1 /* Interprocedural constant propagation
2 Copyright (C) 2005-2018 Free Software Foundation, Inc.
4 Contributed by Razya Ladelsky <RAZYA@il.ibm.com> and Martin Jambor
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* Interprocedural constant propagation (IPA-CP).
25 The goal of this transformation is to
27 1) discover functions which are always invoked with some arguments with the
28 same known constant values and modify the functions so that the
29 subsequent optimizations can take advantage of the knowledge, and
31 2) partial specialization - create specialized versions of functions
32 transformed in this way if some parameters are known constants only in
33 certain contexts but the estimated tradeoff between speedup and cost size
36 The algorithm also propagates types and attempts to perform type based
37 devirtualization. Types are propagated much like constants.
39 The algorithm basically consists of three stages. In the first, functions
40 are analyzed one at a time and jump functions are constructed for all known
41 call-sites. In the second phase, the pass propagates information from the
42 jump functions across the call to reveal what values are available at what
43 call sites, performs estimations of effects of known values on functions and
44 their callees, and finally decides what specialized extra versions should be
45 created. In the third, the special versions materialize and appropriate
48 The algorithm used is to a certain extent based on "Interprocedural Constant
49 Propagation", by David Callahan, Keith D Cooper, Ken Kennedy, Linda Torczon,
50 Comp86, pg 152-161 and "A Methodology for Procedure Cloning" by Keith D
51 Cooper, Mary W. Hall, and Ken Kennedy.
54 First stage - intraprocedural analysis
55 =======================================
57 This phase computes jump_function and modification flags.
59 A jump function for a call-site represents the values passed as an actual
60 arguments of a given call-site. In principle, there are three types of
63 Pass through - the caller's formal parameter is passed as an actual
64 argument, plus an operation on it can be performed.
65 Constant - a constant is passed as an actual argument.
66 Unknown - neither of the above.
68 All jump function types are described in detail in ipa-prop.h, together with
69 the data structures that represent them and methods of accessing them.
71 ipcp_generate_summary() is the main function of the first stage.
73 Second stage - interprocedural analysis
74 ========================================
76 This stage is itself divided into two phases. In the first, we propagate
77 known values over the call graph, in the second, we make cloning decisions.
78 It uses a different algorithm than the original Callahan's paper.
80 First, we traverse the functions topologically from callers to callees and,
81 for each strongly connected component (SCC), we propagate constants
82 according to previously computed jump functions. We also record what known
83 values depend on other known values and estimate local effects. Finally, we
84 propagate cumulative information about these effects from dependent values
85 to those on which they depend.
87 Second, we again traverse the call graph in the same topological order and
88 make clones for functions which we know are called with the same values in
89 all contexts and decide about extra specialized clones of functions just for
90 some contexts - these decisions are based on both local estimates and
91 cumulative estimates propagated from callees.
93 ipcp_propagate_stage() and ipcp_decision_stage() together constitute the
96 Third phase - materialization of clones, call statement updates.
97 ============================================
99 This stage is currently performed by call graph code (mainly in cgraphunit.c
100 and tree-inline.c) according to instructions inserted to the call graph by
105 #include "coretypes.h"
108 #include "gimple-expr.h"
110 #include "alloc-pool.h"
111 #include "tree-pass.h"
113 #include "diagnostic.h"
114 #include "fold-const.h"
115 #include "gimple-fold.h"
116 #include "symbol-summary.h"
117 #include "tree-vrp.h"
118 #include "ipa-prop.h"
119 #include "tree-pretty-print.h"
120 #include "tree-inline.h"
122 #include "ipa-fnsummary.h"
123 #include "ipa-utils.h"
124 #include "tree-ssa-ccp.h"
125 #include "stringpool.h"
128 template <typename valtype
> class ipcp_value
;
130 /* Describes a particular source for an IPA-CP value. */
132 template <typename valtype
>
133 class ipcp_value_source
136 /* Aggregate offset of the source, negative if the source is scalar value of
137 the argument itself. */
138 HOST_WIDE_INT offset
;
139 /* The incoming edge that brought the value. */
141 /* If the jump function that resulted into his value was a pass-through or an
142 ancestor, this is the ipcp_value of the caller from which the described
143 value has been derived. Otherwise it is NULL. */
144 ipcp_value
<valtype
> *val
;
145 /* Next pointer in a linked list of sources of a value. */
146 ipcp_value_source
*next
;
147 /* If the jump function that resulted into his value was a pass-through or an
148 ancestor, this is the index of the parameter of the caller the jump
149 function references. */
153 /* Common ancestor for all ipcp_value instantiations. */
155 class ipcp_value_base
158 /* Time benefit and size cost that specializing the function for this value
159 would bring about in this function alone. */
160 int local_time_benefit
, local_size_cost
;
161 /* Time benefit and size cost that specializing the function for this value
162 can bring about in it's callees (transitively). */
163 int prop_time_benefit
, prop_size_cost
;
166 : local_time_benefit (0), local_size_cost (0),
167 prop_time_benefit (0), prop_size_cost (0) {}
170 /* Describes one particular value stored in struct ipcp_lattice. */
172 template <typename valtype
>
173 class ipcp_value
: public ipcp_value_base
176 /* The actual value for the given parameter. */
178 /* The list of sources from which this value originates. */
179 ipcp_value_source
<valtype
> *sources
;
180 /* Next pointers in a linked list of all values in a lattice. */
182 /* Next pointers in a linked list of values in a strongly connected component
184 ipcp_value
*scc_next
;
185 /* Next pointers in a linked list of SCCs of values sorted topologically
186 according their sources. */
187 ipcp_value
*topo_next
;
188 /* A specialized node created for this value, NULL if none has been (so far)
190 cgraph_node
*spec_node
;
191 /* Depth first search number and low link for topological sorting of
194 /* True if this valye is currently on the topo-sort stack. */
198 : sources (0), next (0), scc_next (0), topo_next (0),
199 spec_node (0), dfs (0), low_link (0), on_stack (false) {}
201 void add_source (cgraph_edge
*cs
, ipcp_value
*src_val
, int src_idx
,
202 HOST_WIDE_INT offset
);
205 /* Lattice describing potential values of a formal parameter of a function, or
206 a part of an aggregate. TOP is represented by a lattice with zero values
207 and with contains_variable and bottom flags cleared. BOTTOM is represented
208 by a lattice with the bottom flag set. In that case, values and
209 contains_variable flag should be disregarded. */
211 template <typename valtype
>
215 /* The list of known values and types in this lattice. Note that values are
216 not deallocated if a lattice is set to bottom because there may be value
217 sources referencing them. */
218 ipcp_value
<valtype
> *values
;
219 /* Number of known values and types in this lattice. */
221 /* The lattice contains a variable component (in addition to values). */
222 bool contains_variable
;
223 /* The value of the lattice is bottom (i.e. variable and unusable for any
227 inline bool is_single_const ();
228 inline bool set_to_bottom ();
229 inline bool set_contains_variable ();
230 bool add_value (valtype newval
, cgraph_edge
*cs
,
231 ipcp_value
<valtype
> *src_val
= NULL
,
232 int src_idx
= 0, HOST_WIDE_INT offset
= -1);
233 void print (FILE * f
, bool dump_sources
, bool dump_benefits
);
236 /* Lattice of tree values with an offset to describe a part of an
239 class ipcp_agg_lattice
: public ipcp_lattice
<tree
>
242 /* Offset that is being described by this lattice. */
243 HOST_WIDE_INT offset
;
244 /* Size so that we don't have to re-compute it every time we traverse the
245 list. Must correspond to TYPE_SIZE of all lat values. */
247 /* Next element of the linked list. */
248 struct ipcp_agg_lattice
*next
;
251 /* Lattice of known bits, only capable of holding one value.
252 Bitwise constant propagation propagates which bits of a
268 In the above case, the param 'x' will always have all
269 the bits (except the bits in lsb) set to 0.
270 Hence the mask of 'x' would be 0xff. The mask
271 reflects that the bits in lsb are unknown.
272 The actual propagated value is given by m_value & ~m_mask. */
274 class ipcp_bits_lattice
277 bool bottom_p () { return m_lattice_val
== IPA_BITS_VARYING
; }
278 bool top_p () { return m_lattice_val
== IPA_BITS_UNDEFINED
; }
279 bool constant_p () { return m_lattice_val
== IPA_BITS_CONSTANT
; }
280 bool set_to_bottom ();
281 bool set_to_constant (widest_int
, widest_int
);
283 widest_int
get_value () { return m_value
; }
284 widest_int
get_mask () { return m_mask
; }
286 bool meet_with (ipcp_bits_lattice
& other
, unsigned, signop
,
287 enum tree_code
, tree
);
289 bool meet_with (widest_int
, widest_int
, unsigned);
294 enum { IPA_BITS_UNDEFINED
, IPA_BITS_CONSTANT
, IPA_BITS_VARYING
} m_lattice_val
;
296 /* Similar to ccp_lattice_t, mask represents which bits of value are constant.
