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 (node
->get_availability () <= AVAIL_INTERPOSABLE
)
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 (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node
->decl
)))
438 /* Ideally we should clone the SIMD clones themselves and create
439 vector copies of them, so IPA-cp and SIMD clones can happily
440 coexist, but that may not be worth the effort. */
441 reason
= "function has SIMD clones";
443 /* Don't clone decls local to a comdat group; it breaks and for C++
444 decloned constructors, inlining is always better anyway. */
445 else if (node
->comdat_local_p ())
446 reason
= "comdat-local function";
448 if (reason
&& dump_file
&& !node
->alias
&& !node
->thunk
.thunk_p
)
449 fprintf (dump_file
, "Function %s/%i is not versionable, reason: %s.\n",
450 node
->name (), node
->order
, reason
);
452 node
->local
.versionable
= (reason
== NULL
);
455 /* Return true if it is at all technically possible to create clones of a
459 ipcp_versionable_function_p (struct cgraph_node
*node
)
461 return node
->local
.versionable
;
464 /* Structure holding accumulated information about callers of a node. */
466 struct caller_statistics
469 int n_calls
, n_hot_calls
, freq_sum
;
472 /* Initialize fields of STAT to zeroes. */
475 init_caller_stats (struct caller_statistics
*stats
)
477 stats
->count_sum
= 0;
479 stats
->n_hot_calls
= 0;
483 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
484 non-thunk incoming edges to NODE. */
487 gather_caller_stats (struct cgraph_node
*node
, void *data
)
489 struct caller_statistics
*stats
= (struct caller_statistics
*) data
;
490 struct cgraph_edge
*cs
;
492 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
493 if (cs
->caller
->thunk
.thunk_p
)
494 cs
->caller
->call_for_symbol_thunks_and_aliases (gather_caller_stats
,
498 stats
->count_sum
+= cs
->count
;
499 stats
->freq_sum
+= cs
->frequency
;
501 if (cs
->maybe_hot_p ())
502 stats
->n_hot_calls
++;
508 /* Return true if this NODE is viable candidate for cloning. */
511 ipcp_cloning_candidate_p (struct cgraph_node
*node
)
513 struct caller_statistics stats
;
515 gcc_checking_assert (node
->has_gimple_body_p ());
517 if (!flag_ipa_cp_clone
)
520 fprintf (dump_file
, "Not considering %s for cloning; "
521 "-fipa-cp-clone disabled.\n",
526 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->decl
)))
529 fprintf (dump_file
, "Not considering %s for cloning; "
530 "optimizing it for size.\n",
535 init_caller_stats (&stats
);
536 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
, false);
538 if (inline_summary (node
)->self_size
< stats
.n_calls
)
541 fprintf (dump_file
, "Considering %s for cloning; code might shrink.\n",
546 /* When profile is available and function is hot, propagate into it even if
547 calls seems cold; constant propagation can improve function's speed
551 if (stats
.count_sum
> node
->count
* 90 / 100)
554 fprintf (dump_file
, "Considering %s for cloning; "
555 "usually called directly.\n",
560 if (!stats
.n_hot_calls
)
563 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
568 fprintf (dump_file
, "Considering %s for cloning.\n",
573 /* Arrays representing a topological ordering of call graph nodes and a stack
574 of noes used during constant propagation. */
578 struct cgraph_node
**order
;
579 struct cgraph_node
**stack
;
580 int nnodes
, stack_top
;
583 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
586 build_toporder_info (struct ipa_topo_info
*topo
)
588 topo
->order
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
589 topo
->stack
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
592 topo
->nnodes
= ipa_reduced_postorder (topo
->order
, true, true, NULL
);
595 /* Free information about strongly connected components and the arrays in
599 free_toporder_info (struct ipa_topo_info
*topo
)
601 ipa_free_postorder_info ();
606 /* Add NODE to the stack in TOPO, unless it is already there. */
609 push_node_to_stack (struct ipa_topo_info
*topo
, struct cgraph_node
*node
)
611 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
612 if (info
->node_enqueued
)
614 info
->node_enqueued
= 1;
615 topo
->stack
[topo
->stack_top
++] = node
;
618 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
621 static struct cgraph_node
*
622 pop_node_from_stack (struct ipa_topo_info
*topo
)
626 struct cgraph_node
*node
;
628 node
= topo
->stack
[topo
->stack_top
];
629 IPA_NODE_REF (node
)->node_enqueued
= 0;
636 /* Set lattice LAT to bottom and return true if it previously was not set as
640 set_lattice_to_bottom (struct ipcp_lattice
*lat
)
642 bool ret
= !lat
->bottom
;
647 /* Mark lattice as containing an unknown value and return true if it previously
648 was not marked as such. */
651 set_lattice_contains_variable (struct ipcp_lattice
*lat
)
653 bool ret
= !lat
->contains_variable
;
654 lat
->contains_variable
= true;
658 /* Set all aggegate lattices in PLATS to bottom and return true if they were
659 not previously set as such. */
662 set_agg_lats_to_bottom (struct ipcp_param_lattices
*plats
)
664 bool ret
= !plats
->aggs_bottom
;
665 plats
->aggs_bottom
= true;
669 /* Mark all aggegate lattices in PLATS as containing an unknown value and
670 return true if they were not previously marked as such. */
673 set_agg_lats_contain_variable (struct ipcp_param_lattices
*plats
)
675 bool ret
= !plats
->aggs_contain_variable
;
676 plats
->aggs_contain_variable
= true;
680 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
681 return true is any of them has not been marked as such so far. */
684 set_all_contains_variable (struct ipcp_param_lattices
*plats
)
686 bool ret
= !plats
->itself
.contains_variable
|| !plats
->aggs_contain_variable
;
687 plats
->itself
.contains_variable
= true;
688 plats
->aggs_contain_variable
= true;
692 /* Initialize ipcp_lattices. */
695 initialize_node_lattices (struct cgraph_node
*node
)
697 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
698 struct cgraph_edge
*ie
;
699 bool disable
= false, variable
= false;
702 gcc_checking_assert (node
->has_gimple_body_p ());
703 if (!node
->local
.local
)
705 /* When cloning is allowed, we can assume that externally visible
706 functions are not called. We will compensate this by cloning
708 if (ipcp_versionable_function_p (node
)
709 && ipcp_cloning_candidate_p (node
))
715 if (disable
|| variable
)
717 for (i
= 0; i
< ipa_get_param_count (info
) ; i
++)
719 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
722 set_lattice_to_bottom (&plats
->itself
);
723 set_agg_lats_to_bottom (plats
);
726 set_all_contains_variable (plats
);
728 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
729 && !node
->alias
&& !node
->thunk
.thunk_p
)
730 fprintf (dump_file
, "Marking all lattices of %s/%i as %s\n",
731 node
->name (), node
->order
,
732 disable
? "BOTTOM" : "VARIABLE");
735 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
736 if (ie
->indirect_info
->polymorphic
737 && ie
->indirect_info
->param_index
>= 0)
739 gcc_checking_assert (ie
->indirect_info
->param_index
>= 0);
740 ipa_get_parm_lattices (info
,
741 ie
->indirect_info
->param_index
)->virt_call
= 1;
745 /* Return the result of a (possibly arithmetic) pass through jump function
746 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
747 determined or be considered an interprocedural invariant. */
750 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
)
754 if (TREE_CODE (input
) == TREE_BINFO
)
756 if (ipa_get_jf_pass_through_type_preserved (jfunc
))
758 gcc_checking_assert (ipa_get_jf_pass_through_operation (jfunc
)
765 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
768 gcc_checking_assert (is_gimple_ip_invariant (input
));
769 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc
))
771 restype
= boolean_type_node
;
773 restype
= TREE_TYPE (input
);
774 res
= fold_binary (ipa_get_jf_pass_through_operation (jfunc
), restype
,
775 input
, ipa_get_jf_pass_through_operand (jfunc
));
777 if (res
&& !is_gimple_ip_invariant (res
))
783 /* Return the result of an ancestor jump function JFUNC on the constant value
784 INPUT. Return NULL_TREE if that cannot be determined. */
787 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
789 if (TREE_CODE (input
) == TREE_BINFO
)
791 if (!ipa_get_jf_ancestor_type_preserved (jfunc
))
793 /* FIXME: At LTO we can't propagate to non-polymorphic type, because
794 we have no ODR equivalency on those. This should be fixed by
795 propagating on types rather than binfos that would make type
796 matching here unnecesary. */
798 && (TREE_CODE (ipa_get_jf_ancestor_type (jfunc
)) != RECORD_TYPE
799 || !TYPE_BINFO (ipa_get_jf_ancestor_type (jfunc
))
800 || !BINFO_VTABLE (TYPE_BINFO (ipa_get_jf_ancestor_type (jfunc
)))))
802 if (!ipa_get_jf_ancestor_offset (jfunc
))
806 return get_binfo_at_offset (input
,
807 ipa_get_jf_ancestor_offset (jfunc
),
808 ipa_get_jf_ancestor_type (jfunc
));
810 else if (TREE_CODE (input
) == ADDR_EXPR
)
812 tree t
= TREE_OPERAND (input
, 0);
813 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
814 ipa_get_jf_ancestor_offset (jfunc
),
815 ipa_get_jf_ancestor_type (jfunc
)
816 ? ipa_get_jf_ancestor_type (jfunc
)
817 : ptr_type_node
, NULL
, false);
818 return build_fold_addr_expr (t
);
824 /* Determine whether JFUNC evaluates to a known value (that is either a
825 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
826 describes the caller node so that pass-through jump functions can be
830 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
)
832 if (jfunc
->type
== IPA_JF_CONST
)
833 return ipa_get_jf_constant (jfunc
);
834 else if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
835 return ipa_binfo_from_known_type_jfunc (jfunc
);
836 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
837 || jfunc
->type
== IPA_JF_ANCESTOR
)
842 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
843 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
845 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
847 if (info
->ipcp_orig_node
)
848 input
= info
->known_vals
[idx
];
851 struct ipcp_lattice
*lat
;
855 gcc_checking_assert (!flag_ipa_cp
);
858 lat
= ipa_get_scalar_lat (info
, idx
);
859 if (!ipa_lat_is_single_const (lat
))
861 input
= lat
->values
->value
;
867 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
868 return ipa_get_jf_pass_through_result (jfunc
, input
);
870 return ipa_get_jf_ancestor_result (jfunc
, input
);
877 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
878 bottom, not containing a variable component and without any known value at
882 ipcp_verify_propagated_values (void)
884 struct cgraph_node
*node
;
886 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
888 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
889 int i
, count
= ipa_get_param_count (info
);
891 for (i
= 0; i
< count
; i
++)
893 struct ipcp_lattice
*lat
= ipa_get_scalar_lat (info
, i
);
896 && !lat
->contains_variable
897 && lat
->values_count
== 0)
901 symtab_node::dump_table (dump_file
);
902 fprintf (dump_file
, "\nIPA lattices after constant "
903 "propagation, before gcc_unreachable:\n");
904 print_all_lattices (dump_file
, true, false);
913 /* Return true iff X and Y should be considered equal values by IPA-CP. */
916 values_equal_for_ipcp_p (tree x
, tree y
)
918 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
923 if (TREE_CODE (x
) == TREE_BINFO
|| TREE_CODE (y
) == TREE_BINFO
)
926 if (TREE_CODE (x
) == ADDR_EXPR
927 && TREE_CODE (y
) == ADDR_EXPR
928 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
929 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
