1 /* Interprocedural constant propagation
2 Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
5 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify it under
11 the terms of the GNU General Public License as published by the Free
12 Software Foundation; either version 3, or (at your option) any later
15 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
24 /* Interprocedural constant propagation (IPA-CP).
26 The goal of this transformation is to
28 1) discover functions which are always invoked with some arguments with the
29 same known constant values and modify the functions so that the
30 subsequent optimizations can take advantage of the knowledge, and
32 2) partial specialization - create specialized versions of functions
33 transformed in this way if some parameters are known constants only in
34 certain contexts but the estimated tradeoff between speedup and cost size
37 The algorithm also propagates types and attempts to perform type based
38 devirtualization. Types are propagated much like constants.
40 The algorithm basically consists of three stages. In the first, functions
41 are analyzed one at a time and jump functions are constructed for all known
42 call-sites. In the second phase, the pass propagates information from the
43 jump functions across the call to reveal what values are available at what
44 call sites, performs estimations of effects of known values on functions and
45 their callees, and finally decides what specialized extra versions should be
46 created. In the third, the special versions materialize and appropriate
49 The algorithm used is to a certain extent based on "Interprocedural Constant
50 Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
51 Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
52 Cooper, Mary W. Hall, and Ken Kennedy.
55 First stage - intraprocedural analysis
56 =======================================
58 This phase computes jump_function and modification flags.
60 A jump function for a call-site represents the values passed as an actual
61 arguments of a given call-site. In principle, there are three types of
64 Pass through - the caller's formal parameter is passed as an actual
65 argument, plus an operation on it can be performed.
66 Constant - a constant is passed as an actual argument.
67 Unknown - neither of the above.
69 All jump function types are described in detail in ipa-prop.h, together with
70 the data structures that represent them and methods of accessing them.
72 ipcp_generate_summary() is the main function of the first stage.
74 Second stage - interprocedural analysis
75 ========================================
77 This stage is itself divided into two phases. In the first, we propagate
78 known values over the call graph, in the second, we make cloning decisions.
79 It uses a different algorithm than the original Callahan's paper.
81 First, we traverse the functions topologically from callers to callees and,
82 for each strongly connected component (SCC), we propagate constants
83 according to previously computed jump functions. We also record what known
84 values depend on other known values and estimate local effects. Finally, we
85 propagate cumulative information about these effects from dependent values
86 to those on which they depend.
88 Second, we again traverse the call graph in the same topological order and
89 make clones for functions which we know are called with the same values in
90 all contexts and decide about extra specialized clones of functions just for
91 some contexts - these decisions are based on both local estimates and
92 cumulative estimates propagated from callees.
94 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
97 Third phase - materialization of clones, call statement updates.
98 ============================================
100 This stage is currently performed by call graph code (mainly in cgraphunit.c
101 and tree-inline.c) according to instructions inserted to the call graph by
106 #include "coretypes.h"
111 #include "ipa-prop.h"
112 #include "tree-flow.h"
113 #include "tree-pass.h"
115 #include "diagnostic.h"
116 #include "tree-pretty-print.h"
117 #include "tree-inline.h"
119 #include "ipa-inline.h"
120 #include "ipa-utils.h"
124 /* Describes a particular source for an IPA-CP value. */
126 struct ipcp_value_source
128 /* Aggregate offset of the source, negative if the source is scalar value of
129 the argument itself. */
130 HOST_WIDE_INT offset
;
131 /* The incoming edge that brought the value. */
132 struct cgraph_edge
*cs
;
133 /* If the jump function that resulted into his value was a pass-through or an
134 ancestor, this is the ipcp_value of the caller from which the described
135 value has been derived. Otherwise it is NULL. */
136 struct ipcp_value
*val
;
137 /* Next pointer in a linked list of sources of a value. */
138 struct ipcp_value_source
*next
;
139 /* If the jump function that resulted into his value was a pass-through or an
140 ancestor, this is the index of the parameter of the caller the jump
141 function references. */
145 /* Describes one particular value stored in struct ipcp_lattice. */
149 /* The actual value for the given parameter. This is either an IPA invariant
150 or a TREE_BINFO describing a type that can be used for
153 /* The list of sources from which this value originates. */
154 struct ipcp_value_source
*sources
;
155 /* Next pointers in a linked list of all values in a lattice. */
156 struct ipcp_value
*next
;
157 /* Next pointers in a linked list of values in a strongly connected component
159 struct ipcp_value
*scc_next
;
160 /* Next pointers in a linked list of SCCs of values sorted topologically
161 according their sources. */
162 struct ipcp_value
*topo_next
;
163 /* A specialized node created for this value, NULL if none has been (so far)
165 struct cgraph_node
*spec_node
;
166 /* Depth first search number and low link for topological sorting of
169 /* Time benefit and size cost that specializing the function for this value
170 would bring about in this function alone. */
171 int local_time_benefit
, local_size_cost
;
172 /* Time benefit and size cost that specializing the function for this value
173 can bring about in it's callees (transitively). */
174 int prop_time_benefit
, prop_size_cost
;
175 /* True if this valye is currently on the topo-sort stack. */
179 /* Lattice describing potential values of a formal parameter of a function, or
180 a part of an aggreagate. TOP is represented by a lattice with zero values
181 and with contains_variable and bottom flags cleared. BOTTOM is represented
182 by a lattice with the bottom flag set. In that case, values and
183 contains_variable flag should be disregarded. */
187 /* The list of known values and types in this lattice. Note that values are
188 not deallocated if a lattice is set to bottom because there may be value
189 sources referencing them. */
190 struct ipcp_value
*values
;
191 /* Number of known values and types in this lattice. */
193 /* The lattice contains a variable component (in addition to values). */
194 bool contains_variable
;
195 /* The value of the lattice is bottom (i.e. variable and unusable for any
200 /* Lattice with an offset to describe a part of an aggregate. */
202 struct ipcp_agg_lattice
: public ipcp_lattice
204 /* Offset that is being described by this lattice. */
205 HOST_WIDE_INT offset
;
206 /* Size so that we don't have to re-compute it every time we traverse the
207 list. Must correspond to TYPE_SIZE of all lat values. */
209 /* Next element of the linked list. */
210 struct ipcp_agg_lattice
*next
;
213 /* Structure containing lattices for a parameter itself and for pieces of
214 aggregates that are passed in the parameter or by a reference in a parameter
215 plus some other useful flags. */
217 struct ipcp_param_lattices
219 /* Lattice describing the value of the parameter itself. */
220 struct ipcp_lattice itself
;
221 /* Lattices describing aggregate parts. */
222 struct ipcp_agg_lattice
*aggs
;
223 /* Number of aggregate lattices */
225 /* True if aggregate data were passed by reference (as opposed to by
228 /* All aggregate lattices contain a variable component (in addition to
230 bool aggs_contain_variable
;
231 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
232 for any propagation). */
235 /* There is a virtual call based on this parameter. */
239 /* Allocation pools for values and their sources in ipa-cp. */
241 alloc_pool ipcp_values_pool
;
242 alloc_pool ipcp_sources_pool
;
243 alloc_pool ipcp_agg_lattice_pool
;
245 /* Maximal count found in program. */
247 static gcov_type max_count
;
249 /* Original overall size of the program. */
251 static long overall_size
, max_new_size
;
253 /* Head of the linked list of topologically sorted values. */
255 static struct ipcp_value
*values_topo
;
257 /* Return the param lattices structure corresponding to the Ith formal
258 parameter of the function described by INFO. */
259 static inline struct ipcp_param_lattices
*
260 ipa_get_parm_lattices (struct ipa_node_params
*info
, int i
)
262 gcc_assert (i
>= 0 && i
< ipa_get_param_count (info
));
263 gcc_checking_assert (!info
->ipcp_orig_node
);
264 gcc_checking_assert (info
->lattices
);
265 return &(info
->lattices
[i
]);
268 /* Return the lattice corresponding to the scalar value of the Ith formal
269 parameter of the function described by INFO. */
270 static inline struct ipcp_lattice
*
271 ipa_get_scalar_lat (struct ipa_node_params
*info
, int i
)
273 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
274 return &plats
->itself
;
277 /* Return whether LAT is a lattice with a single constant and without an
281 ipa_lat_is_single_const (struct ipcp_lattice
*lat
)
284 || lat
->contains_variable
285 || lat
->values_count
!= 1)
291 /* Return true iff the CS is an edge within a strongly connected component as
292 computed by ipa_reduced_postorder. */
295 edge_within_scc (struct cgraph_edge
*cs
)
297 struct ipa_dfs_info
*caller_dfs
= (struct ipa_dfs_info
*) cs
->caller
->symbol
.aux
;
298 struct ipa_dfs_info
*callee_dfs
;
299 struct cgraph_node
*callee
= cgraph_function_node (cs
->callee
, NULL
);
301 callee_dfs
= (struct ipa_dfs_info
*) callee
->symbol
.aux
;
304 && caller_dfs
->scc_no
== callee_dfs
->scc_no
);
307 /* Print V which is extracted from a value in a lattice to F. */
310 print_ipcp_constant_value (FILE * f
, tree v
)
312 if (TREE_CODE (v
) == TREE_BINFO
)
314 fprintf (f
, "BINFO ");
315 print_generic_expr (f
, BINFO_TYPE (v
), 0);
317 else if (TREE_CODE (v
) == ADDR_EXPR
318 && TREE_CODE (TREE_OPERAND (v
, 0)) == CONST_DECL
)
321 print_generic_expr (f
, DECL_INITIAL (TREE_OPERAND (v
, 0)), 0);
324 print_generic_expr (f
, v
, 0);
327 /* Print a lattice LAT to F. */
330 print_lattice (FILE * f
, struct ipcp_lattice
*lat
,
331 bool dump_sources
, bool dump_benefits
)
333 struct ipcp_value
*val
;
338 fprintf (f
, "BOTTOM\n");
342 if (!lat
->values_count
&& !lat
->contains_variable
)
344 fprintf (f
, "TOP\n");
348 if (lat
->contains_variable
)
350 fprintf (f
, "VARIABLE");
356 for (val
= lat
->values
; val
; val
= val
->next
)
358 if (dump_benefits
&& prev
)
360 else if (!dump_benefits
&& prev
)
365 print_ipcp_constant_value (f
, val
->value
);
369 struct ipcp_value_source
*s
;
371 fprintf (f
, " [from:");
372 for (s
= val
