1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2014 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"
107 #include "gimple-fold.h"
108 #include "gimple-expr.h"
110 #include "ipa-prop.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 /* Print V which is extracted from a value in a lattice to F. */
293 print_ipcp_constant_value (FILE * f
, tree v
)
295 if (TREE_CODE (v
) == TREE_BINFO
)
297 fprintf (f
, "BINFO ");
298 print_generic_expr (f
, BINFO_TYPE (v
), 0);
300 else if (TREE_CODE (v
) == ADDR_EXPR
301 && TREE_CODE (TREE_OPERAND (v
, 0)) == CONST_DECL
)
304 print_generic_expr (f
, DECL_INITIAL (TREE_OPERAND (v
, 0)), 0);
307 print_generic_expr (f
, v
, 0);
310 /* Print a lattice LAT to F. */
313 print_lattice (FILE * f
, struct ipcp_lattice
*lat
,
314 bool dump_sources
, bool dump_benefits
)
316 struct ipcp_value
*val
;
321 fprintf (f
, "BOTTOM\n");
325 if (!lat
->values_count
&& !lat
->contains_variable
)
327 fprintf (f
, "TOP\n");
331 if (lat
->contains_variable
)
333 fprintf (f
, "VARIABLE");
339 for (val
= lat
->values
; val
; val
= val
->next
)
341 if (dump_benefits
&& prev
)
343 else if (!dump_benefits
&& prev
)
348 print_ipcp_constant_value (f
, val
->value
);
352 struct ipcp_value_source
*s
;
354 fprintf (f
, " [from:");
355 for (s
= val
->sources
; s
; s
= s
->next
)
356 fprintf (f
, " %i(%i)", s
->cs
->caller
->order
,
362 fprintf (f
, " [loc_time: %i, loc_size: %i, "
363 "prop_time: %i, prop_size: %i]\n",
364 val
->local_time_benefit
, val
->local_size_cost
,
365 val
->prop_time_benefit
, val
->prop_size_cost
);
371 /* Print all ipcp_lattices of all functions to F. */
374 print_all_lattices (FILE * f
, bool dump_sources
, bool dump_benefits
)
376 struct cgraph_node
*node
;
379 fprintf (f
, "\nLattices:\n");
380 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
382 struct ipa_node_params
*info
;
384 info
= IPA_NODE_REF (node
);
385 fprintf (f
, " Node: %s/%i:\n", node
->name (),
387 count
= ipa_get_param_count (info
);
388 for (i
= 0; i
< count
; i
++)
390 struct ipcp_agg_lattice
*aglat
;
391 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
392 fprintf (f
, " param [%d]: ", i
);
393 print_lattice (f
, &plats
->itself
, dump_sources
, dump_benefits
);
395 if (plats
->virt_call
)
396 fprintf (f
, " virt_call flag set\n");
398 if (plats
->aggs_bottom
)
400 fprintf (f
, " AGGS BOTTOM\n");
403 if (plats
->aggs_contain_variable
)
404 fprintf (f
, " AGGS VARIABLE\n");
405 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
407 fprintf (f
, " %soffset " HOST_WIDE_INT_PRINT_DEC
": ",
408 plats
->aggs_by_ref
? "ref " : "", aglat
->offset
);
409 print_lattice (f
, aglat
, dump_sources
, dump_benefits
);
415 /* Determine whether it is at all technically possible to create clones of NODE
416 and store this information in the ipa_node_params structure associated
420 determine_versionability (struct cgraph_node
*node
)
422 const char *reason
= NULL
;
424 /* There are a number of generic reasons functions cannot be versioned. We
425 also cannot remove parameters if there are type attributes such as fnspec
427 if (node
->alias
|| node
->thunk
.thunk_p
)
428 reason
= "alias or thunk";
429 else if (!node
->local
.versionable
)
430 reason
= "not a tree_versionable_function";
431 else if (cgraph_function_body_availability (node
) <= AVAIL_OVERWRITABLE
)
432 reason
= "insufficient body availability";
433 else if (!opt_for_fn (node
->decl
, optimize
)
434 || !opt_for_fn (node
->decl
, flag_ipa_cp
))
435 reason
= "non-optimized function";
436 else if (node
->tm_clone
)
437 reason
= "transactional memory clone";
438 else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node
->decl
)))
440 /* Ideally we should clone the SIMD clones themselves and create
441 vector copies of them, so IPA-cp and SIMD clones can happily
442 coexist, but that may not be worth the effort. */
443 reason
= "function has SIMD clones";
445 /* Don't clone decls local to a comdat group; it breaks and for C++
446 decloned constructors, inlining is always better anyway. */
447 else if (symtab_comdat_local_p (node
))
448 reason
= "comdat-local function";
450 if (reason
&& dump_file
&& !node
->alias
&& !node
->thunk
.thunk_p
)
451 fprintf (dump_file
, "Function %s/%i is not versionable, reason: %s.\n",
452 node
->name (), node
->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",
528 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->decl
)))
531 fprintf (dump_file
, "Not considering %s for cloning; "
532 "optimizing it for size.\n",
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",
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",
562 if (!stats
.n_hot_calls
)
565 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
570 fprintf (dump_file
, "Considering %s for cloning.\n",
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
->alias
&& !node
->thunk
.thunk_p
)
731 fprintf (dump_file
, "Marking all lattices of %s/%i as %s\n",
732 node
->name (), node
->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 be considered an interprocedural invariant. */
751 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
)
755 if (TREE_CODE (input
) == TREE_BINFO
)
757 if (ipa_get_jf_pass_through_type_preserved (jfunc
))
759 gcc_checking_assert (ipa_get_jf_pass_through_operation (jfunc
)
766 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
769 gcc_checking_assert (is_gimple_ip_invariant (input
));
770 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc
))
772 restype
= boolean_type_node
;
774 restype
= TREE_TYPE (input
);
775 res
= fold_binary (ipa_get_jf_pass_through_operation (jfunc
), restype
,
776 input
, ipa_get_jf_pass_through_operand (jfunc
));
778 if (res
&& !is_gimple_ip_invariant (res
))
784 /* Return the result of an ancestor jump function JFUNC on the constant value
785 INPUT. Return NULL_TREE if that cannot be determined. */
788 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
790 if (TREE_CODE (input
) == TREE_BINFO
)
792 if (!ipa_get_jf_ancestor_type_preserved (jfunc
))
794 return get_binfo_at_offset (input
,
795 ipa_get_jf_ancestor_offset (jfunc
),
796 ipa_get_jf_ancestor_type (jfunc
));
798 else if (TREE_CODE (input
) == ADDR_EXPR
)
800 tree t
= TREE_OPERAND (input
, 0);
801 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
802 ipa_get_jf_ancestor_offset (jfunc
),
803 ipa_get_jf_ancestor_type (jfunc
)
804 ? ipa_get_jf_ancestor_type (jfunc
)
805 : ptr_type_node
, NULL
, false);
806 return build_fold_addr_expr (t
);
812 /* Determine whether JFUNC evaluates to a known value (that is either a
813 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
814 describes the caller node so that pass-through jump functions can be
818 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
)
820 if (jfunc
->type
== IPA_JF_CONST
)
821 return ipa_get_jf_constant (jfunc
);
822 else if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
823 return ipa_binfo_from_known_type_jfunc (jfunc
);
824 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
825 || jfunc
->type
== IPA_JF_ANCESTOR
)
830 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
831 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
833 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
835 if (info
->ipcp_orig_node
)
836 input
= info
->known_vals
[idx
];
839 struct ipcp_lattice
*lat
;
843 gcc_checking_assert (!flag_ipa_cp
);
846 lat
= ipa_get_scalar_lat (info
, idx
);
847 if (!ipa_lat_is_single_const (lat
))
849 input
= lat
->values
->value
;
855 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
856 return ipa_get_jf_pass_through_result (jfunc
, input
);
858 return ipa_get_jf_ancestor_result (jfunc
, input
);
865 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
866 bottom, not containing a variable component and without any known value at
870 ipcp_verify_propagated_values (void)
872 struct cgraph_node
*node
;
874 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
876 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
877 int i
, count
= ipa_get_param_count (info
);
879 for (i
= 0; i
< count
; i
++)
881 struct ipcp_lattice
*lat
= ipa_get_scalar_lat (info
, i
);
884 && !lat
->contains_variable
885 && lat
->values_count
== 0)
889 dump_symtab (dump_file
);
890 fprintf (dump_file
, "\nIPA lattices after constant "
891 "propagation, before gcc_unreachable:\n");
892 print_all_lattices (dump_file
, true, false);
901 /* Return true iff X and Y should be considered equal values by IPA-CP. */
904 values_equal_for_ipcp_p (tree x
, tree y
)
906 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
911 if (TREE_CODE (x
) == TREE_BINFO
|| TREE_CODE (y
) == TREE_BINFO
)
914 if (TREE_CODE (x
) == ADDR_EXPR
915 && TREE_CODE (y
) == ADDR_EXPR
916 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
917 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
918 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
919 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
921 return operand_equal_p (x
, y
, 0);
924 /* Add a new value source to VAL, marking that a value comes from edge CS and
925 (if the underlying jump function is a pass-through or an ancestor one) from
926 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
927 is negative if the source was the scalar value of the parameter itself or
928 the offset within an aggregate. */
931 add_value_source (struct ipcp_value
*val
, struct cgraph_edge
*cs
,
932 struct ipcp_value
*src_val
, int src_idx
, HOST_WIDE_INT offset
)
934 struct ipcp_value_source
*src
;
936 src
= (struct ipcp_value_source
*) pool_alloc (ipcp_sources_pool
);
937 src
->offset
= offset
;
940 src
->index
= src_idx
;
942 src
->next
= val
->sources
;
946 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
947 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
948 have the same meaning. */
951 add_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
952 struct cgraph_edge
*cs
, struct ipcp_value
*src_val
,
953 int src_idx
, HOST_WIDE_INT offset
)
955 struct ipcp_value
*val
;
960 for (val
= lat
->values
; val
; val
= val
->next
)
961 if (values_equal_for_ipcp_p (val
->value
, newval
))
963 if (ipa_edge_within_scc (cs
))
965 struct ipcp_value_source
*s
;
966 for (s
= val
->sources
; s
; s
= s
->next
)
973 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
977 if (lat
->values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
979 /* We can only free sources, not the values themselves, because sources
980 of other values in this this SCC might point to them. */
981 for (val
= lat
->values
; val
; val
= val
->next
)
985 struct ipcp_value_source
*src
= val
->sources
;
986 val
->sources
= src
->next
;
987 pool_free (ipcp_sources_pool
, src
);
992 return set_lattice_to_bottom (lat
);
996 val
= (struct ipcp_value
*) pool_alloc (ipcp_values_pool
);
997 memset (val
, 0, sizeof (*val
));
999 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
1000 val
->value
= newval
;
1001 val
->next
= lat
->values
;
1006 /* Like above but passes a special value of offset to distinguish that the
1007 origin is the scalar value of the parameter rather than a part of an
1011 add_scalar_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
1012 struct cgraph_edge
*cs
,
1013 struct ipcp_value
*src_val
, int src_idx
)
1015 return add_value_to_lattice (lat