297 If a bit in mask is set to 0, then the corresponding bit in
298 value is known to be constant. */
299 widest_int m_value
, m_mask
;
301 bool meet_with_1 (widest_int
, widest_int
, unsigned);
302 void get_value_and_mask (tree
, widest_int
*, widest_int
*);
305 /* Lattice of value ranges. */
307 class ipcp_vr_lattice
312 inline bool bottom_p () const;
313 inline bool top_p () const;
314 inline bool set_to_bottom ();
315 bool meet_with (const value_range
*p_vr
);
316 bool meet_with (const ipcp_vr_lattice
&other
);
317 void init () { m_vr
.type
= VR_UNDEFINED
; }
318 void print (FILE * f
);
321 bool meet_with_1 (const value_range
*other_vr
);
324 /* Structure containing lattices for a parameter itself and for pieces of
325 aggregates that are passed in the parameter or by a reference in a parameter
326 plus some other useful flags. */
328 class ipcp_param_lattices
331 /* Lattice describing the value of the parameter itself. */
332 ipcp_lattice
<tree
> itself
;
333 /* Lattice describing the polymorphic contexts of a parameter. */
334 ipcp_lattice
<ipa_polymorphic_call_context
> ctxlat
;
335 /* Lattices describing aggregate parts. */
336 ipcp_agg_lattice
*aggs
;
337 /* Lattice describing known bits. */
338 ipcp_bits_lattice bits_lattice
;
339 /* Lattice describing value range. */
340 ipcp_vr_lattice m_value_range
;
341 /* Number of aggregate lattices */
343 /* True if aggregate data were passed by reference (as opposed to by
346 /* All aggregate lattices contain a variable component (in addition to
348 bool aggs_contain_variable
;
349 /* The value of all aggregate lattices is bottom (i.e. variable and unusable
350 for any propagation). */
353 /* There is a virtual call based on this parameter. */
357 /* Allocation pools for values and their sources in ipa-cp. */
359 object_allocator
<ipcp_value
<tree
> > ipcp_cst_values_pool
360 ("IPA-CP constant values");
362 object_allocator
<ipcp_value
<ipa_polymorphic_call_context
> >
363 ipcp_poly_ctx_values_pool ("IPA-CP polymorphic contexts");
365 object_allocator
<ipcp_value_source
<tree
> > ipcp_sources_pool
366 ("IPA-CP value sources");
368 object_allocator
<ipcp_agg_lattice
> ipcp_agg_lattice_pool
369 ("IPA_CP aggregate lattices");
371 /* Maximal count found in program. */
373 static profile_count max_count
;
375 /* Original overall size of the program. */
377 static long overall_size
, max_new_size
;
379 /* Return the param lattices structure corresponding to the Ith formal
380 parameter of the function described by INFO. */
381 static inline struct ipcp_param_lattices
*
382 ipa_get_parm_lattices (struct ipa_node_params
*info
, int i
)
384 gcc_assert (i
>= 0 && i
< ipa_get_param_count (info
));
385 gcc_checking_assert (!info
->ipcp_orig_node
);
386 gcc_checking_assert (info
->lattices
);
387 return &(info
->lattices
[i
]);
390 /* Return the lattice corresponding to the scalar value of the Ith formal
391 parameter of the function described by INFO. */
392 static inline ipcp_lattice
<tree
> *
393 ipa_get_scalar_lat (struct ipa_node_params
*info
, int i
)
395 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
396 return &plats
->itself
;
399 /* Return the lattice corresponding to the scalar value of the Ith formal
400 parameter of the function described by INFO. */
401 static inline ipcp_lattice
<ipa_polymorphic_call_context
> *
402 ipa_get_poly_ctx_lat (struct ipa_node_params
*info
, int i
)
404 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
405 return &plats
->ctxlat
;
408 /* Return the lattice corresponding to the value range of the Ith formal
409 parameter of the function described by INFO. */
411 static inline ipcp_vr_lattice
*
412 ipa_get_vr_lat (struct ipa_node_params
*info
, int i
)
414 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
415 return &plats
->m_value_range
;
418 /* Return whether LAT is a lattice with a single constant and without an
421 template <typename valtype
>
423 ipcp_lattice
<valtype
>::is_single_const ()
425 if (bottom
|| contains_variable
|| values_count
!= 1)
431 /* Print V which is extracted from a value in a lattice to F. */
434 print_ipcp_constant_value (FILE * f
, tree v
)
436 if (TREE_CODE (v
) == ADDR_EXPR
437 && TREE_CODE (TREE_OPERAND (v
, 0)) == CONST_DECL
)
440 print_generic_expr (f
, DECL_INITIAL (TREE_OPERAND (v
, 0)));
443 print_generic_expr (f
, v
);
446 /* Print V which is extracted from a value in a lattice to F. */
449 print_ipcp_constant_value (FILE * f
, ipa_polymorphic_call_context v
)
454 /* Print a lattice LAT to F. */
456 template <typename valtype
>
458 ipcp_lattice
<valtype
>::print (FILE * f
, bool dump_sources
, bool dump_benefits
)
460 ipcp_value
<valtype
> *val
;
465 fprintf (f
, "BOTTOM\n");
469 if (!values_count
&& !contains_variable
)
471 fprintf (f
, "TOP\n");
475 if (contains_variable
)
477 fprintf (f
, "VARIABLE");
483 for (val
= values
; val
; val
= val
->next
)
485 if (dump_benefits
&& prev
)
487 else if (!dump_benefits
&& prev
)
492 print_ipcp_constant_value (f
, val
->value
);
496 ipcp_value_source
<valtype
> *s
;
498 fprintf (f
, " [from:");
499 for (s
= val
->sources
; s
; s
= s
->next
)
500 fprintf (f
, " %i(%f)", s
->cs
->caller
->order
,
501 s
->cs
->sreal_frequency ().to_double ());
506 fprintf (f
, " [loc_time: %i, loc_size: %i, "
507 "prop_time: %i, prop_size: %i]\n",
508 val
->local_time_benefit
, val
->local_size_cost
,
509 val
->prop_time_benefit
, val
->prop_size_cost
);
516 ipcp_bits_lattice::print (FILE *f
)
519 fprintf (f
, " Bits unknown (TOP)\n");
520 else if (bottom_p ())
521 fprintf (f
, " Bits unusable (BOTTOM)\n");
524 fprintf (f
, " Bits: value = "); print_hex (get_value (), f
);
525 fprintf (f
, ", mask = "); print_hex (get_mask (), f
);
530 /* Print value range lattice to F. */
533 ipcp_vr_lattice::print (FILE * f
)
535 dump_value_range (f
, &m_vr
);
538 /* Print all ipcp_lattices of all functions to F. */
541 print_all_lattices (FILE * f
, bool dump_sources
, bool dump_benefits
)
543 struct cgraph_node
*node
;
546 fprintf (f
, "\nLattices:\n");
547 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
549 struct ipa_node_params
*info
;
551 info
= IPA_NODE_REF (node
);
552 fprintf (f
, " Node: %s:\n", node
->dump_name ());
553 count
= ipa_get_param_count (info
);
554 for (i
= 0; i
< count
; i
++)
556 struct ipcp_agg_lattice
*aglat
;
557 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
558 fprintf (f
, " param [%d]: ", i
);
559 plats
->itself
.print (f
, dump_sources
, dump_benefits
);
560 fprintf (f
, " ctxs: ");
561 plats
->ctxlat
.print (f
, dump_sources
, dump_benefits
);
562 plats
->bits_lattice
.print (f
);
564 plats
->m_value_range
.print (f
);
566 if (plats
->virt_call
)
567 fprintf (f
, " virt_call flag set\n");
569 if (plats
->aggs_bottom
)
571 fprintf (f
, " AGGS BOTTOM\n");
574 if (plats
->aggs_contain_variable
)
575 fprintf (f
, " AGGS VARIABLE\n");
576 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
578 fprintf (f
, " %soffset " HOST_WIDE_INT_PRINT_DEC
": ",
579 plats
->aggs_by_ref
? "ref " : "", aglat
->offset
);
580 aglat
->print (f
, dump_sources
, dump_benefits
);
586 /* Determine whether it is at all technically possible to create clones of NODE
587 and store this information in the ipa_node_params structure associated
591 determine_versionability (struct cgraph_node
*node
,
592 struct ipa_node_params
*info
)
594 const char *reason
= NULL
;
596 /* There are a number of generic reasons functions cannot be versioned. We
597 also cannot remove parameters if there are type attributes such as fnspec
599 if (node
->alias
|| node
->thunk
.thunk_p
)
600 reason
= "alias or thunk";
601 else if (!node
->local
.versionable
)
602 reason
= "not a tree_versionable_function";
603 else if (node
->get_availability () <= AVAIL_INTERPOSABLE
)
604 reason
= "insufficient body availability";
605 else if (!opt_for_fn (node
->decl
, optimize
)
606 || !opt_for_fn (node
->decl
, flag_ipa_cp
))
607 reason
= "non-optimized function";
608 else if (lookup_attribute ("omp declare simd", DECL_ATTRIBUTES (node
->decl
)))
610 /* Ideally we should clone the SIMD clones themselves and create