930 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
931 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
933 return operand_equal_p (x
, y
, 0);
936 /* Add a new value source to VAL, marking that a value comes from edge CS and
937 (if the underlying jump function is a pass-through or an ancestor one) from
938 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
939 is negative if the source was the scalar value of the parameter itself or
940 the offset within an aggregate. */
943 add_value_source (struct ipcp_value
*val
, struct cgraph_edge
*cs
,
944 struct ipcp_value
*src_val
, int src_idx
, HOST_WIDE_INT offset
)
946 struct ipcp_value_source
*src
;
948 src
= (struct ipcp_value_source
*) pool_alloc (ipcp_sources_pool
);
949 src
->offset
= offset
;
952 src
->index
= src_idx
;
954 src
->next
= val
->sources
;
958 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
959 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
960 have the same meaning. */
963 add_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
964 struct cgraph_edge
*cs
, struct ipcp_value
*src_val
,
965 int src_idx
, HOST_WIDE_INT offset
)
967 struct ipcp_value
*val
;
972 for (val
= lat
->values
; val
; val
= val
->next
)
973 if (values_equal_for_ipcp_p (val
->value
, newval
))
975 if (ipa_edge_within_scc (cs
))
977 struct ipcp_value_source
*s
;
978 for (s
= val
->sources
; s
; s
= s
->next
)
985 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
989 if (lat
->values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
991 /* We can only free sources, not the values themselves, because sources
992 of other values in this this SCC might point to them. */
993 for (val
= lat
->values
; val
; val
= val
->next
)
997 struct ipcp_value_source
*src
= val
->sources
;
998 val
->sources
= src
->next
;
999 pool_free (ipcp_sources_pool
, src
);
1004 return set_lattice_to_bottom (lat
);
1007 lat
->values_count
++;
1008 val
= (struct ipcp_value
*) pool_alloc (ipcp_values_pool
);
1009 memset (val
, 0, sizeof (*val
));
1011 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
1012 val
->value
= newval
;
1013 val
->next
= lat
->values
;
1018 /* Like above but passes a special value of offset to distinguish that the
1019 origin is the scalar value of the parameter rather than a part of an
1023 add_scalar_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
1024 struct cgraph_edge
*cs
,
1025 struct ipcp_value
*src_val
, int src_idx
)
1027 return add_value_to_lattice (lat
, newval
, cs
, src_val
, src_idx
, -1);
1030 /* Propagate values through a pass-through jump function JFUNC associated with
1031 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1032 is the index of the source parameter. */
1035 propagate_vals_accross_pass_through (struct cgraph_edge
*cs
,
1036 struct ipa_jump_func
*jfunc
,
1037 struct ipcp_lattice
*src_lat
,
1038 struct ipcp_lattice
*dest_lat
,
1041 struct ipcp_value
*src_val
;
1044 /* Do not create new values when propagating within an SCC because if there
1045 are arithmetic functions with circular dependencies, there is infinite
1046 number of them and we would just make lattices bottom. */
1047 if ((ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1048 && ipa_edge_within_scc (cs
))
1049 ret
= set_lattice_contains_variable (dest_lat
);
1051 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1053 tree cstval
= ipa_get_jf_pass_through_result (jfunc
, src_val
->value
);
1056 ret
|= add_scalar_value_to_lattice (dest_lat
, cstval
, cs
, src_val
,
1059 ret
|= set_lattice_contains_variable (dest_lat
);
1065 /* Propagate values through an ancestor jump function JFUNC associated with
1066 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1067 is the index of the source parameter. */
1070 propagate_vals_accross_ancestor (struct cgraph_edge
*cs
,
1071 struct ipa_jump_func
*jfunc
,
1072 struct ipcp_lattice
*src_lat
,
1073 struct ipcp_lattice
*dest_lat
,
1076 struct ipcp_value
*src_val
;
1079 if (ipa_edge_within_scc (cs
))
1080 return set_lattice_contains_variable (dest_lat
);
1082 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1084 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1087 ret
|= add_scalar_value_to_lattice (dest_lat
, t
, cs
, src_val
, src_idx
);
1089 ret
|= set_lattice_contains_variable (dest_lat
);
1095 /* Propagate scalar values across jump function JFUNC that is associated with
1096 edge CS and put the values into DEST_LAT. */
1099 propagate_scalar_accross_jump_function (struct cgraph_edge
*cs
,
1100 struct ipa_jump_func
*jfunc
,
1101 struct ipcp_lattice
*dest_lat
)
1103 if (dest_lat
->bottom
)
1106 if (jfunc
->type
== IPA_JF_CONST
1107 || jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1111 if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1113 val
= ipa_binfo_from_known_type_jfunc (jfunc
);
1115 return set_lattice_contains_variable (dest_lat
);
1118 val
= ipa_get_jf_constant (jfunc
);
1119 return add_scalar_value_to_lattice (dest_lat
, val
, cs
, NULL
, 0);
1121 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1122 || jfunc
->type
== IPA_JF_ANCESTOR
)
1124 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1125 struct ipcp_lattice
*src_lat
;
1129 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1130 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1132 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1134 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1135 if (src_lat
->bottom
)
1136 return set_lattice_contains_variable (dest_lat
);
1138 /* If we would need to clone the caller and cannot, do not propagate. */
1139 if (!ipcp_versionable_function_p (cs
->caller
)
1140 && (src_lat
->contains_variable
1141 || (src_lat
->values_count
> 1)))
1142 return set_lattice_contains_variable (dest_lat
);
1144 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1145 ret
= propagate_vals_accross_pass_through (cs
, jfunc
, src_lat
,
1148 ret
= propagate_vals_accross_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1151 if (src_lat
->contains_variable
)
1152 ret
|= set_lattice_contains_variable (dest_lat
);
1157 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1158 use it for indirect inlining), we should propagate them too. */
1159 return set_lattice_contains_variable (dest_lat
);
1162 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1163 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1164 other cases, return false). If there are no aggregate items, set
1165 aggs_by_ref to NEW_AGGS_BY_REF. */
1168 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1169 bool new_aggs_by_ref
)
1171 if (dest_plats
->aggs
)
1173 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1175 set_agg_lats_to_bottom (dest_plats
);
1180 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1184 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1185 already existing lattice for the given OFFSET and SIZE, marking all skipped
1186 lattices as containing variable and checking for overlaps. If there is no
1187 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1188 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1189 unless there are too many already. If there are two many, return false. If
1190 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1191 skipped lattices were newly marked as containing variable, set *CHANGE to
1195 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
1196 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
1197 struct ipcp_agg_lattice
***aglat
,
1198 bool pre_existing
, bool *change
)
1200 gcc_checking_assert (offset
>= 0);
1202 while (**aglat
&& (**aglat
)->offset
< offset
)
1204 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
1206 set_agg_lats_to_bottom (dest_plats
);
1209 *change
|= set_lattice_contains_variable (**aglat
);
1210 *aglat
= &(**aglat
)->next
;
1213 if (**aglat
&& (**aglat
)->offset
== offset
)
1215 if ((**aglat
)->size
!= val_size
1217 && (**aglat
)->next
->offset
< offset
+ val_size
))
1219 set_agg_lats_to_bottom (dest_plats
);
1222 gcc_checking_assert (!(**aglat
)->next
1223 || (**aglat
)->next
->offset
>= offset
+ val_size
);
1228 struct ipcp_agg_lattice
*new_al
;
1230 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
1232 set_agg_lats_to_bottom (dest_plats
);
1235 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
1237 dest_plats
->aggs_count
++;
1238 new_al
= (struct ipcp_agg_lattice
*) pool_alloc (ipcp_agg_lattice_pool
);
1239 memset (new_al
, 0, sizeof (*new_al
));
1241 new_al
->offset
= offset
;
1242 new_al
->size
= val_size
;
1243 new_al
->contains_variable
= pre_existing
;
1245 new_al
->next
= **aglat
;
1251 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1252 containing an unknown value. */
1255 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
1260 ret
|= set_lattice_contains_variable (aglat
);
1261 aglat
= aglat
->next
;
1266 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1267 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1268 parameter used for lattice value sources. Return true if DEST_PLATS changed
1272 merge_aggregate_lattices (struct cgraph_edge
*cs
,
1273 struct ipcp_param_lattices
*dest_plats
,
1274 struct ipcp_param_lattices
*src_plats
,
1275 int src_idx
, HOST_WIDE_INT offset_delta
)
1277 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1278 struct ipcp_agg_lattice
**dst_aglat
;
1281 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
1283 if (src_plats
->aggs_bottom
)
1284 return set_agg_lats_contain_variable (dest_plats
);
1285 if (src_plats
->aggs_contain_variable
)
1286 ret
|= set_agg_lats_contain_variable (dest_plats
);
1287 dst_aglat
= &dest_plats
->aggs
;
1289 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
1291 src_aglat
= src_aglat
->next
)
1293 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
1297 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
1298 &dst_aglat
, pre_existing
, &ret
))
1300 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
1302 dst_aglat
= &(*dst_aglat
)->next
;
1303 if (src_aglat
->bottom
)
1305 ret
|= set_lattice_contains_variable (new_al
);
1308 if (src_aglat
->contains_variable
)
1309 ret
|= set_lattice_contains_variable (new_al
);
1310 for (struct ipcp_value
*val
= src_aglat
->values
;
1313 ret
|= add_value_to_lattice (new_al
, val
->value
, cs
, val
, src_idx
,
1316 else if (dest_plats
->aggs_bottom
)
1319 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
1323 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1324 pass-through JFUNC and if so, whether it has conform and conforms to the
1325 rules about propagating values passed by reference. */
1328 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
1329 struct ipa_jump_func
*jfunc
)
1331 return src_plats
->aggs
1332 && (!src_plats
->aggs_by_ref
1333 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
1336 /* Propagate scalar values across jump function JFUNC that is associated with
1337 edge CS and put the values into DEST_LAT. */
1340 propagate_aggs_accross_jump_function (struct cgraph_edge
*cs
,
1341 struct ipa_jump_func
*jfunc
,
1342 struct ipcp_param_lattices
*dest_plats
)
1346 if (dest_plats
->aggs_bottom
)
1349 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1350 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1352 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1353 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1354 struct ipcp_param_lattices
*src_plats
;
1356 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1357 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
1359 /* Currently we do not produce clobber aggregate jump