, newval
, cs
, src_val
, src_idx
, -1);
1018 /* Propagate values through a pass-through jump function JFUNC associated with
1019 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1020 is the index of the source parameter. */
1023 propagate_vals_accross_pass_through (struct cgraph_edge
*cs
,
1024 struct ipa_jump_func
*jfunc
,
1025 struct ipcp_lattice
*src_lat
,
1026 struct ipcp_lattice
*dest_lat
,
1029 struct ipcp_value
*src_val
;
1032 /* Do not create new values when propagating within an SCC because if there
1033 are arithmetic functions with circular dependencies, there is infinite
1034 number of them and we would just make lattices bottom. */
1035 if ((ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1036 && ipa_edge_within_scc (cs
))
1037 ret
= set_lattice_contains_variable (dest_lat
);
1039 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1041 tree cstval
= ipa_get_jf_pass_through_result (jfunc
, src_val
->value
);
1044 ret
|= add_scalar_value_to_lattice (dest_lat
, cstval
, cs
, src_val
,
1047 ret
|= set_lattice_contains_variable (dest_lat
);
1053 /* Propagate values through an ancestor jump function JFUNC associated with
1054 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1055 is the index of the source parameter. */
1058 propagate_vals_accross_ancestor (struct cgraph_edge
*cs
,
1059 struct ipa_jump_func
*jfunc
,
1060 struct ipcp_lattice
*src_lat
,
1061 struct ipcp_lattice
*dest_lat
,
1064 struct ipcp_value
*src_val
;
1067 if (ipa_edge_within_scc (cs
))
1068 return set_lattice_contains_variable (dest_lat
);
1070 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1072 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1075 ret
|= add_scalar_value_to_lattice (dest_lat
, t
, cs
, src_val
, src_idx
);
1077 ret
|= set_lattice_contains_variable (dest_lat
);
1083 /* Propagate scalar values across jump function JFUNC that is associated with
1084 edge CS and put the values into DEST_LAT. */
1087 propagate_scalar_accross_jump_function (struct cgraph_edge
*cs
,
1088 struct ipa_jump_func
*jfunc
,
1089 struct ipcp_lattice
*dest_lat
)
1091 if (dest_lat
->bottom
)
1094 if (jfunc
->type
== IPA_JF_CONST
1095 || jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1099 if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1101 val
= ipa_binfo_from_known_type_jfunc (jfunc
);
1103 return set_lattice_contains_variable (dest_lat
);
1106 val
= ipa_get_jf_constant (jfunc
);
1107 return add_scalar_value_to_lattice (dest_lat
, val
, cs
, NULL
, 0);
1109 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1110 || jfunc
->type
== IPA_JF_ANCESTOR
)
1112 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1113 struct ipcp_lattice
*src_lat
;
1117 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1118 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1120 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1122 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1123 if (src_lat
->bottom
)
1124 return set_lattice_contains_variable (dest_lat
);
1126 /* If we would need to clone the caller and cannot, do not propagate. */
1127 if (!ipcp_versionable_function_p (cs
->caller
)
1128 && (src_lat
->contains_variable
1129 || (src_lat
->values_count
> 1)))
1130 return set_lattice_contains_variable (dest_lat
);
1132 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1133 ret
= propagate_vals_accross_pass_through (cs
, jfunc
, src_lat
,
1136 ret
= propagate_vals_accross_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1139 if (src_lat
->contains_variable
)
1140 ret
|= set_lattice_contains_variable (dest_lat
);
1145 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1146 use it for indirect inlining), we should propagate them too. */
1147 return set_lattice_contains_variable (dest_lat
);
1150 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1151 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1152 other cases, return false). If there are no aggregate items, set
1153 aggs_by_ref to NEW_AGGS_BY_REF. */
1156 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1157 bool new_aggs_by_ref
)
1159 if (dest_plats
->aggs
)
1161 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1163 set_agg_lats_to_bottom (dest_plats
);
1168 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1172 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1173 already existing lattice for the given OFFSET and SIZE, marking all skipped
1174 lattices as containing variable and checking for overlaps. If there is no
1175 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1176 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1177 unless there are too many already. If there are two many, return false. If
1178 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1179 skipped lattices were newly marked as containing variable, set *CHANGE to
1183 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
1184 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
1185 struct ipcp_agg_lattice
***aglat
,
1186 bool pre_existing
, bool *change
)
1188 gcc_checking_assert (offset
>= 0);
1190 while (**aglat
&& (**aglat
)->offset
< offset
)
1192 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
1194 set_agg_lats_to_bottom (dest_plats
);
1197 *change
|= set_lattice_contains_variable (**aglat
);
1198 *aglat
= &(**aglat
)->next
;
1201 if (**aglat
&& (**aglat
)->offset
== offset
)
1203 if ((**aglat
)->size
!= val_size
1205 && (**aglat
)->next
->offset
< offset
+ val_size
))
1207 set_agg_lats_to_bottom (dest_plats
);
1210 gcc_checking_assert (!(**aglat
)->next
1211 || (**aglat
)->next
->offset
>= offset
+ val_size
);
1216 struct ipcp_agg_lattice
*new_al
;
1218 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
1220 set_agg_lats_to_bottom (dest_plats
);
1223 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
1225 dest_plats
->aggs_count
++;
1226 new_al
= (struct ipcp_agg_lattice
*) pool_alloc (ipcp_agg_lattice_pool
);
1227 memset (new_al
, 0, sizeof (*new_al
));
1229 new_al
->offset
= offset
;
1230 new_al
->size
= val_size
;
1231 new_al
->contains_variable
= pre_existing
;
1233 new_al
->next
= **aglat
;
1239 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1240 containing an unknown value. */
1243 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
1248 ret
|= set_lattice_contains_variable (aglat
);
1249 aglat
= aglat
->next
;
1254 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1255 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1256 parameter used for lattice value sources. Return true if DEST_PLATS changed
1260 merge_aggregate_lattices (struct cgraph_edge
*cs
,
1261 struct ipcp_param_lattices
*dest_plats
,
1262 struct ipcp_param_lattices
*src_plats
,
1263 int src_idx
, HOST_WIDE_INT offset_delta
)
1265 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1266 struct ipcp_agg_lattice
**dst_aglat
;
1269 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
1271 if (src_plats
->aggs_bottom
)
1272 return set_agg_lats_contain_variable (dest_plats
);
1273 if (src_plats
->aggs_contain_variable
)
1274 ret
|= set_agg_lats_contain_variable (dest_plats
);
1275 dst_aglat
= &dest_plats
->aggs
;
1277 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
1279 src_aglat
= src_aglat
->next
)
1281 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
1285 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
1286 &dst_aglat
, pre_existing
, &ret
))
1288 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
1290 dst_aglat
= &(*dst_aglat
)->next
;
1291 if (src_aglat
->bottom
)
1293 ret
|= set_lattice_contains_variable (new_al
);
1296 if (src_aglat
->contains_variable
)
1297 ret
|= set_lattice_contains_variable (new_al
);
1298 for (struct ipcp_value
*val
= src_aglat
->values
;
1301 ret
|= add_value_to_lattice (new_al
, val
->value
, cs
, val
, src_idx
,
1304 else if (dest_plats
->aggs_bottom
)
1307 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
1311 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1312 pass-through JFUNC and if so, whether it has conform and conforms to the
1313 rules about propagating values passed by reference. */
1316 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
1317 struct ipa_jump_func
*jfunc
)
1319 return src_plats
->aggs
1320 && (!src_plats
->aggs_by_ref
1321 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
1324 /* Propagate scalar values across jump function JFUNC that is associated with
1325 edge CS and put the values into DEST_LAT. */
1328 propagate_aggs_accross_jump_function (struct cgraph_edge
*cs
,
1329 struct ipa_jump_func
*jfunc
,
1330 struct ipcp_param_lattices
*dest_plats
)
1334 if (dest_plats
->aggs_bottom
)
1337 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1338 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1340 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1341 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1342 struct ipcp_param_lattices
*src_plats
;
1344 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1345 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
1347 /* Currently we do not produce clobber aggregate jump
1348 functions, replace with merging when we do. */
1349 gcc_assert (!jfunc
->agg
.items
);
1350 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
1354 ret
|= set_agg_lats_contain_variable (dest_plats
);
1356 else if (jfunc
->type
== IPA_JF_ANCESTOR
1357 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
1359 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1360 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1361 struct ipcp_param_lattices
*src_plats
;
1363 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1364 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
1366 /* Currently we do not produce clobber aggregate jump
1367 functions, replace with merging when we do. */
1368 gcc_assert (!jfunc
->agg
.items
);
1369 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
1370 ipa_get_jf_ancestor_offset (jfunc
));
1372 else if (!src_plats
->aggs_by_ref
)
1373 ret
|= set_agg_lats_to_bottom (dest_plats
);
1375 ret
|= set_agg_lats_contain_variable (dest_plats
);
1377 else if (jfunc
->agg
.items
)
1379 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1380 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
1381 struct ipa_agg_jf_item
*item
;
1384 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
1387 FOR_EACH_VEC_ELT (*jfunc
->agg
.items
, i
, item
)
1389 HOST_WIDE_INT val_size
;
1391 if (item
->offset
< 0)
1393 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
1394 val_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item
->value
)));
1396 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
1397 &aglat
, pre_existing
, &ret
))
1399 ret
|= add_value_to_lattice (*aglat
, item
->value
, cs
, NULL
, 0, 0);
1400 aglat
= &(*aglat
)->next
;
1402 else if (dest_plats
->aggs_bottom
)
1406 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
1409 ret
|= set_agg_lats_contain_variable (dest_plats
);
1414 /* Propagate constants from the caller to the callee of CS. INFO describes the
1418 propagate_constants_accross_call (struct cgraph_edge
*cs
)
1420 struct ipa_node_params
*callee_info
;
1421 enum availability availability
;
1422 struct cgraph_node
*callee
, *alias_or_thunk
;
1423 struct ipa_edge_args