->sources
; s
; s
= s
->next
)
373 fprintf (f
, " %i(%i)", s
->cs
->caller
->uid
,s
->cs
->frequency
);
378 fprintf (f
, " [loc_time: %i, loc_size: %i, "
379 "prop_time: %i, prop_size: %i]\n",
380 val
->local_time_benefit
, val
->local_size_cost
,
381 val
->prop_time_benefit
, val
->prop_size_cost
);
387 /* Print all ipcp_lattices of all functions to F. */
390 print_all_lattices (FILE * f
, bool dump_sources
, bool dump_benefits
)
392 struct cgraph_node
*node
;
395 fprintf (f
, "\nLattices:\n");
396 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
398 struct ipa_node_params
*info
;
400 info
= IPA_NODE_REF (node
);
401 fprintf (f
, " Node: %s/%i:\n", cgraph_node_name (node
), node
->uid
);
402 count
= ipa_get_param_count (info
);
403 for (i
= 0; i
< count
; i
++)
405 struct ipcp_agg_lattice
*aglat
;
406 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
407 fprintf (f
, " param [%d]: ", i
);
408 print_lattice (f
, &plats
->itself
, dump_sources
, dump_benefits
);
410 if (plats
->virt_call
)
411 fprintf (f
, " virt_call flag set\n");
413 if (plats
->aggs_bottom
)
415 fprintf (f
, " AGGS BOTTOM\n");
418 if (plats
->aggs_contain_variable
)
419 fprintf (f
, " AGGS VARIABLE\n");
420 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
422 fprintf (f
, " %soffset " HOST_WIDE_INT_PRINT_DEC
": ",
423 plats
->aggs_by_ref
? "ref " : "", aglat
->offset
);
424 print_lattice (f
, aglat
, dump_sources
, dump_benefits
);
430 /* Determine whether it is at all technically possible to create clones of NODE
431 and store this information in the ipa_node_params structure associated
435 determine_versionability (struct cgraph_node
*node
)
437 const char *reason
= NULL
;
439 /* There are a number of generic reasons functions cannot be versioned. We
440 also cannot remove parameters if there are type attributes such as fnspec
442 if (node
->alias
|| node
->thunk
.thunk_p
)
443 reason
= "alias or thunk";
444 else if (!node
->local
.versionable
)
445 reason
= "not a tree_versionable_function";
446 else if (cgraph_function_body_availability (node
) <= AVAIL_OVERWRITABLE
)
447 reason
= "insufficient body availability";
449 if (reason
&& dump_file
&& !node
->alias
&& !node
->thunk
.thunk_p
)
450 fprintf (dump_file
, "Function %s/%i is not versionable, reason: %s.\n",
451 cgraph_node_name (node
), node
->uid
, reason
);
453 node
->local
.versionable
= (reason
== NULL
);
456 /* Return true if it is at all technically possible to create clones of a
460 ipcp_versionable_function_p (struct cgraph_node
*node
)
462 return node
->local
.versionable
;
465 /* Structure holding accumulated information about callers of a node. */
467 struct caller_statistics
470 int n_calls
, n_hot_calls
, freq_sum
;
473 /* Initialize fields of STAT to zeroes. */
476 init_caller_stats (struct caller_statistics
*stats
)
478 stats
->count_sum
= 0;
480 stats
->n_hot_calls
= 0;
484 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
485 non-thunk incoming edges to NODE. */
488 gather_caller_stats (struct cgraph_node
*node
, void *data
)
490 struct caller_statistics
*stats
= (struct caller_statistics
*) data
;
491 struct cgraph_edge
*cs
;
493 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
494 if (cs
->caller
->thunk
.thunk_p
)
495 cgraph_for_node_and_aliases (cs
->caller
, gather_caller_stats
,
499 stats
->count_sum
+= cs
->count
;
500 stats
->freq_sum
+= cs
->frequency
;
502 if (cgraph_maybe_hot_edge_p (cs
))
503 stats
->n_hot_calls
++;
509 /* Return true if this NODE is viable candidate for cloning. */
512 ipcp_cloning_candidate_p (struct cgraph_node
*node
)
514 struct caller_statistics stats
;
516 gcc_checking_assert (cgraph_function_with_gimple_body_p (node
));
518 if (!flag_ipa_cp_clone
)
521 fprintf (dump_file
, "Not considering %s for cloning; "
522 "-fipa-cp-clone disabled.\n",
523 cgraph_node_name (node
));
527 if (!optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
)))
530 fprintf (dump_file
, "Not considering %s for cloning; "
531 "optimizing it for size.\n",
532 cgraph_node_name (node
));
536 init_caller_stats (&stats
);
537 cgraph_for_node_and_aliases (node
, gather_caller_stats
, &stats
, false);
539 if (inline_summary (node
)->self_size
< stats
.n_calls
)
542 fprintf (dump_file
, "Considering %s for cloning; code might shrink.\n",
543 cgraph_node_name (node
));
547 /* When profile is available and function is hot, propagate into it even if
548 calls seems cold; constant propagation can improve function's speed
552 if (stats
.count_sum
> node
->count
* 90 / 100)
555 fprintf (dump_file
, "Considering %s for cloning; "
556 "usually called directly.\n",
557 cgraph_node_name (node
));
561 if (!stats
.n_hot_calls
)
564 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
565 cgraph_node_name (node
));
569 fprintf (dump_file
, "Considering %s for cloning.\n",
570 cgraph_node_name (node
));
574 /* Arrays representing a topological ordering of call graph nodes and a stack
575 of noes used during constant propagation. */
579 struct cgraph_node
**order
;
580 struct cgraph_node
**stack
;
581 int nnodes
, stack_top
;
584 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
587 build_toporder_info (struct topo_info
*topo
)
589 topo
->order
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
590 topo
->stack
= XCNEWVEC (struct cgraph_node
*, cgraph_n_nodes
);
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 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 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 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 (cgraph_function_with_gimple_body_p (node
));
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 cgraph_node_name (node
), node
->uid
,
732 disable
? "BOTTOM" : "VARIABLE");
735 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
736 if (ie
->indirect_info
->polymorphic
)
738 gcc_checking_assert (ie
->indirect_info
->param_index
>= 0);
739 ipa_get_parm_lattices (info
,
740 ie
->indirect_info
->param_index
)->virt_call
= 1;
744 /* Return the result of a (possibly arithmetic) pass through jump function
745 JFUNC on the constant value INPUT. Return NULL_TREE if that cannot be
746 determined or itself is considered an interprocedural invariant. */
749 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
)
753 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
755 else if (TREE_CODE (input
) == TREE_BINFO
)
758 gcc_checking_assert (is_gimple_ip_invariant (input
));
759 if (TREE_CODE_CLASS (ipa_get_jf_pass_through_operation (jfunc
))
761 restype
= boolean_type_node
;
763 restype
= TREE_TYPE (input
);
764 res
= fold_binary (ipa_get_jf_pass_through_operation (jfunc
), restype
,
765 input
, ipa_get_jf_pass_through_operand (jfunc
));
767 if (res
&& !is_gimple_ip_invariant (res
))
773 /* Return the result of an ancestor jump function JFUNC on the constant value
774 INPUT. Return NULL_TREE if that cannot be determined. */
777 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
779 if (TREE_CODE (input
) == TREE_BINFO
)
780 return get_binfo_at_offset (input
,
781 ipa_get_jf_ancestor_offset (jfunc
),
782 ipa_get_jf_ancestor_type (jfunc
));
783 else if (TREE_CODE (input
) == ADDR_EXPR
)
785 tree t
= TREE_OPERAND (input
, 0);
786 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
787 ipa_get_jf_ancestor_offset (jfunc
),
788 ipa_get_jf_ancestor_type (jfunc
), NULL
, false);
789 return build_fold_addr_expr (t
);
795 /* Extract the acual BINFO being described by JFUNC which must be a known type
799 ipa_value_from_known_type_jfunc (struct ipa_jump_func
*jfunc
)
801 tree base_binfo
= TYPE_BINFO (ipa_get_jf_known_type_base_type (jfunc
));
804 return get_binfo_at_offset (base_binfo
,
805 ipa_get_jf_known_type_offset (jfunc
),
806 ipa_get_jf_known_type_component_type (jfunc
));
809 /* Determine whether JFUNC evaluates to a known value (that is either a
810 constant or a binfo) and if so, return it. Otherwise return NULL. INFO
811 describes the caller node so that pass-through jump functions can be
815 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
)
817 if (jfunc
->type
== IPA_JF_CONST
)
818 return ipa_get_jf_constant (jfunc
);
819 else if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
820 return ipa_value_from_known_type_jfunc (jfunc
);
821 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
822 || jfunc
->type
== IPA_JF_ANCESTOR
)
827 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
828 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
830 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
832 if (info
->ipcp_orig_node
)
833 input
= VEC_index (tree
, info
->known_vals
, idx
);
836 struct ipcp_lattice
*lat
;
840 gcc_checking_assert (!flag_ipa_cp
);
843 lat
= ipa_get_scalar_lat (info
, idx
);
844 if (!ipa_lat_is_single_const (lat
))
846 input
= lat
->values
->value
;
852 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
853 return ipa_get_jf_pass_through_result (jfunc
, input
);
855 return ipa_get_jf_ancestor_result (jfunc
, input
);
862 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
863 bottom, not containing a variable component and without any known value at
867 ipcp_verify_propagated_values (void)
869 struct cgraph_node
*node
;
871 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
873 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
874 int i
, count
= ipa_get_param_count (info
);
876 for (i
= 0; i
< count
; i
++)
878 struct ipcp_lattice
*lat
= ipa_get_scalar_lat (info
, i
);
881 && !lat
->contains_variable
882 && lat
->values_count
== 0)
886 fprintf (dump_file
, "\nIPA lattices after constant "
888 print_all_lattices (dump_file
, true, false);
897 /* Return true iff X and Y should be considered equal values by IPA-CP. */
900 values_equal_for_ipcp_p (tree x
, tree y
)
902 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
907 if (TREE_CODE (x
) == TREE_BINFO
|| TREE_CODE (y
) == TREE_BINFO
)
910 if (TREE_CODE (x
) == ADDR_EXPR
911 && TREE_CODE (y
) == ADDR_EXPR
912 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
913 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
914 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
915 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
917 return operand_equal_p (x
, y
, 0);
920 /* Add a new value source to VAL, marking that a value comes from edge CS and
921 (if the underlying jump function is a pass-through or an ancestor one) from
922 a caller value SRC_VAL of a caller parameter described by SRC_INDEX. OFFSET
923 is negative if the source was the scalar value of the parameter itself or
924 the offset within an aggregate. */
927 add_value_source (struct ipcp_value
*val
, struct cgraph_edge
*cs
,
928 struct ipcp_value
*src_val
, int src_idx
, HOST_WIDE_INT offset
)
930 struct ipcp_value_source
*src
;
932 src
= (struct ipcp_value_source
*) pool_alloc (ipcp_sources_pool
);
933 src
->offset
= offset
;
936 src
->index
= src_idx
;
938 src
->next
= val
->sources
;
942 /* Try to add NEWVAL to LAT, potentially creating a new struct ipcp_value for
943 it. CS, SRC_VAL SRC_INDEX and OFFSET are meant for add_value_source and
944 have the same meaning. */
947 add_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
948 struct cgraph_edge
*cs
, struct ipcp_value
*src_val
,
949 int src_idx
, HOST_WIDE_INT offset
)
951 struct ipcp_value
*val
;
956 for (val
= lat
->values
