INT
) cs
->count
);
2524 fprintf (dump_file
, " setting count of the original node to "
2525 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) orig_node
->count
);
2526 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2527 fprintf (dump_file
, " edge to %s is left with "
2528 HOST_WIDE_INT_PRINT_DEC
"\n",
2529 cgraph_node_name (cs
->callee
), (HOST_WIDE_INT
) cs
->count
);
2532 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2533 their profile information to reflect this. */
2536 update_profiling_info (struct cgraph_node
*orig_node
,
2537 struct cgraph_node
*new_node
)
2539 struct cgraph_edge
*cs
;
2540 struct caller_statistics stats
;
2541 gcov_type new_sum
, orig_sum
;
2542 gcov_type remainder
, orig_node_count
= orig_node
->count
;
2544 if (orig_node_count
== 0)
2547 init_caller_stats (&stats
);
2548 cgraph_for_node_and_aliases (orig_node
, gather_caller_stats
, &stats
, false);
2549 orig_sum
= stats
.count_sum
;
2550 init_caller_stats (&stats
);
2551 cgraph_for_node_and_aliases (new_node
, gather_caller_stats
, &stats
, false);
2552 new_sum
= stats
.count_sum
;
2554 if (orig_node_count
< orig_sum
+ new_sum
)
2557 fprintf (dump_file
, " Problem: node %s/%i has too low count "
2558 HOST_WIDE_INT_PRINT_DEC
" while the sum of incoming "
2559 "counts is " HOST_WIDE_INT_PRINT_DEC
"\n",
2560 cgraph_node_name (orig_node
), orig_node
->symbol
.order
,
2561 (HOST_WIDE_INT
) orig_node_count
,
2562 (HOST_WIDE_INT
) (orig_sum
+ new_sum
));
2564 orig_node_count
= (orig_sum
+ new_sum
) * 12 / 10;
2566 fprintf (dump_file
, " proceeding by pretending it was "
2567 HOST_WIDE_INT_PRINT_DEC
"\n",
2568 (HOST_WIDE_INT
) orig_node_count
);
2571 new_node
->count
= new_sum
;
2572 remainder
= orig_node_count
- new_sum
;
2573 orig_node
->count
= remainder
;
2575 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2577 cs
->count
= apply_probability (cs
->count
,
2578 GCOV_COMPUTE_SCALE (new_sum
,
2583 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2584 cs
->count
= apply_probability (cs
->count
,
2585 GCOV_COMPUTE_SCALE (remainder
,
2589 dump_profile_updates (orig_node
, new_node
);
2592 /* Update the respective profile of specialized NEW_NODE and the original
2593 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2594 have been redirected to the specialized version. */
2597 update_specialized_profile (struct cgraph_node
*new_node
,
2598 struct cgraph_node
*orig_node
,
2599 gcov_type redirected_sum
)
2601 struct cgraph_edge
*cs
;
2602 gcov_type new_node_count
, orig_node_count
= orig_node
->count
;
2605 fprintf (dump_file
, " the sum of counts of redirected edges is "
2606 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) redirected_sum
);
2607 if (orig_node_count
== 0)
2610 gcc_assert (orig_node_count
>= redirected_sum
);
2612 new_node_count
= new_node
->count
;
2613 new_node
->count
+= redirected_sum
;
2614 orig_node
->count
-= redirected_sum
;
2616 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2618 cs
->count
+= apply_probability (cs
->count
,
2619 GCOV_COMPUTE_SCALE (redirected_sum
,
2624 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2626 gcov_type dec
= apply_probability (cs
->count
,
2627 GCOV_COMPUTE_SCALE (redirected_sum
,
2629 if (dec
< cs
->count
)
2636 dump_profile_updates (orig_node
, new_node
);
2639 /* Create a specialized version of NODE with known constants and types of
2640 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2642 static struct cgraph_node
*
2643 create_specialized_node (struct cgraph_node
*node
,
2644 vec
<tree
> known_vals
,
2645 struct ipa_agg_replacement_value
*aggvals
,
2646 vec
<cgraph_edge_p
> callers
)
2648 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
2649 vec
<ipa_replace_map_p