611 vector copies of them, so IPA-cp and SIMD clones can happily
612 coexist, but that may not be worth the effort. */
613 reason
= "function has SIMD clones";
615 else if (lookup_attribute ("target_clones", DECL_ATTRIBUTES (node
->decl
)))
617 /* Ideally we should clone the target clones themselves and create
618 copies of them, so IPA-cp and target clones can happily
619 coexist, but that may not be worth the effort. */
620 reason
= "function target_clones attribute";
622 /* Don't clone decls local to a comdat group; it breaks and for C++
623 decloned constructors, inlining is always better anyway. */
624 else if (node
->comdat_local_p ())
625 reason
= "comdat-local function";
626 else if (node
->calls_comdat_local
)
628 /* TODO: call is versionable if we make sure that all
629 callers are inside of a comdat group. */
630 reason
= "calls comdat-local function";
633 /* Functions calling BUILT_IN_VA_ARG_PACK and BUILT_IN_VA_ARG_PACK_LEN
634 work only when inlined. Cloning them may still lead to better code
635 because ipa-cp will not give up on cloning further. If the function is
636 external this however leads to wrong code because we may end up producing
637 offline copy of the function. */
638 if (DECL_EXTERNAL (node
->decl
))
639 for (cgraph_edge
*edge
= node
->callees
; !reason
&& edge
;
640 edge
= edge
->next_callee
)
641 if (fndecl_built_in_p (edge
->callee
->decl
, BUILT_IN_NORMAL
))
643 if (DECL_FUNCTION_CODE (edge
->callee
->decl
) == BUILT_IN_VA_ARG_PACK
)
644 reason
= "external function which calls va_arg_pack";
645 if (DECL_FUNCTION_CODE (edge
->callee
->decl
)
646 == BUILT_IN_VA_ARG_PACK_LEN
)
647 reason
= "external function which calls va_arg_pack_len";
650 if (reason
&& dump_file
&& !node
->alias
&& !node
->thunk
.thunk_p
)
651 fprintf (dump_file
, "Function %s is not versionable, reason: %s.\n",
652 node
->dump_name (), reason
);
654 info
->versionable
= (reason
== NULL
);
657 /* Return true if it is at all technically possible to create clones of a
661 ipcp_versionable_function_p (struct cgraph_node
*node
)
663 return IPA_NODE_REF (node
)->versionable
;
666 /* Structure holding accumulated information about callers of a node. */
668 struct caller_statistics
670 profile_count count_sum
;
671 int n_calls
, n_hot_calls
, freq_sum
;
674 /* Initialize fields of STAT to zeroes. */
677 init_caller_stats (struct caller_statistics
*stats
)
679 stats
->count_sum
= profile_count::zero ();
681 stats
->n_hot_calls
= 0;
685 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
686 non-thunk incoming edges to NODE. */
689 gather_caller_stats (struct cgraph_node
*node
, void *data
)
691 struct caller_statistics
*stats
= (struct caller_statistics
*) data
;
692 struct cgraph_edge
*cs
;
694 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
695 if (!cs
->caller
->thunk
.thunk_p
)
697 if (cs
->count
.ipa ().initialized_p ())
698 stats
->count_sum
+= cs
->count
.ipa ();
699 stats
->freq_sum
+= cs
->frequency ();
701 if (cs
->maybe_hot_p ())
702 stats
->n_hot_calls
++;
708 /* Return true if this NODE is viable candidate for cloning. */
711 ipcp_cloning_candidate_p (struct cgraph_node
*node
)
713 struct caller_statistics stats
;
715 gcc_checking_assert (node
->has_gimple_body_p ());
717 if (!opt_for_fn (node
->decl
, flag_ipa_cp_clone
))
720 fprintf (dump_file
, "Not considering %s for cloning; "
721 "-fipa-cp-clone disabled.\n",
726 if (node
->optimize_for_size_p ())
729 fprintf (dump_file
, "Not considering %s for cloning; "
730 "optimizing it for size.\n",
735 init_caller_stats (&stats
);
736 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
, false);
738 if (ipa_fn_summaries
->get (node
)->self_size
< stats
.n_calls
)
741 fprintf (dump_file
, "Considering %s for cloning; code might shrink.\n",
746 /* When profile is available and function is hot, propagate into it even if
747 calls seems cold; constant propagation can improve function's speed
749 if (max_count
> profile_count::zero ())
751 if (stats
.count_sum
> node
->count
.ipa ().apply_scale (90, 100))
754 fprintf (dump_file
, "Considering %s for cloning; "
755 "usually called directly.\n",
760 if (!stats
.n_hot_calls
)
763 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
768 fprintf (dump_file
, "Considering %s for cloning.\n",
773 template <typename valtype
>
774 class value_topo_info
777 /* Head of the linked list of topologically sorted values. */
778 ipcp_value
<valtype
> *values_topo
;
779 /* Stack for creating SCCs, represented by a linked list too. */
780 ipcp_value
<valtype
> *stack
;
781 /* Counter driving the algorithm in add_val_to_toposort. */
784 value_topo_info () : values_topo (NULL
), stack (NULL
), dfs_counter (0)
786 void add_val (ipcp_value
<valtype
> *cur_val
);
787 void propagate_effects ();
790 /* Arrays representing a topological ordering of call graph nodes and a stack
791 of nodes used during constant propagation and also data required to perform
792 topological sort of values and propagation of benefits in the determined
798 /* Array with obtained topological order of cgraph nodes. */
799 struct cgraph_node
**order
;
800 /* Stack of cgraph nodes used during propagation within SCC until all values
801 in the SCC stabilize. */
802 struct cgraph_node
**stack
;
803 int nnodes
, stack_top
;
805 value_topo_info
<tree
> constants
;
806 value_topo_info
<ipa_polymorphic_call_context
> contexts
;
808 ipa_topo_info () : order(NULL
), stack(NULL
), nnodes(0), stack_top(0),
813 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
816 build_toporder_info (struct ipa_topo_info
*topo
)
818 topo
->order
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
819 topo
->stack
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
821 gcc_checking_assert (topo
->stack_top
== 0);
822 topo
->nnodes
= ipa_reduced_postorder (topo
->order
, true, true, NULL
);
825 /* Free information about strongly connected components and the arrays in
829 free_toporder_info (struct ipa_topo_info
*topo
)
831 ipa_free_postorder_info ();
836 /* Add NODE to the stack in TOPO, unless it is already there. */
839 push_node_to_stack (struct ipa_topo_info
*topo
, struct cgraph_node
*node
)
841 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
842 if (info
->node_enqueued
)
844 info
->node_enqueued
= 1;
845 topo
->stack
[topo
->stack_top
++] = node
;
848 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
851 static struct cgraph_node
*
852 pop_node_from_stack (struct ipa_topo_info
*topo
)
856 struct cgraph_node
*node
;
858 node
= topo
->stack
[topo
->stack_top
];
859 IPA_NODE_REF (node
)->node_enqueued
= 0;
866 /* Set lattice LAT to bottom and return true if it previously was not set as
869 template <typename valtype
>
871 ipcp_lattice
<valtype
>::set_to_bottom ()
878 /* Mark lattice as containing an unknown value and return true if it previously
879 was not marked as such. */
881 template <typename valtype
>
883 ipcp_lattice
<valtype
>::set_contains_variable ()
885 bool ret
= !contains_variable
;
886 contains_variable
= true;
890 /* Set all aggegate lattices in PLATS to bottom and return true if they were
891 not previously set as such. */
894 set_agg_lats_to_bottom (struct ipcp_param_lattices
*plats
)
896 bool ret
= !plats
->aggs_bottom
;
897 plats
->aggs_bottom
= true;
901 /* Mark all aggegate lattices in PLATS as containing an unknown value and
902 return true if they were not previously marked as such. */
905 set_agg_lats_contain_variable (struct ipcp_param_lattices
*plats
)
907 bool ret
= !plats
->aggs_contain_variable
;
908 plats
->aggs_contain_variable
= true;
913 ipcp_vr_lattice::meet_with (const ipcp_vr_lattice
&other
)
915 return meet_with_1 (&other
.m_vr
);
918 /* Meet the current value of the lattice with value ranfge described by VR
922 ipcp_vr_lattice::meet_with (const value_range
*p_vr
)
924 return meet_with_1 (p_vr
);
927 /* Meet the current value of the lattice with value ranfge described by
931 ipcp_vr_lattice::meet_with_1 (const value_range
*other_vr
)
933 tree min
= m_vr
.min
, max
= m_vr
.max
;
934 value_range_type type
= m_vr
.type
;
939 if (other_vr
->type