; val
; val
= val
->next
)
957 if (values_equal_for_ipcp_p (val
->value
, newval
))
959 if (edge_within_scc (cs
))
961 struct ipcp_value_source
*s
;
962 for (s
= val
->sources
; s
; s
= s
->next
)
969 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
973 if (lat
->values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
975 /* We can only free sources, not the values themselves, because sources
976 of other values in this this SCC might point to them. */
977 for (val
= lat
->values
; val
; val
= val
->next
)
981 struct ipcp_value_source
*src
= val
->sources
;
982 val
->sources
= src
->next
;
983 pool_free (ipcp_sources_pool
, src
);
988 return set_lattice_to_bottom (lat
);
992 val
= (struct ipcp_value
*) pool_alloc (ipcp_values_pool
);
993 memset (val
, 0, sizeof (*val
));
995 add_value_source (val
, cs
, src_val
, src_idx
, offset
);
997 val
->next
= lat
->values
;
1002 /* Like above but passes a special value of offset to distinguish that the
1003 origin is the scalar value of the parameter rather than a part of an
1007 add_scalar_value_to_lattice (struct ipcp_lattice
*lat
, tree newval
,
1008 struct cgraph_edge
*cs
,
1009 struct ipcp_value
*src_val
, int src_idx
)
1011 return add_value_to_lattice (lat
, newval
, cs
, src_val
, src_idx
, -1);
1014 /* Propagate values through a pass-through jump function JFUNC associated with
1015 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1016 is the index of the source parameter. */
1019 propagate_vals_accross_pass_through (struct cgraph_edge
*cs
,
1020 struct ipa_jump_func
*jfunc
,
1021 struct ipcp_lattice
*src_lat
,
1022 struct ipcp_lattice
*dest_lat
,
1025 struct ipcp_value
*src_val
;
1028 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1029 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1030 ret
|= add_scalar_value_to_lattice (dest_lat
, src_val
->value
, cs
,
1032 /* Do not create new values when propagating within an SCC because if there
1033 are arithmetic functions with circular dependencies, there is infinite
1034 number of them and we would just make lattices bottom. */
1035 else if (edge_within_scc (cs
))
1036 ret
= set_lattice_contains_variable (dest_lat
);
1038 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1040 tree cstval
= src_val
->value
;
1042 if (TREE_CODE (cstval
) == TREE_BINFO
)
1044 ret
|= set_lattice_contains_variable (dest_lat
);
1047 cstval
= ipa_get_jf_pass_through_result (jfunc
, cstval
);
1050 ret
|= add_scalar_value_to_lattice (dest_lat
, cstval
, cs
, src_val
,
1053 ret
|= set_lattice_contains_variable (dest_lat
);
1059 /* Propagate values through an ancestor jump function JFUNC associated with
1060 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1061 is the index of the source parameter. */
1064 propagate_vals_accross_ancestor (struct cgraph_edge
*cs
,
1065 struct ipa_jump_func
*jfunc
,
1066 struct ipcp_lattice
*src_lat
,
1067 struct ipcp_lattice
*dest_lat
,
1070 struct ipcp_value
*src_val
;
1073 if (edge_within_scc (cs
))
1074 return set_lattice_contains_variable (dest_lat
);
1076 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1078 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1081 ret
|= add_scalar_value_to_lattice (dest_lat
, t
, cs
, src_val
, src_idx
);
1083 ret
|= set_lattice_contains_variable (dest_lat
);
1089 /* Propagate scalar values across jump function JFUNC that is associated with
1090 edge CS and put the values into DEST_LAT. */
1093 propagate_scalar_accross_jump_function (struct cgraph_edge
*cs
,
1094 struct ipa_jump_func
*jfunc
,
1095 struct ipcp_lattice
*dest_lat
)
1097 if (dest_lat
->bottom
)
1100 if (jfunc
->type
== IPA_JF_CONST
1101 || jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1105 if (jfunc
->type
== IPA_JF_KNOWN_TYPE
)
1107 val
= ipa_value_from_known_type_jfunc (jfunc
);
1109 return set_lattice_contains_variable (dest_lat
);
1112 val
= ipa_get_jf_constant (jfunc
);
1113 return add_scalar_value_to_lattice (dest_lat
, val
, cs
, NULL
, 0);
1115 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1116 || jfunc
->type
== IPA_JF_ANCESTOR
)
1118 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1119 struct ipcp_lattice
*src_lat
;
1123 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1124 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1126 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1128 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1129 if (src_lat
->bottom
)
1130 return set_lattice_contains_variable (dest_lat
);
1132 /* If we would need to clone the caller and cannot, do not propagate. */
1133 if (!ipcp_versionable_function_p (cs
->caller
)
1134 && (src_lat
->contains_variable
1135 || (src_lat
->values_count
> 1)))
1136 return set_lattice_contains_variable (dest_lat
);
1138 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1139 ret
= propagate_vals_accross_pass_through (cs
, jfunc
, src_lat
,
1142 ret
= propagate_vals_accross_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1145 if (src_lat
->contains_variable
)
1146 ret
|= set_lattice_contains_variable (dest_lat
);
1151 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1152 use it for indirect inlining), we should propagate them too. */
1153 return set_lattice_contains_variable (dest_lat
);
1156 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1157 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1158 other cases, return false). If there are no aggregate items, set
1159 aggs_by_ref to NEW_AGGS_BY_REF. */
1162 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1163 bool new_aggs_by_ref
)
1165 if (dest_plats
->aggs
)
1167 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1169 set_agg_lats_to_bottom (dest_plats
);
1174 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1178 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1179 already existing lattice for the given OFFSET and SIZE, marking all skipped
1180 lattices as containing variable and checking for overlaps. If there is no
1181 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1182 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1183 unless there are too many already. If there are two many, return false. If
1184 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
1185 skipped lattices were newly marked as containing variable, set *CHANGE to
1189 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
1190 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
1191 struct ipcp_agg_lattice
***aglat
,
1192 bool pre_existing
, bool *change
)
1194 gcc_checking_assert (offset
>= 0);
1196 while (**aglat
&& (**aglat
)->offset
< offset
)
1198 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
1200 set_agg_lats_to_bottom (dest_plats
);
1203 *change
|= set_lattice_contains_variable (**aglat
);
1204 *aglat
= &(**aglat
)->next
;
1207 if (**aglat
&& (**aglat
)->offset
== offset
)
1209 if ((**aglat
)->size
!= val_size
1211 && (**aglat
)->next
->offset
< offset
+ val_size
))
1213 set_agg_lats_to_bottom (dest_plats
);
1216 gcc_checking_assert (!(**aglat
)->next
1217 || (**aglat
)->next
->offset
>= offset
+ val_size
);
1222 struct ipcp_agg_lattice
*new_al
;
1224 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
1226 set_agg_lats_to_bottom (dest_plats
);
1229 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
1231 dest_plats
->aggs_count
++;
1232 new_al
= (struct ipcp_agg_lattice
*) pool_alloc (ipcp_agg_lattice_pool
);
1233 memset (new_al
, 0, sizeof (*new_al
));
1235 new_al
->offset
= offset
;
1236 new_al
->size
= val_size
;
1237 new_al
->contains_variable
= pre_existing
;
1239 new_al
->next
= **aglat
;
1245 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
1246 containing an unknown value. */
1249 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
1254 ret
|= set_lattice_contains_variable (aglat
);
1255 aglat
= aglat
->next
;
1260 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
1261 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
1262 parameter used for lattice value sources. Return true if DEST_PLATS changed
1266 merge_aggregate_lattices (struct cgraph_edge
*cs
,
1267 struct ipcp_param_lattices
*dest_plats
,
1268 struct ipcp_param_lattices
*src_plats
,
1269 int src_idx
, HOST_WIDE_INT offset_delta
)
1271 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1272 struct ipcp_agg_lattice
**dst_aglat
;
1275 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
1277 if (src_plats
->aggs_bottom
)
1278 return set_agg_lats_contain_variable (dest_plats
);
1279 dst_aglat
= &dest_plats
->aggs
;
1281 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
1283 src_aglat
= src_aglat
->next
)
1285 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
1289 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
1290 &dst_aglat
, pre_existing
, &ret
))
1292 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
1294 dst_aglat
= &(*dst_aglat
)->next
;
1295 if (src_aglat
->bottom
)
1297 ret
|= set_lattice_contains_variable (new_al
);
1300 if (src_aglat
->contains_variable
)
1301 ret
|= set_lattice_contains_variable (new_al
);
1302 for (struct ipcp_value
*val
= src_aglat
->values
;
1305 ret
|= add_value_to_lattice (new_al
, val
->value
, cs
, val
, src_idx
,
1308 else if (dest_plats
->aggs_bottom
)
1311 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
1315 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
1316 pass-through JFUNC and if so, whether it has conform and conforms to the
1317 rules about propagating values passed by reference. */
1320 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
1321 struct ipa_jump_func
*jfunc
)
1323 return src_plats
->aggs
1324 && (!src_plats
->aggs_by_ref
1325 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
1328 /* Propagate scalar values across jump function JFUNC that is associated with
1329 edge CS and put the values into DEST_LAT. */
1332 propagate_aggs_accross_jump_function (struct cgraph_edge
*cs
,
1333 struct ipa_jump_func
*jfunc
,
1334 struct ipcp_param_lattices
*dest_plats
)
1338 if (dest_plats
->aggs_bottom
)
1341 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1342 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1344 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1345 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1346 struct ipcp_param_lattices
*src_plats
;
1348 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1349 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
1351 /* Currently we do not produce clobber aggregate jump
1352 functions, replace with merging when we do. */
1353 gcc_assert (!jfunc
->agg
.items
);
1354 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
1358 ret
|= set_agg_lats_contain_variable (dest_plats
);
1360 else if (jfunc
->type
== IPA_JF_ANCESTOR
1361 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
1363 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1364 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1365 struct ipcp_param_lattices
*src_plats
;
1367 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
1368 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
1370 /* Currently we do not produce clobber aggregate jump
1371 functions, replace with merging when we do. */
1372 gcc_assert (!jfunc
->agg
.items
);
1373 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
1374 ipa_get_jf_ancestor_offset (jfunc
));
1376 else if (!src_plats