1360 functions, replace with merging when we do. */
1361 gcc_assert (!jfunc
->agg
.items
);
1362 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
1366 ret
|= set_agg_lats_contain_variable (dest_plats
);
1368 else if (jfunc
->type
== IPA_JF_ANCESTOR
1369 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
1371 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1372 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1373 struct ipcp_param_lattices
*src_plats
;
1375 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1376 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
1378 /* Currently we do not produce clobber aggregate jump
1379 functions, replace with merging when we do. */
1380 gcc_assert (!jfunc
->agg
.items
);
1381 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
1382 ipa_get_jf_ancestor_offset (jfunc
));
1384 else if (!src_plats
->aggs_by_ref
)
1385 ret
|= set_agg_lats_to_bottom (dest_plats
);
1387 ret
|= set_agg_lats_contain_variable (dest_plats
);
1389 else if (jfunc
->agg
.items
)
1391 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1392 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
1393 struct ipa_agg_jf_item
*item
;
1396 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
1399 FOR_EACH_VEC_ELT (*jfunc
->agg
.items
, i
, item
)
1401 HOST_WIDE_INT val_size
;
1403 if (item
->offset
< 0)
1405 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
1406 val_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item
->value
)));
1408 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
1409 &aglat
, pre_existing
, &ret
))
1411 ret
|= add_value_to_lattice (*aglat
, item
->value
, cs
, NULL
, 0, 0);
1412 aglat
= &(*aglat
)->next
;
1414 else if (dest_plats
->aggs_bottom
)
1418 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
1421 ret
|= set_agg_lats_contain_variable (dest_plats
);
1426 /* Propagate constants from the caller to the callee of CS. INFO describes the
1430 propagate_constants_accross_call (struct cgraph_edge
*cs
)
1432 struct ipa_node_params
*callee_info
;
1433 enum availability availability
;
1434 struct cgraph_node
*callee
, *alias_or_thunk
;
1435 struct ipa_edge_args
*args
;
1437 int i
, args_count
, parms_count
;
1439 callee
= cs
->callee
->function_symbol (&availability
);
1440 if (!callee
->definition
)
1442 gcc_checking_assert (callee
->has_gimple_body_p ());
1443 callee_info
= IPA_NODE_REF (callee
);
1445 args
= IPA_EDGE_REF (cs
);
1446 args_count
= ipa_get_cs_argument_count (args
);
1447 parms_count
= ipa_get_param_count (callee_info
);
1448 if (parms_count
== 0)
1451 /* If this call goes through a thunk we must not propagate to the first (0th)
1452 parameter. However, we might need to uncover a thunk from below a series
1453 of aliases first. */
1454 alias_or_thunk
= cs
->callee
;
1455 while (alias_or_thunk
->alias
)
1456 alias_or_thunk
= alias_or_thunk
->get_alias_target ();
1457 if (alias_or_thunk
->thunk
.thunk_p
)
1459 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
1466 for (; (i
< args_count
) && (i
< parms_count
); i
++)
1468 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
1469 struct ipcp_param_lattices
*dest_plats
;
1471 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
1472 if (availability
== AVAIL_INTERPOSABLE
)
1473 ret
|= set_all_contains_variable (dest_plats
);
1476 ret
|= propagate_scalar_accross_jump_function (cs
, jump_func
,
1477 &dest_plats
->itself
);
1478 ret
|= propagate_aggs_accross_jump_function (cs
, jump_func
,
1482 for (; i
< parms_count
; i
++)
1483 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
1488 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1489 (which can contain both constants and binfos), KNOWN_BINFOS, KNOWN_AGGS or
1490 AGG_REPS return the destination. The latter three can be NULL. If AGG_REPS
1491 is not NULL, KNOWN_AGGS is ignored. */
1494 ipa_get_indirect_edge_target_1 (struct cgraph_edge
*ie
,
1495 vec
<tree
> known_vals
,
1496 vec
<tree
> known_binfos
,
1497 vec
<ipa_agg_jump_function_p
> known_aggs
,
1498 struct ipa_agg_replacement_value
*agg_reps
)
1500 int param_index
= ie
->indirect_info
->param_index
;
1501 HOST_WIDE_INT token
, anc_offset
;
1506 if (param_index
== -1
1507 || known_vals
.length () <= (unsigned int) param_index
)
1510 if (!ie
->indirect_info
->polymorphic
)
1514 if (ie
->indirect_info
->agg_contents
)
1521 if (agg_reps
->index
== param_index
1522 && agg_reps
->offset
== ie
->indirect_info
->offset
1523 && agg_reps
->by_ref
== ie
->indirect_info
->by_ref
)
1525 t
= agg_reps
->value
;
1528 agg_reps
= agg_reps
->next
;
1531 else if (known_aggs
.length () > (unsigned int) param_index
)
1533 struct ipa_agg_jump_function
*agg
;
1534 agg
= known_aggs
[param_index
];
1535 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1536 ie
->indirect_info
->by_ref
);
1542 t
= known_vals
[param_index
];
1545 TREE_CODE (t
) == ADDR_EXPR
1546 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
1547 return TREE_OPERAND (t
, 0);
1552 if (!flag_devirtualize
)
1555 gcc_assert (!ie
->indirect_info
->agg_contents
);
1556 token
= ie
->indirect_info
->otr_token
;
1557 anc_offset
= ie
->indirect_info
->offset
;
1558 otr_type
= ie
->indirect_info
->otr_type
;
1562 /* Try to work out value of virtual table pointer value in replacemnets. */
1563 if (!t
&& agg_reps
&& !ie
->indirect_info
->by_ref
)
1567 if (agg_reps
->index
== param_index
1568 && agg_reps
->offset
== ie
->indirect_info
->offset
1569 && agg_reps
->by_ref
)
1571 t
= agg_reps
->value
;
1574 agg_reps
= agg_reps
->next
;
1578 /* Try to work out value of virtual table pointer value in known
1579 aggregate values. */
1580 if (!t
&& known_aggs
.length () > (unsigned int) param_index
1581 && !ie
->indirect_info
->by_ref
)
1583 struct ipa_agg_jump_function
*agg
;
1584 agg
= known_aggs
[param_index
];
1585 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1589 /* If we found the virtual table pointer, lookup the target. */
1593 unsigned HOST_WIDE_INT offset
;
1594 if (vtable_pointer_value_to_vtable (t
, &vtable
, &offset
))
1596 target
= gimple_get_virt_method_for_vtable (ie
->indirect_info
->otr_token
,
1600 if ((TREE_CODE (TREE_TYPE (target
)) == FUNCTION_TYPE
1601 && DECL_FUNCTION_CODE (target
) == BUILT_IN_UNREACHABLE
)
1602 || !possible_polymorphic_call_target_p
1603 (ie
, cgraph_node::get (target
)))
1604 target
= ipa_impossible_devirt_target (ie
, target
);
1610 /* Did we work out BINFO via type propagation? */
1611 if (!t
&& known_binfos
.length () > (unsigned int) param_index
)
1612 t
= known_binfos
[param_index
];
1613 /* Or do we know the constant value of pointer? */
1615 t
= known_vals
[param_index
];
1619 if (TREE_CODE (t
) != TREE_BINFO
)
1621 ipa_polymorphic_call_context
context (t
, ie
->indirect_info
->otr_type
,
1623 vec
<cgraph_node
*>targets
;
1626 targets
= possible_polymorphic_call_targets
1627 (ie
->indirect_info
->otr_type
,
1628 ie
->indirect_info
->otr_token
,
1630 if (!final
|| targets
.length () > 1)
1632 if (targets
.length () == 1)
1633 target
= targets
[0]->decl
;
1635 target
= ipa_impossible_devirt_target (ie
, NULL_TREE
);
1641 binfo
= get_binfo_at_offset (t
, anc_offset
, otr_type
);
1644 target
= gimple_get_virt_method_for_binfo (token
, binfo
);
1647 if (target
&& !possible_polymorphic_call_target_p (ie
,
1648 cgraph_node::get (target
)))
1649 target
= ipa_impossible_devirt_target (ie
, target
);
1655 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1656 (which can contain both constants and binfos), KNOWN_BINFOS (which can be
1657 NULL) or KNOWN_AGGS (which also can be NULL) return the destination. */
1660 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
1661 vec
<tree
> known_vals
,
1662 vec
<tree
> known_binfos
,
1663 vec
<ipa_agg_jump_function_p
> known_aggs
)
1665 return ipa_get_indirect_edge_target_1 (ie
, known_vals
, known_binfos
,
1669 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1670 and KNOWN_BINFOS. */
1673 devirtualization_time_bonus (struct cgraph_node
*node
,
1674 vec
<tree
> known_csts
,
1675 vec
<tree
> known_binfos
,
1676 vec
<ipa_agg_jump_function_p
> known_aggs
)
1678 struct cgraph_edge
*ie
;
1681 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1683 struct cgraph_node
*callee
;
1684 struct inline_summary
*isummary
;
1685 enum availability avail
;
1688 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_binfos
,
1693 /* Only bare minimum benefit for clearly un-inlineable targets. */
1695 callee
= cgraph_node::get (target
);
1696 if (!callee
|| !callee
->definition
)
1698 callee
= callee
->function_symbol (&avail
);
1699 if (avail
< AVAIL_AVAILABLE
)
1701 isummary
= inline_summary (callee
);
1702 if (!isummary
->inlinable
)
1705 /* FIXME: The values below need re-considering and perhaps also
1706 integrating into the cost metrics, at lest in some very basic way. */
1707 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
1709 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
1711 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
1712 || DECL_DECLARED_INLINE_P (callee
->decl
))
1719 /* Return time bonus incurred because of HINTS. */
1722 hint_time_bonus (inline_hints hints
)
1725 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
1726 result
+= PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
1727 if (hints
& INLINE_HINT_array_index
)
1728 result
+= PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS
);
1732 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1733 and SIZE_COST and with the sum of frequencies of incoming edges to the
1734 potential new clone in FREQUENCIES. */
1737 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
1738 int freq_sum
, gcov_type count_sum
, int size_cost
)
1740 if (time_benefit
== 0
1741 || !flag_ipa_cp_clone
1742 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->decl
)))
1745 gcc_assert (size_cost
> 0);
1749 int factor
= (count_sum
* 1000) / max_count
;
1750 int64_t evaluation
= (((int64_t) time_benefit
* factor
)
1753 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1754 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1755 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1756 ") -> evaluation: " "%"PRId64
1757 ", threshold: %i\n",
1758 time_benefit
, size_cost
, (HOST_WIDE_INT
) count_sum
,
1759 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1761 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1765 int64_t evaluation
= (((int64_t) time_benefit
* freq_sum
)
1768 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1769 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1770 "size: %i, freq_sum: %i) -> evaluation: "
1771 "%"PRId64
", threshold: %i\n",
1772 time_benefit
, size_cost
, freq_sum
, evaluation
,
1773 PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
1775 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1779 /* Return all context independent values from aggregate lattices in PLATS in a
1780 vector. Return NULL if there are none. */
1782 static vec
<ipa_agg_jf_item
, va_gc
> *
1783 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
1785 vec
<ipa_agg_jf_item
, va_gc
> *res
= NULL
;
1787 if (plats
->aggs_bottom
1788 || plats
->aggs_contain_variable
1789 || plats
->aggs_count
== 0)
1792 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
1794 aglat
= aglat
->next
)
1795 if (ipa_lat_is_single_const (aglat
))
1797 struct ipa_agg_jf_item item
;
1798 item
.offset
= aglat
->offset
;
1799 item
.value
= aglat
->values
->value
;
1800 vec_safe_push (res
, item
);
1805 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1806 them with values of parameters that are known independent of the context.