, va_gc
> *replace_trees
= NULL
;
2650 struct ipa_agg_replacement_value
*av
;
2651 struct cgraph_node
*new_node
;
2652 int i
, count
= ipa_get_param_count (info
);
2653 bitmap args_to_skip
;
2655 gcc_assert (!info
->ipcp_orig_node
);
2657 if (node
->local
.can_change_signature
)
2659 args_to_skip
= BITMAP_GGC_ALLOC ();
2660 for (i
= 0; i
< count
; i
++)
2662 tree t
= known_vals
[i
];
2664 if ((t
&& TREE_CODE (t
) != TREE_BINFO
)
2665 || !ipa_is_param_used (info
, i
))
2666 bitmap_set_bit (args_to_skip
, i
);
2671 args_to_skip
= NULL
;
2672 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2673 fprintf (dump_file
, " cannot change function signature\n");
2676 for (i
= 0; i
< count
; i
++)
2678 tree t
= known_vals
[i
];
2679 if (t
&& TREE_CODE (t
) != TREE_BINFO
)
2681 struct ipa_replace_map
*replace_map
;
2683 replace_map
= get_replacement_map (info
, t
, i
);
2685 vec_safe_push (replace_trees
, replace_map
);
2689 new_node
= cgraph_create_virtual_clone (node
, callers
, replace_trees
,
2690 args_to_skip
, "constprop");
2691 ipa_set_node_agg_value_chain (new_node
, aggvals
);
2692 for (av
= aggvals
; av
; av
= av
->next
)
2693 ipa_maybe_record_reference ((symtab_node
) new_node
, av
->value
,
2694 IPA_REF_ADDR
, NULL
);
2696 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2698 fprintf (dump_file
, " the new node is %s/%i.\n",
2699 cgraph_node_name (new_node
), new_node
->symbol
.order
);
2701 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
2703 gcc_checking_assert (ipa_node_params_vector
.exists ()
2704 && (ipa_node_params_vector
.length ()
2705 > (unsigned) cgraph_max_uid
));
2706 update_profiling_info (node
, new_node
);
2707 new_info
= IPA_NODE_REF (new_node
);
2708 new_info
->ipcp_orig_node
= node
;
2709 new_info
->known_vals
= known_vals
;
2711 ipcp_discover_new_direct_edges (new_node
, known_vals
, aggvals
);
2717 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2718 KNOWN_VALS with constants and types that are also known for all of the
2722 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
2723 vec
<tree
> known_vals
,
2724 vec
<cgraph_edge_p
> callers
)
2726 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2727 int i
, count
= ipa_get_param_count (info
);
2729 for (i
= 0; i
< count
; i
++)
2731 struct cgraph_edge
*cs
;
2732 tree newval
= NULL_TREE
;
2735 if (ipa_get_scalar_lat (info
, i
)->bottom
|| known_vals
[i
])
2738 FOR_EACH_VEC_ELT (callers
, j
, cs
)
2740 struct ipa_jump_func
*jump_func
;
2743 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
2748 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
2749 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
);
2752 && !values_equal_for_ipcp_p (t
, newval
)))
2763 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2765 fprintf (dump_file
, " adding an extra known scalar value ");
2766 print_ipcp_constant_value (dump_file
, newval
);
2767 fprintf (dump_file
, " for ");
2768 ipa_dump_param (dump_file
, info
, i
);
2769 fprintf (dump_file
, "\n");
2772 known_vals
[i
] = newval
;
2777 /* Go through PLATS and create a vector of values consisting of values and
2778 offsets (minus OFFSET) of lattices that contain only a single value. */
2780 static vec
<ipa_agg_jf_item_t
>
2781 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
2783 vec
<ipa_agg_jf_item_t
> res
= vNULL
;
2785 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2788 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2789 if (ipa_lat_is_single_const (aglat
))
2791 struct ipa_agg_jf_item ti
;
2792 ti
.offset
= aglat
->offset
- offset
;
2793 ti
.value
= aglat
->values
->value
;
2799 /* Intersect all values in INTER with single value lattices in PLATS (while
2800 subtracting OFFSET). */