*args
;
1425 int i
, args_count
, parms_count
;
1427 callee
= cgraph_function_node (cs
->callee
, &availability
);
1428 if (!callee
->definition
)
1430 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee
));
1431 callee_info
= IPA_NODE_REF (callee
);
1433 args
= IPA_EDGE_REF (cs
);
1434 args_count
= ipa_get_cs_argument_count (args
);
1435 parms_count
= ipa_get_param_count (callee_info
);
1436 if (parms_count
== 0)
1439 /* If this call goes through a thunk we must not propagate to the first (0th)
1440 parameter. However, we might need to uncover a thunk from below a series
1441 of aliases first. */
1442 alias_or_thunk
= cs
->callee
;
1443 while (alias_or_thunk
->alias
)
1444 alias_or_thunk
= cgraph_alias_target (alias_or_thunk
);
1445 if (alias_or_thunk
->thunk
.thunk_p
)
1447 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
1454 for (; (i
< args_count
) && (i
< parms_count
); i
++)
1456 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
1457 struct ipcp_param_lattices
*dest_plats
;
1459 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
1460 if (availability
== AVAIL_OVERWRITABLE
)
1461 ret
|= set_all_contains_variable (dest_plats
);
1464 ret
|= propagate_scalar_accross_jump_function (cs
, jump_func
,
1465 &dest_plats
->itself
);
1466 ret
|= propagate_aggs_accross_jump_function (cs
, jump_func
,
1470 for (; i
< parms_count
; i
++)
1471 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
1476 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1477 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1478 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1479 is not NULL, KNOWN_AGGS is ignored. */
1482 ipa_get_indirect_edge_target_1 (struct cgraph_edge
*ie
,
1483 vec
<tree
> known_vals
,
1484 vec
<tree
> known_binfos
,
1485 vec
<ipa_agg_jump_function_p
> known_aggs
,
1486 struct ipa_agg_replacement_value
*agg_reps
)
1488 int param_index
= ie
->indirect_info
->param_index
;
1489 HOST_WIDE_INT token
, anc_offset
;
1494 if (param_index
== -1
1495 || known_vals
.length () <= (unsigned int) param_index
)
1498 if (!ie
->indirect_info
->polymorphic
)
1502 if (ie
->indirect_info
->agg_contents
)
1509 if (agg_reps
->index
== param_index
1510 && agg_reps
->offset
== ie
->indirect_info
->offset
1511 && agg_reps
->by_ref
== ie
->indirect_info
->by_ref
)
1513 t
= agg_reps
->value
;
1516 agg_reps
= agg_reps
->next
;
1519 else if (known_aggs
.length () > (unsigned int) param_index
)
1521 struct ipa_agg_jump_function
*agg
;
1522 agg
= known_aggs
[param_index
];
1523 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1524 ie
->indirect_info
->by_ref
);
1530 t
= known_vals
[param_index
];
1533 TREE_CODE (t
) == ADDR_EXPR
1534 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
1535 return TREE_OPERAND (t
, 0);
1540 if (!flag_devirtualize
)
1543 gcc_assert (!ie
->indirect_info
->agg_contents
);
1544 token
= ie
->indirect_info
->otr_token
;
1545 anc_offset
= ie
->indirect_info
->offset
;
1546 otr_type
= ie
->indirect_info
->otr_type
;
1550 /* Try to work out value of virtual table pointer value in replacemnets. */
1551 if (!t
&& agg_reps
&& !ie
->indirect_info
->by_ref
)
1555 if (agg_reps
->index
== param_index
1556 && agg_reps
->offset
== ie
->indirect_info
->offset
1557 && agg_reps
->by_ref
)
1559 t
= agg_reps
->value
;
1562 agg_reps
= agg_reps
->next
;
1566 /* Try to work out value of virtual table pointer value in known
1567 aggregate values. */
1568 if (!t
&& known_aggs
.length () > (unsigned int) param_index
1569 && !ie
->indirect_info
->by_ref
)
1571 struct ipa_agg_jump_function
*agg
;
1572 agg
= known_aggs
[param_index
];
1573 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1577 /* If we found the virtual table pointer, lookup the target. */
1581 unsigned HOST_WIDE_INT offset
;
1582 if (vtable_pointer_value_to_vtable (t
, &vtable
, &offset
))
1584 target
= gimple_get_virt_method_for_vtable (ie
->indirect_info
->otr_token
,
1588 if ((TREE_CODE (TREE_TYPE (target
)) == FUNCTION_TYPE
1589 && DECL_FUNCTION_CODE (target
) == BUILT_IN_UNREACHABLE
)
1590 || !possible_polymorphic_call_target_p
1591 (ie
, cgraph_get_node (target
)))
1592 target
= ipa_impossible_devirt_target (ie
, target
);
1598 /* Did we work out BINFO via type propagation? */
1599 if (!t
&& known_binfos
.length () > (unsigned int) param_index
)
1600 t
= known_binfos
[param_index
];
1601 /* Or do we know the constant value of pointer? */
1603 t
= known_vals
[param_index
];
1607 if (TREE_CODE (t
) != TREE_BINFO
)
1609 ipa_polymorphic_call_context context
;
1610 vec
<cgraph_node
*>targets
;
1613 if (!get_polymorphic_call_info_from_invariant
1614 (&context
, t
, ie
->indirect_info
->otr_type
,
1617 targets
= possible_polymorphic_call_targets
1618 (ie
->indirect_info
->otr_type
,
1619 ie
->indirect_info
->otr_token
,
1621 if (!final
|| targets
.length () > 1)
1623 if (targets
.length () == 1)
1624 target
= targets
[0]->decl
;
1626 target
= ipa_impossible_devirt_target (ie
, NULL_TREE
);
1632 binfo
= get_binfo_at_offset (t
, anc_offset
, otr_type
);
1635 target
= gimple_get_virt_method_for_binfo (token
, binfo
);
1638 if (target
&& !possible_polymorphic_call_target_p (ie
,
1639 cgraph_get_node (target
)))
1640 target
= ipa_impossible_devirt_target (ie
, target
);
1646 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1647 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1648 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1651 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
1652 vec
<tree
> known_vals
,
1653 vec
<tree
> known_binfos
,
1654 vec
<ipa_agg_jump_function_p
> known_aggs
)
1656 return ipa_get_indirect_edge_target_1 (ie
, known_vals
, known_binfos
,
1660 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1661 and KNOWN_BINFOS. */
1664 devirtualization_time_bonus (struct cgraph_node
*node
,
1665 vec
<tree
> known_csts
,
1666 vec
<tree
> known_binfos
,
1667 vec
<ipa_agg_jump_function_p
> known_aggs
)
1669 struct cgraph_edge
*ie
;
1672 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1674 struct cgraph_node
*callee
;
1675 struct inline_summary
*isummary
;
1678 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_binfos
,
1683 /* Only bare minimum benefit for clearly un-inlineable targets. */
1685 callee
= cgraph_get_node (target
);
1686 if (!callee
|| !callee
->definition
)
1688 isummary
= inline_summary (callee
);
1689 if (!isummary
->inlinable
)
1692 /* FIXME: The values below need re-considering and perhaps also
1693 integrating into the cost metrics, at lest in some very basic way. */
1694 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
1696 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
1698 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
1699 || DECL_DECLARED_INLINE_P (callee
->decl
))
1706 /* Return time bonus incurred because of HINTS. */
1709 hint_time_bonus (inline_hints hints
)
1712 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
1713 result
+= PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
1714 if (hints
& INLINE_HINT_array_index
)
1715 result
+= PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS
);
1719 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1720 and SIZE_COST and with the sum of frequencies of incoming edges to the
1721 potential new clone in FREQUENCIES. */
1724 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
1725 int freq_sum
, gcov_type count_sum
, int size_cost
)
1727 if (time_benefit
== 0
1728 || !flag_ipa_cp_clone
1729 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->decl
)))
1732 gcc_assert (size_cost
> 0);
1736 int factor
= (count_sum
* 1000) / max_count
;
1737 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* factor
)
1740 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1741 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1742 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1743 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1744 ", threshold: %i\n",
1745 time_benefit
, size_cost
, (HOST_WIDE_INT
) count_sum
,
1746 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1748 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1752 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* freq_sum
)
1755 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1756 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1757 "size: %i, freq_sum: %i) -> evaluation: "
1758 HOST_WIDEST_INT_PRINT_DEC
", threshold: %i\n",
1759 time_benefit
, size_cost
, freq_sum
, evaluation
,
1760 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1762 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1766 /* Return all context independent values from aggregate lattices in PLATS in a
1767 vector. Return NULL if there are none. */
1769 static vec
<ipa_agg_jf_item
, va_gc
> *
1770 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
1772 vec
<ipa_agg_jf_item
, va_gc
> *res
= NULL
;
1774 if (plats
->aggs_bottom
1775 || plats
->aggs_contain_variable
1776 || plats
->aggs_count
== 0)
1779 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
1781 aglat
= aglat
->next
)
1782 if (ipa_lat_is_single_const (aglat
))
1784 struct ipa_agg_jf_item item
;
1785 item
.offset
= aglat
->offset
;
1786 item
.value
= aglat
->values
->value
;
1787 vec_safe_push (res
, item
);
1792 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1793 them with values of parameters that are known independent of the context.
1794 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1795 movement cost of all removable parameters will be stored in it. */
1798 gather_context_independent_values (struct ipa_node_params
*info
,
1799 vec
<tree
> *known_csts
,
1800 vec
<tree
> *known_binfos
,
1801 vec
<ipa_agg_jump_function
> *known_aggs
,
1802 int *removable_params_cost
)
1804 int i
, count
= ipa_get_param_count (info
);
1807 known_csts
->create (0);
1808 known_binfos
->create (0);
1809 known_csts
->safe_grow_cleared (count
);
1810 known_binfos
->safe_grow_cleared (count
);
1813 known_aggs
->create (0);
1814 known_aggs
->safe_grow_cleared (count
);
1817 if (removable_params_cost
)
1818 *removable_params_cost
= 0;
1820 for (i
= 0; i
< count
; i
++)
1822 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1823 struct ipcp_lattice
*lat
= &plats
->itself
;
1825 if (ipa_lat_is_single_const (lat
))
1827 struct ipcp_value
*val
= lat
->values
;
1828 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1830 (*known_csts
)[i
] = val
->value
;
1831 if (removable_params_cost
)
1832 *removable_params_cost
1833 += estimate_move_cost (TREE_TYPE (val
->value
));
1836 else if (plats
->virt_call
)
1838 (*known_binfos
)[i
] = val
->value
;
1841 else if (removable_params_cost
1842 && !ipa_is_param_used (info
, i
))
1843 *removable_params_cost
+= ipa_get_param_move_cost (info
, i
);
1845 else if (removable_params_cost
1846 && !ipa_is_param_used (info
, i
))
1847 *removable_params_cost
1848 += ipa_get_param_move_cost (info
, i
);
1852 vec
<ipa_agg_jf_item
, va_gc
> *agg_items
;
1853 struct ipa_agg_jump_function
*ajf
;
1855 agg_items
= context_independent_aggregate_values (plats
);
1856 ajf
= &(*known_aggs
)[i
];
1857 ajf
->items
= agg_items
;
1858 ajf
->by_ref
= plats
->aggs_by_ref
;
1859 ret
|= agg_items
!= NULL
;
1866 /* The current interface in ipa-inline-analysis requires a pointer vector.