1807 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1808 movement cost of all removable parameters will be stored in it. */
1811 gather_context_independent_values (struct ipa_node_params
*info
,
1812 vec
<tree
> *known_csts
,
1813 vec
<tree
> *known_binfos
,
1814 vec
<ipa_agg_jump_function
> *known_aggs
,
1815 int *removable_params_cost
)
1817 int i
, count
= ipa_get_param_count (info
);
1820 known_csts
->create (0);
1821 known_binfos
->create (0);
1822 known_csts
->safe_grow_cleared (count
);
1823 known_binfos
->safe_grow_cleared (count
);
1826 known_aggs
->create (0);
1827 known_aggs
->safe_grow_cleared (count
);
1830 if (removable_params_cost
)
1831 *removable_params_cost
= 0;
1833 for (i
= 0; i
< count
; i
++)
1835 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1836 struct ipcp_lattice
*lat
= &plats
->itself
;
1838 if (ipa_lat_is_single_const (lat
))
1840 struct ipcp_value
*val
= lat
->values
;
1841 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1843 (*known_csts
)[i
] = val
->value
;
1844 if (removable_params_cost
)
1845 *removable_params_cost
1846 += estimate_move_cost (TREE_TYPE (val
->value
), false);
1849 else if (plats
->virt_call
)
1851 (*known_binfos
)[i
] = val
->value
;
1854 else if (removable_params_cost
1855 && !ipa_is_param_used (info
, i
))
1856 *removable_params_cost
+= ipa_get_param_move_cost (info
, i
);
1858 else if (removable_params_cost
1859 && !ipa_is_param_used (info
, i
))
1860 *removable_params_cost
1861 += ipa_get_param_move_cost (info
, i
);
1865 vec
<ipa_agg_jf_item
, va_gc
> *agg_items
;
1866 struct ipa_agg_jump_function
*ajf
;
1868 agg_items
= context_independent_aggregate_values (plats
);
1869 ajf
= &(*known_aggs
)[i
];
1870 ajf
->items
= agg_items
;
1871 ajf
->by_ref
= plats
->aggs_by_ref
;
1872 ret
|= agg_items
!= NULL
;
1879 /* The current interface in ipa-inline-analysis requires a pointer vector.
1882 FIXME: That interface should be re-worked, this is slightly silly. Still,
1883 I'd like to discuss how to change it first and this demonstrates the
1886 static vec
<ipa_agg_jump_function_p
>
1887 agg_jmp_p_vec_for_t_vec (vec
<ipa_agg_jump_function
> known_aggs
)
1889 vec
<ipa_agg_jump_function_p
> ret
;
1890 struct ipa_agg_jump_function
*ajf
;
1893 ret
.create (known_aggs
.length ());
1894 FOR_EACH_VEC_ELT (known_aggs
, i
, ajf
)
1895 ret
.quick_push (ajf
);
1899 /* Iterate over known values of parameters of NODE and estimate the local
1900 effects in terms of time and size they have. */
1903 estimate_local_effects (struct cgraph_node
*node
)
1905 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1906 int i
, count
= ipa_get_param_count (info
);
1907 vec
<tree
> known_csts
, known_binfos
;
1908 vec
<ipa_agg_jump_function
> known_aggs
;
1909 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
;
1911 int base_time
= inline_summary (node
)->time
;
1912 int removable_params_cost
;
1914 if (!count
|| !ipcp_versionable_function_p (node
))
1917 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1918 fprintf (dump_file
, "\nEstimating effects for %s/%i, base_time: %i.\n",
1919 node
->name (), node
->order
, base_time
);
1921 always_const
= gather_context_independent_values (info
, &known_csts
,
1922 &known_binfos
, &known_aggs
,
1923 &removable_params_cost
);
1924 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
1927 struct caller_statistics stats
;
1931 init_caller_stats (&stats
);
1932 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
1934 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1935 known_aggs_ptrs
, &size
, &time
, &hints
);
1936 time
-= devirtualization_time_bonus (node
, known_csts
, known_binfos
,
1938 time
-= hint_time_bonus (hints
);
1939 time
-= removable_params_cost
;
1940 size
-= stats
.n_calls
* removable_params_cost
;
1943 fprintf (dump_file
, " - context independent values, size: %i, "
1944 "time_benefit: %i\n", size
, base_time
- time
);
1947 || node
->will_be_removed_from_program_if_no_direct_calls_p ())
1949 info
->do_clone_for_all_contexts
= true;
1953 fprintf (dump_file
, " Decided to specialize for all "
1954 "known contexts, code not going to grow.\n");
1956 else if (good_cloning_opportunity_p (node
, base_time
- time
,
1957 stats
.freq_sum
, stats
.count_sum
,
1960 if (size
+ overall_size
<= max_new_size
)
1962 info
->do_clone_for_all_contexts
= true;
1964 overall_size
+= size
;
1967 fprintf (dump_file
, " Decided to specialize for all "
1968 "known contexts, growth deemed beneficial.\n");
1970 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1971 fprintf (dump_file
, " Not cloning for all contexts because "
1972 "max_new_size would be reached with %li.\n",
1973 size
+ overall_size
);
1977 for (i
= 0; i
< count
; i
++)
1979 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1980 struct ipcp_lattice
*lat
= &plats
->itself
;
1981 struct ipcp_value
*val
;
1990 for (val
= lat
->values
; val
; val
= val
->next
)
1992 int time
, size
, time_benefit
;
1995 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1997 known_csts
[i
] = val
->value
;
1998 known_binfos
[i
] = NULL_TREE
;
1999 emc
= estimate_move_cost (TREE_TYPE (val
->value
), true);
2001 else if (plats
->virt_call
)
2003 known_csts
[i
] = NULL_TREE
;
2004 known_binfos
[i
] = val
->value
;
2010 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
2011 known_aggs_ptrs
, &size
, &time
,
2013 time_benefit
= base_time
- time
2014 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
2016 + hint_time_bonus (hints
)
2017 + removable_params_cost
+ emc
;
2019 gcc_checking_assert (size
>=0);
2020 /* The inliner-heuristics based estimates may think that in certain
2021 contexts some functions do not have any size at all but we want
2022 all specializations to have at least a tiny cost, not least not to
2027 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2029 fprintf (dump_file
, " - estimates for value ");
2030 print_ipcp_constant_value (dump_file
, val
->value
);
2031 fprintf (dump_file
, " for ");
2032 ipa_dump_param (dump_file
, info
, i
);
2033 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
2034 time_benefit
, size
);
2037 val
->local_time_benefit
= time_benefit
;
2038 val
->local_size_cost
= size
;
2040 known_binfos
[i
] = NULL_TREE
;
2041 known_csts
[i
] = NULL_TREE
;
2044 for (i
= 0; i
< count
; i
++)
2046 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2047 struct ipa_agg_jump_function
*ajf
;
2048 struct ipcp_agg_lattice
*aglat
;
2050 if (plats
->aggs_bottom
|| !plats
->aggs
)
2053 ajf
= &known_aggs
[i
];
2054 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2056 struct ipcp_value
*val
;
2057 if (aglat
->bottom
|| !aglat
->values
2058 /* If the following is true, the one value is in known_aggs. */
2059 || (!plats
->aggs_contain_variable
2060 && ipa_lat_is_single_const (aglat
)))
2063 for (val
= aglat
->values
; val
; val
= val
->next
)
2065 int time
, size
, time_benefit
;
2066 struct ipa_agg_jf_item item
;
2069 item
.offset
= aglat
->offset
;
2070 item
.value
= val
->value
;
2071 vec_safe_push (ajf
->items
, item
);
2073 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
2074 known_aggs_ptrs
, &size
, &time
,
2076 time_benefit
= base_time
- time
2077 + devirtualization_time_bonus (node
, known_csts
, known_binfos
,
2079 + hint_time_bonus (hints
);
2080 gcc_checking_assert (size
>=0);
2084 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2086 fprintf (dump_file
, " - estimates for value ");
2087 print_ipcp_constant_value (dump_file
, val
->value
);
2088 fprintf (dump_file
, " for ");
2089 ipa_dump_param (dump_file
, info
, i
);
2090 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
2091 "]: time_benefit: %i, size: %i\n",
2092 plats
->aggs_by_ref
? "ref " : "",
2093 aglat
->offset
, time_benefit
, size
);
2096 val
->local_time_benefit
= time_benefit
;
2097 val
->local_size_cost
= size
;
2103 for (i
= 0; i
< count
; i
++)
2104 vec_free (known_aggs
[i
].items
);
2106 known_csts
.release ();
2107 known_binfos
.release ();
2108 known_aggs
.release ();
2109 known_aggs_ptrs
.release ();
2113 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
2114 topological sort of values. */
2117 add_val_to_toposort (struct ipcp_value
*cur_val
)
2119 static int dfs_counter
= 0;
2120 static struct ipcp_value
*stack
;
2121 struct ipcp_value_source
*src
;
2127 cur_val
->dfs
= dfs_counter
;
2128 cur_val
->low_link
= dfs_counter
;
2130 cur_val
->topo_next
= stack
;
2132 cur_val
->on_stack
= true;
2134 for (src
= cur_val
->sources
; src
; src
= src
->next
)
2137 if (src
->val
->dfs
== 0)
2139 add_val_to_toposort (src
->val
);
2140 if (src
->val
->low_link
< cur_val
->low_link
)
2141 cur_val
->low_link
= src
->val
->low_link
;
2143 else if (src
->val
->on_stack
2144 && src
->val
->dfs
< cur_val
->low_link
)
2145 cur_val
->low_link
= src
->val
->dfs
;
2148 if (cur_val
->dfs
== cur_val
->low_link
)
2150 struct ipcp_value
*v
, *scc_list
= NULL
;
2155 stack
= v
->topo_next
;
2156 v
->on_stack
= false;
2158 v
->scc_next
= scc_list
;
2161 while (v
!= cur_val
);
2163 cur_val
->topo_next
= values_topo
;
2164 values_topo
= cur_val
;
2168 /* Add all values in lattices associated with NODE to the topological sort if
2169 they are not there yet. */
2172 add_all_node_vals_to_toposort (struct cgraph_node
*node
)
2174 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2175 int i
, count
= ipa_get_param_count (info
);
2177 for (i
= 0; i
< count
; i
++)
2179 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2180 struct ipcp_lattice
*lat
= &plats
->itself
;
2181 struct ipcp_agg_lattice