->aggs_by_ref
)
1377 ret
|= set_agg_lats_to_bottom (dest_plats
);
1379 ret
|= set_agg_lats_contain_variable (dest_plats
);
1381 else if (jfunc
->agg
.items
)
1383 bool pre_existing
= dest_plats
->aggs
!= NULL
;
1384 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
1385 struct ipa_agg_jf_item
*item
;
1388 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
1391 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, jfunc
->agg
.items
, i
, item
)
1393 HOST_WIDE_INT val_size
;
1395 if (item
->offset
< 0)
1397 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
1398 val_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (item
->value
)), 1);
1400 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
1401 &aglat
, pre_existing
, &ret
))
1403 ret
|= add_value_to_lattice (*aglat
, item
->value
, cs
, NULL
, 0, 0);
1404 aglat
= &(*aglat
)->next
;
1406 else if (dest_plats
->aggs_bottom
)
1410 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
1413 ret
|= set_agg_lats_contain_variable (dest_plats
);
1418 /* Propagate constants from the caller to the callee of CS. INFO describes the
1422 propagate_constants_accross_call (struct cgraph_edge
*cs
)
1424 struct ipa_node_params
*callee_info
;
1425 enum availability availability
;
1426 struct cgraph_node
*callee
, *alias_or_thunk
;
1427 struct ipa_edge_args
*args
;
1429 int i
, args_count
, parms_count
;
1431 callee
= cgraph_function_node (cs
->callee
, &availability
);
1432 if (!callee
->analyzed
)
1434 gcc_checking_assert (cgraph_function_with_gimple_body_p (callee
));
1435 callee_info
= IPA_NODE_REF (callee
);
1437 args
= IPA_EDGE_REF (cs
);
1438 args_count
= ipa_get_cs_argument_count (args
);
1439 parms_count
= ipa_get_param_count (callee_info
);
1441 /* If this call goes through a thunk we must not propagate to the first (0th)
1442 parameter. However, we might need to uncover a thunk from below a series
1443 of aliases first. */
1444 alias_or_thunk
= cs
->callee
;
1445 while (alias_or_thunk
->alias
)
1446 alias_or_thunk
= cgraph_alias_aliased_node (alias_or_thunk
);
1447 if (alias_or_thunk
->thunk
.thunk_p
)
1449 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
1456 for (; (i
< args_count
) && (i
< parms_count
); i
++)
1458 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
1459 struct ipcp_param_lattices
*dest_plats
;
1461 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
1462 if (availability
== AVAIL_OVERWRITABLE
)
1463 ret
|= set_all_contains_variable (dest_plats
);
1466 ret
|= propagate_scalar_accross_jump_function (cs
, jump_func
,
1467 &dest_plats
->itself
);
1468 ret
|= propagate_aggs_accross_jump_function (cs
, jump_func
,
1472 for (; i
< parms_count
; i
++)
1473 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
1478 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
1479 (which can contain both constants and binfos) or KNOWN_BINFOS (which can be
1480 NULL) return the destination. */
1483 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
1484 VEC (tree
, heap
) *known_vals
,
1485 VEC (tree
, heap
) *known_binfos
,
1486 VEC (ipa_agg_jump_function_p
, heap
) *known_aggs
)
1488 int param_index
= ie
->indirect_info
->param_index
;
1489 HOST_WIDE_INT token
, anc_offset
;
1493 if (param_index
== -1)
1496 if (!ie
->indirect_info
->polymorphic
)
1500 if (ie
->indirect_info
->agg_contents
)
1502 if (VEC_length (ipa_agg_jump_function_p
, known_aggs
)
1503 > (unsigned int) param_index
)
1505 struct ipa_agg_jump_function
*agg
;
1506 agg
= VEC_index (ipa_agg_jump_function_p
, known_aggs
,
1508 t
= ipa_find_agg_cst_for_param (agg
, ie
->indirect_info
->offset
,
1509 ie
->indirect_info
->by_ref
);
1515 t
= (VEC_length (tree
, known_vals
) > (unsigned int) param_index
1516 ? VEC_index (tree
, known_vals
, param_index
) : NULL
);
1519 TREE_CODE (t
) == ADDR_EXPR
1520 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
1521 return TREE_OPERAND (t
, 0);
1526 gcc_assert (!ie
->indirect_info
->agg_contents
);
1527 token
= ie
->indirect_info
->otr_token
;
1528 anc_offset
= ie
->indirect_info
->offset
;
1529 otr_type
= ie
->indirect_info
->otr_type
;
1531 t
= VEC_index (tree
, known_vals
, param_index
);
1532 if (!t
&& known_binfos
1533 && VEC_length (tree
, known_binfos
) > (unsigned int) param_index
)
1534 t
= VEC_index (tree
, known_binfos
, param_index
);
1538 if (TREE_CODE (t
) != TREE_BINFO
)
1541 binfo
= gimple_extract_devirt_binfo_from_cst (t
);
1544 binfo
= get_binfo_at_offset (binfo
, anc_offset
, otr_type
);
1547 return gimple_get_virt_method_for_binfo (token
, binfo
);
1553 binfo
= get_binfo_at_offset (t
, anc_offset
, otr_type
);
1556 return gimple_get_virt_method_for_binfo (token
, binfo
);
1560 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
1561 and KNOWN_BINFOS. */
1564 devirtualization_time_bonus (struct cgraph_node
*node
,
1565 VEC (tree
, heap
) *known_csts
,
1566 VEC (tree
, heap
) *known_binfos
)
1568 struct cgraph_edge
*ie
;
1571 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1573 struct cgraph_node
*callee
;
1574 struct inline_summary
*isummary
;
1577 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_binfos
,
1582 /* Only bare minimum benefit for clearly un-inlineable targets. */
1584 callee
= cgraph_get_node (target
);
1585 if (!callee
|| !callee
->analyzed
)
1587 isummary
= inline_summary (callee
);
1588 if (!isummary
->inlinable
)
1591 /* FIXME: The values below need re-considering and perhaps also
1592 integrating into the cost metrics, at lest in some very basic way. */
1593 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
1595 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
1597 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
1598 || DECL_DECLARED_INLINE_P (callee
->symbol
.decl
))
1605 /* Return time bonus incurred because of HINTS. */
1608 hint_time_bonus (inline_hints hints
)
1610 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
1611 return PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
1615 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
1616 and SIZE_COST and with the sum of frequencies of incoming edges to the
1617 potential new clone in FREQUENCIES. */
1620 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
1621 int freq_sum
, gcov_type count_sum
, int size_cost
)
1623 if (time_benefit
== 0
1624 || !flag_ipa_cp_clone
1625 || !optimize_function_for_speed_p (DECL_STRUCT_FUNCTION (node
->symbol
.decl
)))
1628 gcc_assert (size_cost
> 0);
1632 int factor
= (count_sum
* 1000) / max_count
;
1633 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* factor
)
1636 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1637 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1638 "size: %i, count_sum: " HOST_WIDE_INT_PRINT_DEC
1639 ") -> evaluation: " HOST_WIDEST_INT_PRINT_DEC
1640 ", threshold: %i\n",
1641 time_benefit
, size_cost
, (HOST_WIDE_INT
) count_sum
,
1644 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1648 HOST_WIDEST_INT evaluation
= (((HOST_WIDEST_INT
) time_benefit
* freq_sum
)
1651 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1652 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
1653 "size: %i, freq_sum: %i) -> evaluation: "
1654 HOST_WIDEST_INT_PRINT_DEC
", threshold: %i\n",
1655 time_benefit
, size_cost
, freq_sum
, evaluation
,
1656 CGRAPH_FREQ_BASE
/2);
1658 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
1662 /* Return all context independent values from aggregate lattices in PLATS in a
1663 vector. Return NULL if there are none. */
1665 static VEC (ipa_agg_jf_item_t
, gc
) *
1666 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
1668 VEC (ipa_agg_jf_item_t
, gc
) *res
= NULL
;
1670 if (plats
->aggs_bottom
1671 || plats
->aggs_contain_variable
1672 || plats
->aggs_count
== 0)
1675 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
1677 aglat
= aglat
->next
)
1678 if (ipa_lat_is_single_const (aglat
))
1680 struct ipa_agg_jf_item item
;
1681 item
.offset
= aglat
->offset
;
1682 item
.value
= aglat
->values
->value
;
1683 VEC_safe_push (ipa_agg_jf_item_t
, gc
, res
, item
);
1688 /* Allocate KNOWN_CSTS, KNOWN_BINFOS and, if non-NULL, KNOWN_AGGS and populate
1689 them with values of parameters that are known independent of the context.
1690 INFO describes the function. If REMOVABLE_PARAMS_COST is non-NULL, the
1691 movement cost of all removable parameters will be stored in it. */
1694 gather_context_independent_values (struct ipa_node_params
*info
,
1695 VEC (tree
, heap
) **known_csts
,
1696 VEC (tree
, heap
) **known_binfos
,
1697 VEC (ipa_agg_jump_function_t
, heap
) **known_aggs
,
1698 int *removable_params_cost
)
1700 int i
, count
= ipa_get_param_count (info
);
1704 *known_binfos
= NULL
;
1705 VEC_safe_grow_cleared (tree
, heap
, *known_csts
, count
);
1706 VEC_safe_grow_cleared (tree
, heap
, *known_binfos
, count
);
1710 VEC_safe_grow_cleared (ipa_agg_jump_function_t
, heap
, *known_aggs
, count
);
1713 if (removable_params_cost
)
1714 *removable_params_cost
= 0;
1716 for (i
= 0; i
< count
; i
++)
1718 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1719 struct ipcp_lattice
*lat
= &plats
->itself
;
1721 if (ipa_lat_is_single_const (lat
))
1723 struct ipcp_value
*val
= lat
->values
;
1724 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1726 VEC_replace (tree
, *known_csts
, i
, val
->value
);
1727 if (removable_params_cost
)
1728 *removable_params_cost
1729 += estimate_move_cost (TREE_TYPE (val
->value
));
1732 else if (plats
->virt_call
)
1734 VEC_replace (tree
, *known_binfos
, i
, val
->value
);
1737 else if (removable_params_cost
1738 && !ipa_is_param_used (info
, i
))
1739 *removable_params_cost
1740 += estimate_move_cost (TREE_TYPE (ipa_get_param (info
, i
)));
1742 else if (removable_params_cost
1743 && !ipa_is_param_used (info
, i
))
1744 *removable_params_cost
1745 += estimate_move_cost (TREE_TYPE (ipa_get_param (info
, i
)));
1749 VEC (ipa_agg_jf_item_t
, gc
) *agg_items
;
1750 struct ipa_agg_jump_function
*ajf
;
1752 agg_items
= context_independent_aggregate_values (plats
);
1753 ajf
= &VEC_index (ipa_agg_jump_function_t
, *known_aggs
, i
);
1754 ajf
->items
= agg_items
;
1755 ajf
->by_ref
= plats
->aggs_by_ref
;
1756 ret
|= agg_items
!= NULL
;
1763 /* The current interface in ipa-inline-analysis requires a pointer vector.
1766 FIXME: That interface should be re-worked, this is slightly silly. Still,
1767 I'd like to discuss how to change it first and this demonstrates the
1770 static VEC (ipa_agg_jump_function_p
, heap
) *
1771 agg_jmp_p_vec_for_t_vec (VEC (ipa_agg_jump_function_t
, heap
) *known_aggs
)
1773 VEC (ipa_agg_jump_function_p
, heap
) *ret
;
1774 struct ipa_agg_jump_function
*ajf
;
1777 ret
= VEC_alloc (ipa_agg_jump_function_p
, heap
,
1778 VEC_length (ipa_agg_jump_function_t
, known_aggs
));
1779 FOR_EACH_VEC_ELT (ipa_agg_jump_function_t
, known_aggs
, i
, ajf
)
1780 VEC_quick_push (ipa_agg_jump_function_p
, ret
, ajf
);
1784 /* Iterate over known values of parameters of NODE and estimate the local
1785 effects in terms of time and size they have. */
1788 estimate_local_effects (struct cgraph_node
*node
)
1790 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1791 int i
, count
= ipa_get_param_count (info
);
1792 VEC (tree
, heap
) *known_csts
, *known_binfos
;
1793 VEC (ipa_agg_jump_function_t
, heap
) *known_aggs
;
1794 VEC (ipa_agg_jump_function_p