1869 FIXME: That interface should be re-worked, this is slightly silly. Still,
1870 I'd like to discuss how to change it first and this demonstrates the
1873 static vec
<ipa_agg_jump_function_p
>
1874 agg_jmp_p_vec_for_t_vec (vec
<ipa_agg_jump_function
> known_aggs
)
1876 vec
<ipa_agg_jump_function_p
> ret
;
1877 struct ipa_agg_jump_function
*ajf
;
1880 ret
.create (known_aggs
.length ());
1881 FOR_EACH_VEC_ELT (known_aggs
, i
, ajf
)
1882 ret
.quick_push (ajf
);
1886 /* Iterate over known values of parameters of NODE and estimate the local
1887 effects in terms of time and size they have. */
1890 estimate_local_effects (struct cgraph_node
*node
)
1892 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1893 int i
, count
= ipa_get_param_count (info
);
1894 vec
<tree
> known_csts
, known_binfos
;
1895 vec
<ipa_agg_jump_function
> known_aggs
;
1896 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
;
1898 int base_time
= inline_summary (node
)->time
;
1899 int removable_params_cost
;
1901 if (!count
|| !ipcp_versionable_function_p (node
))
1904 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1905 fprintf (dump_file
, "\nEstimating effects for %s/%i, base_time: %i.\n",
1906 node
->name (), node
->order
, base_time
);
1908 always_const
= gather_context_independent_values (info
, &known_csts
,
1909 &known_binfos
, &known_aggs
,
1910 &removable_params_cost
);
1911 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
1914 struct caller_statistics stats
;
1918 init_caller_stats (&stats
);
1919 cgraph_for_node_and_aliases (node
, gather_caller_stats
, &stats
, false);
1920 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1921 known_aggs_ptrs
, &size
, &time
, &hints
);
1922 time
-= devirtualization_time_bonus (node
, known_csts
, known_binfos
,
1924 time
-= hint_time_bonus (hints
);
1925 time
-= removable_params_cost
;
1926 size
-= stats
.n_calls
* removable_params_cost
;
1929 fprintf (dump_file
, " - context independent values, size: %i, "
1930 "time_benefit: %i\n", size
, base_time
- time
);
1933 || cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
1935 info
->do_clone_for_all_contexts
= true;
1939 fprintf (dump_file
, " Decided to specialize for all "
1940 "known contexts, code not going to grow.\n");
1942 else if (good_cloning_opportunity_p (node
, base_time
- time
,
1943 stats
.freq_sum
, stats
.count_sum
,
1946 if (size
+ overall_size
<= max_new_size
)
1948 info
->do_clone_for_all_contexts
= true;
1950 overall_size
+= size
;
1953 fprintf (dump_file
, " Decided to specialize for all "
1954 "known contexts, growth deemed beneficial.\n");
1956 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1957 fprintf (dump_file
, " Not cloning for all contexts because "
1958 "max_new_size would be reached with %li.\n",
1959 size
+ overall_size
);
1963 for (i
= 0; i
< count
; i
++)
1965 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1966 struct ipcp_lattice
*lat
= &plats
->itself
;
1967 struct ipcp_value
*val
;
1976 for (val
= lat
->values
; val
; val
= val
->next
)
1978 int time
, size
, time_benefit
;
1981 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1983 known_csts
[i
] = val
->value
;
1984 known_binfos
[i
] = NULL_TREE
;
1985 emc
= estimate_move_cost (TREE_TYPE (val
->value
));
1987 else if (plats
->virt_call
)
1989 known_csts
[i
] = NULL_TREE
;
1990 known_binfos
[i
] = val
->value
;
1996 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1997 known_aggs_ptrs
, &size
, &time
,
1999 time_benefit
= base_time
- time
2000 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
2002 + hint_time_bonus (hints
)
2003 + removable_params_cost
+ emc
;
2005 gcc_checking_assert (size
>=0);
2006 /* The inliner-heuristics based estimates may think that in certain
2007 contexts some functions do not have any size at all but we want
2008 all specializations to have at least a tiny cost, not least not to
2013 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2015 fprintf (dump_file
, " - estimates for value ");
2016 print_ipcp_constant_value (dump_file
, val
->value
);
2017 fprintf (dump_file
, " for ");
2018 ipa_dump_param (dump_file
, info
, i
);
2019 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
2020 time_benefit
, size
);
2023 val
->local_time_benefit
= time_benefit
;
2024 val
->local_size_cost
= size
;
2026 known_binfos
[i
] = NULL_TREE
;
2027 known_csts
[i
] = NULL_TREE
;
2030 for (i
= 0; i
< count
; i
++)
2032 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2033 struct ipa_agg_jump_function
*ajf
;
2034 struct ipcp_agg_lattice
*aglat
;
2036 if (plats
->aggs_bottom
|| !plats
->aggs
)
2039 ajf
= &known_aggs
[i
];
2040 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2042 struct ipcp_value
*val
;
2043 if (aglat
->bottom
|| !aglat
->values
2044 /* If the following is true, the one value is in known_aggs. */
2045 || (!plats
->aggs_contain_variable
2046 && ipa_lat_is_single_const (aglat
)))
2049 for (val
= aglat
->values
; val
; val
= val
->next
)
2051 int time
, size
, time_benefit
;
2052 struct ipa_agg_jf_item item
;
2055 item
.offset
= aglat
->offset
;
2056 item
.value
= val
->value
;
2057 vec_safe_push (ajf
->items
, item
);
2059 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
2060 known_aggs_ptrs
, &size
, &time
,
2062 time_benefit
= base_time
- time
2063 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
2065 + hint_time_bonus (hints
);
2066 gcc_checking_assert (size
>=0);
2070 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2072 fprintf (dump_file
, " - estimates for value ");
2073 print_ipcp_constant_value (dump_file
, val
->value
);
2074 fprintf (dump_file
, " for ");
2075 ipa_dump_param (dump_file
, info
, i
);
2076 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2077 "]: time_benefit: %i, size: %i\n",
2078 plats
->aggs_by_ref
? "ref " : "",
2079 aglat
->offset
, time_benefit
, size
);
2082 val
->local_time_benefit
= time_benefit
;
2083 val
->local_size_cost
= size
;
2089 for (i
= 0; i
< count
; i
++)
2090 vec_free (known_aggs
[i
].items
);
2092 known_csts
.release ();
2093 known_binfos
.release ();
2094 known_aggs
.release ();
2095 known_aggs_ptrs
.release ();
2099 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2100 topological sort of values. */
2103 add_val_to_toposort (struct ipcp_value
*cur_val
)
2105 static int dfs_counter
= 0;
2106 static struct ipcp_value
*stack
;
2107 struct ipcp_value_source
*src
;
2113 cur_val
->dfs
= dfs_counter
;
2114 cur_val
->low_link
= dfs_counter
;
2116 cur_val
->topo_next
= stack
;
2118 cur_val
->on_stack
= true;
2120 for (src
= cur_val
->sources
; src
; src
= src
->next
)
2123 if (src
->val
->dfs
== 0)
2125 add_val_to_toposort (src
->val
);
2126 if (src
->val
->low_link
< cur_val
->low_link
)
2127 cur_val
->low_link
= src
->val
->low_link
;
2129 else if (src
->val
->on_stack
2130 && src
->val
->dfs
< cur_val
->low_link
)
2131 cur_val
->low_link
= src
->val
->dfs
;
2134 if (cur_val
->dfs
== cur_val
->low_link
)
2136 struct ipcp_value
*v
, *scc_list
= NULL
;
2141 stack
= v
->topo_next
;
2142 v
->on_stack
= false;
2144 v
->scc_next
= scc_list
;
2147 while (v
!= cur_val
);
2149 cur_val
->topo_next
= values_topo
;
2150 values_topo
= cur_val
;
2154 /* Add all values in lattices associated with NODE to the topological sort if
2155 they are not there yet. */
2158 add_all_node_vals_to_toposort (struct cgraph_node
*node
)
2160 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2161 int i
, count
= ipa_get_param_count (info
);
2163 for (i
= 0; i
< count
; i
++)
2165 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2166 struct ipcp_lattice
*lat
= &plats
->itself
;
2167 struct ipcp_agg_lattice
*aglat
;
2168 struct ipcp_value
*val
;
2171 for (val
= lat
->values
; val
; val
= val
->next
)
2172 add_val_to_toposort (val
);
2174 if (!plats
->aggs_bottom
)
2175 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2177 for (val
= aglat
->values
; val
; val
= val
->next
)
2178 add_val_to_toposort (val
);
2182 /* One pass of constants propagation along the call graph edges, from callers
2183 to callees (requires topological ordering in TOPO), iterate over strongly
2184 connected components. */
2187 propagate_constants_topo (struct topo_info
*topo
)
2191 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
2194 struct cgraph_node
*v
, *node
= topo
->order
[i
];
2195 vec
<cgraph_node_ptr
> cycle_nodes
= ipa_get_nodes_in_cycle (node
);
2197 /* First, iteratively propagate within the strongly connected component
2198 until all lattices stabilize. */
2199 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
2200 if (cgraph_function_with_gimple_body_p (v
))
2201 push_node_to_stack (topo
, v
);
2203 v
= pop_node_from_stack (topo
);
2206 struct cgraph_edge
*cs
;
2208 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2209 if (ipa_edge_within_scc (cs
)
2210 && propagate_constants_accross_call (cs
))
2211 push_node_to_stack (topo
, cs
->callee
);
2212 v
= pop_node_from_stack (topo
);
2215 /* Afterwards, propagate along edges leading out of the SCC, calculates
2216 the local effects of the discovered constants and all valid values to