== VR_VARYING
)
940 return set_to_bottom ();
942 vrp_meet (&m_vr
, other_vr
);
943 if (type
!= m_vr
.type
951 /* Return true if value range information in the lattice is yet unknown. */
954 ipcp_vr_lattice::top_p () const
956 return m_vr
.type
== VR_UNDEFINED
;
959 /* Return true if value range information in the lattice is known to be
963 ipcp_vr_lattice::bottom_p () const
965 return m_vr
.type
== VR_VARYING
;
968 /* Set value range information in the lattice to bottom. Return true if it
969 previously was in a different state. */
972 ipcp_vr_lattice::set_to_bottom ()
974 if (m_vr
.type
== VR_VARYING
)
976 m_vr
.type
= VR_VARYING
;
980 /* Set lattice value to bottom, if it already isn't the case. */
983 ipcp_bits_lattice::set_to_bottom ()
987 m_lattice_val
= IPA_BITS_VARYING
;
993 /* Set to constant if it isn't already. Only meant to be called
994 when switching state from TOP. */
997 ipcp_bits_lattice::set_to_constant (widest_int value
, widest_int mask
)
999 gcc_assert (top_p ());
1000 m_lattice_val
= IPA_BITS_CONSTANT
;
1006 /* Convert operand to value, mask form. */
1009 ipcp_bits_lattice::get_value_and_mask (tree operand
, widest_int
*valuep
, widest_int
*maskp
)
1011 wide_int
get_nonzero_bits (const_tree
);
1013 if (TREE_CODE (operand
) == INTEGER_CST
)
1015 *valuep
= wi::to_widest (operand
);
1025 /* Meet operation, similar to ccp_lattice_meet, we xor values
1026 if this->value, value have different values at same bit positions, we want
1027 to drop that bit to varying. Return true if mask is changed.
1028 This function assumes that the lattice value is in CONSTANT state */
1031 ipcp_bits_lattice::meet_with_1 (widest_int value
, widest_int mask
,
1034 gcc_assert (constant_p ());
1036 widest_int old_mask
= m_mask
;
1037 m_mask
= (m_mask
| mask
) | (m_value
^ value
);
1039 if (wi::sext (m_mask
, precision
) == -1)
1040 return set_to_bottom ();
1042 return m_mask
!= old_mask
;
1045 /* Meet the bits lattice with operand
1046 described by <value, mask, sgn, precision. */
1049 ipcp_bits_lattice::meet_with (widest_int value
, widest_int mask
,
1057 if (wi::sext (mask
, precision
) == -1)
1058 return set_to_bottom ();
1059 return set_to_constant (value
, mask
);
1062 return meet_with_1 (value
, mask
, precision
);
1065 /* Meet bits lattice with the result of bit_value_binop (other, operand)
1066 if code is binary operation or bit_value_unop (other) if code is unary op.
1067 In the case when code is nop_expr, no adjustment is required. */
1070 ipcp_bits_lattice::meet_with (ipcp_bits_lattice
& other
, unsigned precision
,
1071 signop sgn
, enum tree_code code
, tree operand
)
1073 if (other
.bottom_p ())
1074 return set_to_bottom ();
1076 if (bottom_p () || other
.top_p ())
1079 widest_int adjusted_value
, adjusted_mask
;
1081 if (TREE_CODE_CLASS (code
) == tcc_binary
)
1083 tree type
= TREE_TYPE (operand
);
1084 gcc_assert (INTEGRAL_TYPE_P (type
));
1085 widest_int o_value
, o_mask
;
1086 get_value_and_mask (operand
, &o_value
, &o_mask
);
1088 bit_value_binop (code
, sgn
, precision
, &adjusted_value
, &adjusted_mask
,
1089 sgn
, precision
, other
.get_value (), other
.get_mask (),
1090 TYPE_SIGN (type
), TYPE_PRECISION (type
), o_value
, o_mask
);
1092 if (wi::sext (adjusted_mask
, precision
) == -1)
1093 return set_to_bottom ();
1096 else if (TREE_CODE_CLASS (code
) == tcc_unary
)
1098 bit_value_unop (code
, sgn
, precision
, &adjusted_value
,
1099 &adjusted_mask
, sgn
, precision
, other
.get_value (),
1102 if (wi::sext (adjusted_mask
, precision
) == -1)
1103 return set_to_bottom ();
1107 return set_to_bottom ();
1111 if (wi::sext (adjusted_mask
, precision
) == -1)
1112 return set_to_bottom ();
1113 return set_to_constant (adjusted_value
, adjusted_mask
);
1116 return meet_with_1 (adjusted_value
, adjusted_mask
, precision
);
1119 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
1120 return true is any of them has not been marked as such so far. */
1123 set_all_contains_variable (struct ipcp_param_lattices
*plats
)
1126 ret
= plats
->itself
.set_contains_variable ();
1127 ret
|= plats
->ctxlat
.set_contains_variable ();
1128 ret
|= set_agg_lats_contain_variable (plats
);
1129 ret
|= plats
->bits_lattice
.set_to_bottom ();
1130 ret
|= plats
->m_value_range
.set_to_bottom ();
1134 /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA
1135 points to by the number of callers to NODE. */
1138 count_callers (cgraph_node
*node
, void *data
)
1140 int *caller_count
= (int *) data
;
1142 for (cgraph_edge
*cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
1143 /* Local thunks can be handled transparently, but if the thunk can not
1144 be optimized out, count it as a real use. */
1145 if (!cs
->caller
->thunk
.thunk_p
|| !cs
->caller
->local
.local
)
1150 /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on
1151 the one caller of some other node. Set the caller's corresponding flag. */
1154 set_single_call_flag (cgraph_node
*node
, void *)
1156 cgraph_edge
*cs
= node
->callers
;
1157 /* Local thunks can be handled transparently, skip them. */
1158 while (cs
&& cs
->caller
->thunk
.thunk_p
&& cs
->caller
->local
.local
)
1159 cs
= cs
->next_caller
;
1162 IPA_NODE_REF (cs
->caller
)->node_calling_single_call
= true;
1168 /* Initialize ipcp_lattices. */
1171 initialize_node_lattices (struct cgraph_node
*node
)
1173 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1174 struct cgraph_edge
*ie
;
1175 bool disable
= false, variable
= false;
1178 gcc_checking_assert (node
->has_gimple_body_p ());
1179 if (node
->local
.local
)
1181 int caller_count
= 0;
1182 node
->call_for_symbol_thunks_and_aliases (count_callers
, &caller_count
,
1184 gcc_checking_assert (caller_count
> 0);
1185 if (caller_count
== 1)
1186 node
->call_for_symbol_thunks_and_aliases (set_single_call_flag
,
1191 /* When cloning is allowed, we can assume that externally visible
1192 functions are not called. We will compensate this by cloning
1194 if (ipcp_versionable_function_p (node
)
1195 && ipcp_cloning_candidate_p (node
))
1201 for (i
= 0; i
< ipa_get_param_count (info
); i
++)
1203 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1204 plats
->m_value_range
.init ();
1207 if (disable
|| variable
)
1209 for (i
= 0; i
< ipa_get_param_count (info
); i
++)
1211 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1214 plats
->itself
.set_to_bottom ();
1215 plats
->ctxlat
.set_to_bottom ();
1216 set_agg_lats_to_bottom (plats
);
1217 plats
->bits_lattice
.set_to_bottom ();
1218 plats
->m_value_range
.set_to_bottom ();
1221 set_all_contains_variable (plats
);
1223 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
1224 && !node
->alias
&& !node
->thunk
.thunk_p
)
1225 fprintf (dump_file
, "Marking all lattices of %s as %s\n",
1226 node
->dump_name (), disable
? "BOTTOM" : "VARIABLE");
1229 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1230 if (ie
->indirect_info
->polymorphic
1231 && ie
->indirect_info
->param_index
>= 0)
1233 gcc_checking_assert (ie
->indirect_info
->param_index
>= 0);
1234 ipa_get_parm_lattices (info
,
1235 ie
->indirect_info
->param_index
)->virt_call
= 1;
1239 /* Return the result of a (possibly arithmetic) pass through jump function
1240 JFUNC on the constant value INPUT. RES_TYPE is the type of the parameter
1241 to which the result is passed. Return NULL_TREE if that cannot be
1242 determined or be considered an interprocedural invariant. */
1245 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
,
1250 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1252 if (!is_gimple_ip_invariant (input
))
1255 tree_code opcode
= ipa_get_jf_pass_through_operation (jfunc
);
1258 if (TREE_CODE_CLASS (opcode
) == tcc_comparison
)
1259 res_type
= boolean_type_node
;
1260 else if (expr_type_first_operand_type_p (opcode
))
1261 res_type
= TREE_TYPE (input
);
1266 if (TREE_CODE_CLASS (opcode
) == tcc_unary
)
1267 res
= fold_unary (opcode
, res_type
, input
);
1269 res
= fold_binary (opcode
, res_type
, input
,
1270 ipa_get_jf_pass_through_operand (jfunc
));
1272 if (res
&& !is_gimple_ip_invariant (res
))