2803 intersect_with_plats (struct ipcp_param_lattices
*plats
,
2804 vec
<ipa_agg_jf_item_t
> *inter
,
2805 HOST_WIDE_INT offset
)
2807 struct ipcp_agg_lattice
*aglat
;
2808 struct ipa_agg_jf_item
*item
;
2811 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2817 aglat
= plats
->aggs
;
2818 FOR_EACH_VEC_ELT (*inter
, k
, item
)
2825 if (aglat
->offset
- offset
> item
->offset
)
2827 if (aglat
->offset
- offset
== item
->offset
)
2829 gcc_checking_assert (item
->value
);
2830 if (values_equal_for_ipcp_p (item
->value
, aglat
->values
->value
))
2834 aglat
= aglat
->next
;
2837 item
->value
= NULL_TREE
;
2841 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2842 vector result while subtracting OFFSET from the individual value offsets. */
2844 static vec
<ipa_agg_jf_item_t
>
2845 agg_replacements_to_vector (struct cgraph_node
*node
, int index
,
2846 HOST_WIDE_INT offset
)
2848 struct ipa_agg_replacement_value
*av
;
2849 vec
<ipa_agg_jf_item_t
> res
= vNULL
;
2851 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
2852 if (av
->index
== index
2853 && (av
->offset
- offset
) >= 0)
2855 struct ipa_agg_jf_item item
;
2856 gcc_checking_assert (av
->value
);
2857 item
.offset
= av
->offset
- offset
;
2858 item
.value
= av
->value
;
2859 res
.safe_push (item
);
2865 /* Intersect all values in INTER with those that we have already scheduled to
2866 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2867 (while subtracting OFFSET). */
2870 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
2871 vec
<ipa_agg_jf_item_t
> *inter
,
2872 HOST_WIDE_INT offset
)
2874 struct ipa_agg_replacement_value
*srcvals
;
2875 struct ipa_agg_jf_item
*item
;
2878 srcvals
= ipa_get_agg_replacements_for_node (node
);
2885 FOR_EACH_VEC_ELT (*inter
, i
, item
)
2887 struct ipa_agg_replacement_value
*av
;
2891 for (av
= srcvals
; av
; av
= av
->next
)
2893 gcc_checking_assert (av
->value
);
2894 if (av
->index
== index
2895 && av
->offset
- offset
== item
->offset
)
2897 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
2903 item
->value
= NULL_TREE
;
2907 /* Intersect values in INTER with aggregate values that come along edge CS to
2908 parameter number INDEX and return it. If INTER does not actually exist yet,
2909 copy all incoming values to it. If we determine we ended up with no values
2910 whatsoever, return a released vector. */
2912 static vec
<ipa_agg_jf_item_t
>
2913 intersect_aggregates_with_edge (struct cgraph_edge
*cs
, int index
,
2914 vec
<ipa_agg_jf_item_t
> inter
)
2916 struct ipa_jump_func
*jfunc
;
2917 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), index
);
2918 if (jfunc
->type
== IPA_JF_PASS_THROUGH
2919 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
2921 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2922 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
2924 if (caller_info
->ipcp_orig_node
)
2926 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
2927 struct ipcp_param_lattices
*orig_plats
;
2928 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
2930 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
2932 if (!inter
.exists ())
2933 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, 0);
2935 intersect_with_agg_replacements (cs
->caller
, src_idx
,
2941 struct ipcp_param_lattices
*src_plats
;
2942 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2943 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
2945 /* Currently we do not produce clobber aggregate jump
2946 functions, adjust when we do. */
2947 gcc_checking_assert (!jfunc
->agg
.items
);
2948 if (!inter
.exists ())
2949 inter
= copy_plats_to_inter (src_plats
, 0);
2951 intersect_with_plats (src_plats
, &inter
, 0);
2955 else if (jfunc