2217 their topological sort. */
2218 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
2219 if (cgraph_function_with_gimple_body_p (v
))
2221 struct cgraph_edge
*cs
;
2223 estimate_local_effects (v
);
2224 add_all_node_vals_to_toposort (v
);
2225 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2226 if (!ipa_edge_within_scc (cs
))
2227 propagate_constants_accross_call (cs
);
2229 cycle_nodes
.release ();
2234 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2235 the bigger one if otherwise. */
2238 safe_add (int a
, int b
)
2240 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
2241 return a
> b
? a
: b
;
2247 /* Propagate the estimated effects of individual values along the topological
2248 from the dependent values to those they depend on. */
2251 propagate_effects (void)
2253 struct ipcp_value
*base
;
2255 for (base
= values_topo
; base
; base
= base
->topo_next
)
2257 struct ipcp_value_source
*src
;
2258 struct ipcp_value
*val
;
2259 int time
= 0, size
= 0;
2261 for (val
= base
; val
; val
= val
->scc_next
)
2263 time
= safe_add (time
,
2264 val
->local_time_benefit
+ val
->prop_time_benefit
);
2265 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
2268 for (val
= base
; val
; val
= val
->scc_next
)
2269 for (src
= val
->sources
; src
; src
= src
->next
)
2271 && cgraph_maybe_hot_edge_p (src
->cs
))
2273 src
->val
->prop_time_benefit
= safe_add (time
,
2274 src
->val
->prop_time_benefit
);
2275 src
->val
->prop_size_cost
= safe_add (size
,
2276 src
->val
->prop_size_cost
);
2282 /* Propagate constants, binfos and their effects from the summaries
2283 interprocedurally. */
2286 ipcp_propagate_stage (struct topo_info
*topo
)
2288 struct cgraph_node
*node
;
2291 fprintf (dump_file
, "\n Propagating constants:\n\n");
2294 ipa_update_after_lto_read ();
2297 FOR_EACH_DEFINED_FUNCTION (node
)
2299 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2301 determine_versionability (node
);
2302 if (cgraph_function_with_gimple_body_p (node
))
2304 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
2305 ipa_get_param_count (info
));
2306 initialize_node_lattices (node
);
2308 if (node
->definition
&& !node
->alias
)
2309 overall_size
+= inline_summary (node
)->self_size
;
2310 if (node
->count
> max_count
)
2311 max_count
= node
->count
;
2314 max_new_size
= overall_size
;
2315 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
2316 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
2317 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
2320 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
2321 overall_size
, max_new_size
);
2323 propagate_constants_topo (topo
);
2324 #ifdef ENABLE_CHECKING
2325 ipcp_verify_propagated_values ();
2327 propagate_effects ();
2331 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
2332 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
2336 /* Discover newly direct outgoing edges from NODE which is a new clone with
2337 known KNOWN_VALS and make them direct. */
2340 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
2341 vec
<tree
> known_vals
,
2342 struct ipa_agg_replacement_value
*aggvals
)
2344 struct cgraph_edge
*ie
, *next_ie
;
2347 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
2351 next_ie
= ie
->next_callee
;
2352 target
= ipa_get_indirect_edge_target_1 (ie
, known_vals
, vNULL
, vNULL
,
2356 bool agg_contents
= ie
->indirect_info
->agg_contents
;
2357 bool polymorphic
= ie
->indirect_info
->polymorphic
;
2358 int param_index
= ie
->indirect_info
->param_index
;
2359 struct cgraph_edge
*cs
= ipa_make_edge_direct_to_target (ie
, target
);
2362 if (cs
&& !agg_contents
&& !polymorphic
)
2364 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2365 int c
= ipa_get_controlled_uses (info
, param_index
);
2366 if (c
!= IPA_UNDESCRIBED_USE
)
2368 struct ipa_ref
*to_del
;
2371 ipa_set_controlled_uses (info
, param_index
, c
);
2372 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2373 fprintf (dump_file
, " controlled uses count of param "
2374 "%i bumped down to %i\n", param_index
, c
);
2376 && (to_del
= ipa_find_reference (node
,
2380 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2381 fprintf (dump_file
, " and even removing its "
2382 "cloning-created reference\n");
2383 ipa_remove_reference (to_del
);
2389 /* Turning calls to direct calls will improve overall summary. */
2391 inline_update_overall_summary (node
);
2394 /* Vector of pointers which for linked lists of clones of an original crgaph
2397 static vec
<cgraph_edge_p
> next_edge_clone
;
2398 static vec
<cgraph_edge_p
> prev_edge_clone
;
2401 grow_edge_clone_vectors (void)
2403 if (next_edge_clone
.length ()
2404 <= (unsigned) cgraph_edge_max_uid
)
2405 next_edge_clone
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
2406 if (prev_edge_clone
.length ()
2407 <= (unsigned) cgraph_edge_max_uid
)
2408 prev_edge_clone
.safe_grow_cleared (cgraph_edge_max_uid
+ 1);
2411 /* Edge duplication hook to grow the appropriate linked list in
2415 ipcp_edge_duplication_hook (struct cgraph_edge
*src
, struct cgraph_edge
*dst
,
2418 grow_edge_clone_vectors ();
2420 struct cgraph_edge
*old_next
= next_edge_clone
[src
->uid
];
2422 prev_edge_clone
[old_next
->uid
] = dst
;
2423 prev_edge_clone
[dst
->uid
] = src
;
2425 next_edge_clone
[dst
->uid
] = old_next
;
2426 next_edge_clone
[src
->uid
] = dst
;
2429 /* Hook that is called by cgraph.c when an edge is removed. */
2432 ipcp_edge_removal_hook (struct cgraph_edge
*cs
, void *)
2434 grow_edge_clone_vectors ();
2436 struct cgraph_edge
*prev
= prev_edge_clone
[cs
->uid
];
2437 struct cgraph_edge
*next
= next_edge_clone
[cs
->uid
];
2439 next_edge_clone
[prev
->uid
] = next
;
2441 prev_edge_clone
[next
->uid
] = prev
;
2444 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2445 parameter with the given INDEX. */
2448 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDEST_INT offset
,
2451 struct ipa_agg_replacement_value
*aggval
;
2453 aggval
= ipa_get_agg_replacements_for_node (node
);
2456 if (aggval
->offset
== offset
2457 && aggval
->index
== index
)
2458 return aggval
->value
;
2459 aggval
= aggval
->next
;
2464 /* Return true if edge CS does bring about the value described by SRC. */
2467 cgraph_edge_brings_value_p (struct cgraph_edge
*cs
,
2468 struct ipcp_value_source
*src
)
2470 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2471 cgraph_node
*real_dest
= cgraph_function_node (cs
->callee
);
2472 struct ipa_node_params
*dst_info
= IPA_NODE_REF (real_dest
);
2474 if ((dst_info
->ipcp_orig_node
&& !dst_info
->is_all_contexts_clone
)
2475 || caller_info
->node_dead
)
2480 if (caller_info
->ipcp_orig_node
)
2483 if (src
->offset
== -1)
2484 t
= caller_info
->known_vals
[src
->index
];
2486 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
2487 return (t
!= NULL_TREE
2488 && values_equal_for_ipcp_p (src
->val
->value
, t
));
2492 struct ipcp_agg_lattice
*aglat
;
2493 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
2495 if (src
->offset
== -1)
2496 return (ipa_lat_is_single_const (&plats
->itself
)
2497 && values_equal_for_ipcp_p (src
->val
->value
,
2498 plats
->itself
.values
->value
));
2501 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
2503 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2504 if (aglat
->offset
== src
->offset
)
2505 return (ipa_lat_is_single_const (aglat
)
2506 && values_equal_for_ipcp_p (src
->val
->value
,
2507 aglat
->values
->value
));
2513 /* Get the next clone in the linked list of clones of an edge. */
2515 static inline struct cgraph_edge
*
2516 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
2518 return next_edge_clone
[cs
->uid
];
2521 /* Given VAL, iterate over all its sources and if they still hold, add their
2522 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2526 get_info_about_necessary_edges (struct ipcp_value
*val
, int *freq_sum
,
2527 gcov_type
*count_sum
, int *caller_count
)
2529 struct ipcp_value_source
*src
;
2530 int freq
= 0, count
= 0;
2534 for (src
= val
->sources
; src
; src
= src
->next
)
2536 struct cgraph_edge
*cs
= src
->cs
;
2539 if (cgraph_edge_brings_value_p (cs
, src
))
2542 freq
+= cs
->frequency
;
2544 hot
|= cgraph_maybe_hot_edge_p (cs
);
2546 cs
= get_next_cgraph_edge_clone (cs
);
2552 *caller_count
= count
;
2556 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2557 their number is known and equal to CALLER_COUNT. */
2559 static vec
<cgraph_edge_p
>
2560 gather_edges_for_value (struct ipcp_value
*val
, int caller_count
)
2562 struct ipcp_value_source
*src
;
2563 vec
<cgraph_edge_p
> ret
;
2565 ret
.create (caller_count
);
2566 for (src
= val
->sources
; src
; src
= src
->next
)
2568 struct cgraph_edge
*cs
= src
->cs
;
2571 if (cgraph_edge_brings_value_p (cs
, src
))
2572 ret
.quick_push (cs
);
2573 cs
= get_next_cgraph_edge_clone (cs
);
2580 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2581 Return it or NULL if for some reason it cannot be created. */
2583 static struct ipa_replace_map
*
2584 get_replacement_map (struct ipa_node_params
*info