1278 /* Return the result of an ancestor jump function JFUNC on the constant value
1279 INPUT. Return NULL_TREE if that cannot be determined. */
1282 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
1284 gcc_checking_assert (TREE_CODE (input
) != TREE_BINFO
);
1285 if (TREE_CODE (input
) == ADDR_EXPR
)
1287 tree t
= TREE_OPERAND (input
, 0);
1288 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
1289 ipa_get_jf_ancestor_offset (jfunc
), false,
1290 ptr_type_node
, NULL
, false);
1291 return build_fold_addr_expr (t
);
1297 /* Determine whether JFUNC evaluates to a single known constant value and if
1298 so, return it. Otherwise return NULL. INFO describes the caller node or
1299 the one it is inlined to, so that pass-through jump functions can be
1300 evaluated. PARM_TYPE is the type of the parameter to which the result is
1304 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
,
1307 if (jfunc
->type
== IPA_JF_CONST
)
1308 return ipa_get_jf_constant (jfunc
);
1309 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1310 || jfunc
->type
== IPA_JF_ANCESTOR
)
1315 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1316 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1318 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1320 if (info
->ipcp_orig_node
)
1321 input
= info
->known_csts
[idx
];
1324 ipcp_lattice
<tree
> *lat
;
1327 || idx
>= ipa_get_param_count (info
))
1329 lat
= ipa_get_scalar_lat (info
, idx
);
1330 if (!lat
->is_single_const ())
1332 input
= lat
->values
->value
;
1338 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1339 return ipa_get_jf_pass_through_result (jfunc
, input
, parm_type
);
1341 return ipa_get_jf_ancestor_result (jfunc
, input
);
1347 /* Determie whether JFUNC evaluates to single known polymorphic context, given
1348 that INFO describes the caller node or the one it is inlined to, CS is the
1349 call graph edge corresponding to JFUNC and CSIDX index of the described
1352 ipa_polymorphic_call_context
1353 ipa_context_from_jfunc (ipa_node_params
*info
, cgraph_edge
*cs
, int csidx
,
1354 ipa_jump_func
*jfunc
)
1356 ipa_edge_args
*args
= IPA_EDGE_REF (cs
);
1357 ipa_polymorphic_call_context ctx
;
1358 ipa_polymorphic_call_context
*edge_ctx
1359 = cs
? ipa_get_ith_polymorhic_call_context (args
, csidx
) : NULL
;
1361 if (edge_ctx
&& !edge_ctx
->useless_p ())
1364 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1365 || jfunc
->type
== IPA_JF_ANCESTOR
)
1367 ipa_polymorphic_call_context srcctx
;
1369 bool type_preserved
= true;
1370 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1372 if (ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1374 type_preserved
= ipa_get_jf_pass_through_type_preserved (jfunc
);
1375 srcidx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1379 type_preserved
= ipa_get_jf_ancestor_type_preserved (jfunc
);
1380 srcidx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1382 if (info
->ipcp_orig_node
)
1384 if (info
->known_contexts
.exists ())
1385 srcctx
= info
->known_contexts
[srcidx
];
1390 || srcidx
>= ipa_get_param_count (info
))
1392 ipcp_lattice
<ipa_polymorphic_call_context
> *lat
;
1393 lat
= ipa_get_poly_ctx_lat (info
, srcidx
);
1394 if (!lat
->is_single_const ())
1396 srcctx
= lat
->values
->value
;
1398 if (srcctx
.useless_p ())
1400 if (jfunc
->type
== IPA_JF_ANCESTOR
)
1401 srcctx
.offset_by (ipa_get_jf_ancestor_offset (jfunc
));
1402 if (!type_preserved
)
1403 srcctx
.possible_dynamic_type_change (cs
->in_polymorphic_cdtor
);
1404 srcctx
.combine_with (ctx
);
1411 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
1412 bottom, not containing a variable component and without any known value at
1416 ipcp_verify_propagated_values (void)
1418 struct cgraph_node
*node
;
1420 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
1422 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1423 int i
, count
= ipa_get_param_count (info
);
1425 for (i
= 0; i
< count
; i
++)
1427 ipcp_lattice
<tree
> *lat
= ipa_get_scalar_lat (info
, i
);
1430 && !lat
->contains_variable
1431 && lat
->values_count
== 0)
1435 symtab
->dump (dump_file
);
1436 fprintf (dump_file
, "\nIPA lattices after constant "
1437 "propagation, before gcc_unreachable:\n");
1438 print_all_lattices (dump_file
, true, false);
1447 /* Return true iff X and Y should be considered equal values by IPA-CP. */
1450 values_equal_for_ipcp_p (tree x
, tree y
)
1452 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
1457 if (TREE_CODE (x
) == ADDR_EXPR
1458 && TREE_CODE (y
) == ADDR_EXPR
1459 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
1460 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
1461 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
1462 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
1464 return operand_equal_p (x
, y
, 0);
1467 /* Return true iff X and Y should be considered equal contexts by IPA-CP. */
1470 values_equal_for_ipcp_p (ipa_polymorphic_call_context x
,
1471 ipa_polymorphic_call_context y
)
1473 return x
.equal_to (y
);
1477 /* Add a new value source to the value represented by THIS, marking that a
1478 value comes from edge CS and (if the underlying jump function is a
1479 pass-through or an ancestor one) from a caller value SRC_VAL of a caller
1480 parameter described by SRC_INDEX. OFFSET is negative if the source was the
1481 scalar value of the parameter itself or the offset within an aggregate. */
1483 template <typename valtype
>
1485 ipcp_value
<valtype
>::add_source (cgraph_edge
*cs
, ipcp_value
*src_val
,
1486 int src_idx
, HOST_WIDE_INT offset
)
1488 ipcp_value_source
<valtype
> *src
;
1490 src
= new (ipcp_sources_pool
.allocate ()) ipcp_value_source
<valtype
>;
1491 src
->offset
= offset
;
1494 src
->index
= src_idx
;
1496 src
->next
= sources
;
1500 /* Allocate a new ipcp_value holding a tree constant, initialize its value to
1501 SOURCE and clear all other fields. */
1503 static ipcp_value
<tree
> *
1504 allocate_and_init_ipcp_value (tree source
)
1506 ipcp_value
<tree
> *val
;
1508 val
= new (ipcp_cst_values_pool
.allocate ()) ipcp_value
<tree
>();
1509 val
->value
= source
;
1513 /* Allocate a new ipcp_value holding a polymorphic context, initialize its
1514 value to SOURCE and clear all other fields. */
1516 static ipcp_value
<ipa_polymorphic_call_context
> *
1517 allocate_and_init_ipcp_value (ipa_polymorphic_call_context source
)
1519 ipcp_value
<ipa_polymorphic_call_context
> *val
;
1522 val
= new (ipcp_poly_ctx_values_pool
.allocate ())
1523 ipcp_value
<ipa_polymorphic_call_context
>();
1524 val
->value
= source
;
1528 /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS,
1529 SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same
1530 meaning. OFFSET -1 means the source is scalar and not a part of an
1533 template <typename valtype
>
1535 ipcp_lattice
<valtype
>::add_value (valtype newval
, cgraph_edge
*cs
,
1536 ipcp_value
<valtype
> *src_val
,
1537 int src_idx
, HOST_WIDE_INT offset
)
1539 ipcp_value
<valtype
> *val
;
1544 for (val
= values
; val
; val
= val
->next
)
1545 if (values_equal_for_ipcp_p (val
->value
, newval
))
1547 if (ipa_edge_within_scc (cs
))
1549 ipcp_value_source
<valtype
> *s
;
1550 for (s
= val
->sources
; s
; s
= s
->next
)
1557 val
->add_source (cs
, src_val
, src_idx
, offset
);
1561 if (values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
1563 /* We can only free sources, not the values themselves, because sources
1564 of other values in this SCC might point to them. */
1565 for (val
= values
; val
; val
= val
->next
)
1567 while (val
->sources
)
1569 ipcp_value_source
<valtype
> *src
= val
->sources
;
1570 val
->sources
= src
->next
;
1571 ipcp_sources_pool
.remove ((ipcp_value_source
<tree
>*)src
);
1576 return set_to_bottom ();
1580 val
= allocate_and_init_ipcp_value (newval
);
1581 val
->add_source (cs
, src_val
, src_idx
, offset
);
1587 /* Propagate values through a pass-through jump function JFUNC associated with
1588 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1589 is the index of the source parameter. PARM_TYPE is the type of the
1590 parameter to which the result is passed. */