, heap
) *known_aggs_ptrs
;
1796 int base_time
= inline_summary (node
)->time
;
1797 int removable_params_cost
;
1799 if (!count
|| !ipcp_versionable_function_p (node
))
1802 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1803 fprintf (dump_file
, "\nEstimating effects for %s/%i, base_time: %i.\n",
1804 cgraph_node_name (node
), node
->uid
, base_time
);
1806 always_const
= gather_context_independent_values (info
, &known_csts
,
1807 &known_binfos
, &known_aggs
,
1808 &removable_params_cost
);
1809 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
1812 struct caller_statistics stats
;
1816 init_caller_stats (&stats
);
1817 cgraph_for_node_and_aliases (node
, gather_caller_stats
, &stats
, false);
1818 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1819 known_aggs_ptrs
, &size
, &time
, &hints
);
1820 time
-= devirtualization_time_bonus (node
, known_csts
, known_binfos
);
1821 time
-= hint_time_bonus (hints
);
1822 time
-= removable_params_cost
;
1823 size
-= stats
.n_calls
* removable_params_cost
;
1826 fprintf (dump_file
, " - context independent values, size: %i, "
1827 "time_benefit: %i\n", size
, base_time
- time
);
1830 || cgraph_will_be_removed_from_program_if_no_direct_calls (node
))
1832 info
->clone_for_all_contexts
= true;
1836 fprintf (dump_file
, " Decided to specialize for all "
1837 "known contexts, code not going to grow.\n");
1839 else if (good_cloning_opportunity_p (node
, base_time
- time
,
1840 stats
.freq_sum
, stats
.count_sum
,
1843 if (size
+ overall_size
<= max_new_size
)
1845 info
->clone_for_all_contexts
= true;
1847 overall_size
+= size
;
1850 fprintf (dump_file
, " Decided to specialize for all "
1851 "known contexts, growth deemed beneficial.\n");
1853 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1854 fprintf (dump_file
, " Not cloning for all contexts because "
1855 "max_new_size would be reached with %li.\n",
1856 size
+ overall_size
);
1860 for (i
= 0; i
< count
; i
++)
1862 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1863 struct ipcp_lattice
*lat
= &plats
->itself
;
1864 struct ipcp_value
*val
;
1869 || VEC_index (tree
, known_csts
, i
)
1870 || VEC_index (tree
, known_binfos
, i
))
1873 for (val
= lat
->values
; val
; val
= val
->next
)
1875 int time
, size
, time_benefit
;
1878 if (TREE_CODE (val
->value
) != TREE_BINFO
)
1880 VEC_replace (tree
, known_csts
, i
, val
->value
);
1881 VEC_replace (tree
, known_binfos
, i
, NULL_TREE
);
1882 emc
= estimate_move_cost (TREE_TYPE (val
->value
));
1884 else if (plats
->virt_call
)
1886 VEC_replace (tree
, known_csts
, i
, NULL_TREE
);
1887 VEC_replace (tree
, known_binfos
, i
, val
->value
);
1893 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1894 known_aggs_ptrs
, &size
, &time
,
1896 time_benefit
= base_time
- time
1897 + devirtualization_time_bonus (node
, known_csts
, known_binfos
)
1898 + hint_time_bonus (hints
)
1899 + removable_params_cost
+ emc
;
1901 gcc_checking_assert (size
>=0);
1902 /* The inliner-heuristics based estimates may think that in certain
1903 contexts some functions do not have any size at all but we want
1904 all specializations to have at least a tiny cost, not least not to
1909 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1911 fprintf (dump_file
, " - estimates for value ");
1912 print_ipcp_constant_value (dump_file
, val
->value
);
1913 fprintf (dump_file
, " for parameter ");
1914 print_generic_expr (dump_file
, ipa_get_param (info
, i
), 0);
1915 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
1916 time_benefit
, size
);
1919 val
->local_time_benefit
= time_benefit
;
1920 val
->local_size_cost
= size
;
1922 VEC_replace (tree
, known_binfos
, i
, NULL_TREE
);
1923 VEC_replace (tree
, known_csts
, i
, NULL_TREE
);
1926 for (i
= 0; i
< count
; i
++)
1928 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1929 struct ipa_agg_jump_function
*ajf
;
1930 struct ipcp_agg_lattice
*aglat
;
1932 if (plats
->aggs_bottom
|| !plats
->aggs
)
1935 ajf
= &VEC_index (ipa_agg_jump_function_t
, known_aggs
, i
);
1936 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
1938 struct ipcp_value
*val
;
1939 if (aglat
->bottom
|| !aglat
->values
1940 /* If the following is true, the one value is in known_aggs. */
1941 || (!plats
->aggs_contain_variable
1942 && ipa_lat_is_single_const (aglat
)))
1945 for (val
= aglat
->values
; val
; val
= val
->next
)
1947 int time
, size
, time_benefit
;
1948 struct ipa_agg_jf_item item
;
1951 item
.offset
= aglat
->offset
;
1952 item
.value
= val
->value
;
1953 VEC_safe_push (ipa_agg_jf_item_t
, gc
, ajf
->items
, item
);
1955 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_binfos
,
1956 known_aggs_ptrs
, &size
, &time
,
1958 time_benefit
= base_time
- time
1959 + devirtualization_time_bonus (node
, known_csts
, known_binfos
)
1960 + hint_time_bonus (hints
);
1961 gcc_checking_assert (size
>=0);
1965 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1967 fprintf (dump_file
, " - estimates for value ");
1968 print_ipcp_constant_value (dump_file
, val
->value
);
1969 fprintf (dump_file
, " for parameter ");
1970 print_generic_expr (dump_file
, ipa_get_param (info
, i
), 0);
1971 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
1972 "]: time_benefit: %i, size: %i\n",
1973 plats
->aggs_by_ref
? "ref " : "",
1974 aglat
->offset
, time_benefit
, size
);
1977 val
->local_time_benefit
= time_benefit
;
1978 val
->local_size_cost
= size
;
1979 VEC_pop (ipa_agg_jf_item_t
, ajf
->items
);
1984 for (i
= 0; i
< count
; i
++)
1986 VEC_free (ipa_agg_jf_item_t
, gc
,
1987 VEC_index (ipa_agg_jump_function_t
, known_aggs
, i
).items
);
1988 VEC_index (ipa_agg_jump_function_t
, known_aggs
, i
).items
= NULL
;
1991 VEC_free (tree
, heap
, known_csts
);
1992 VEC_free (tree
, heap
, known_binfos
);
1993 VEC_free (ipa_agg_jump_function_t
, heap
, known_aggs
);
1994 VEC_free (ipa_agg_jump_function_p
, heap
, known_aggs_ptrs
);
1998 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
1999 topological sort of values. */
2002 add_val_to_toposort (struct ipcp_value
*cur_val
)
2004 static int dfs_counter
= 0;
2005 static struct ipcp_value
*stack
;
2006 struct ipcp_value_source
*src
;
2012 cur_val
->dfs
= dfs_counter
;
2013 cur_val
->low_link
= dfs_counter
;
2015 cur_val
->topo_next
= stack
;
2017 cur_val
->on_stack
= true;
2019 for (src
= cur_val
->sources
; src
; src
= src
->next
)
2022 if (src
->val
->dfs
== 0)
2024 add_val_to_toposort (src
->val
);
2025 if (src
->val
->low_link
< cur_val
->low_link
)
2026 cur_val
->low_link
= src
->val
->low_link
;
2028 else if (src
->val
->on_stack
2029 && src
->val
->dfs
< cur_val
->low_link
)
2030 cur_val
->low_link
= src
->val
->dfs
;
2033 if (cur_val
->dfs
== cur_val
->low_link
)
2035 struct ipcp_value
*v
, *scc_list
= NULL
;
2040 stack
= v
->topo_next
;
2041 v
->on_stack
= false;
2043 v
->scc_next
= scc_list
;
2046 while (v
!= cur_val
);
2048 cur_val
->topo_next
= values_topo
;
2049 values_topo
= cur_val
;
2053 /* Add all values in lattices associated with NODE to the topological sort if
2054 they are not there yet. */
2057 add_all_node_vals_to_toposort (struct cgraph_node
*node
)
2059 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2060 int i
, count
= ipa_get_param_count (info
);
2062 for (i
= 0; i
< count
; i
++)
2064 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2065 struct ipcp_lattice
*lat
= &plats
->itself
;
2066 struct ipcp_agg_lattice
*aglat
;
2067 struct ipcp_value
*val
;
2070 for (val
= lat
->values
; val
; val
= val
->next
)
2071 add_val_to_toposort (val
);
2073 if (!plats
->aggs_bottom
)
2074 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2076 for (val
= aglat
->values
; val
; val
= val
->next
)
2077 add_val_to_toposort (val
);
2081 /* One pass of constants propagation along the call graph edges, from callers
2082 to callees (requires topological ordering in TOPO), iterate over strongly
2083 connected components. */
2086 propagate_constants_topo (struct topo_info
*topo
)
2090 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
2092 struct cgraph_node
*v
, *node
= topo
->order
[i
];
2093 struct ipa_dfs_info
*node_dfs_info
;
2095 if (!cgraph_function_with_gimple_body_p (node
))
2098 node_dfs_info
= (struct ipa_dfs_info
*) node
->symbol
.aux
;
2099 /* First, iteratively propagate within the strongly connected component
2100 until all lattices stabilize. */
2101 v
= node_dfs_info
->next_cycle
;
2104 push_node_to_stack (topo
, v
);
2105 v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
;
2111 struct cgraph_edge
*cs
;
2113 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2114 if (edge_within_scc (cs
)
2115 && propagate_constants_accross_call (cs
))
2116 push_node_to_stack (topo
, cs
->callee
);
2117 v
= pop_node_from_stack (topo
);
2120 /* Afterwards, propagate along edges leading out of the SCC, calculates
2121 the local effects of the discovered constants and all valid values to
2122 their topological sort. */
2126 struct cgraph_edge
*cs
;
2128 estimate_local_effects (v
);
2129 add_all_node_vals_to_toposort (v
);
2130 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
2131 if (!edge_within_scc (cs
))
2132 propagate_constants_accross_call (cs
);
2134 v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
;
2140 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
2141 the bigger one if otherwise. */
2144 safe_add (int a
, int b
)
2146 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
2147 return a
> b
? a
: b
;
2153 /* Propagate the estimated effects of individual values along the topological
2154 from the dependent values to those they depend on. */
2157 propagate_effects (void)
2159 struct ipcp_value
*base
;
2161 for (base
= values_topo
; base
; base
= base
->topo_next
)
2163 struct ipcp_value_source
*src
;
2164 struct ipcp_value
*val
;
2165 int time
= 0, size
= 0;
2167 for (val
= base
; val
; val
= val
->scc_next
)
2169 time
= safe_add (time
,
2170 val
->local_time_benefit
+ val
->prop_time_benefit
);
2171 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
2174 for (val
= base
; val
; val
= val
->scc_next
)
2175 for (src
= val
->sources
; src
; src
= src
->next
)
2177 && cgraph_maybe_hot_edge_p (src
->cs
))
2179 src
->val
->prop_time_benefit
= safe_add (time
,
2180 src
->val
->prop_time_benefit
);
2181 src
->val
->prop_size_cost
= safe_add (size
,
2182 src
->val
->prop_size_cost
);
2188 /* Propagate constants, binfos and their effects from the summaries
2189 interprocedurally. */
2192 ipcp_propagate_stage (struct topo_info
*topo
)
2194 struct cgraph_node
*node
;
2197 fprintf (dump_file
, "\n Propagating constants:\n\n");
2200 ipa_update_after_lto_read ();
2203 FOR_EACH_DEFINED_FUNCTION (node
)
2205 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2207 determine_versionability (node
);
2208 if (cgraph_function_with_gimple_body_p (node
))
2210 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
2211 ipa_get_param_count (info
));
2212 initialize_node_lattices (node
);
2214 if (node
->count
> max_count
)
2215 max_count
= node
->count
;
2216 overall_size
+= inline_summary (node
)->self_size
;
2219 max_new_size
= overall_size
;
2220 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
2221 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
2222 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
2225 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