, tree value
, int parm_num
)
2586 struct ipa_replace_map
*replace_map
;
2589 replace_map
= ggc_alloc_ipa_replace_map ();
2592 fprintf (dump_file
, " replacing ");
2593 ipa_dump_param (dump_file
, info
, parm_num
);
2595 fprintf (dump_file
, " with const ");
2596 print_generic_expr (dump_file
, value
, 0);
2597 fprintf (dump_file
, "\n");
2599 replace_map
->old_tree
= NULL
;
2600 replace_map
->parm_num
= parm_num
;
2601 replace_map
->new_tree
= value
;
2602 replace_map
->replace_p
= true;
2603 replace_map
->ref_p
= false;
2608 /* Dump new profiling counts */
2611 dump_profile_updates (struct cgraph_node
*orig_node
,
2612 struct cgraph_node
*new_node
)
2614 struct cgraph_edge
*cs
;
2616 fprintf (dump_file
, " setting count of the specialized node to "
2617 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) new_node
->count
);
2618 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2619 fprintf (dump_file
, " edge to %s has count "
2620 HOST_WIDE_INT_PRINT_DEC
"\n",
2621 cs
->callee
->name (), (HOST_WIDE_INT
) cs
->count
);
2623 fprintf (dump_file
, " setting count of the original node to "
2624 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) orig_node
->count
);
2625 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2626 fprintf (dump_file
, " edge to %s is left with "
2627 HOST_WIDE_INT_PRINT_DEC
"\n",
2628 cs
->callee
->name (), (HOST_WIDE_INT
) cs
->count
);
2631 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2632 their profile information to reflect this. */
2635 update_profiling_info (struct cgraph_node
*orig_node
,
2636 struct cgraph_node
*new_node
)
2638 struct cgraph_edge
*cs
;
2639 struct caller_statistics stats
;
2640 gcov_type new_sum
, orig_sum
;
2641 gcov_type remainder
, orig_node_count
= orig_node
->count
;
2643 if (orig_node_count
== 0)
2646 init_caller_stats (&stats
);
2647 cgraph_for_node_and_aliases (orig_node
, gather_caller_stats
, &stats
, false);
2648 orig_sum
= stats
.count_sum
;
2649 init_caller_stats (&stats
);
2650 cgraph_for_node_and_aliases (new_node
, gather_caller_stats
, &stats
, false);
2651 new_sum
= stats
.count_sum
;
2653 if (orig_node_count
< orig_sum
+ new_sum
)
2656 fprintf (dump_file
, " Problem: node %s/%i has too low count "
2657 HOST_WIDE_INT_PRINT_DEC
" while the sum of incoming "
2658 "counts is " HOST_WIDE_INT_PRINT_DEC
"\n",
2659 orig_node
->name (), orig_node
->order
,
2660 (HOST_WIDE_INT
) orig_node_count
,
2661 (HOST_WIDE_INT
) (orig_sum
+ new_sum
));
2663 orig_node_count
= (orig_sum
+ new_sum
) * 12 / 10;
2665 fprintf (dump_file
, " proceeding by pretending it was "
2666 HOST_WIDE_INT_PRINT_DEC
"\n",
2667 (HOST_WIDE_INT
) orig_node_count
);
2670 new_node
->count
= new_sum
;
2671 remainder
= orig_node_count
- new_sum
;
2672 orig_node
->count
= remainder
;
2674 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2676 cs
->count
= apply_probability (cs
->count
,
2677 GCOV_COMPUTE_SCALE (new_sum
,
2682 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2683 cs
->count
= apply_probability (cs
->count
,
2684 GCOV_COMPUTE_SCALE (remainder
,
2688 dump_profile_updates (orig_node
, new_node
);
2691 /* Update the respective profile of specialized NEW_NODE and the original
2692 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2693 have been redirected to the specialized version. */
2696 update_specialized_profile (struct cgraph_node
*new_node
,
2697 struct cgraph_node
*orig_node
,
2698 gcov_type redirected_sum
)
2700 struct cgraph_edge
*cs
;
2701 gcov_type new_node_count
, orig_node_count
= orig_node
->count
;
2704 fprintf (dump_file
, " the sum of counts of redirected edges is "
2705 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) redirected_sum
);
2706 if (orig_node_count
== 0)
2709 gcc_assert (orig_node_count
>= redirected_sum
);
2711 new_node_count
= new_node
->count
;
2712 new_node
->count
+= redirected_sum
;
2713 orig_node
->count
-= redirected_sum
;
2715 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2717 cs
->count
+= apply_probability (cs
->count
,
2718 GCOV_COMPUTE_SCALE (redirected_sum
,
2723 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2725 gcov_type dec
= apply_probability (cs
->count
,
2726 GCOV_COMPUTE_SCALE (redirected_sum
,
2728 if (dec
< cs
->count
)
2735 dump_profile_updates (orig_node
, new_node
);
2738 /* Create a specialized version of NODE with known constants and types of
2739 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2741 static struct cgraph_node
*
2742 create_specialized_node (struct cgraph_node
*node
,
2743 vec
<tree
> known_vals
,
2744 struct ipa_agg_replacement_value
*aggvals
,
2745 vec
<cgraph_edge_p
> callers
)
2747 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
2748 vec
<ipa_replace_map_p
, va_gc
> *replace_trees
= NULL
;
2749 struct ipa_agg_replacement_value
*av
;
2750 struct cgraph_node
*new_node
;
2751 int i
, count
= ipa_get_param_count (info
);
2752 bitmap args_to_skip
;
2754 gcc_assert (!info
->ipcp_orig_node
);
2756 if (node
->local
.can_change_signature
)
2758 args_to_skip
= BITMAP_GGC_ALLOC ();
2759 for (i
= 0; i
< count
; i
++)
2761 tree t
= known_vals
[i
];
2763 if ((t
&& TREE_CODE (t
) != TREE_BINFO
)
2764 || !ipa_is_param_used (info
, i
))
2765 bitmap_set_bit (args_to_skip
, i
);
2770 args_to_skip
= NULL
;
2771 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2772 fprintf (dump_file
, " cannot change function signature\n");
2775 for (i
= 0; i
< count
; i
++)
2777 tree t
= known_vals
[i
];
2778 if (t
&& TREE_CODE (t
) != TREE_BINFO
)
2780 struct ipa_replace_map
*replace_map
;
2782 replace_map
= get_replacement_map (info
, t
, i
);
2784 vec_safe_push (replace_trees
, replace_map
);
2788 new_node
= cgraph_create_virtual_clone (node
, callers
, replace_trees
,
2789 args_to_skip
, "constprop");
2790 ipa_set_node_agg_value_chain (new_node
, aggvals
);
2791 for (av
= aggvals
; av
; av
= av
->next
)
2792 ipa_maybe_record_reference (new_node
, av
->value
,
2793 IPA_REF_ADDR
, NULL
);
2795 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2797 fprintf (dump_file
, " the new node is %s/%i.\n",
2798 new_node
->name (), new_node
->order
);
2800 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
2802 ipa_check_create_node_params ();
2803 update_profiling_info (node
, new_node
);
2804 new_info
= IPA_NODE_REF (new_node
);
2805 new_info
->ipcp_orig_node
= node
;
2806 new_info
->known_vals
= known_vals
;
2808 ipcp_discover_new_direct_edges (new_node
, known_vals
, aggvals
);
2814 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2815 KNOWN_VALS with constants and types that are also known for all of the
2819 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
2820 vec
<tree
> known_vals
,
2821 vec
<cgraph_edge_p
> callers
)
2823 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2824 int i
, count
= ipa_get_param_count (info
);
2826 for (i
= 0; i
< count
; i
++)
2828 struct cgraph_edge
*cs
;
2829 tree newval
= NULL_TREE
;
2832 if (ipa_get_scalar_lat (info
, i
)->bottom
|| known_vals
[i
])
2835 FOR_EACH_VEC_ELT (callers
, j
, cs
)
2837 struct ipa_jump_func
*jump_func
;
2840 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
2845 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
2846 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
);
2849 && !values_equal_for_ipcp_p (t
, newval
)))
2860 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2862 fprintf (dump_file
, " adding an extra known scalar value ");
2863 print_ipcp_constant_value (dump_file
, newval
);
2864 fprintf (dump_file
, " for ");
2865 ipa_dump_param (dump_file
, info
, i
);
2866 fprintf (dump_file
, "\n");
2869 known_vals
[i
] = newval
;
2874 /* Go through PLATS and create a vector of values consisting of values and
2875 offsets (minus OFFSET) of lattices that contain only a single value. */
2877 static vec
<ipa_agg_jf_item
>
2878 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
2880 vec
<ipa_agg_jf_item
> res
= vNULL
;
2882 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2885 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2886 if (ipa_lat_is_single_const (aglat
))
2888 struct ipa_agg_jf_item ti
;
2889 ti
.offset
= aglat
->offset
- offset
;
2890 ti
.value
= aglat
->values
->value
;
2896 /* Intersect all values in INTER with single value lattices in PLATS (while
2897 subtracting OFFSET). */
2900 intersect_with_plats (struct ipcp_param_lattices
*plats
,
2901 vec
<ipa_agg_jf_item
> *inter
,
2902 HOST_WIDE_INT offset
)
2904 struct ipcp_agg_lattice
*aglat
;
2905 struct ipa_agg_jf_item
*item
;
2908 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2914 aglat
= plats
->aggs
;
2915 FOR_EACH_VEC_ELT (*inter
, k
, item
)
2922 if (aglat
->offset
- offset
> item
->offset
)
2924 if (aglat
->offset
- offset
== item
->offset
)
2926 gcc_checking_assert (item
->value
);
2927 if (values_equal_for_ipcp_p (item
->value
, aglat
->values
->value
))
2931 aglat
= aglat
->next
;
2934 item
->value
= NULL_TREE
;
2938 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2939 vector result while subtracting OFFSET from the individual value offsets. */
2941 static vec