1593 propagate_vals_across_pass_through (cgraph_edge
*cs
, ipa_jump_func
*jfunc
,
1594 ipcp_lattice
<tree
> *src_lat
,
1595 ipcp_lattice
<tree
> *dest_lat
, int src_idx
,
1598 ipcp_value
<tree
> *src_val
;
1601 /* Do not create new values when propagating within an SCC because if there
1602 are arithmetic functions with circular dependencies, there is infinite
1603 number of them and we would just make lattices bottom. If this condition
1604 is ever relaxed we have to detect self-feeding recursive calls in
1605 cgraph_edge_brings_value_p in a smarter way. */
1606 if ((ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1607 && ipa_edge_within_scc (cs
))
1608 ret
= dest_lat
->set_contains_variable ();
1610 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1612 tree cstval
= ipa_get_jf_pass_through_result (jfunc
, src_val
->value
,
1616 ret
|= dest_lat
->add_value (cstval
, cs
, src_val
, src_idx
);
1618 ret
|= dest_lat
->set_contains_variable ();
1624 /* Propagate values through an ancestor jump function JFUNC associated with
1625 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1626 is the index of the source parameter. */
1629 propagate_vals_across_ancestor (struct cgraph_edge
*cs
,
1630 struct ipa_jump_func
*jfunc
,
1631 ipcp_lattice
<tree
> *src_lat
,
1632 ipcp_lattice
<tree
> *dest_lat
, int src_idx
)
1634 ipcp_value
<tree
> *src_val
;
1637 if (ipa_edge_within_scc (cs
))
1638 return dest_lat
->set_contains_variable ();
1640 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1642 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1645 ret
|= dest_lat
->add_value (t
, cs
, src_val
, src_idx
);
1647 ret
|= dest_lat
->set_contains_variable ();
1653 /* Propagate scalar values across jump function JFUNC that is associated with
1654 edge CS and put the values into DEST_LAT. PARM_TYPE is the type of the
1655 parameter to which the result is passed. */
1658 propagate_scalar_across_jump_function (struct cgraph_edge
*cs
,
1659 struct ipa_jump_func
*jfunc
,
1660 ipcp_lattice
<tree
> *dest_lat
,
1663 if (dest_lat
->bottom
)
1666 if (jfunc
->type
== IPA_JF_CONST
)
1668 tree val
= ipa_get_jf_constant (jfunc
);
1669 return dest_lat
->add_value (val
, cs
, NULL
, 0);
1671 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1672 || jfunc
->type
== IPA_JF_ANCESTOR
)
1674 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1675 ipcp_lattice
<tree
> *src_lat
;
1679 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1680 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1682 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1684 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1685 if (src_lat
->bottom
)
1686 return dest_lat
->set_contains_variable ();
1688 /* If we would need to clone the caller and cannot, do not propagate. */
1689 if (!ipcp_versionable_function_p (cs
->caller
)
1690 && (src_lat
->contains_variable
1691 || (src_lat
->values_count
> 1)))
1692 return dest_lat
->set_contains_variable ();
1694 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1695 ret
= propagate_vals_across_pass_through (cs
, jfunc
, src_lat
,
1696 dest_lat
, src_idx
, param_type
);
1698 ret
= propagate_vals_across_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1701 if (src_lat
->contains_variable
)
1702 ret
|= dest_lat
->set_contains_variable ();
1707 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1708 use it for indirect inlining), we should propagate them too. */
1709 return dest_lat
->set_contains_variable ();
1712 /* Propagate scalar values across jump function JFUNC that is associated with
1713 edge CS and describes argument IDX and put the values into DEST_LAT. */
1716 propagate_context_across_jump_function (cgraph_edge
*cs
,
1717 ipa_jump_func
*jfunc
, int idx
,
1718 ipcp_lattice
<ipa_polymorphic_call_context
> *dest_lat
)
1720 ipa_edge_args
*args
= IPA_EDGE_REF (cs
);
1721 if (dest_lat
->bottom
)
1724 bool added_sth
= false;
1725 bool type_preserved
= true;
1727 ipa_polymorphic_call_context edge_ctx
, *edge_ctx_ptr
1728 = ipa_get_ith_polymorhic_call_context (args
, idx
);
1731 edge_ctx
= *edge_ctx_ptr
;
1733 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1734 || jfunc
->type
== IPA_JF_ANCESTOR
)
1736 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1738 ipcp_lattice
<ipa_polymorphic_call_context
> *src_lat
;
1740 /* TODO: Once we figure out how to propagate speculations, it will
1741 probably be a good idea to switch to speculation if type_preserved is
1742 not set instead of punting. */
1743 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1745 if (ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1747 type_preserved
= ipa_get_jf_pass_through_type_preserved (jfunc
);
1748 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1752 type_preserved
= ipa_get_jf_ancestor_type_preserved (jfunc
);
1753 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1756 src_lat
= ipa_get_poly_ctx_lat (caller_info
, src_idx
);
1757 /* If we would need to clone the caller and cannot, do not propagate. */
1758 if (!ipcp_versionable_function_p (cs
->caller
)
1759 && (src_lat
->contains_variable
1760 || (src_lat
->values_count
> 1)))
1763 ipcp_value
<ipa_polymorphic_call_context
> *src_val
;
1764 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1766 ipa_polymorphic_call_context cur
= src_val
->value
;
1768 if (!type_preserved
)
1769 cur
.possible_dynamic_type_change (cs
->in_polymorphic_cdtor
);
1770 if (jfunc
->type
== IPA_JF_ANCESTOR
)
1771 cur
.offset_by (ipa_get_jf_ancestor_offset (jfunc
));
1772 /* TODO: In cases we know how the context is going to be used,
1773 we can improve the result by passing proper OTR_TYPE. */
1774 cur
.combine_with (edge_ctx
);
1775 if (!cur
.useless_p ())
1777 if (src_lat
->contains_variable
1778 && !edge_ctx
.equal_to (cur
))
1779 ret
|= dest_lat
->set_contains_variable ();
1780 ret
|= dest_lat
->add_value (cur
, cs
, src_val
, src_idx
);
1790 if (!edge_ctx
.useless_p ())
1791 ret
|= dest_lat
->add_value (edge_ctx
, cs
);
1793 ret
|= dest_lat
->set_contains_variable ();
1799 /* Propagate bits across jfunc that is associated with
1800 edge cs and update dest_lattice accordingly. */
1803 propagate_bits_across_jump_function (cgraph_edge
*cs
, int idx
,
1804 ipa_jump_func
*jfunc
,
1805 ipcp_bits_lattice
*dest_lattice
)
1807 if (dest_lattice
->bottom_p ())
1810 enum availability availability
;
1811 cgraph_node
*callee
= cs
->callee
->function_symbol (&availability
);
1812 struct ipa_node_params
*callee_info
= IPA_NODE_REF (callee
);
1813 tree parm_type
= ipa_get_type (callee_info
, idx
);
1815 /* For K&R C programs, ipa_get_type() could return NULL_TREE. Avoid the
1816 transform for these cases. Similarly, we can have bad type mismatches
1817 with LTO, avoid doing anything with those too. */
1819 || (!INTEGRAL_TYPE_P (parm_type
) && !POINTER_TYPE_P (parm_type
)))
1821 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1822 fprintf (dump_file
, "Setting dest_lattice to bottom, because type of "
1823 "param %i of %s is NULL or unsuitable for bits propagation\n",
1824 idx
, cs
->callee
->name ());
1826 return dest_lattice
->set_to_bottom ();
1829 unsigned precision
= TYPE_PRECISION (parm_type
);
1830 signop sgn
= TYPE_SIGN (parm_type
);
1832 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1833 || jfunc
->type
== IPA_JF_ANCESTOR
)
1835 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1836 tree operand
= NULL_TREE
;
1837 enum tree_code code
;
1840 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1842 code
= ipa_get_jf_pass_through_operation (jfunc
);
1843 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1844 if (code
!= NOP_EXPR
)
1845 operand
= ipa_get_jf_pass_through_operand (jfunc
);
1849 code
= POINTER_PLUS_EXPR
;
1850 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1851 unsigned HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
) / BITS_PER_UNIT
;
1852 operand
= build_int_cstu (size_type_node
, offset
);
1855 struct ipcp_param_lattices
*src_lats
1856 = ipa_get_parm_lattices (caller_info
, src_idx
);
1858 /* Try to propagate bits if src_lattice is bottom, but jfunc is known.