*aglat
;
2182 struct ipcp_value
*val
;
2185 for (val
= lat
->values
; val
; val
= val
->next
)
2186 add_val_to_toposort (val
);
2188 if (!plats
->aggs_bottom
)
2189 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2191 for (val
= aglat
->values
; val
; val
= val
->next
)
2192 add_val_to_toposort (val
);
2196 /* One pass of constants propagation along the call graph edges, from callers
2197 to callees (requires topological ordering in TOPO), iterate over strongly
2198 connected components. */
2201 propagate_constants_topo (struct ipa_topo_info
*topo
)
2205 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
2208 struct cgraph_node
*v
, *node
= topo
->order
[i
];
2209 vec
<cgraph_node
*> cycle_nodes
= ipa_get_nodes_in_cycle (node
);
2211 /* First, iteratively propagate within the strongly connected component
2212 until all lattices stabilize. */
2213 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
2214 if (v
->has_gimple_body_p ())
2215 push_node_to_stack (topo
, v
);
2217 v
= pop_node_from_stack (topo
);
2220 struct cgraph_edge
*cs
;
2222 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2223 if (ipa_edge_within_scc (cs
)
2224 && propagate_constants_accross_call (cs
))
2225 push_node_to_stack (topo
, cs
->callee
);
2226 v
= pop_node_from_stack (topo
);
2229 /* Afterwards, propagate along edges leading out of the SCC, calculates
2230 the local effects of the discovered constants and all valid values to
2231 their topological sort. */
2232 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
2233 if (v
->has_gimple_body_p ())
2235 struct cgraph_edge
*cs
;
2237 estimate_local_effects (v
);
2238 add_all_node_vals_to_toposort (v
);
2239 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2240 if (!ipa_edge_within_scc (cs
))
2241 propagate_constants_accross_call (cs
);
2243 cycle_nodes
.release ();
2248 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2249 the bigger one if otherwise. */
2252 safe_add (int a
, int b
)
2254 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
2255 return a
> b
? a
: b
;
2261 /* Propagate the estimated effects of individual values along the topological
2262 from the dependent values to those they depend on. */
2265 propagate_effects (void)
2267 struct ipcp_value
*base
;
2269 for (base
= values_topo
; base
; base
= base
->topo_next
)
2271 struct ipcp_value_source
*src
;
2272 struct ipcp_value
*val
;
2273 int time
= 0, size
= 0;
2275 for (val
= base
; val
; val
= val
->scc_next
)
2277 time
= safe_add (time
,
2278 val
->local_time_benefit
+ val
->prop_time_benefit
);
2279 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
2282 for (val
= base
; val
; val
= val
->scc_next
)
2283 for (src
= val
->sources
; src
; src
= src
->next
)
2285 && src
->cs
->maybe_hot_p ())
2287 src
->val
->prop_time_benefit
= safe_add (time
,
2288 src
->val
->prop_time_benefit
);
2289 src
->val
->prop_size_cost
= safe_add (size
,
2290 src
->val
->prop_size_cost
);
2296 /* Propagate constants, binfos and their effects from the summaries
2297 interprocedurally. */
2300 ipcp_propagate_stage (struct ipa_topo_info
*topo
)
2302 struct cgraph_node
*node
;
2305 fprintf (dump_file
, "\n Propagating constants:\n\n");
2308 ipa_update_after_lto_read ();
2311 FOR_EACH_DEFINED_FUNCTION (node
)
2313 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2315 determine_versionability (node
);
2316 if (node
->has_gimple_body_p ())
2318 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
2319 ipa_get_param_count (info
));
2320 initialize_node_lattices (node
);
2322 if (node
->definition
&& !node
->alias
)
2323 overall_size
+= inline_summary (node
)->self_size
;
2324 if (node
->count
> max_count
)
2325 max_count
= node
->count
;
2328 max_new_size
= overall_size
;
2329 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
2330 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
2331 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
2334 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
2335 overall_size
, max_new_size
);
2337 propagate_constants_topo (topo
);
2338 #ifdef ENABLE_CHECKING
2339 ipcp_verify_propagated_values ();
2341 propagate_effects ();
2345 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
2346 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
2350 /* Discover newly direct outgoing edges from NODE which is a new clone with
2351 known KNOWN_VALS and make them direct. */
2354 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
2355 vec
<tree
> known_vals
,
2356 struct ipa_agg_replacement_value
*aggvals
)
2358 struct cgraph_edge
*ie
, *next_ie
;
2361 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
2365 next_ie
= ie
->next_callee
;
2366 target
= ipa_get_indirect_edge_target_1 (ie
, known_vals
, vNULL
, vNULL
,
2370 bool agg_contents
= ie
->indirect_info
->agg_contents
;
2371 bool polymorphic
= ie
->indirect_info
->polymorphic
;
2372 int param_index
= ie
->indirect_info
->param_index
;
2373 struct cgraph_edge
*cs
= ipa_make_edge_direct_to_target (ie
, target
);
2376 if (cs
&& !agg_contents
&& !polymorphic
)
2378 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2379 int c
= ipa_get_controlled_uses (info
, param_index
);
2380 if (c
!= IPA_UNDESCRIBED_USE
)
2382 struct ipa_ref
*to_del
;
2385 ipa_set_controlled_uses (info
, param_index
, c
);
2386 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2387 fprintf (dump_file
, " controlled uses count of param "
2388 "%i bumped down to %i\n", param_index
, c
);
2390 && (to_del
= node
->find_reference (cs
->callee
, NULL
, 0)))
2392 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2393 fprintf (dump_file
, " and even removing its "
2394 "cloning-created reference\n");
2395 to_del
->remove_reference ();
2401 /* Turning calls to direct calls will improve overall summary. */
2403 inline_update_overall_summary (node
);
2406 /* Vector of pointers which for linked lists of clones of an original crgaph
2409 static vec
<cgraph_edge
*> next_edge_clone
;
2410 static vec
<cgraph_edge
*> prev_edge_clone
;
2413 grow_edge_clone_vectors (void)
2415 if (next_edge_clone
.length ()
2416 <= (unsigned) symtab
->edges_max_uid
)
2417 next_edge_clone
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
2418 if (prev_edge_clone
.length ()
2419 <= (unsigned) symtab
->edges_max_uid
)
2420 prev_edge_clone
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
2423 /* Edge duplication hook to grow the appropriate linked list in
2427 ipcp_edge_duplication_hook (struct cgraph_edge
*src
, struct cgraph_edge
*dst
,
2430 grow_edge_clone_vectors ();
2432 struct cgraph_edge
*old_next
= next_edge_clone
[src
->uid
];
2434 prev_edge_clone
[old_next
->uid
] = dst
;
2435 prev_edge_clone
[dst
->uid
] = src
;
2437 next_edge_clone
[dst
->uid
] = old_next
;
2438 next_edge_clone
[src
->uid
] = dst
;
2441 /* Hook that is called by cgraph.c when an edge is removed. */
2444 ipcp_edge_removal_hook (struct cgraph_edge
*cs
, void *)
2446 grow_edge_clone_vectors ();
2448 struct cgraph_edge
*prev
= prev_edge_clone
[cs
->uid
];
2449 struct cgraph_edge
*next
= next_edge_clone
[cs
->uid
];
2451 next_edge_clone
[prev
->uid
] = next
;
2453 prev_edge_clone
[next
->uid
] = prev
;
2456 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2457 parameter with the given INDEX. */
2460 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDE_INT offset
,
2463 struct ipa_agg_replacement_value
*aggval
;
2465 aggval
= ipa_get_agg_replacements_for_node (node
);
2468 if (aggval
->offset
== offset
2469 && aggval
->index
== index
)
2470 return aggval
->value
;
2471 aggval
= aggval
->next
;
2476 /* Return true if edge CS does bring about the value described by SRC. */
2479 cgraph_edge_brings_value_p (struct cgraph_edge
*cs
,
2480 struct ipcp_value_source
*src
)
2482 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2483 cgraph_node
*real_dest
= cs
->callee
->function_symbol ();
2484 struct ipa_node_params
*dst_info
= IPA_NODE_REF (real_dest
);
2486 if ((dst_info
->ipcp_orig_node
&& !dst_info
->is_all_contexts_clone
)
2487 || caller_info
->node_dead
)
2492 if (caller_info
->ipcp_orig_node
)
2495 if (src
->offset
== -1)
2496 t
= caller_info
->known_vals
[src
->index
];
2498 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
2499 return (t
!= NULL_TREE
2500 && values_equal_for_ipcp_p (src
->val
->value
, t
));
2504 struct ipcp_agg_lattice
*aglat
;
2505 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
2507 if (src
->offset
== -1)
2508 return (ipa_lat_is_single_const (&plats
->itself
)
2509 && values_equal_for_ipcp_p (src
->val
->value
,
2510 plats
->itself
.values
->value
));
2513 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
2515 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2516 if (aglat
->offset
== src
->offset
)
2517 return (ipa_lat_is_single_const (aglat
)
2518 && values_equal_for_ipcp_p (src
->val
->value
,
2519 aglat
->values
->value
));
2525 /* Get the next clone in the linked list of clones of an edge. */
2527 static inline struct cgraph_edge
*
2528 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
2530 return next_edge_clone
[cs
->uid
];
2533 /* Given VAL, iterate over all its sources and if they still hold, add their
2534 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2538 get_info_about_necessary_edges (struct ipcp_value
*val
, int *freq_sum
,
2539 gcov_type
*count_sum
, int *caller_count
)
2541 struct ipcp_value_source
*src
;
2542 int freq
= 0, count
= 0;
2546 for (src
= val
->sources
; src
; src
= src
->next
)
2548 struct cgraph_edge
*cs
= src
->cs