->type
== IPA_JF_ANCESTOR
2956 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
2958 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2959 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
2960 struct ipcp_param_lattices
*src_plats
;
2961 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
2963 if (caller_info
->ipcp_orig_node
)
2965 if (!inter
.exists ())
2966 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, delta
);
2968 intersect_with_agg_replacements (cs
->caller
, src_idx
, &inter
,
2973 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);;
2974 /* Currently we do not produce clobber aggregate jump
2975 functions, adjust when we do. */
2976 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
2977 if (!inter
.exists ())
2978 inter
= copy_plats_to_inter (src_plats
, delta
);
2980 intersect_with_plats (src_plats
, &inter
, delta
);
2983 else if (jfunc
->agg
.items
)
2985 struct ipa_agg_jf_item
*item
;
2988 if (!inter
.exists ())
2989 for (unsigned i
= 0; i
< jfunc
->agg
.items
->length (); i
++)
2990 inter
.safe_push ((*jfunc
->agg
.items
)[i
]);
2992 FOR_EACH_VEC_ELT (inter
, k
, item
)
2995 bool found
= false;;
3000 while ((unsigned) l
< jfunc
->agg
.items
->length ())
3002 struct ipa_agg_jf_item
*ti
;
3003 ti
= &(*jfunc
->agg
.items
)[l
];
3004 if (ti
->offset
> item
->offset
)
3006 if (ti
->offset
== item
->offset
)
3008 gcc_checking_assert (ti
->value
);
3009 if (values_equal_for_ipcp_p (item
->value
,
3023 return vec
<ipa_agg_jf_item_t
>();
3028 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3029 from all of them. */
3031 static struct ipa_agg_replacement_value
*
3032 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
3033 vec
<cgraph_edge_p
> callers
)
3035 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3036 struct ipa_agg_replacement_value
*res
= NULL
;
3037 struct cgraph_edge
*cs
;
3038 int i
, j
, count
= ipa_get_param_count (dest_info
);
3040 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3042 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3047 for (i
= 0; i
< count
; i
++)
3049 struct cgraph_edge
*cs
;
3050 vec
<ipa_agg_jf_item_t
> inter
= vNULL
;
3051 struct ipa_agg_jf_item
*item
;
3052 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (dest_info
, i
);
3055 /* Among other things, the following check should deal with all by_ref
3057 if (plats
->aggs_bottom
)
3060 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3062 inter
= intersect_aggregates_with_edge (cs
, i
, inter
);
3064 if (!inter
.exists ())
3068 FOR_EACH_VEC_ELT (inter
, j
, item
)
3070 struct ipa_agg_replacement_value
*v
;
3075 v
= ggc_alloc_ipa_agg_replacement_value ();
3077 v
->offset
= item
->offset
;
3078 v
->value
= item
->value
;
3079 v
->by_ref
= plats
->aggs_by_ref
;
3085 if (inter
.exists ())
3091 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3093 static struct ipa_agg_replacement_value
*
3094 known_aggs_to_agg_replacement_list (vec
<ipa_agg_jump_function_t
> known_aggs
)
3096 struct ipa_agg_replacement_value
*res
= NULL
;
3097 struct ipa_agg_jump_function
*aggjf
;
3098 struct ipa_agg_jf_item
*item
;
3101 FOR_EACH_VEC_ELT (known_aggs
, i
, aggjf
)
3102 FOR_EACH_VEC_SAFE_ELT (aggjf
->items
, j
, item
)
3104 struct ipa_agg_replacement_value
*v
;
3105 v
= ggc_alloc_ipa_agg_replacement_value ();
3107 v
->offset
= item
->offset
;
3108 v
->value
= item
->value
;
3109 v
->by_ref
= aggjf
->by_ref
;
3116 /* Determine whether CS also brings all scalar values that the NODE is
3120 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
3121 struct cgraph_node
*node
)
3123 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3124 int count
= ipa_get_param_count (dest_info
);
3125 struct ipa_node_params