2226 overall_size
, max_new_size
);
2228 propagate_constants_topo (topo
);
2229 #ifdef ENABLE_CHECKING
2230 ipcp_verify_propagated_values ();
2232 propagate_effects ();
2236 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
2237 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
2241 /* Discover newly direct outgoing edges from NODE which is a new clone with
2242 known KNOWN_VALS and make them direct. */
2245 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
2246 VEC (tree
, heap
) *known_vals
)
2248 struct cgraph_edge
*ie
, *next_ie
;
2251 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
2255 next_ie
= ie
->next_callee
;
2256 target
= ipa_get_indirect_edge_target (ie
, known_vals
, NULL
, NULL
);
2259 ipa_make_edge_direct_to_target (ie
, target
);
2263 /* Turning calls to direct calls will improve overall summary. */
2265 inline_update_overall_summary (node
);
2268 /* Vector of pointers which for linked lists of clones of an original crgaph
2271 static VEC (cgraph_edge_p
, heap
) *next_edge_clone
;
2274 grow_next_edge_clone_vector (void)
2276 if (VEC_length (cgraph_edge_p
, next_edge_clone
)
2277 <= (unsigned) cgraph_edge_max_uid
)
2278 VEC_safe_grow_cleared (cgraph_edge_p
, heap
, next_edge_clone
,
2279 cgraph_edge_max_uid
+ 1);
2282 /* Edge duplication hook to grow the appropriate linked list in
2286 ipcp_edge_duplication_hook (struct cgraph_edge
*src
, struct cgraph_edge
*dst
,
2287 __attribute__((unused
)) void *data
)
2289 grow_next_edge_clone_vector ();
2290 VEC_replace (cgraph_edge_p
, next_edge_clone
, dst
->uid
,
2291 VEC_index (cgraph_edge_p
, next_edge_clone
, src
->uid
));
2292 VEC_replace (cgraph_edge_p
, next_edge_clone
, src
->uid
, dst
);
2295 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
2296 parameter with the given INDEX. */
2299 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDEST_INT offset
,
2302 struct ipa_agg_replacement_value
*aggval
;
2304 aggval
= ipa_get_agg_replacements_for_node (node
);
2307 if (aggval
->offset
== offset
2308 && aggval
->index
== index
)
2309 return aggval
->value
;
2310 aggval
= aggval
->next
;
2315 /* Return true if edge CS does bring about the value described by SRC. */
2318 cgraph_edge_brings_value_p (struct cgraph_edge
*cs
,
2319 struct ipcp_value_source
*src
)
2321 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2323 if (IPA_NODE_REF (cs
->callee
)->ipcp_orig_node
2324 || caller_info
->node_dead
)
2329 if (caller_info
->ipcp_orig_node
)
2332 if (src
->offset
== -1)
2333 t
= VEC_index (tree
, caller_info
->known_vals
, src
->index
);
2335 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
2336 return (t
!= NULL_TREE
2337 && values_equal_for_ipcp_p (src
->val
->value
, t
));
2341 struct ipcp_agg_lattice
*aglat
;
2342 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
2344 if (src
->offset
== -1)
2345 return (ipa_lat_is_single_const (&plats
->itself
)
2346 && values_equal_for_ipcp_p (src
->val
->value
,
2347 plats
->itself
.values
->value
));
2350 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
2352 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2353 if (aglat
->offset
== src
->offset
)
2354 return (ipa_lat_is_single_const (aglat
)
2355 && values_equal_for_ipcp_p (src
->val
->value
,
2356 aglat
->values
->value
));
2362 /* Get the next clone in the linked list of clones of an edge. */
2364 static inline struct cgraph_edge
*
2365 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
2367 return VEC_index (cgraph_edge_p
, next_edge_clone
, cs
->uid
);
2370 /* Given VAL, iterate over all its sources and if they still hold, add their
2371 edge frequency and their number into *FREQUENCY and *CALLER_COUNT
2375 get_info_about_necessary_edges (struct ipcp_value
*val
, int *freq_sum
,
2376 gcov_type
*count_sum
, int *caller_count
)
2378 struct ipcp_value_source
*src
;
2379 int freq
= 0, count
= 0;
2383 for (src
= val
->sources
; src
; src
= src
->next
)
2385 struct cgraph_edge
*cs
= src
->cs
;
2388 if (cgraph_edge_brings_value_p (cs
, src
))
2391 freq
+= cs
->frequency
;
2393 hot
|= cgraph_maybe_hot_edge_p (cs
);
2395 cs
= get_next_cgraph_edge_clone (cs
);
2401 *caller_count
= count
;
2405 /* Return a vector of incoming edges that do bring value VAL. It is assumed
2406 their number is known and equal to CALLER_COUNT. */
2408 static VEC (cgraph_edge_p
,heap
) *
2409 gather_edges_for_value (struct ipcp_value
*val
, int caller_count
)
2411 struct ipcp_value_source
*src
;
2412 VEC (cgraph_edge_p
,heap
) *ret
;
2414 ret
= VEC_alloc (cgraph_edge_p
, heap
, caller_count
);
2415 for (src
= val
->sources
; src
; src
= src
->next
)
2417 struct cgraph_edge
*cs
= src
->cs
;
2420 if (cgraph_edge_brings_value_p (cs
, src
))
2421 VEC_quick_push (cgraph_edge_p
, ret
, cs
);
2422 cs
= get_next_cgraph_edge_clone (cs
);
2429 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
2430 Return it or NULL if for some reason it cannot be created. */
2432 static struct ipa_replace_map
*
2433 get_replacement_map (tree value
, tree parm
)
2435 tree req_type
= TREE_TYPE (parm
);
2436 struct ipa_replace_map
*replace_map
;
2438 if (!useless_type_conversion_p (req_type
, TREE_TYPE (value
)))
2440 if (fold_convertible_p (req_type
, value
))
2441 value
= fold_build1 (NOP_EXPR
, req_type
, value
);
2442 else if (TYPE_SIZE (req_type
) == TYPE_SIZE (TREE_TYPE (value
)))
2443 value
= fold_build1 (VIEW_CONVERT_EXPR
, req_type
, value
);
2448 fprintf (dump_file
, " const ");
2449 print_generic_expr (dump_file
, value
, 0);
2450 fprintf (dump_file
, " can't be converted to param ");
2451 print_generic_expr (dump_file
, parm
, 0);
2452 fprintf (dump_file
, "\n");
2458 replace_map
= ggc_alloc_ipa_replace_map ();
2461 fprintf (dump_file
, " replacing param ");
2462 print_generic_expr (dump_file
, parm
, 0);
2463 fprintf (dump_file
, " with const ");
2464 print_generic_expr (dump_file
, value
, 0);
2465 fprintf (dump_file
, "\n");
2467 replace_map
->old_tree
= parm
;
2468 replace_map
->new_tree
= value
;
2469 replace_map
->replace_p
= true;
2470 replace_map
->ref_p
= false;
2475 /* Dump new profiling counts */
2478 dump_profile_updates (struct cgraph_node
*orig_node
,
2479 struct cgraph_node
*new_node
)
2481 struct cgraph_edge
*cs
;
2483 fprintf (dump_file
, " setting count of the specialized node to "
2484 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) new_node
->count
);
2485 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2486 fprintf (dump_file
, " edge to %s has count "
2487 HOST_WIDE_INT_PRINT_DEC
"\n",
2488 cgraph_node_name (cs
->callee
), (HOST_WIDE_INT
) cs
->count
);
2490 fprintf (dump_file
, " setting count of the original node to "
2491 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) orig_node
->count
);
2492 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2493 fprintf (dump_file
, " edge to %s is left with "
2494 HOST_WIDE_INT_PRINT_DEC
"\n",
2495 cgraph_node_name (cs
->callee
), (HOST_WIDE_INT
) cs
->count
);
2498 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
2499 their profile information to reflect this. */
2502 update_profiling_info (struct cgraph_node
*orig_node
,
2503 struct cgraph_node
*new_node
)
2505 struct cgraph_edge
*cs
;
2506 struct caller_statistics stats
;
2507 gcov_type new_sum
, orig_sum
;
2508 gcov_type remainder
, orig_node_count
= orig_node
->count
;
2510 if (orig_node_count
== 0)
2513 init_caller_stats (&stats
);
2514 cgraph_for_node_and_aliases (orig_node
, gather_caller_stats
, &stats
, false);
2515 orig_sum
= stats
.count_sum
;
2516 init_caller_stats (&stats
);
2517 cgraph_for_node_and_aliases (new_node
, gather_caller_stats
, &stats
, false);
2518 new_sum
= stats
.count_sum
;
2520 if (orig_node_count
< orig_sum
+ new_sum
)
2523 fprintf (dump_file
, " Problem: node %s/%i has too low count "
2524 HOST_WIDE_INT_PRINT_DEC
" while the sum of incoming "
2525 "counts is " HOST_WIDE_INT_PRINT_DEC
"\n",
2526 cgraph_node_name (orig_node
), orig_node
->uid
,
2527 (HOST_WIDE_INT
) orig_node_count
,
2528 (HOST_WIDE_INT
) (orig_sum
+ new_sum
));
2530 orig_node_count
= (orig_sum
+ new_sum
) * 12 / 10;
2532 fprintf (dump_file
, " proceeding by pretending it was "
2533 HOST_WIDE_INT_PRINT_DEC
"\n",
2534 (HOST_WIDE_INT
) orig_node_count
);
2537 new_node
->count
= new_sum
;
2538 remainder
= orig_node_count
- new_sum
;
2539 orig_node
->count
= remainder
;
2541 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2543 cs
->count
= cs
->count
* (new_sum
* REG_BR_PROB_BASE
2544 / orig_node_count
) / REG_BR_PROB_BASE
;
2548 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2549 cs
->count
= cs
->count
* (remainder
* REG_BR_PROB_BASE
2550 / orig_node_count
) / REG_BR_PROB_BASE
;
2553 dump_profile_updates (orig_node
, new_node
);
2556 /* Update the respective profile of specialized NEW_NODE and the original
2557 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
2558 have been redirected to the specialized version. */
2561 update_specialized_profile (struct cgraph_node
*new_node
,
2562 struct cgraph_node
*orig_node
,
2563 gcov_type redirected_sum
)
2565 struct cgraph_edge
*cs
;
2566 gcov_type new_node_count
, orig_node_count
= orig_node
->count
;
2569 fprintf (dump_file
, " the sum of counts of redirected edges is "
2570 HOST_WIDE_INT_PRINT_DEC
"\n", (HOST_WIDE_INT
) redirected_sum
);
2571 if (orig_node_count
== 0)
2574 gcc_assert (orig_node_count
>= redirected_sum
);
2576 new_node_count
= new_node
->count
;
2577 new_node
->count
+= redirected_sum
;
2578 orig_node
->count
-= redirected_sum
;
2580 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
2582 cs
->count
+= cs
->count
* redirected_sum
/ new_node_count
;
2586 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
2588 gcov_type dec
= cs
->count
* (redirected_sum
* REG_BR_PROB_BASE
2589 / orig_node_count
) / REG_BR_PROB_BASE
;
2590 if (dec
< cs
->count
)
2597 dump_profile_updates (orig_node
, new_node
);
2600 /* Create a specialized version of NODE with known constants and types of
2601 parameters in KNOWN_VALS and redirect all edges in CALLERS to it. */
2603 static struct cgraph_node
*
2604 create_specialized_node (struct cgraph_node
*node
,
2605 VEC (tree
, heap
) *known_vals
,
2606 struct ipa_agg_replacement_value
*aggvals
,
2607 VEC (cgraph_edge_p
,heap
) *callers
)
2609 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
2610 VEC (ipa_replace_map_p
,gc
)* replace_trees
= NULL
;
2611 struct cgraph_node
*new_node
;
2612 int i
, count
= ipa_get_param_count (info
);
2613 bitmap args_to_skip
;
2615 gcc_assert (!info
->ipcp_orig_node
);
2617 if (node
->local
.can_change_signature
)
2619 args_to_skip
= BITMAP_GGC_ALLOC ();
2620 for (i
= 0; i
< count
; i
++)
2622 tree t
= VEC_index (tree
, known_vals
, i
);
2624 if ((t
&& TREE_CODE (t
) != TREE_BINFO
)
2625 || !ipa_is_param_used (info
, i
))
2626 bitmap_set_bit (args_to_skip
, i
);
2631 args_to_skip
= NULL
;
2632 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2633 fprintf (dump_file
, " cannot change function signature\n");
2636 for (i
= 0; i
< count
; i
++)
2638 tree t
= VEC_index (tree
, known_vals
, i
);
2639 if (t
&& TREE_CODE (t
) != TREE_BINFO
)
2641 struct ipa_replace_map
*replace_map
;
2643 replace_map
= get_replacement_map (t
, ipa_get_param (info
, i
));
2645 VEC_safe_push (ipa_replace_map_p
, gc
, replace_trees
, replace_map
);