<ipa_agg_jf_item
>
2942 agg_replacements_to_vector (struct cgraph_node
*node
, int index
,
2943 HOST_WIDE_INT offset
)
2945 struct ipa_agg_replacement_value
*av
;
2946 vec
<ipa_agg_jf_item
> res
= vNULL
;
2948 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
2949 if (av
->index
== index
2950 && (av
->offset
- offset
) >= 0)
2952 struct ipa_agg_jf_item item
;
2953 gcc_checking_assert (av
->value
);
2954 item
.offset
= av
->offset
- offset
;
2955 item
.value
= av
->value
;
2956 res
.safe_push (item
);
2962 /* Intersect all values in INTER with those that we have already scheduled to
2963 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2964 (while subtracting OFFSET). */
2967 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
2968 vec
<ipa_agg_jf_item
> *inter
,
2969 HOST_WIDE_INT offset
)
2971 struct ipa_agg_replacement_value
*srcvals
;
2972 struct ipa_agg_jf_item
*item
;
2975 srcvals
= ipa_get_agg_replacements_for_node (node
);
2982 FOR_EACH_VEC_ELT (*inter
, i
, item
)
2984 struct ipa_agg_replacement_value
*av
;
2988 for (av
= srcvals
; av
; av
= av
->next
)
2990 gcc_checking_assert (av
->value
);
2991 if (av
->index
== index
2992 && av
->offset
- offset
== item
->offset
)
2994 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
3000 item
->value
= NULL_TREE
;
3004 /* Intersect values in INTER with aggregate values that come along edge CS to
3005 parameter number INDEX and return it. If INTER does not actually exist yet,
3006 copy all incoming values to it. If we determine we ended up with no values
3007 whatsoever, return a released vector. */
3009 static vec
<ipa_agg_jf_item
>
3010 intersect_aggregates_with_edge (struct cgraph_edge
*cs
, int index
,
3011 vec
<ipa_agg_jf_item
> inter
)
3013 struct ipa_jump_func
*jfunc
;
3014 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), index
);
3015 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3016 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3018 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3019 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
3021 if (caller_info
->ipcp_orig_node
)
3023 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
3024 struct ipcp_param_lattices
*orig_plats
;
3025 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
3027 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
3029 if (!inter
.exists ())
3030 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, 0);
3032 intersect_with_agg_replacements (cs
->caller
, src_idx
,
3043 struct ipcp_param_lattices
*src_plats
;
3044 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
3045 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
3047 /* Currently we do not produce clobber aggregate jump
3048 functions, adjust when we do. */
3049 gcc_checking_assert (!jfunc
->agg
.items
);
3050 if (!inter
.exists ())
3051 inter
= copy_plats_to_inter (src_plats
, 0);
3053 intersect_with_plats (src_plats
, &inter
, 0);
3062 else if (jfunc
->type
== IPA_JF_ANCESTOR
3063 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
3065 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3066 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
3067 struct ipcp_param_lattices
*src_plats
;
3068 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
3070 if (caller_info
->ipcp_orig_node
)
3072 if (!inter
.exists ())
3073 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, delta
);
3075 intersect_with_agg_replacements (cs
->caller
, src_idx
, &inter
,
3080 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);;
3081 /* Currently we do not produce clobber aggregate jump
3082 functions, adjust when we do. */
3083 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
3084 if (!inter
.exists ())
3085 inter
= copy_plats_to_inter (src_plats
, delta
);
3087 intersect_with_plats (src_plats
, &inter
, delta
);
3090 else if (jfunc
->agg
.items
)
3092 struct ipa_agg_jf_item
*item
;
3095 if (!inter
.exists ())
3096 for (unsigned i
= 0; i
< jfunc
->agg
.items
->length (); i
++)
3097 inter
.safe_push ((*jfunc
->agg
.items
)[i
]);
3099 FOR_EACH_VEC_ELT (inter
, k
, item
)
3102 bool found
= false;;
3107 while ((unsigned) l
< jfunc
->agg
.items
->length ())
3109 struct ipa_agg_jf_item
*ti
;
3110 ti
= &(*jfunc
->agg
.items
)[l
];
3111 if (ti
->offset
> item
->offset
)
3113 if (ti
->offset
== item
->offset
)
3115 gcc_checking_assert (ti
->value
);
3116 if (values_equal_for_ipcp_p (item
->value
,
3130 return vec
<ipa_agg_jf_item
>();
3135 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3136 from all of them. */
3138 static struct ipa_agg_replacement_value
*
3139 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
3140 vec
<cgraph_edge_p
> callers
)
3142 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3143 struct ipa_agg_replacement_value
*res
;
3144 struct ipa_agg_replacement_value
**tail
= &res
;
3145 struct cgraph_edge
*cs
;
3146 int i
, j
, count
= ipa_get_param_count (dest_info
);
3148 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3150 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3155 for (i
= 0; i
< count
; i
++)
3157 struct cgraph_edge
*cs
;
3158 vec
<ipa_agg_jf_item
> inter
= vNULL
;
3159 struct ipa_agg_jf_item
*item
;
3160 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (dest_info
, i
);
3163 /* Among other things, the following check should deal with all by_ref
3165 if (plats
->aggs_bottom
)
3168 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3170 inter
= intersect_aggregates_with_edge (cs
, i
, inter
);
3172 if (!inter
.exists ())
3176 FOR_EACH_VEC_ELT (inter
, j
, item
)
3178 struct ipa_agg_replacement_value
*v
;
3183 v
= ggc_alloc_ipa_agg_replacement_value ();
3185 v
->offset
= item
->offset
;
3186 v
->value
= item
->value
;
3187 v
->by_ref
= plats
->aggs_by_ref
;
3193 if (inter
.exists ())
3200 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3202 static struct ipa_agg_replacement_value
*
3203 known_aggs_to_agg_replacement_list (vec
<ipa_agg_jump_function
> known_aggs
)
3205 struct ipa_agg_replacement_value
*res
;
3206 struct ipa_agg_replacement_value
**tail
= &res
;
3207 struct ipa_agg_jump_function
*aggjf
;
3208 struct ipa_agg_jf_item
*item
;
3211 FOR_EACH_VEC_ELT (known_aggs
, i
, aggjf
)
3212 FOR_EACH_VEC_SAFE_ELT (aggjf
->items
, j
, item
)
3214 struct ipa_agg_replacement_value
*v
;
3215 v
= ggc_alloc_ipa_agg_replacement_value ();
3217 v
->offset
= item
->offset
;
3218 v
->value
= item
->value
;
3219 v
->by_ref
= aggjf
->by_ref
;
3227 /* Determine whether CS also brings all scalar values that the NODE is
3231 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
3232 struct cgraph_node
*node
)
3234 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3235 int count
= ipa_get_param_count (dest_info
);
3236 struct ipa_node_params
*caller_info
;
3237 struct ipa_edge_args
*args
;
3240 caller_info
= IPA_NODE_REF (cs
->caller
);
3241 args
= IPA_EDGE_REF (cs
);
3242 for (i
= 0; i
< count
; i
++)
3244 struct ipa_jump_func
*jump_func
;
3247 val
= dest_info
->known_vals
[i
];
3251 if (i
>= ipa_get_cs_argument_count (args
))
3253 jump_func
= ipa_get_ith_jump_func (args
, i
);
3254 t
= ipa_value_from_jfunc (caller_info
, jump_func
);
3255 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
3261 /* Determine whether CS also brings all aggregate values that NODE is
3264 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
3265 struct cgraph_node
*node
)
3267 struct ipa_node_params
*orig_caller_info
= IPA_NODE_REF (cs
->caller
);
3268 struct ipa_node_params
*orig_node_info
;
3269 struct ipa_agg_replacement_value
*aggval
;
3272 aggval
= ipa_get_agg_replacements_for_node (node
);
3276 count
= ipa_get_param_count (IPA_NODE_REF (node
));
3277 ec
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3279 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3280 if (aggval
->index
>= ec
)
3283 orig_node_info
= IPA_NODE_REF (IPA_NODE_REF (node
)->ipcp_orig_node
);
3284 if (orig_caller_info
->ipcp_orig_node
)
3285 orig_caller_info
= IPA_NODE_REF (orig_caller_info
->ipcp_orig_node
);
3287 for (i
= 0; i
< count
; i
++)
3289 static vec
<ipa_agg_jf_item
> values
= vec
<ipa_agg_jf_item
>();
3290 struct ipcp_param_lattices
*plats
;
3291 bool interesting
= false;
3292 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3293 if (aggval
->index
== i
)
3301 plats
= ipa_get_parm_lattices (orig_node_info
, aggval
->index
);
3302 if (plats
->aggs_bottom
)
3305 values
= intersect_aggregates_with_edge (cs
, i
, values
);
3306 if (!values
.exists ())
3309 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3310 if (aggval
->index
== i
)
3312 struct ipa_agg_jf_item
*item
;
3315 FOR_EACH_VEC_ELT (values
, j
, item
)
3317 && item
->offset
== av
->offset
3318 && values_equal_for_ipcp_p (item
->value
, av
->value
))
3333 /* Given an original NODE and a VAL for which we have already created a
3334 specialized clone, look whether there are incoming edges that still lead
3335 into the old node but now also bring the requested value and also conform to
3336 all other criteria such that they can be redirected the the special node.