*caller_info
;
3126 struct ipa_edge_args
*args
;
3129 caller_info
= IPA_NODE_REF (cs
->caller
);
3130 args
= IPA_EDGE_REF (cs
);
3131 for (i
= 0; i
< count
; i
++)
3133 struct ipa_jump_func
*jump_func
;
3136 val
= dest_info
->known_vals
[i
];
3140 if (i
>= ipa_get_cs_argument_count (args
))
3142 jump_func
= ipa_get_ith_jump_func (args
, i
);
3143 t
= ipa_value_from_jfunc (caller_info
, jump_func
);
3144 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
3150 /* Determine whether CS also brings all aggregate values that NODE is
3153 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
3154 struct cgraph_node
*node
)
3156 struct ipa_node_params
*orig_caller_info
= IPA_NODE_REF (cs
->caller
);
3157 struct ipa_agg_replacement_value
*aggval
;
3160 aggval
= ipa_get_agg_replacements_for_node (node
);
3164 count
= ipa_get_param_count (IPA_NODE_REF (node
));
3165 ec
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3167 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3168 if (aggval
->index
>= ec
)
3171 if (orig_caller_info
->ipcp_orig_node
)
3172 orig_caller_info
= IPA_NODE_REF (orig_caller_info
->ipcp_orig_node
);
3174 for (i
= 0; i
< count
; i
++)
3176 static vec
<ipa_agg_jf_item_t
> values
= vec
<ipa_agg_jf_item_t
>();
3177 struct ipcp_param_lattices
*plats
;
3178 bool interesting
= false;
3179 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3180 if (aggval
->index
== i
)
3188 plats
= ipa_get_parm_lattices (orig_caller_info
, aggval
->index
);
3189 if (plats
->aggs_bottom
)
3192 values
= intersect_aggregates_with_edge (cs
, i
, values
);
3193 if (!values
.exists())
3196 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3197 if (aggval
->index
== i
)
3199 struct ipa_agg_jf_item
*item
;
3202 FOR_EACH_VEC_ELT (values
, j
, item
)
3204 && item
->offset
== av
->offset
3205 && values_equal_for_ipcp_p (item
->value
, av
->value
))
3220 /* Given an original NODE and a VAL for which we have already created a
3221 specialized clone, look whether there are incoming edges that still lead
3222 into the old node but now also bring the requested value and also conform to
3223 all other criteria such that they can be redirected the the special node.
3224 This function can therefore redirect the final edge in a SCC. */
3227 perhaps_add_new_callers (struct cgraph_node
*node
, struct ipcp_value
*val
)
3229 struct ipcp_value_source
*src
;
3230 gcov_type redirected_sum
= 0;
3232 for (src
= val
->sources
; src
; src
= src
->next
)
3234 struct cgraph_edge
*cs
= src
->cs
;
3237 enum availability availability
;
3238 struct cgraph_node
*dst
= cgraph_function_node (cs
->callee
,
3240 if ((dst
== node
|| IPA_NODE_REF (dst
)->is_all_contexts_clone
)
3241 && availability
> AVAIL_OVERWRITABLE
3242 && cgraph_edge_brings_value_p (cs
, src
))
3244 if (cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
3245 && cgraph_edge_brings_all_agg_vals_for_node (cs
,
3249 fprintf (dump_file
, " - adding an extra caller %s/%i"
3251 xstrdup (cgraph_node_name (cs
->caller
)),
3252 cs
->caller
->symbol
.order
,
3253 xstrdup (cgraph_node_name (val
->spec_node
)),
3254 val
->spec_node
->symbol
.order
);
3256 cgraph_redirect_edge_callee (cs
, val
->spec_node
);
3257 redirected_sum
+= cs
->count
;
3260 cs
= get_next_cgraph_edge_clone (cs
);
3265 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
3269 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3272 move_binfos_to_values (vec
<tree
> known_vals
,
3273 vec
<tree
> known_binfos
)
3278 for (i
= 0; known_binfos
.iterate (i
, &t
); i
++)
3283 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3284 among those in the AGGVALS list. */
3287 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value
*aggvals
,
3288 int index