;
2551 if (cgraph_edge_brings_value_p (cs
, src
))
2554 freq
+= cs
->frequency
;
2556 hot
|= cs
->maybe_hot_p ();
2558 cs
= get_next_cgraph_edge_clone (cs
);
2564 *caller_count
= count
;
2568 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2569 their number is known and equal to CALLER_COUNT. */
2571 static vec
<cgraph_edge
*>
2572 gather_edges_for_value (struct ipcp_value
*val
, int caller_count
)
2574 struct ipcp_value_source
*src
;
2575 vec
<cgraph_edge
*> ret
;
2577 ret
.create (caller_count
);
2578 for (src
= val
->sources
; src
; src
= src
->next
)
2580 struct cgraph_edge
*cs
= src
->cs
;
2583 if (cgraph_edge_brings_value_p (cs
, src
))
2584 ret
.quick_push (cs
);
2585 cs
= get_next_cgraph_edge_clone (cs
);
2592 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2593 Return it or NULL if for some reason it cannot be created. */
2595 static struct ipa_replace_map
*
2596 get_replacement_map (struct ipa_node_params
*info
, tree value
, int parm_num
)
2598 struct ipa_replace_map
*replace_map
;
2601 replace_map
= ggc_alloc
<ipa_replace_map
> ();
2604 fprintf (dump_file
, " replacing ");
2605 ipa_dump_param (dump_file
, info
, parm_num
);
2607 fprintf (dump_file
, " with const ");
2608 print_generic_expr (dump_file
, value
, 0);
2609 fprintf (dump_file
, "\n");
2611 replace_map
->old_tree
= NULL
;
2612 replace_map
->parm_num
= parm_num
;
2613 replace_map
->new_tree
= value
;
2614 replace_map
->replace_p
= true;
2615 replace_map
->ref_p
= false;
2620 /* Dump new profiling counts */
2623 dump_profile_updates (struct cgraph_node
*orig_node
,
2624 struct cgraph_node
*new_node
)
2626 struct cgraph_edge
*cs
;
2628 fprintf (dump_file
, " setting count of the specialized node to "
2629 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) new_node
->count
);
2630 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2631 fprintf (dump_file
, " edge to %s has count "
2632 HOST_WIDE_INT_PRINT_DEC
"\n",
2633 cs
->callee
->name (), (HOST_WIDE_INT
) cs
->count
);
2635 fprintf (dump_file
, " setting count of the original node to "
2636 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) orig_node
->count
);
2637 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2638 fprintf (dump_file
, " edge to %s is left with "
2639 HOST_WIDE_INT_PRINT_DEC
"\n",
2640 cs
->callee
->name (), (HOST_WIDE_INT
) cs
->count
);
2643 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2644 their profile information to reflect this. */
2647 update_profiling_info (struct cgraph_node
*orig_node
,
2648 struct cgraph_node
*new_node
)
2650 struct cgraph_edge
*cs
;
2651 struct caller_statistics stats
;
2652 gcov_type new_sum
, orig_sum
;
2653 gcov_type remainder
, orig_node_count
= orig_node
->count
;
2655 if (orig_node_count
== 0)
2658 init_caller_stats (&stats
);
2659 orig_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
2661 orig_sum
= stats
.count_sum
;
2662 init_caller_stats (&stats
);
2663 new_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
2665 new_sum
= stats
.count_sum
;
2667 if (orig_node_count
< orig_sum
+ new_sum
)
2670 fprintf (dump_file
, " Problem: node %s/%i has too low count "
2671 HOST_WIDE_INT_PRINT_DEC
" while the sum of incoming "
2672 "counts is " HOST_WIDE_INT_PRINT_DEC
"\n",
2673 orig_node
->name (), orig_node
->order
,
2674 (HOST_WIDE_INT
) orig_node_count
,
2675 (HOST_WIDE_INT
) (orig_sum
+ new_sum
));
2677 orig_node_count
= (orig_sum
+ new_sum
) * 12 / 10;
2679 fprintf (dump_file
, " proceeding by pretending it was "
2680 HOST_WIDE_INT_PRINT_DEC
"\n",
2681 (HOST_WIDE_INT
) orig_node_count
);
2684 new_node
->count
= new_sum
;
2685 remainder
= orig_node_count
- new_sum
;
2686 orig_node
->count
= remainder
;
2688 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2690 cs
->count
= apply_probability (cs
->count
,
2691 GCOV_COMPUTE_SCALE (new_sum
,
2696 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2697 cs
->count
= apply_probability (cs
->count
,
2698 GCOV_COMPUTE_SCALE (remainder
,
2702 dump_profile_updates (orig_node
, new_node
);
2705 /* Update the respective profile of specialized NEW_NODE and the original
2706 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2707 have been redirected to the specialized version. */
2710 update_specialized_profile (struct cgraph_node
*new_node
,
2711 struct cgraph_node
*orig_node
,
2712 gcov_type redirected_sum
)
2714 struct cgraph_edge
*cs
;
2715 gcov_type new_node_count
, orig_node_count
= orig_node
->count
;
2718 fprintf (dump_file
, " the sum of counts of redirected edges is "
2719 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) redirected_sum
);
2720 if (orig_node_count
== 0)
2723 gcc_assert (orig_node_count
>= redirected_sum
);
2725 new_node_count
= new_node
->count
;
2726 new_node
->count
+= redirected_sum
;
2727 orig_node
->count
-= redirected_sum
;
2729 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2731 cs
->count
+= apply_probability (cs
->count
,
2732 GCOV_COMPUTE_SCALE (redirected_sum
,
2737 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2739 gcov_type dec
= apply_probability (cs
->count
,
2740 GCOV_COMPUTE_SCALE (redirected_sum
,
2742 if (dec
< cs
->count
)
2749 dump_profile_updates (orig_node
, new_node
);
2752 /* Create a specialized version of NODE with known constants and types of
2753 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2755 static struct cgraph_node
*
2756 create_specialized_node (struct cgraph_node
*node
,
2757 vec
<tree
> known_vals
,
2758 struct ipa_agg_replacement_value
*aggvals
,
2759 vec
<cgraph_edge
*> callers
)
2761 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
2762 vec
<ipa_replace_map
*, va_gc
> *replace_trees
= NULL
;
2763 struct ipa_agg_replacement_value
*av
;
2764 struct cgraph_node
*new_node
;
2765 int i
, count
= ipa_get_param_count (info
);
2766 bitmap args_to_skip
;
2768 gcc_assert (!info
->ipcp_orig_node
);
2770 if (node
->local
.can_change_signature
)
2772 args_to_skip
= BITMAP_GGC_ALLOC ();
2773 for (i
= 0; i
< count
; i
++)
2775 tree t
= known_vals
[i
];
2777 if ((t
&& TREE_CODE (t
) != TREE_BINFO
)
2778 || !ipa_is_param_used (info
, i
))
2779 bitmap_set_bit (args_to_skip
, i
);
2784 args_to_skip
= NULL
;
2785 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2786 fprintf (dump_file
, " cannot change function signature\n");
2789 for (i
= 0; i
< count
; i
++)
2791 tree t
= known_vals
[i
];
2792 if (t
&& TREE_CODE (t
) != TREE_BINFO
)
2794 struct ipa_replace_map
*replace_map
;
2796 replace_map
= get_replacement_map (info
, t
, i
);
2798 vec_safe_push (replace_trees
, replace_map
);
2802 new_node
= node
->create_virtual_clone (callers
, replace_trees
,
2803 args_to_skip
, "constprop");
2804 ipa_set_node_agg_value_chain (new_node
, aggvals
);
2805 for (av
= aggvals
; av
; av
= av
->next
)
2806 new_node
->maybe_create_reference (av
->value
, IPA_REF_ADDR
, NULL
);
2808 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2810 fprintf (dump_file
, " the new node is %s/%i.\n",
2811 new_node
->name (), new_node
->order
);
2813 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
2815 ipa_check_create_node_params ();
2816 update_profiling_info (node
, new_node
);
2817 new_info
= IPA_NODE_REF (new_node
);
2818 new_info
->ipcp_orig_node
= node
;
2819 new_info
->known_vals
= known_vals
;
2821 ipcp_discover_new_direct_edges (new_node
, known_vals
, aggvals
);
2827 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2828 KNOWN_VALS with constants and types that are also known for all of the
2832 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
2833 vec
<tree
> known_vals
,
2834 vec
<cgraph_edge
*> callers
)
2836 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2837 int i
, count
= ipa_get_param_count (info
);
2839 for (i
= 0; i
< count
; i
++)
2841 struct cgraph_edge
*cs
;
2842 tree newval
= NULL_TREE
;
2845 if (ipa_get_scalar_lat (info
, i
)->bottom
|| known_vals
[i
])
2848 FOR_EACH_VEC_ELT (callers
, j
, cs
)
2850 struct ipa_jump_func
*jump_func
;
2853 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
2858 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
2859 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
);
2862 && !values_equal_for_ipcp_p (t
, newval
)))
2873 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2875 fprintf (dump_file
, " adding an extra known scalar value ");
2876 print_ipcp_constant_value (dump_file
, newval
);
2877 fprintf (dump_file
, " for ");
2878 ipa_dump_param (dump_file
, info
, i
);
2879 fprintf (dump_file
, "\n");
2882 known_vals
[i
] = newval
;
2887 /* Go through PLATS and create a vector of values consisting of values and
2888 offsets (minus OFFSET) of lattices that contain only a single value. */
2890 static vec
<ipa_agg_jf_item
>
2891 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
2893 vec
<ipa_agg_jf_item
> res
= vNULL
;
2895 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2898 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2899 if (ipa_lat_is_single_const (aglat
))
2901 struct ipa_agg_jf_item ti
;
2902 ti
.offset
= aglat
->offset
- offset
;
2903 ti
.value
= aglat
->values
->value
;
2909 /* Intersect all values in INTER with single value lattices in PLATS (while
2910 subtracting OFFSET). */
2913 intersect_with_plats (struct ipcp_param_lattices
*plats
,
2914 vec
<ipa_agg_jf_item
> *inter
,
2915 HOST_WIDE_INT offset
)
2917 struct ipcp_agg_lattice
*aglat
;
2918 struct ipa_agg_jf_item
*item
;
2921 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2927 aglat
= plats
->aggs
;
2928 FOR_EACH_VEC_ELT (*inter
, k
, item