1864 Assume lattice for x is bottom, however we can still propagate
1865 result of x & 0xff == 0xff, which gets computed during ccp1 pass
1866 and we store it in jump function during analysis stage. */
1868 if (src_lats
->bits_lattice
.bottom_p ()
1870 return dest_lattice
->meet_with (jfunc
->bits
->value
, jfunc
->bits
->mask
,
1873 return dest_lattice
->meet_with (src_lats
->bits_lattice
, precision
, sgn
,
1877 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
1878 return dest_lattice
->set_to_bottom ();
1879 else if (jfunc
->bits
)
1880 return dest_lattice
->meet_with (jfunc
->bits
->value
, jfunc
->bits
->mask
,
1883 return dest_lattice
->set_to_bottom ();
1886 /* Emulate effects of unary OPERATION and/or conversion from SRC_TYPE to
1887 DST_TYPE on value range in SRC_VR and store it to DST_VR. Return true if
1888 the result is a range or an anti-range. */
1891 ipa_vr_operation_and_type_effects (value_range
*dst_vr
, value_range
*src_vr
,
1892 enum tree_code operation
,
1893 tree dst_type
, tree src_type
)
1895 memset (dst_vr
, 0, sizeof (*dst_vr
));
1896 extract_range_from_unary_expr (dst_vr
, operation
, dst_type
, src_vr
, src_type
);
1897 if (dst_vr
->type
== VR_RANGE
|| dst_vr
->type
== VR_ANTI_RANGE
)
1903 /* Propagate value range across jump function JFUNC that is associated with
1904 edge CS with param of callee of PARAM_TYPE and update DEST_PLATS
1908 propagate_vr_across_jump_function (cgraph_edge
*cs
, ipa_jump_func
*jfunc
,
1909 struct ipcp_param_lattices
*dest_plats
,
1912 ipcp_vr_lattice
*dest_lat
= &dest_plats
->m_value_range
;
1914 if (dest_lat
->bottom_p ())
1918 || (!INTEGRAL_TYPE_P (param_type
)
1919 && !POINTER_TYPE_P (param_type
)))
1920 return dest_lat
->set_to_bottom ();
1922 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1924 enum tree_code operation
= ipa_get_jf_pass_through_operation (jfunc
);
1926 if (TREE_CODE_CLASS (operation
) == tcc_unary
)
1928 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1929 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1930 tree operand_type
= ipa_get_type (caller_info
, src_idx
);
1931 struct ipcp_param_lattices
*src_lats
1932 = ipa_get_parm_lattices (caller_info
, src_idx
);
1934 if (src_lats
->m_value_range
.bottom_p ())
1935 return dest_lat
->set_to_bottom ();
1937 if (ipa_vr_operation_and_type_effects (&vr
,
1938 &src_lats
->m_value_range
.m_vr
,
1939 operation
, param_type
,
1941 return dest_lat
->meet_with (&vr
);
1944 else if (jfunc
->type
== IPA_JF_CONST
)
1946 tree val
= ipa_get_jf_constant (jfunc
);
1947 if (TREE_CODE (val
) == INTEGER_CST
)
1949 val
= fold_convert (param_type
, val
);
1950 if (TREE_OVERFLOW_P (val
))
1951 val
= drop_tree_overflow (val
);
1954 memset (&tmpvr
, 0, sizeof (tmpvr
));
1955 tmpvr
.type
= VR_RANGE
;
1958 return dest_lat
->meet_with (&tmpvr
);
1964 && ipa_vr_operation_and_type_effects (&vr
, jfunc
->m_vr
, NOP_EXPR
,
1966 TREE_TYPE (jfunc
->m_vr
->min
)))
1967 return dest_lat
->meet_with (&vr
);
1969 return dest_lat
->set_to_bottom ();
1972 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1973 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1974 other cases, return false). If there are no aggregate items, set
1975 aggs_by_ref to NEW_AGGS_BY_REF. */
1978 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1979 bool new_aggs_by_ref
)
1981 if (dest_plats
->aggs
)
1983 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1985 set_agg_lats_to_bottom (dest_plats
);
1990 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1994 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1995 already existing lattice for the given OFFSET and SIZE, marking all skipped
1996 lattices as containing variable and checking for overlaps. If there is no
1997 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1998 it with offset, size and contains_variable to PRE_EXISTING, and return true,
1999 unless there are too many already. If there are two many, return false. If
2000 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
2001 skipped lattices were newly marked as containing variable, set *CHANGE to
2005 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
2006 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
2007 struct ipcp_agg_lattice
***aglat
,
2008 bool pre_existing
, bool *change
)
2010 gcc_checking_assert (offset
>= 0);
2012 while (**aglat
&& (**aglat
)->offset
< offset
)
2014 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
2016 set_agg_lats_to_bottom (dest_plats
);
2019 *change
|= (**aglat
)->set_contains_variable ();
2020 *aglat
= &(**aglat
)->next
;
2023 if (**aglat
&& (**aglat
)->offset
== offset
)
2025 if ((**aglat
)->size
!= val_size
2027 && (**aglat
)->next
->offset
< offset
+ val_size
))
2029 set_agg_lats_to_bottom (dest_plats
);
2032 gcc_checking_assert (!(**aglat
)->next
2033 || (**aglat
)->next
->offset
>= offset
+ val_size
);
2038 struct ipcp_agg_lattice
*new_al
;
2040 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
2042 set_agg_lats_to_bottom (dest_plats
);
2045 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
2047 dest_plats
->aggs_count
++;
2048 new_al
= ipcp_agg_lattice_pool
.allocate ();
2049 memset (new_al
, 0, sizeof (*new_al
));
2051 new_al
->offset
= offset
;
2052 new_al
->size
= val_size
;
2053 new_al
->contains_variable
= pre_existing
;
2055 new_al
->next
= **aglat
;
2061 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
2062 containing an unknown value. */
2065 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
2070 ret
|= aglat
->set_contains_variable ();
2071 aglat
= aglat
->next
;
2076 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
2077 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
2078 parameter used for lattice value sources. Return true if DEST_PLATS changed
2082 merge_aggregate_lattices (struct cgraph_edge
*cs
,
2083 struct ipcp_param_lattices
*dest_plats
,
2084 struct ipcp_param_lattices
*src_plats
,
2085 int src_idx
, HOST_WIDE_INT offset_delta
)
2087 bool pre_existing
= dest_plats
->aggs
!= NULL
;
2088 struct ipcp_agg_lattice
**dst_aglat
;
2091 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
2093 if (src_plats
->aggs_bottom
)
2094 return set_agg_lats_contain_variable (dest_plats
);
2095 if (src_plats
->aggs_contain_variable
)
2096 ret
|= set_agg_lats_contain_variable (dest_plats
);
2097 dst_aglat
= &dest_plats
->aggs
;
2099 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
2101 src_aglat
= src_aglat
->next
)
2103 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
2107 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
2108 &dst_aglat
, pre_existing
, &ret
))
2110 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
2112 dst_aglat
= &(*dst_aglat
)->next
;
2113 if (src_aglat
->bottom
)
2115 ret
|= new_al
->set_contains_variable ();
2118 if (src_aglat
->contains_variable
)
2119 ret
|= new_al
->set_contains_variable ();
2120 for (ipcp_value
<tree
> *val
= src_aglat
->values
;
2123 ret
|= new_al
->add_value (val
->value
, cs
, val
, src_idx
,
2126 else if (dest_plats
->aggs_bottom
)
2129 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
2133 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
2134 pass-through JFUNC and if so, whether it has conform and conforms to the
2135 rules about propagating values passed by reference. */
2138 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
2139 struct ipa_jump_func
*jfunc
)
2141 return src_plats
->aggs
2142 && (!src_plats
->aggs_by_ref
2143 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