2649 new_node
= cgraph_create_virtual_clone (node
, callers
, replace_trees
,
2650 args_to_skip
, "constprop");
2651 ipa_set_node_agg_value_chain (new_node
, aggvals
);
2652 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2654 fprintf (dump_file
, " the new node is %s/%i.\n",
2655 cgraph_node_name (new_node
), new_node
->uid
);
2657 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
2659 gcc_checking_assert (ipa_node_params_vector
2660 && (VEC_length (ipa_node_params_t
,
2661 ipa_node_params_vector
)
2662 > (unsigned) cgraph_max_uid
));
2663 update_profiling_info (node
, new_node
);
2664 new_info
= IPA_NODE_REF (new_node
);
2665 new_info
->ipcp_orig_node
= node
;
2666 new_info
->known_vals
= known_vals
;
2668 ipcp_discover_new_direct_edges (new_node
, known_vals
);
2670 VEC_free (cgraph_edge_p
, heap
, callers
);
2674 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
2675 KNOWN_VALS with constants and types that are also known for all of the
2679 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
2680 VEC (tree
, heap
) *known_vals
,
2681 VEC (cgraph_edge_p
,heap
) *callers
)
2683 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2684 int i
, count
= ipa_get_param_count (info
);
2686 for (i
= 0; i
< count
; i
++)
2688 struct cgraph_edge
*cs
;
2689 tree newval
= NULL_TREE
;
2692 if (ipa_get_scalar_lat (info
, i
)->bottom
2693 || VEC_index (tree
, known_vals
, i
))
2696 FOR_EACH_VEC_ELT (cgraph_edge_p
, callers
, j
, cs
)
2698 struct ipa_jump_func
*jump_func
;
2701 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
2706 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
2707 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
);
2710 && !values_equal_for_ipcp_p (t
, newval
)))
2721 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2723 fprintf (dump_file
, " adding an extra known scalar value ");
2724 print_ipcp_constant_value (dump_file
, newval
);
2725 fprintf (dump_file
, " for parameter ");
2726 print_generic_expr (dump_file
, ipa_get_param (info
, i
), 0);
2727 fprintf (dump_file
, "\n");
2730 VEC_replace (tree
, known_vals
, i
, newval
);
2735 /* Go through PLATS and create a vector of values consisting of values and
2736 offsets (minus OFFSET) of lattices that contain only a single value. */
2738 static VEC (ipa_agg_jf_item_t
, heap
) *
2739 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
2741 VEC (ipa_agg_jf_item_t
, heap
) *res
= NULL
;
2743 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2746 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
2747 if (ipa_lat_is_single_const (aglat
))
2749 struct ipa_agg_jf_item ti
;
2750 ti
.offset
= aglat
->offset
- offset
;
2751 ti
.value
= aglat
->values
->value
;
2752 VEC_safe_push (ipa_agg_jf_item_t
, heap
, res
, ti
);
2757 /* Intersect all values in INTER with single value lattices in PLATS (while
2758 subtracting OFFSET). */
2761 intersect_with_plats (struct ipcp_param_lattices
*plats
,
2762 VEC (ipa_agg_jf_item_t
, heap
) **inter
,
2763 HOST_WIDE_INT offset
)
2765 struct ipcp_agg_lattice
*aglat
;
2766 struct ipa_agg_jf_item
*item
;
2769 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
2771 VEC_free (ipa_agg_jf_item_t
, heap
, *inter
);
2776 aglat
= plats
->aggs
;
2777 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, *inter
, k
, item
)
2784 if (aglat
->offset
- offset
> item
->offset
)
2786 if (aglat
->offset
- offset
== item
->offset
)
2788 gcc_checking_assert (item
->value
);
2789 if (values_equal_for_ipcp_p (item
->value
, aglat
->values
->value
))
2793 aglat
= aglat
->next
;
2796 item
->value
= NULL_TREE
;
2800 /* Copy agggregate replacement values of NODE (which is an IPA-CP clone) to the
2801 vector result while subtracting OFFSET from the individual value offsets. */
2803 static VEC (ipa_agg_jf_item_t
, heap
) *
2804 agg_replacements_to_vector (struct cgraph_node
*node
, HOST_WIDE_INT offset
)
2806 struct ipa_agg_replacement_value
*av
;
2807 VEC (ipa_agg_jf_item_t
, heap
) *res
= NULL
;
2809 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
2811 struct ipa_agg_jf_item item
;
2812 gcc_checking_assert (av
->value
);
2813 item
.offset
= av
->offset
- offset
;
2814 item
.value
= av
->value
;
2815 VEC_safe_push (ipa_agg_jf_item_t
, heap
, res
, item
);
2821 /* Intersect all values in INTER with those that we have already scheduled to
2822 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
2823 (while subtracting OFFSET). */
2826 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
2827 VEC (ipa_agg_jf_item_t
, heap
) **inter
,
2828 HOST_WIDE_INT offset
)
2830 struct ipa_agg_replacement_value
*srcvals
;
2831 struct ipa_agg_jf_item
*item
;
2834 srcvals
= ipa_get_agg_replacements_for_node (node
);
2837 VEC_free (ipa_agg_jf_item_t
, heap
, *inter
);
2842 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, *inter
, i
, item
)
2844 struct ipa_agg_replacement_value
*av
;
2848 for (av
= srcvals
; av
; av
= av
->next
)
2850 gcc_checking_assert (av
->value
);
2851 if (av
->index
== index
2852 && av
->offset
- offset
== item
->offset
)
2854 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
2860 item
->value
= NULL_TREE
;
2864 /* Look at edges in CALLERS and collect all known aggregate values that arrive
2865 from all of them. */
2867 static struct ipa_agg_replacement_value
*
2868 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
2869 VEC (cgraph_edge_p
,heap
) *callers
)
2871 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2872 struct ipa_agg_replacement_value
*res
= NULL
;
2873 struct cgraph_edge
*cs
;
2874 int i
, j
, count
= ipa_get_param_count (info
);
2876 FOR_EACH_VEC_ELT (cgraph_edge_p
, callers
, j
, cs
)
2878 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
2883 for (i
= 0; i
< count
; i
++)
2885 struct cgraph_edge
*cs
;
2886 VEC (ipa_agg_jf_item_t
, heap
) *inter
= NULL
;
2887 struct ipa_agg_jf_item
*item
;
2890 /* Among other things, the following check should deal with all by_ref
2892 if (ipa_get_parm_lattices (info
, i
)->aggs_bottom
)
2895 FOR_EACH_VEC_ELT (cgraph_edge_p
, callers
, j
, cs
)
2897 struct ipa_jump_func
*jfunc
;
2898 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
2899 if (jfunc
->type
== IPA_JF_PASS_THROUGH
2900 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
2902 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2903 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
2905 if (caller_info
->ipcp_orig_node
)
2907 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
2908 struct ipcp_param_lattices
*orig_plats
;
2909 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
2911 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
2914 inter
= agg_replacements_to_vector (cs
->caller
, 0);
2916 intersect_with_agg_replacements (cs
->caller
, src_idx
,
2922 struct ipcp_param_lattices
*src_plats
;
2923 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2924 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
2926 /* Currently we do not produce clobber aggregate jump
2927 functions, adjust when we do. */
2928 gcc_checking_assert (!jfunc
->agg
.items
);
2930 inter
= copy_plats_to_inter (src_plats
, 0);
2932 intersect_with_plats (src_plats
, &inter
, 0);
2936 else if (jfunc
->type
== IPA_JF_ANCESTOR
2937 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
2939 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2940 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
2941 struct ipcp_param_lattices
*src_plats
;
2942 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
2944 if (info
->ipcp_orig_node
)
2947 inter
= agg_replacements_to_vector (cs
->caller
, delta
);
2949 intersect_with_agg_replacements (cs
->caller
, i
, &inter
,
2954 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);;
2955 /* Currently we do not produce clobber aggregate jump
2956 functions, adjust when we do. */
2957 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
2959 inter
= copy_plats_to_inter (src_plats
, delta
);
2961 intersect_with_plats (src_plats
, &inter
, delta
);
2964 else if (jfunc
->agg
.items
)
2969 inter
= VEC_copy (ipa_agg_jf_item
, heap
, jfunc
->agg
.items
);
2971 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, inter
, k
, item
)
2974 bool found
= false;;
2979 while ((unsigned) l
< VEC_length (ipa_agg_jf_item_t
,
2982 struct ipa_agg_jf_item
*ti
;
2983 ti
= &VEC_index (ipa_agg_jf_item_t
,
2984 jfunc
->agg
.items
, l
);
2985 if (ti
->offset
> item
->offset
)
2987 if (ti
->offset
== item
->offset
)
2989 gcc_checking_assert (ti
->value
);
2990 if (values_equal_for_ipcp_p (item
->value
,
3008 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, inter
, j
, item
)
3010 struct ipa_agg_replacement_value
*v
;
3015 v
= ggc_alloc_ipa_agg_replacement_value ();
3017 v
->offset
= item
->offset
;
3018 v
->value
= item
->value
;
3025 VEC_free (ipa_agg_jf_item
, heap
, inter
);
3030 /* Turn KNOWN_AGGS into a list of aggreate replacement values. */
3032 static struct ipa_agg_replacement_value
*
3033 known_aggs_to_agg_replacement_list (VEC (ipa_agg_jump_function_t
,
3036 struct ipa_agg_replacement_value
*res
= NULL
;
3037 struct ipa_agg_jump_function
*aggjf
;
3038 struct ipa_agg_jf_item
*item
;
3041 FOR_EACH_VEC_ELT (ipa_agg_jump_function_t
, known_aggs
, i
, aggjf
)
3042 FOR_EACH_VEC_ELT (ipa_agg_jf_item_t
, aggjf
->items
, j
, item
)
3044 struct ipa_agg_replacement_value
*v
;
3045 v
= ggc_alloc_ipa_agg_replacement_value ();
3047 v
->offset
= item
->offset
;
3048 v
->value
= item
->value
;
3055 /* Determine whether CS also brings all scalar values that the NODE is
3059 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
3060 struct cgraph_node
*node
)
3062 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
3063 int count
= ipa_get_param_count (dest_info
);
3064 struct ipa_node_params
*caller_info
;
3065 struct ipa_edge_args
*args
;
3068 caller_info
= IPA_NODE_REF (cs
->caller
);
3069 args
= IPA_EDGE_REF (cs
);
3070 for (i
= 0; i
< count
; i
++)
3072 struct ipa_jump_func
*jump_func
;
3075 val
= VEC_index (tree
, dest_info
->known_vals
, i
);
3079 if (i
>= ipa_get_cs_argument_count (args
))
3081 jump_func
= ipa_get_ith_jump_func (args
, i
);
3082 t
= ipa_value_from_jfunc (caller_info
, jump_func
);
3083 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
3089 /* Determine whether CS also brings all aggregate values that NODE is
3092 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
3093 struct cgraph_node
*node
)
3095 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3096 struct ipa_agg_replacement_value
*aggval
;
3098 aggval
= ipa_get_agg_replacements_for_node (node
);
3102 struct ipcp_param_lattices
*plats
;
3103 plats
= ipa_get_parm_lattices (caller_info
, aggval
->index
);
3104 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
3106 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
3108 aglat
= aglat
->next
)
3109 if (aglat
->offset
== aggval
->offset
)
3111 if (ipa_lat_is_single_const (aglat
)
3112 && values_equal_for_ipcp_p (aggval
->value
,
3113 aglat
->values
->value
))
3125 aggval
= aggval
->next
;
3130 /* Given an original NODE and a VAL for which we have already created a
3131 specialized clone, look whether there are incoming edges that still lead
3132 into the old node but now also bring the requested value and also conform to
3133 all other criteria such that they can be redirected the the special node.