)
2935 if (aglat
->offset
- offset
> item
->offset
)
2937 if (aglat
->offset
- offset
== item
->offset
)
2939 gcc_checking_assert (item
->value
);
2940 if (values_equal_for_ipcp_p (item
->value
, aglat
->values
->value
))
2944 aglat
= aglat
->next
;
2947 item
->value
= NULL_TREE
;
2951 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2952 vector result while subtracting OFFSET from the individual value offsets. */
2954 static vec
<ipa_agg_jf_item
>
2955 agg_replacements_to_vector (struct cgraph_node
*node
, int index
,
2956 HOST_WIDE_INT offset
)
2958 struct ipa_agg_replacement_value
*av
;
2959 vec
<ipa_agg_jf_item
> res
= vNULL
;
2961 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
2962 if (av
->index
== index
2963 && (av
->offset
- offset
) >= 0)
2965 struct ipa_agg_jf_item item
;
2966 gcc_checking_assert (av
->value
);
2967 item
.offset
= av
->offset
- offset
;
2968 item
.value
= av
->value
;
2969 res
.safe_push (item
);
2975 /* Intersect all values in INTER with those that we have already scheduled to
2976 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2977 (while subtracting OFFSET). */
2980 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
2981 vec
<ipa_agg_jf_item
> *inter
,
2982 HOST_WIDE_INT offset
)
2984 struct ipa_agg_replacement_value
*srcvals
;
2985 struct ipa_agg_jf_item
*item
;
2988 srcvals
= ipa_get_agg_replacements_for_node (node
);
2995 FOR_EACH_VEC_ELT (*inter
, i
, item
)
2997 struct ipa_agg_replacement_value
*av
;
3001 for (av
= srcvals
; av
; av
= av
->next
)
3003 gcc_checking_assert (av
->value
);
3004 if (av
->index
== index
3005 && av
->offset
- offset
== item
->offset
)
3007 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
3013 item
->value
= NULL_TREE
;
3017 /* Intersect values in INTER with aggregate values that come along edge CS to
3018 parameter number INDEX and return it. If INTER does not actually exist yet,
3019 copy all incoming values to it. If we determine we ended up with no values
3020 whatsoever, return a released vector. */
3022 static vec
<ipa_agg_jf_item
>
3023 intersect_aggregates_with_edge (struct cgraph_edge
*cs
, int index
,
3024 vec
<ipa_agg_jf_item
> inter
)
3026 struct ipa_jump_func
*jfunc
;
3027 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), index
);
3028 if (jfunc
->type
== IPA_JF_PASS_THROUGH
3029 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
3031 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3032 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
3034 if (caller_info
->ipcp_orig_node
)
3036 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
3037 struct ipcp_param_lattices
*orig_plats
;
3038 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
3040 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
3042 if (!inter
.exists ())
3043 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, 0);
3045 intersect_with_agg_replacements (cs
->caller
, src_idx
,
3056 struct ipcp_param_lattices
*src_plats
;
3057 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
3058 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
3060 /* Currently we do not produce clobber aggregate jump
3061 functions, adjust when we do. */
3062 gcc_checking_assert (!jfunc
->agg
.items
);
3063 if (!inter
.exists ())
3064 inter
= copy_plats_to_inter (src_plats
, 0);
3066 intersect_with_plats (src_plats
, &inter
, 0);
3075 else if (jfunc
->type
== IPA_JF_ANCESTOR
3076 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
3078 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3079 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
3080 struct ipcp_param_lattices
*src_plats
;
3081 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
3083 if (caller_info
->ipcp_orig_node
)
3085 if (!inter
.exists ())
3086 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, delta
);
3088 intersect_with_agg_replacements (cs
->caller
, src_idx
, &inter
,
3093 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);;
3094 /* Currently we do not produce clobber aggregate jump
3095 functions, adjust when we do. */
3096 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
3097 if (!inter
.exists ())
3098 inter
= copy_plats_to_inter (src_plats
, delta
);
3100 intersect_with_plats (src_plats
, &inter
, delta
);
3103 else if (jfunc
->agg
.items
)
3105 struct ipa_agg_jf_item
*item
;
3108 if (!inter
.exists ())
3109 for (unsigned i
= 0; i
< jfunc
->agg
.items
->length (); i
++)
3110 inter
.safe_push ((*jfunc
->agg
.items
)[i
]);
3112 FOR_EACH_VEC_ELT (inter
, k
, item
)
3115 bool found
= false;;
3120 while ((unsigned) l
< jfunc
->agg
.items
->length ())
3122 struct ipa_agg_jf_item
*ti
;
3123 ti
= &(*jfunc
->agg
.items
)[l
];
3124 if (ti
->offset
> item
->offset
)
3126 if (ti
->offset
== item
->offset
)
3128 gcc_checking_assert (ti
->value
);
3129 if (values_equal_for_ipcp_p (item
->value
,
3143 return vec
<ipa_agg_jf_item
>();
3148 /* Look at edges in CALLERS and collect all known aggregate values that arrive
3149 from all of them. */
3151 static struct ipa_agg_replacement_value
*
3152 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
3153 vec
<cgraph_edge
*> callers
)
3155 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3156 struct ipa_agg_replacement_value
*res
;
3157 struct ipa_agg_replacement_value
**tail
= &res
;
3158 struct cgraph_edge
*cs
;
3159 int i
, j
, count
= ipa_get_param_count (dest_info
);
3161 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3163 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3168 for (i
= 0; i
< count
; i
++)
3170 struct cgraph_edge
*cs
;
3171 vec
<ipa_agg_jf_item
> inter
= vNULL
;
3172 struct ipa_agg_jf_item
*item
;
3173 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (dest_info
, i
);
3176 /* Among other things, the following check should deal with all by_ref
3178 if (plats
->aggs_bottom
)
3181 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3183 inter
= intersect_aggregates_with_edge (cs
, i
, inter
);
3185 if (!inter
.exists ())
3189 FOR_EACH_VEC_ELT (inter
, j
, item
)
3191 struct ipa_agg_replacement_value
*v
;
3196 v
= ggc_alloc
<ipa_agg_replacement_value
> ();
3198 v
->offset
= item
->offset
;
3199 v
->value
= item
->value
;
3200 v
->by_ref
= plats
->aggs_by_ref
;
3206 if (inter
.exists ())
3213 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3215 static struct ipa_agg_replacement_value
*
3216 known_aggs_to_agg_replacement_list (vec
<ipa_agg_jump_function
> known_aggs
)
3218 struct ipa_agg_replacement_value
*res
;
3219 struct ipa_agg_replacement_value
**tail
= &res
;
3220 struct ipa_agg_jump_function
*aggjf
;
3221 struct ipa_agg_jf_item
*item
;
3224 FOR_EACH_VEC_ELT (known_aggs
, i
, aggjf
)
3225 FOR_EACH_VEC_SAFE_ELT (aggjf
->items
, j
, item
)
3227 struct ipa_agg_replacement_value
*v
;
3228 v
= ggc_alloc
<ipa_agg_replacement_value
> ();
3230 v
->offset
= item
->offset
;
3231 v
->value
= item
->value
;
3232 v
->by_ref
= aggjf
->by_ref
;
3240 /* Determine whether CS also brings all scalar values that the NODE is
3244 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
3245 struct cgraph_node
*node
)
3247 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3248 int count
= ipa_get_param_count (dest_info
);
3249 struct ipa_node_params
*caller_info
;
3250 struct ipa_edge_args
*args
;
3253 caller_info
= IPA_NODE_REF (cs
->caller
);
3254 args
= IPA_EDGE_REF (cs
);
3255 for (i
= 0; i
< count
; i
++)
3257 struct ipa_jump_func
*jump_func
;
3260 val
= dest_info
->known_vals
[i
];
3264 if (i
>= ipa_get_cs_argument_count (args
))
3266 jump_func
= ipa_get_ith_jump_func (args
, i
);
3267 t
= ipa_value_from_jfunc (caller_info
, jump_func
);
3268 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
3274 /* Determine whether CS also brings all aggregate values that NODE is
3277 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
3278 struct cgraph_node
*node
)
3280 struct ipa_node_params
*orig_caller_info
= IPA_NODE_REF (cs
->caller
);
3281 struct ipa_node_params
*orig_node_info
;
3282 struct ipa_agg_replacement_value
*aggval
;
3285 aggval
= ipa_get_agg_replacements_for_node (node
);
3289 count
= ipa_get_param_count (IPA_NODE_REF (node
));
3290 ec
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
3292 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3293 if (aggval
->index
>= ec
)
3296 orig_node_info
= IPA_NODE_REF (IPA_NODE_REF (node
)->ipcp_orig_node
);
3297 if (orig_caller_info
->ipcp_orig_node
)
3298 orig_caller_info
= IPA_NODE_REF (orig_caller_info
->ipcp_orig_node
);
3300 for (i
= 0; i
< count
; i
++)
3302 static vec
<ipa_agg_jf_item
> values
= vec
<ipa_agg_jf_item
>();
3303 struct ipcp_param_lattices
*plats
;
3304 bool interesting
= false;
3305 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3306 if (aggval
->index
== i
)
3314 plats
= ipa_get_parm_lattices (orig_node_info
, aggval
->index
);
3315 if (plats
->aggs_bottom
)
3318 values
= intersect_aggregates_with_edge (cs
, i
, values
);
3319 if (!values
.exists ())
3322 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
3323 if (aggval
->index
== i
)
3325 struct ipa_agg_jf_item
*item
;
3328 FOR_EACH_VEC_ELT (values
, j
, item
)
3330 && item
->offset
== av
->offset
3331 && values_equal_for_ipcp_p (item
->value
, av
->value
))
3346 /* Given an original NODE and a VAL for which we have already created a
3347 specialized clone, look whether there are incoming edges that still lead
3348 into the old node but now also bring the requested value and also conform to
3349 all other criteria such that they can be redirected the the special node.