3337 This function can therefore redirect the final edge in a SCC. */
3340 perhaps_add_new_callers (struct cgraph_node
*node
, struct ipcp_value
*val
)
3342 struct ipcp_value_source
*src
;
3343 gcov_type redirected_sum
= 0;
3345 for (src
= val
->sources
; src
; src
= src
->next
)
3347 struct cgraph_edge
*cs
= src
->cs
;
3350 enum availability availability
;
3351 struct cgraph_node
*dst
= cgraph_function_node (cs
->callee
,
3353 if ((dst
== node
|| IPA_NODE_REF (dst
)->is_all_contexts_clone
)
3354 && availability
> AVAIL_OVERWRITABLE
3355 && cgraph_edge_brings_value_p (cs
, src
))
3357 if (cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
3358 && cgraph_edge_brings_all_agg_vals_for_node (cs
,
3362 fprintf (dump_file
, " - adding an extra caller %s/%i"
3364 xstrdup (cs
->caller
->name ()),
3366 xstrdup (val
->spec_node
->name ()),
3367 val
->spec_node
->order
);
3369 cgraph_redirect_edge_callee (cs
, val
->spec_node
);
3370 redirected_sum
+= cs
->count
;
3373 cs
= get_next_cgraph_edge_clone (cs
);
3378 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
3382 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3385 move_binfos_to_values (vec
<tree
> known_vals
,
3386 vec
<tree
> known_binfos
)
3391 for (i
= 0; known_binfos
.iterate (i
, &t
); i
++)
3396 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3397 among those in the AGGVALS list. */
3400 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value
*aggvals
,
3401 int index
, HOST_WIDE_INT offset
, tree value
)
3405 if (aggvals
->index
== index
3406 && aggvals
->offset
== offset
3407 && values_equal_for_ipcp_p (aggvals
->value
, value
))
3409 aggvals
= aggvals
->next
;
3414 /* Decide wheter to create a special version of NODE for value VAL of parameter
3415 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3416 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3417 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3420 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
3421 struct ipcp_value
*val
, vec
<tree
> known_csts
,
3422 vec
<tree
> known_binfos
)
3424 struct ipa_agg_replacement_value
*aggvals
;
3425 int freq_sum
, caller_count
;
3426 gcov_type count_sum
;
3427 vec
<cgraph_edge_p
> callers
;
3432 perhaps_add_new_callers (node
, val
);
3435 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
3437 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3438 fprintf (dump_file
, " Ignoring candidate value because "
3439 "max_new_size would be reached with %li.\n",
3440 val
->local_size_cost
+ overall_size
);
3443 else if (!get_info_about_necessary_edges (val
, &freq_sum
, &count_sum
,
3447 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3449 fprintf (dump_file
, " - considering value ");
3450 print_ipcp_constant_value (dump_file
, val
->value
);
3451 fprintf (dump_file
, " for ");
3452 ipa_dump_param (dump_file
, IPA_NODE_REF (node
), index
);
3454 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
3455 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
3458 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
3459 freq_sum
, count_sum
,
3460 val
->local_size_cost
)
3461 && !good_cloning_opportunity_p (node
,
3462 val
->local_time_benefit
3463 + val
->prop_time_benefit
,
3464 freq_sum
, count_sum
,
3465 val
->local_size_cost
3466 + val
->prop_size_cost
))
3470 fprintf (dump_file
, " Creating a specialized node of %s/%i.\n",
3471 node
->name (), node
->order
);
3473 callers
= gather_edges_for_value (val
, caller_count
);
3474 kv
= known_csts
.copy ();
3475 move_binfos_to_values (kv
, known_binfos
);
3477 kv
[index
] = val
->value
;
3478 find_more_scalar_values_for_callers_subset (node
, kv
, callers
);
3479 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
3480 gcc_checking_assert (offset
== -1
3481 || ipcp_val_in_agg_replacements_p (aggvals
, index
,
3482 offset
, val
->value
));
3483 val
->spec_node
= create_specialized_node (node
, kv
, aggvals
, callers
);
3484 overall_size
+= val
->local_size_cost
;
3486 /* TODO: If for some lattice there is only one other known value
3487 left, make a special node for it too. */
3492 /* Decide whether and what specialized clones of NODE should be created. */
3495 decide_whether_version_node (struct cgraph_node
*node
)
3497 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3498 int i
, count
= ipa_get_param_count (info
);
3499 vec
<tree
> known_csts
, known_binfos
;
3500 vec
<ipa_agg_jump_function
> known_aggs
= vNULL
;
3506 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3507 fprintf (dump_file
, "\nEvaluating opportunities for %s/%i.\n",
3508 node
->name (), node
->order
);
3510 gather_context_independent_values (info
, &known_csts
, &known_binfos
,
3511 info
->do_clone_for_all_contexts
? &known_aggs
3514 for (i
= 0; i
< count
;i
++)
3516 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3517 struct ipcp_lattice
*lat
= &plats
->itself
;
3518 struct ipcp_value
*val
;
3522 && !known_binfos
[i
])
3523 for (val
= lat
->values
; val
; val
= val
->next
)
3524 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
3527 if (!plats
->aggs_bottom
)
3529 struct ipcp_agg_lattice
*aglat
;
3530 struct ipcp_value
*val
;
3531 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3532 if (!aglat
->bottom
&& aglat
->values
3533 /* If the following is false, the one value is in
3535 && (plats
->aggs_contain_variable
3536 || !ipa_lat_is_single_const (aglat
)))
3537 for (val
= aglat
->values
; val
; val
= val
->next
)
3538 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
3539 known_csts
, known_binfos
);
3541 info
= IPA_NODE_REF (node
);
3544 if (info
->do_clone_for_all_contexts
)
3546 struct cgraph_node
*clone
;
3547 vec
<cgraph_edge_p
> callers
;
3550 fprintf (dump_file
, " - Creating a specialized node of %s/%i "
3551 "for all known contexts.\n", node
->name (),
3554 callers
= collect_callers_of_node (node
);
3555 move_binfos_to_values (known_csts
, known_binfos
);
3556 clone
= create_specialized_node (node
, known_csts
,
3557 known_aggs_to_agg_replacement_list (known_aggs
),
3559 info
= IPA_NODE_REF (node
);
3560 info
->do_clone_for_all_contexts
= false;
3561 IPA_NODE_REF (clone
)->is_all_contexts_clone
= true;
3562 for (i
= 0; i
< count
; i
++)
3563 vec_free (known_aggs
[i
].items
);
3564 known_aggs
.release ();
3568 known_csts
.release ();
3570 known_binfos
.release ();
3574 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3577 spread_undeadness (struct cgraph_node
*node
)
3579 struct cgraph_edge
*cs
;
3581 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
3582 if (ipa_edge_within_scc (cs
))
3584 struct cgraph_node
*callee
;
3585 struct ipa_node_params
*info
;
3587 callee
= cgraph_function_node (cs
->callee
, NULL
);
3588 info
= IPA_NODE_REF (callee
);
3590 if (info
->node_dead
)
3592 info
->node_dead
= 0;
3593 spread_undeadness (callee
);
3598 /* Return true if NODE has a caller from outside of its SCC that is not
3599 dead. Worker callback for cgraph_for_node_and_aliases. */
3602 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
3603 void *data ATTRIBUTE_UNUSED
)
3605 struct cgraph_edge
*cs
;
3607 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
3608 if (cs
->caller
->thunk
.thunk_p
3609 && cgraph_for_node_and_aliases (cs
->caller
,
3610 has_undead_caller_from_outside_scc_p
,
3613 else if (!ipa_edge_within_scc (cs
)
3614 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
3620 /* Identify nodes within the same SCC as NODE which are no longer needed
3621 because of new clones and will be removed as unreachable. */
3624 identify_dead_nodes (struct cgraph_node
*node
)
3626 struct cgraph_node
*v
;
3627 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3628 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v
)
3629 && !cgraph_for_node_and_aliases (v
,
3630 has_undead_caller_from_outside_scc_p
,
3632 IPA_NODE_REF (v
)->node_dead
= 1;
3634 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3635 if (!IPA_NODE_REF (v
)->node_dead
)
3636 spread_undeadness (v
);
3638 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3640 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3641 if (IPA_NODE_REF (v
)->node_dead
)
3642 fprintf (dump_file
, " Marking node as dead: %s/%i.\n",
3643 v
->name (), v
->order
);
3647 /* The decision stage. Iterate over the topological order of call graph nodes
3648 TOPO and make specialized clones if deemed beneficial. */
3651 ipcp_decision_stage (struct topo_info
*topo
)
3656 fprintf (dump_file
, "\nIPA decision stage:\n\n");
3658 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3660 struct cgraph_node
*node
= topo
->order
[i
];
3661 bool change
= false, iterate
= true;
3665 struct cgraph_node
*v
;
3667 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3668 if (cgraph_function_with_gimple_body_p (v
)
3669 && ipcp_versionable_function_p (v
))
3670 iterate
|= decide_whether_version_node (v
);
3675 identify_dead_nodes (node
);
3679 /* The IPCP driver. */
3684 struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
3685 struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
3686 struct topo_info topo
;
3688 ipa_check_create_node_params ();
3689 ipa_check_create_edge_args ();
3690 grow_edge_clone_vectors ();
3691 edge_duplication_hook_holder
=
3692 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook
, NULL
);
3693 edge_removal_hook_holder
=
3694 cgraph_add_edge_removal_hook (&ipcp_edge_removal_hook
, NULL
);
3696 ipcp_values_pool
= create_alloc_pool ("IPA-CP values",
3697 sizeof (struct ipcp_value
), 32);
3698 ipcp_sources_pool
= create_alloc_pool ("IPA-CP value sources",
3699 sizeof (struct ipcp_value_source
), 64);
3700 ipcp_agg_lattice_pool
= create_alloc_pool ("IPA_CP aggregate lattices",
3701 sizeof (struct ipcp_agg_lattice
),
3705 fprintf (dump_file
, "\nIPA structures before propagation:\n");
3706 if (dump_flags
& TDF_DETAILS
)
3707 ipa_print_all_params (dump_file
);
3708 ipa_print_all_jump_functions (dump_file
);
3711 /* Topological sort. */
3712 build_toporder_info (&topo
);
3713 /* Do the interprocedural propagation. */
3714 ipcp_propagate_stage (&topo
);
3715 /* Decide what constant propagation and cloning should be performed. */
3716 ipcp_decision_stage (&topo
);
3718 /* Free all IPCP structures. */
3719 free_toporder_info (&topo
);
3720 next_edge_clone
.release ();
3721 cgraph_remove_edge_removal_hook (edge_removal_hook_holder
);
3722 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder
);
3723 ipa_free_all_structures_after_ipa_cp ();
3725 fprintf (dump_file
, "\nIPA constant propagation end\n");
3729 /* Initialization and computation of IPCP data structures. This is the initial
3730 intraprocedural analysis of functions, which gathers information to be
3731 propagated later on. */
3734 ipcp_generate_summary (void)
3736 struct cgraph_node
*node
;
3739 fprintf (dump_file
, "\nIPA constant propagation start:\n");
3740 ipa_register_cgraph_hooks ();
3742 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
3744 node
->local
.versionable
3745 = tree_versionable_function_p (node
->decl
);
3746 ipa_analyze_node (node
);
3750 /* Write ipcp summary for nodes in SET. */
3753 ipcp_write_summary (void)
3755 ipa_prop_write_jump_functions ();
3758 /* Read ipcp summary. */
3761 ipcp_read_summary (void)
3763 ipa_prop_read_jump_functions ();
3766 /* Gate for IPCP optimization. */
3769 cgraph_gate_cp (void)
3771 /* FIXME: We should remove the optimize check after we ensure we never run
3772 IPA passes when not optimizing. */
3773 return flag_ipa_cp
&& optimize
;
3778 const pass_data pass_data_ipa_cp
=
3780 IPA_PASS
, /* type */
3782 OPTGROUP_NONE
, /* optinfo_flags */
3783 true, /* has_gate */
3784 true, /* has_execute */
3785 TV_IPA_CONSTANT_PROP
, /* tv_id */
3786 0, /* properties_required */
3787 0, /* properties_provided */
3788 0, /* properties_destroyed */
3789 0, /* todo_flags_start */
3790 ( TODO_dump_symtab
| TODO_remove_functions
), /* todo_flags_finish */
3793 class pass_ipa_cp
: public ipa_opt_pass_d
3796 pass_ipa_cp (gcc::context
*ctxt
)
3797 : ipa_opt_pass_d (pass_data_ipa_cp
, ctxt
,
3798 ipcp_generate_summary
, /* generate_summary */
3799 ipcp_write_summary
, /* write_summary */
3800 ipcp_read_summary
, /* read_summary */
3801 ipa_prop_write_all_agg_replacement
, /*
3802 write_optimization_summary */
3803 ipa_prop_read_all_agg_replacement
, /*
3804 read_optimization_summary */
3805 NULL
, /* stmt_fixup */
3806 0, /* function_transform_todo_flags_start */
3807 ipcp_transform_function
, /* function_transform */
3808 NULL
) /* variable_transform */
3811 /* opt_pass methods: */
3812 bool gate () { return cgraph_gate_cp (); }
3813 unsigned int execute () { return ipcp_driver (); }
3815 }; // class pass_ipa_cp
3820 make_pass_ipa_cp (gcc::context
*ctxt
)
3822 return new pass_ipa_cp (ctxt
);