, HOST_WIDE_INT offset
, tree value
)
3292 if (aggvals
->index
== index
3293 && aggvals
->offset
== offset
3294 && values_equal_for_ipcp_p (aggvals
->value
, value
))
3296 aggvals
= aggvals
->next
;
3301 /* Decide wheter to create a special version of NODE for value VAL of parameter
3302 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3303 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3304 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3307 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
3308 struct ipcp_value
*val
, vec
<tree
> known_csts
,
3309 vec
<tree
> known_binfos
)
3311 struct ipa_agg_replacement_value
*aggvals
;
3312 int freq_sum
, caller_count
;
3313 gcov_type count_sum
;
3314 vec
<cgraph_edge_p
> callers
;
3319 perhaps_add_new_callers (node
, val
);
3322 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
3324 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3325 fprintf (dump_file
, " Ignoring candidate value because "
3326 "max_new_size would be reached with %li.\n",
3327 val
->local_size_cost
+ overall_size
);
3330 else if (!get_info_about_necessary_edges (val
, &freq_sum
, &count_sum
,
3334 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3336 fprintf (dump_file
, " - considering value ");
3337 print_ipcp_constant_value (dump_file
, val
->value
);
3338 fprintf (dump_file
, " for ");
3339 ipa_dump_param (dump_file
, IPA_NODE_REF (node
), index
);
3341 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
3342 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
3345 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
3346 freq_sum
, count_sum
,
3347 val
->local_size_cost
)
3348 && !good_cloning_opportunity_p (node
,
3349 val
->local_time_benefit
3350 + val
->prop_time_benefit
,
3351 freq_sum
, count_sum
,
3352 val
->local_size_cost
3353 + val
->prop_size_cost
))
3357 fprintf (dump_file
, " Creating a specialized node of %s/%i.\n",
3358 cgraph_node_name (node
), node
->symbol
.order
);
3360 callers
= gather_edges_for_value (val
, caller_count
);
3361 kv
= known_csts
.copy ();
3362 move_binfos_to_values (kv
, known_binfos
);
3364 kv
[index
] = val
->value
;
3365 find_more_scalar_values_for_callers_subset (node
, kv
, callers
);
3366 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
3367 gcc_checking_assert (offset
== -1
3368 || ipcp_val_in_agg_replacements_p (aggvals
, index
,
3369 offset
, val
->value
));
3370 val
->spec_node
= create_specialized_node (node
, kv
, aggvals
, callers
);
3371 overall_size
+= val
->local_size_cost
;
3373 /* TODO: If for some lattice there is only one other known value
3374 left, make a special node for it too. */
3379 /* Decide whether and what specialized clones of NODE should be created. */
3382 decide_whether_version_node (struct cgraph_node
*node
)
3384 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3385 int i
, count
= ipa_get_param_count (info
);
3386 vec
<tree
> known_csts
, known_binfos
;
3387 vec
<ipa_agg_jump_function_t
> known_aggs
= vNULL
;
3393 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3394 fprintf (dump_file
, "\nEvaluating opportunities for %s/%i.\n",
3395 cgraph_node_name (node
), node
->symbol
.order
);
3397 gather_context_independent_values (info
, &known_csts
, &known_binfos
,
3398 info
->do_clone_for_all_contexts
? &known_aggs
3401 for (i
= 0; i
< count
;i
++)
3403 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3404 struct ipcp_lattice
*lat
= &plats
->itself
;
3405 struct ipcp_value
*val
;
3409 && !known_binfos
[i
])
3410 for (val
= lat
->values
; val
; val
= val
->next
)
3411 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
3414 if (!plats
->aggs_bottom
)
3416 struct ipcp_agg_lattice
*aglat
;
3417 struct ipcp_value