3134 This function can therefore redirect the final edge in a SCC. */
3137 perhaps_add_new_callers (struct cgraph_node
*node
, struct ipcp_value
*val
)
3139 struct ipcp_value_source
*src
;
3140 gcov_type redirected_sum
= 0;
3142 for (src
= val
->sources
; src
; src
= src
->next
)
3144 struct cgraph_edge
*cs
= src
->cs
;
3147 enum availability availability
;
3149 if (cgraph_function_node (cs
->callee
, &availability
) == node
3150 && availability
> AVAIL_OVERWRITABLE
3151 && cgraph_edge_brings_value_p (cs
, src
))
3153 if (cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
3154 && cgraph_edge_brings_all_agg_vals_for_node (cs
,
3158 fprintf (dump_file
, " - adding an extra caller %s/%i"
3160 xstrdup (cgraph_node_name (cs
->caller
)),
3162 xstrdup (cgraph_node_name (val
->spec_node
)),
3163 val
->spec_node
->uid
);
3165 cgraph_redirect_edge_callee (cs
, val
->spec_node
);
3166 redirected_sum
+= cs
->count
;
3169 cs
= get_next_cgraph_edge_clone (cs
);
3174 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
3178 /* Copy KNOWN_BINFOS to KNOWN_VALS. */
3181 move_binfos_to_values (VEC (tree
, heap
) *known_vals
,
3182 VEC (tree
, heap
) *known_binfos
)
3187 for (i
= 0; VEC_iterate (tree
, known_binfos
, i
, t
); i
++)
3189 VEC_replace (tree
, known_vals
, i
, t
);
3192 /* Return true if there is a replacement equivalent to VALUE, INDEX and OFFSET
3193 among those in the AGGVALS list. */
3196 ipcp_val_in_agg_replacements_p (struct ipa_agg_replacement_value
*aggvals
,
3197 int index
, HOST_WIDE_INT offset
, tree value
)
3201 if (aggvals
->index
== index
3202 && aggvals
->offset
== offset
3203 && values_equal_for_ipcp_p (aggvals
->value
, value
))
3205 aggvals
= aggvals
->next
;
3210 /* Decide wheter to create a special version of NODE for value VAL of parameter
3211 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
3212 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
3213 KNOWN_BINFOS and KNOWN_AGGS describe the other already known values. */
3216 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
3217 struct ipcp_value
*val
, VEC (tree
, heap
) *known_csts
,
3218 VEC (tree
, heap
) *known_binfos
)
3220 struct ipa_agg_replacement_value
*aggvals
;
3221 int freq_sum
, caller_count
;
3222 gcov_type count_sum
;
3223 VEC (cgraph_edge_p
, heap
) *callers
;
3224 VEC (tree
, heap
) *kv
;
3228 perhaps_add_new_callers (node
, val
);
3231 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
3233 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3234 fprintf (dump_file
, " Ignoring candidate value because "
3235 "max_new_size would be reached with %li.\n",
3236 val
->local_size_cost
+ overall_size
);
3239 else if (!get_info_about_necessary_edges (val
, &freq_sum
, &count_sum
,
3243 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3245 fprintf (dump_file
, " - considering value ");
3246 print_ipcp_constant_value (dump_file
, val
->value
);
3247 fprintf (dump_file
, " for parameter ");
3248 print_generic_expr (dump_file
, ipa_get_param (IPA_NODE_REF (node
),
3251 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
3252 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
3255 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
3256 freq_sum
, count_sum
,
3257 val
->local_size_cost
)
3258 && !good_cloning_opportunity_p (node
,
3259 val
->local_time_benefit
3260 + val
->prop_time_benefit
,
3261 freq_sum
, count_sum
,
3262 val
->local_size_cost
3263 + val
->prop_size_cost
))
3267 fprintf (dump_file
, " Creating a specialized node of %s/%i.\n",
3268 cgraph_node_name (node
), node
->uid
);
3270 callers
= gather_edges_for_value (val
, caller_count
);
3271 kv
= VEC_copy (tree
, heap
, known_csts
);
3272 move_binfos_to_values (kv
, known_binfos
);
3274 VEC_replace (tree
, kv
, index
, val
->value
);
3275 find_more_scalar_values_for_callers_subset (node
, kv
, callers
);
3276 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
3277 gcc_checking_assert (offset
== -1
3278 || ipcp_val_in_agg_replacements_p (aggvals
, index
,
3279 offset
, val
->value
));
3280 val
->spec_node
= create_specialized_node (node
, kv
, aggvals
, callers
);
3281 overall_size
+= val
->local_size_cost
;
3283 /* TODO: If for some lattice there is only one other known value
3284 left, make a special node for it too. */
3289 /* Decide whether and what specialized clones of NODE should be created. */
3292 decide_whether_version_node (struct cgraph_node
*node
)
3294 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3295 int i
, count
= ipa_get_param_count (info
);
3296 VEC (tree
, heap
) *known_csts
, *known_binfos
;
3297 VEC (ipa_agg_jump_function_t
, heap
) *known_aggs
= NULL
;
3303 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3304 fprintf (dump_file
, "\nEvaluating opportunities for %s/%i.\n",
3305 cgraph_node_name (node
), node
->uid
);
3307 gather_context_independent_values (info
, &known_csts
, &known_binfos
,
3308 info
->clone_for_all_contexts
? &known_aggs
3311 for (i
= 0; i
< count
;i
++)
3313 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3314 struct ipcp_lattice
*lat
= &plats
->itself
;
3315 struct ipcp_value
*val
;
3318 && !VEC_index (tree
, known_csts
, i
)
3319 && !VEC_index (tree
, known_binfos
, i
))
3320 for (val
= lat
->values
; val
; val
= val
->next
)
3321 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
3324 if (!plats
->aggs_bottom
|| !plats
->aggs
)
3326 struct ipcp_agg_lattice
*aglat
;
3327 struct ipcp_value
*val
;
3328 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3329 if (!aglat
->bottom
&& aglat
->values
3330 /* If the following is false, the one value is in
3332 && (plats
->aggs_contain_variable
3333 || !ipa_lat_is_single_const (aglat
)))
3334 for (val
= aglat
->values
; val
; val
= val
->next
)
3335 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
3336 known_csts
, known_binfos
);
3338 info
= IPA_NODE_REF (node
);
3341 if (info
->clone_for_all_contexts
)
3343 VEC (cgraph_edge_p
, heap
) *callers
;
3346 fprintf (dump_file
, " - Creating a specialized node of %s/%i "
3347 "for all known contexts.\n", cgraph_node_name (node
),
3350 callers
= collect_callers_of_node (node
);
3351 move_binfos_to_values (known_csts
, known_binfos
);
3352 create_specialized_node (node
, known_csts
,
3353 known_aggs_to_agg_replacement_list (known_aggs
),
3355 info
= IPA_NODE_REF (node
);
3356 info
->clone_for_all_contexts
= false;
3360 VEC_free (tree
, heap
, known_csts
);
3362 VEC_free (tree
, heap
, known_binfos
);
3366 /* Transitively mark all callees of NODE within the same SCC as not dead. */
3369 spread_undeadness (struct cgraph_node
*node
)
3371 struct cgraph_edge
*cs
;
3373 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
3374 if (edge_within_scc (cs
))
3376 struct cgraph_node
*callee
;
3377 struct ipa_node_params
*info
;
3379 callee
= cgraph_function_node (cs
->callee
, NULL
);
3380 info
= IPA_NODE_REF (callee
);
3382 if (info
->node_dead
)
3384 info
->node_dead
= 0;
3385 spread_undeadness (callee
);
3390 /* Return true if NODE has a caller from outside of its SCC that is not
3391 dead. Worker callback for cgraph_for_node_and_aliases. */
3394 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
3395 void *data ATTRIBUTE_UNUSED
)
3397 struct cgraph_edge
*cs
;
3399 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
3400 if (cs
->caller
->thunk
.thunk_p
3401 && cgraph_for_node_and_aliases (cs
->caller
,
3402 has_undead_caller_from_outside_scc_p
,
3405 else if (!edge_within_scc (cs
)
3406 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
3412 /* Identify nodes within the same SCC as NODE which are no longer needed
3413 because of new clones and will be removed as unreachable. */
3416 identify_dead_nodes (struct cgraph_node
*node
)
3418 struct cgraph_node
*v
;
3419 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3420 if (cgraph_will_be_removed_from_program_if_no_direct_calls (v
)
3421 && !cgraph_for_node_and_aliases (v
,
3422 has_undead_caller_from_outside_scc_p
,
3424 IPA_NODE_REF (v
)->node_dead
= 1;
3426 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3427 if (!IPA_NODE_REF (v
)->node_dead
)
3428 spread_undeadness (v
);
3430 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3432 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3433 if (IPA_NODE_REF (v
)->node_dead
)
3434 fprintf (dump_file
, " Marking node as dead: %s/%i.\n",
3435 cgraph_node_name (v
), v
->uid
);
3439 /* The decision stage. Iterate over the topological order of call graph nodes
3440 TOPO and make specialized clones if deemed beneficial. */
3443 ipcp_decision_stage (struct topo_info
*topo
)
3448 fprintf (dump_file
, "\nIPA decision stage:\n\n");
3450 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3452 struct cgraph_node
*node
= topo
->order
[i
];
3453 bool change
= false, iterate
= true;
3457 struct cgraph_node
*v
;
3459 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->symbol
.aux
)->next_cycle
)
3460 if (cgraph_function_with_gimple_body_p (v
)
3461 && ipcp_versionable_function_p (v
))
3462 iterate
|= decide_whether_version_node (v
);
3467 identify_dead_nodes (node
);
3471 /* The IPCP driver. */
3476 struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
3477 struct topo_info topo
;
3479 ipa_check_create_node_params ();
3480 ipa_check_create_edge_args ();
3481 grow_next_edge_clone_vector ();
3482 edge_duplication_hook_holder
=
3483 cgraph_add_edge_duplication_hook (&ipcp_edge_duplication_hook
, NULL
);
3484 ipcp_values_pool
= create_alloc_pool ("IPA-CP values",
3485 sizeof (struct ipcp_value
), 32);
3486 ipcp_sources_pool
= create_alloc_pool ("IPA-CP value sources",
3487 sizeof (struct ipcp_value_source
), 64);
3488 ipcp_agg_lattice_pool
= create_alloc_pool ("IPA_CP aggregate lattices",
3489 sizeof (struct ipcp_agg_lattice
),
3493 fprintf (dump_file
, "\nIPA structures before propagation:\n");
3494 if (dump_flags
& TDF_DETAILS
)
3495 ipa_print_all_params (dump_file
);
3496 ipa_print_all_jump_functions (dump_file
);
3499 /* Topological sort. */
3500 build_toporder_info (&topo
);
3501 /* Do the interprocedural propagation. */
3502 ipcp_propagate_stage (&topo
);
3503 /* Decide what constant propagation and cloning should be performed. */
3504 ipcp_decision_stage (&topo
);
3506 /* Free all IPCP structures. */
3507 free_toporder_info (&topo
);
3508 VEC_free (cgraph_edge_p
, heap
, next_edge_clone
);
3509 cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder
);
3510 ipa_free_all_structures_after_ipa_cp ();
3512 fprintf (dump_file
, "\nIPA constant propagation end\n");
3516 /* Initialization and computation of IPCP data structures. This is the initial
3517 intraprocedural analysis of functions, which gathers information to be
3518 propagated later on. */
3521 ipcp_generate_summary (void)
3523 struct cgraph_node
*node
;
3526 fprintf (dump_file
, "\nIPA constant propagation start:\n");
3527 ipa_register_cgraph_hooks ();
3529 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
3531 node
->local
.versionable
3532 = tree_versionable_function_p (node
->symbol
.decl
);
3533 ipa_analyze_node (node
);
3537 /* Write ipcp summary for nodes in SET. */
3540 ipcp_write_summary (void)
3542 ipa_prop_write_jump_functions ();
3545 /* Read ipcp summary. */
3548 ipcp_read_summary (void)
3550 ipa_prop_read_jump_functions ();
3553 /* Gate for IPCP optimization. */
3556 cgraph_gate_cp (void)
3558 /* FIXME: We should remove the optimize check after we ensure we never run
3559 IPA passes when not optimizing. */
3560 return flag_ipa_cp
&& optimize
;
3563 struct ipa_opt_pass_d pass_ipa_cp
=
3568 OPTGROUP_NONE
, /* optinfo_flags */
3569 cgraph_gate_cp
, /* gate */
3570 ipcp_driver
, /* execute */
3573 0, /* static_pass_number */
3574 TV_IPA_CONSTANT_PROP
, /* tv_id */
3575 0, /* properties_required */
3576 0, /* properties_provided */
3577 0, /* properties_destroyed */
3578 0, /* todo_flags_start */
3580 TODO_remove_functions
| TODO_ggc_collect
/* todo_flags_finish */
3582 ipcp_generate_summary
, /* generate_summary */
3583 ipcp_write_summary
, /* write_summary */
3584 ipcp_read_summary
, /* read_summary */
3585 ipa_prop_write_all_agg_replacement
, /* write_optimization_summary */
3586 ipa_prop_read_all_agg_replacement
, /* read_optimization_summary */
3587 NULL
, /* stmt_fixup */
3589 ipcp_transform_function
, /* function_transform */
3590 NULL
, /* variable_transform */