3350 This function can therefore redirect the final edge in a SCC. */
3353 perhaps_add_new_callers (struct cgraph_node
*node
, struct ipcp_value
*val
)
3355 struct ipcp_value_source
*src
;
3356 gcov_type redirected_sum
= 0;
3358 for (src
= val
->sources
; src
; src
= src
->next
)
3360 struct cgraph_edge
*cs
= src
->cs
;
3363 enum availability availability
;
3364 struct cgraph_node
*dst
= cs
->callee
->function_symbol (&availability
);
3365 if ((dst
== node
|| IPA_NODE_REF (dst
)->is_all_contexts_clone
)
3366 && availability
> AVAIL_INTERPOSABLE
3367 && cgraph_edge_brings_value_p (cs
, src
))
3369 if (cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
3370 && cgraph_edge_brings_all_agg_vals_for_node (cs
,
3374 fprintf (dump_file
, " - adding an extra caller %s/%i"
3376 xstrdup (cs
->caller
->name ()),
3378 xstrdup (val
->spec_node
->name ()),
3379 val
->spec_node
->order
);
3381 cs
->redirect_callee (val
->spec_node
);
3382 redirected_sum
+= cs
->count
;
3385 cs
= get_next_cgraph_edge_clone (cs
);
3390 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
3394 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3397 move_binfos_to_values (vec
<tree
> known_vals
,
3398 vec
<tree
> known_binfos
)
3403 for (i
= 0; known_binfos
.iterate (i
, &t
); i
++)
3408 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3409 among those in the AGGVALS list. */
3412 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value
*aggvals
,
3413 int index
, HOST_WIDE_INT offset
, tree value
)
3417 if (aggvals
->index
== index
3418 && aggvals
->offset
== offset
3419 && values_equal_for_ipcp_p (aggvals
->value
, value
))
3421 aggvals
= aggvals
->next
;
3426 /* Decide wheter to create a special version of NODE for value VAL of parameter
3427 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3428 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3429 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3432 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
3433 struct ipcp_value
*val
, vec
<tree
> known_csts
,
3434 vec
<tree
> known_binfos
)
3436 struct ipa_agg_replacement_value
*aggvals
;
3437 int freq_sum
, caller_count
;
3438 gcov_type count_sum
;
3439 vec
<cgraph_edge
*> callers
;
3444 perhaps_add_new_callers (node
, val
);
3447 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
3449 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3450 fprintf (dump_file
, " Ignoring candidate value because "
3451 "max_new_size would be reached with %li.\n",
3452 val
->local_size_cost
+ overall_size
);
3455 else if (!get_info_about_necessary_edges (val
, &freq_sum
, &count_sum
,
3459 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3461 fprintf (dump_file
, " - considering value ");
3462 print_ipcp_constant_value (dump_file
, val
->value
);
3463 fprintf (dump_file
, " for ");
3464 ipa_dump_param (dump_file
, IPA_NODE_REF (node
), index
);
3466 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
3467 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
3470 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
3471 freq_sum
, count_sum
,
3472 val
->local_size_cost
)
3473 && !good_cloning_opportunity_p (node
,
3474 val
->local_time_benefit
3475 + val
->prop_time_benefit
,
3476 freq_sum
, count_sum
,
3477 val
->local_size_cost
3478 + val
->prop_size_cost
))
3482 fprintf (dump_file
, " Creating a specialized node of %s/%i.\n",
3483 node
->name (), node
->order
);
3485 callers
= gather_edges_for_value (val
, caller_count
);
3486 kv
= known_csts
.copy ();
3487 move_binfos_to_values (kv
, known_binfos
);
3489 kv
[index
] = val
->value
;
3490 find_more_scalar_values_for_callers_subset (node
, kv
, callers
);
3491 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
3492 gcc_checking_assert (offset
== -1
3493 || ipcp_val_in_agg_replacements_p (aggvals
, index
,
3494 offset
, val
->value
));
3495 val
->spec_node
= create_specialized_node (node
, kv
, aggvals
, callers
);
3496 overall_size
+= val
->local_size_cost
;
3498 /* TODO: If for some lattice there is only one other known value
3499 left, make a special node for it too. */
3504 /* Decide whether and what specialized clones of NODE should be created. */
3507 decide_whether_version_node (struct cgraph_node
*node
)
3509 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3510 int i
, count
= ipa_get_param_count (info
);
3511 vec
<tree
> known_csts
, known_binfos
;
3512 vec
<ipa_agg_jump_function
> known_aggs
= vNULL
;
3518 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3519 fprintf (dump_file
, "\nEvaluating opportunities for %s/%i.\n",
3520 node
->name (), node
->order
);
3522 gather_context_independent_values (info
, &known_csts
, &known_binfos
,
3523 info
->do_clone_for_all_contexts
? &known_aggs
3526 for (i
= 0; i
< count
;i
++)
3528 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3529 struct ipcp_lattice
*lat
= &plats
->itself
;
3530 struct ipcp_value
*val
;
3534 && !known_binfos
[i
])
3535 for (val
= lat
->values
; val
; val
= val
->next
)
3536 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
3539 if (!plats
->aggs_bottom
)
3541 struct ipcp_agg_lattice
*aglat
;
3542 struct ipcp_value
*val
;
3543 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3544 if (!aglat
->bottom
&& aglat
->values
3545 /* If the following is false, the one value is in
3547 && (plats
->aggs_contain_variable
3548 || !ipa_lat_is_single_const (aglat
)))
3549 for (val
= aglat
->values
; val
; val
= val
->next
)
3550 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
3551 known_csts
, known_binfos
);
3553 info
= IPA_NODE_REF (node
);
3556 if (info
->do_clone_for_all_contexts
)
3558 struct cgraph_node
*clone
;
3559 vec
<cgraph_edge
*> callers
;
3562 fprintf (dump_file
, " - Creating a specialized node of %s/%i "
3563 "for all known contexts.\n", node
->name (),
3566 callers
= node
->collect_callers ();
3567 move_binfos_to_values (known_csts
, known_binfos
);
3568 clone
= create_specialized_node (node
, known_csts
,
3569 known_aggs_to_agg_replacement_list (known_aggs
),
3571 info
= IPA_NODE_REF (node
);
3572 info
->do_clone_for_all_contexts
= false;
3573 IPA_NODE_REF (clone
)->is_all_contexts_clone
= true;
3574 for (i
= 0; i
< count
; i
++)
3575 vec_free (known_aggs
[i
].items
);
3576 known_aggs
.release ();
3580 known_csts
.release ();
3582 known_binfos
.release ();
3586 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3589 spread_undeadness (struct cgraph_node
*node
)
3591 struct cgraph_edge
*cs
;
3593 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
3594 if (ipa_edge_within_scc (cs
))
3596 struct cgraph_node
*callee
;
3597 struct ipa_node_params
*info
;
3599 callee
= cs
->callee
->function_symbol (NULL
);
3600 info
= IPA_NODE_REF (callee
);
3602 if (info
->node_dead
)
3604 info
->node_dead
= 0;
3605 spread_undeadness (callee
);
3610 /* Return true if NODE has a caller from outside of its SCC that is not
3611 dead. Worker callback for cgraph_for_node_and_aliases. */
3614 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
3615 void *data ATTRIBUTE_UNUSED
)
3617 struct cgraph_edge
*cs
;
3619 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
3620 if (cs
->caller
->thunk
.thunk_p
3621 && cs
->caller
->call_for_symbol_thunks_and_aliases
3622 (has_undead_caller_from_outside_scc_p
, NULL
, true))
3624 else if (!ipa_edge_within_scc (cs
)
3625 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
3631 /* Identify nodes within the same SCC as NODE which are no longer needed
3632 because of new clones and will be removed as unreachable. */
3635 identify_dead_nodes (struct cgraph_node
*node
)
3637 struct cgraph_node
*v
;
3638 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3639 if (v
->will_be_removed_from_program_if_no_direct_calls_p ()
3640 && !v
->call_for_symbol_thunks_and_aliases
3641 (has_undead_caller_from_outside_scc_p
, NULL
, true))
3642 IPA_NODE_REF (v
)->node_dead
= 1;
3644 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3645 if (!IPA_NODE_REF (v
)->node_dead
)
3646 spread_undeadness (v
);
3648 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3650 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3651 if (IPA_NODE_REF (v
)->node_dead
)
3652 fprintf (dump_file
, " Marking node as dead: %s/%i.\n",
3653 v
->name (), v
->order
);
3657 /* The decision stage. Iterate over the topological order of call graph nodes
3658 TOPO and make specialized clones if deemed beneficial. */
3661 ipcp_decision_stage (struct ipa_topo_info
*topo
)
3666 fprintf (dump_file
, "\nIPA decision stage:\n\n");
3668 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3670 struct cgraph_node
*node
= topo
->order
[i
];
3671 bool change
= false, iterate
= true;
3675 struct cgraph_node
*v
;
3677 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
3678 if (v
->has_gimple_body_p ()
3679 && ipcp_versionable_function_p (v
))
3680 iterate
|= decide_whether_version_node (v
);
3685 identify_dead_nodes (node
);
3689 /* The IPCP driver. */
3694 struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
3695 struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
3696 struct ipa_topo_info topo
;
3698 ipa_check_create_node_params ();
3699 ipa_check_create_edge_args ();
3700 grow_edge_clone_vectors ();
3701 edge_duplication_hook_holder
=
3702 symtab
->add_edge_duplication_hook (&ipcp_edge_duplication_hook
, NULL
);
3703 edge_removal_hook_holder
=
3704 symtab
->add_edge_removal_hook (&ipcp_edge_removal_hook
, NULL
);
3706 ipcp_values_pool
= create_alloc_pool ("IPA-CP values",
3707 sizeof (struct ipcp_value
), 32);
3708 ipcp_sources_pool
= create_alloc_pool ("IPA-CP value sources",
3709 sizeof (struct ipcp_value_source
), 64);
3710 ipcp_agg_lattice_pool
= create_alloc_pool ("IPA_CP aggregate lattices",
3711 sizeof (struct ipcp_agg_lattice
),
3715 fprintf (dump_file
, "\nIPA structures before propagation:\n");
3716 if (dump_flags
& TDF_DETAILS
)
3717 ipa_print_all_params (dump_file
);
3718 ipa_print_all_jump_functions (dump_file
);
3721 /* Topological sort. */
3722 build_toporder_info (&topo
);
3723 /* Do the interprocedural propagation. */
3724 ipcp_propagate_stage (&topo
);
3725 /* Decide what constant propagation and cloning should be performed. */
3726 ipcp_decision_stage (&topo
);
3728 /* Free all IPCP structures. */
3729 free_toporder_info (&topo
);
3730 next_edge_clone
.release ();
3731 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
3732 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
3733 ipa_free_all_structures_after_ipa_cp ();
3735 fprintf (dump_file
, "\nIPA constant propagation end\n");
3739 /* Initialization and computation of IPCP data structures. This is the initial
3740 intraprocedural analysis of functions, which gathers information to be
3741 propagated later on. */
3744 ipcp_generate_summary (void)
3746 struct cgraph_node
*node
;
3749 fprintf (dump_file
, "\nIPA constant propagation start:\n");
3750 ipa_register_cgraph_hooks ();
3752 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
3754 node
->local
.versionable
3755 = tree_versionable_function_p (node
->decl
);
3756 ipa_analyze_node (node
);
3760 /* Write ipcp summary for nodes in SET. */
3763 ipcp_write_summary (void)
3765 ipa_prop_write_jump_functions ();
3768 /* Read ipcp summary. */
3771 ipcp_read_summary (void)
3773 ipa_prop_read_jump_functions ();
3778 const pass_data pass_data_ipa_cp
=
3780 IPA_PASS
, /* type */
3782 OPTGROUP_NONE
, /* optinfo_flags */
3783 TV_IPA_CONSTANT_PROP
, /* tv_id */
3784 0, /* properties_required */
3785 0, /* properties_provided */
3786 0, /* properties_destroyed */
3787 0, /* todo_flags_start */
3788 ( TODO_dump_symtab
| TODO_remove_functions
), /* todo_flags_finish */
3791 class pass_ipa_cp
: public ipa_opt_pass_d
3794 pass_ipa_cp (gcc::context
*ctxt
)
3795 : ipa_opt_pass_d (pass_data_ipa_cp
, ctxt
,
3796 ipcp_generate_summary
, /* generate_summary */
3797 ipcp_write_summary
, /* write_summary */
3798 ipcp_read_summary
, /* read_summary */
3799 ipa_prop_write_all_agg_replacement
, /*
3800 write_optimization_summary */
3801 ipa_prop_read_all_agg_replacement
, /*
3802 read_optimization_summary */
3803 NULL
, /* stmt_fixup */
3804 0, /* function_transform_todo_flags_start */
3805 ipcp_transform_function
, /* function_transform */
3806 NULL
) /* variable_transform */
3809 /* opt_pass methods: */
3810 virtual bool gate (function
*)
3812 /* FIXME: We should remove the optimize check after we ensure we never run
3813 IPA passes when not optimizing. */
3814 return flag_ipa_cp
&& optimize
;
3817 virtual unsigned int execute (function
*) { return ipcp_driver (); }
3819 }; // class pass_ipa_cp
3824 make_pass_ipa_cp (gcc::context
*ctxt
)
3826 return new pass_ipa_cp (ctxt
);