*val
;
3418 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3419 if (!aglat
->bottom
&& aglat
->values
3420 /* If the following is false, the one value is in
3422 && (plats
->aggs_contain_variable
3423 || !ipa_lat_is_single_const (aglat
)))
3424 for (val
= aglat
->values
; val
; val
= val
->next
)
3425 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
3426 known_csts
, known_binfos
);
3428 info
= IPA_NODE_REF (node
);
3431 if (info
->do_clone_for_all_contexts
)
3433 struct cgraph_node
*clone
;
3434 vec
<cgraph_edge_p
> callers
;
3437 fprintf (dump_file
, " - Creating a specialized node of %s/%i "
3438 "for all known contexts.\n", cgraph_node_name (node
),
3439 node
->symbol
.order
);
3441 callers
= collect_callers_of_node (node
);
3442 move_binfos_to_values (known_csts
, known_binfos
);
3443 clone
= create_specialized_node (node
, known_csts
,
3444 known_aggs_to_agg_replacement_list (known_aggs
),
3446 info
= IPA_NODE_REF (node
);
3447 info
->do_clone_for_all_contexts
= false;
3448 IPA_NODE_REF (clone
)->is_all_contexts_clone
= true;
3449 for (i
= 0; i
< count
; i
++)
3450 vec_free (known_aggs
[i
].items
);
3451 known_aggs
.release ();
3455 known_csts
.release ();
3457 known_binfos
.release ();
3461 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3464 spread_undeadness (struct cgraph_node
*node
)
3466 struct cgraph_edge
*cs
;
3468 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
3469 if (edge_within_scc (cs
))
3471 struct cgraph_node
*callee
;
3472 struct ipa_node_params
*info
;
3474 callee
= cgraph_function_node (cs
->callee
, NULL
);
3475 info
= IPA_NODE_REF (callee
);
3477 if (info
->node_dead
)
3479 info
->node_dead
= 0;
3480 spread_undeadness (callee
);
3485 /* Return true if NODE has a caller from outside of its SCC that is not
3486 dead. Worker callback for cgraph_for_node_and_aliases. */
3489 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
3490 void *data ATTRIBUTE_UNUSED
)
3492 struct cgraph_edge
*cs
;
3494 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
3495 if (cs
->caller
->thunk
.thunk_p
3496 && cgraph_for_node_and_aliases (cs
->caller
,
3497 has_undead_caller_from_outside_scc_p
,
3500 else if (!edge_within_scc (cs
)
3501 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
3507 /* Identify nodes within the same SCC as NODE which are no longer needed
3508 because of new clones and will be removed as unreachable. */
3511 identify_dead_nodes (struct cgraph_node
*node
)
3513 struct cgraph_node
*v
;
3514 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3515 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v
)
3516 && !cgraph_for_node_and_aliases (v
,
3517 has_undead_caller_from_outside_scc_p
,
3519 IPA_NODE_REF (v
)->node_dead
= 1;
3521 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3522 if (!IPA_NODE_REF (v
)->node_dead
)
3523 spread_undeadness (v
);
3525 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3527 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3528 if (IPA_NODE_REF (v
)->node_dead
)
3529 fprintf (dump_file
, " Marking node as dead: %s/%i.\n",
3530 cgraph_node_name (v
), v
->symbol
.order
);
3534 /* The decision stage. Iterate over the topological order of call graph nodes
3535 TOPO and make specialized clones if deemed beneficial. */
3538 ipcp_decision_stage (struct topo_info
*topo
)
3543 fprintf (dump_file
, "\nIPA decision stage:\n\n");
3545 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3547 struct cgraph_node
*node
= topo
->order
[i
];
3548 bool change
= false, iterate
= true;
3552 struct cgraph_node
*v
;
3554 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3555 if (cgraph_function_with_gimple_body_p (v
)
3556 && ipcp_versionable_function_p (v
))
3557 iterate
|= decide_whether_version_node (v
);
3562 identify_dead_nodes (node
);
3566 /* The IPCP driver. */
3571 struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
3572 struct topo_info topo
;
3574 ipa_check_create_node_params ();
3575 ipa_check_create_edge_args ();
3576 grow_next_edge_clone_vector ();
3577 edge_duplication_hook_holder
=
3578 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook
, NULL
);
3579 ipcp_values_pool
= create_alloc_pool ("IPA-CP values",
3580 sizeof (struct ipcp_value
), 32);
3581 ipcp_sources_pool
= create_alloc_pool ("IPA-CP value sources",
3582 sizeof (struct ipcp_value_source
), 64);
3583 ipcp_agg_lattice_pool
= create_alloc_pool ("IPA_CP aggregate lattices",
3584 sizeof (struct ipcp_agg_lattice
),
3588 fprintf (dump_file
, "\nIPA structures before propagation:\n");
3589 if (dump_flags
& TDF_DETAILS
)
3590 ipa_print_all_params (dump_file
);
3591 ipa_print_all_jump_functions (dump_file
);
3594 /* Topological sort. */
3595 build_toporder_info (&topo
);
3596 /* Do the interprocedural propagation. */
3597 ipcp_propagate_stage (&topo
);
3598 /* Decide what constant propagation and cloning should be performed. */
3599 ipcp_decision_stage (&topo
);
3601 /* Free all IPCP structures. */
3602 free_toporder_info (&topo
);
3603 next_edge_clone
.release ();
3604 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder
);
3605 ipa_free_all_structures_after_ipa_cp ();
3607 fprintf (dump_file
, "\nIPA constant propagation end\n");
3611 /* Initialization and computation of IPCP data structures. This is the initial
3612 intraprocedural analysis of functions, which gathers information to be
3613 propagated later on. */
3616 ipcp_generate_summary (void)
3618 struct cgraph_node
*node
;
3621 fprintf (dump_file
, "\nIPA constant propagation start:\n");
3622 ipa_register_cgraph_hooks ();
3624 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
3626 node
->local
.versionable
3627 = tree_versionable_function_p (node
->symbol
.decl
);
3628 ipa_analyze_node (node
);
3632 /* Write ipcp summary for nodes in SET. */
3635 ipcp_write_summary (void)
3637 ipa_prop_write_jump_functions ();
3640 /* Read ipcp summary. */
3643 ipcp_read_summary (void)
3645 ipa_prop_read_jump_functions ();
3648 /* Gate for IPCP optimization. */
3651 cgraph_gate_cp (void)
3653 /* FIXME: We should remove the optimize check after we ensure we never run
3654 IPA passes when not optimizing. */
3655 return flag_ipa_cp
&& optimize
;
3660 const pass_data pass_data_ipa_cp
=
3662 IPA_PASS
, /* type */
3664 OPTGROUP_NONE
, /* optinfo_flags */
3665 true, /* has_gate */
3666 true, /* has_execute */
3667 TV_IPA_CONSTANT_PROP
, /* tv_id */
3668 0, /* properties_required */
3669 0, /* properties_provided */
3670 0, /* properties_destroyed */
3671 0, /* todo_flags_start */
3672 ( TODO_dump_symtab
| TODO_remove_functions
), /* todo_flags_finish */
3675 class pass_ipa_cp
: public ipa_opt_pass_d
3678 pass_ipa_cp(gcc::context
*ctxt
)
3679 : ipa_opt_pass_d(pass_data_ipa_cp
, ctxt
,
3680 ipcp_generate_summary
, /* generate_summary */
3681 ipcp_write_summary
, /* write_summary */
3682 ipcp_read_summary
, /* read_summary */
3683 ipa_prop_write_all_agg_replacement
, /*
3684 write_optimization_summary */
3685 ipa_prop_read_all_agg_replacement
, /*
3686 read_optimization_summary */
3687 NULL
, /* stmt_fixup */
3688 0, /* function_transform_todo_flags_start */
3689 ipcp_transform_function
, /* function_transform */
3690 NULL
) /* variable_transform */
3693 /* opt_pass methods: */
3694 bool gate () { return cgraph_gate_cp (); }
3695 unsigned int execute () { return ipcp_driver (); }
3697 }; // class pass_ipa_cp
3702 make_pass_ipa_cp (gcc::context
*ctxt
)
3704 return new pass_ipa_cp (ctxt
);