2146 /* Propagate scalar values across jump function JFUNC that is associated with
2147 edge CS and put the values into DEST_LAT. */
2150 propagate_aggs_across_jump_function (struct cgraph_edge
*cs
,
2151 struct ipa_jump_func
*jfunc
,
2152 struct ipcp_param_lattices
*dest_plats
)
2156 if (dest_plats
->aggs_bottom
)
2159 if (jfunc
->type
== IPA_JF_PASS_THROUGH
2160 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
2162 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2163 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
2164 struct ipcp_param_lattices
*src_plats
;
2166 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2167 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
2169 /* Currently we do not produce clobber aggregate jump
2170 functions, replace with merging when we do. */
2171 gcc_assert (!jfunc
->agg
.items
);
2172 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
2176 ret
|= set_agg_lats_contain_variable (dest_plats
);
2178 else if (jfunc
->type
== IPA_JF_ANCESTOR
2179 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
2181 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2182 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
2183 struct ipcp_param_lattices
*src_plats
;
2185 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2186 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
2188 /* Currently we do not produce clobber aggregate jump
2189 functions, replace with merging when we do. */
2190 gcc_assert (!jfunc
->agg
.items
);
2191 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
2192 ipa_get_jf_ancestor_offset (jfunc
));
2194 else if (!src_plats
->aggs_by_ref
)
2195 ret
|= set_agg_lats_to_bottom (dest_plats
);
2197 ret
|= set_agg_lats_contain_variable (dest_plats
);
2199 else if (jfunc
->agg
.items
)
2201 bool pre_existing
= dest_plats
->aggs
!= NULL
;
2202 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
2203 struct ipa_agg_jf_item
*item
;
2206 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
2209 FOR_EACH_VEC_ELT (*jfunc
->agg
.items
, i
, item
)
2211 HOST_WIDE_INT val_size
;
2213 if (item
->offset
< 0)
2215 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
2216 val_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item
->value
)));
2218 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
2219 &aglat
, pre_existing
, &ret
))
2221 ret
|= (*aglat
)->add_value (item
->value
, cs
, NULL
, 0, 0);
2222 aglat
= &(*aglat
)->next
;
2224 else if (dest_plats
->aggs_bottom
)
2228 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
2231 ret
|= set_agg_lats_contain_variable (dest_plats
);
2236 /* Return true if on the way cfrom CS->caller to the final (non-alias and
2237 non-thunk) destination, the call passes through a thunk. */
2240 call_passes_through_thunk_p (cgraph_edge
*cs
)
2242 cgraph_node
*alias_or_thunk
= cs
->callee
;
2243 while (alias_or_thunk
->alias
)
2244 alias_or_thunk
= alias_or_thunk
->get_alias_target ();
2245 return alias_or_thunk
->thunk
.thunk_p
;
2248 /* Propagate constants from the caller to the callee of CS. INFO describes the
2252 propagate_constants_across_call (struct cgraph_edge
*cs
)
2254 struct ipa_node_params
*callee_info
;
2255 enum availability availability
;
2256 cgraph_node
*callee
;
2257 struct ipa_edge_args
*args
;
2259 int i
, args_count
, parms_count
;
2261 callee
= cs
->callee
->function_symbol (&availability
);
2262 if (!callee
->definition
)
2264 gcc_checking_assert (callee
->has_gimple_body_p ());
2265 callee_info
= IPA_NODE_REF (callee
);
2267 args
= IPA_EDGE_REF (cs
);
2268 args_count
= ipa_get_cs_argument_count (args
);
2269 parms_count
= ipa_get_param_count (callee_info
);
2270 if (parms_count
== 0)
2273 /* If this call goes through a thunk we must not propagate to the first (0th)
2274 parameter. However, we might need to uncover a thunk from below a series
2275 of aliases first. */
2276 if (call_passes_through_thunk_p (cs
))
2278 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
2285 for (; (i
< args_count
) && (i
< parms_count
); i
++)
2287 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
2288 struct ipcp_param_lattices
*dest_plats
;
2289 tree param_type
= ipa_get_type (callee_info
, i
);
2291 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
2292 if (availability
== AVAIL_INTERPOSABLE
)
2293 ret
|= set_all_contains_variable (dest_plats
);
2296 ret
|= propagate_scalar_across_jump_function (cs
, jump_func
,
2297 &dest_plats
->itself
,
2299 ret
|= propagate_context_across_jump_function (cs
, jump_func
, i
,
2300 &dest_plats
->ctxlat
);
2302 |= propagate_bits_across_jump_function (cs
, i
, jump_func
,
2303 &dest_plats
->bits_lattice
);
2304 ret
|= propagate_aggs_across_jump_function (cs
, jump_func
,
2306 if (opt_for_fn (callee
->decl
, flag_ipa_vrp
))
2307 ret
|= propagate_vr_across_jump_function (cs
, jump_func
,
2308 dest_plats
, param_type
);
2310 ret
|= dest_plats
->m_value_range
.set_to_bottom ();
2313 for (; i
< parms_count
; i
++)
2314 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
2319 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
2320 KNOWN_CONTEXTS, KNOWN_AGGS or AGG_REPS return the destination. The latter
2321 three can be NULL. If AGG_REPS is not NULL, KNOWN_AGGS is ignored. */
2324 ipa_get_indirect_edge_target_1 (struct cgraph_edge
*ie
,
2325 vec
<tree
> known_csts
,
2326 vec
<ipa_polymorphic_call_context
> known_contexts
,
2327 vec
<ipa_agg_jump_function_p
> known_aggs
,
2328 struct ipa_agg_replacement_value
*agg_reps
,
2331 int param_index
= ie
->indirect_info
->param_index
;
2332 HOST_WIDE_INT anc_offset
;
2336 *speculative
= false;
2338 if (param_index
== -1
2339 || known_csts
.length () <= (unsigned int) param_index
)
2342 if (!ie
->indirect_info
->polymorphic
)
2346 if (ie
->indirect_info
->agg_contents
)
2349 if (agg_reps
&& ie
->indirect_info
->guaranteed_unmodified
)
2353 if (agg_reps
->index
== param_index
2354 && agg_reps
->offset
== ie
->indirect_info
->offset
2355 && agg_reps
->by_ref
== ie
->indirect_info
->by_ref
)
2357 t
= agg_reps
->value
;
2360 agg_reps
= agg_reps
->next
;
2365 struct ipa_agg_jump_function
*agg
;
2366 if (known_aggs
.length () > (unsigned int) param_index
)
2367 agg
= known_aggs
[param_index
];
2370 bool from_global_constant
;
2371 t
= ipa_find_agg_cst_for_param (agg
, known_csts
[param_index
],
2372 ie
->indirect_info
->offset
,
2373 ie
->indirect_info
->by_ref
,
2374 &from_global_constant
);
2376 && !from_global_constant
2377 && !ie
->indirect_info
->guaranteed_unmodified
)
2382 t
= known_csts
[param_index
];
2385 && TREE_CODE (t
) == ADDR_EXPR
2386 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
2387 return TREE_OPERAND (t
, 0);
2392 if (!opt_for_fn (ie
->caller
->decl
, flag_devirtualize
))
2395 gcc_assert (!ie
->indirect_info
->agg_contents
);
2396 anc_offset
= ie
->indirect_info
->offset
;
2400 /* Try to work out value of virtual table pointer value in replacemnets. */
2401 if (!t
&& agg_reps
&& !ie
->indirect_info
->by_ref
)
2405 if (agg_reps
->index
== param_index
2406 && agg_reps
->offset
== ie
->indirect_info
->offset
2407 && agg_reps
->by_ref
)
2409 t
= agg_reps
->value
;
2412 agg_reps
= agg_reps
->next
;
2416 /* Try to work out value of virtual table pointer value in known
2417 aggregate values. */
2418 if (!t
&& known_aggs
.length () > (unsigned int) param_index
2419 && !ie
->indirect_info
->by_ref
)
2421 struct ipa_agg_jump_function
*agg
;
2422 agg
= known_aggs
[param_index
];
2423 t
= ipa_find_agg_cst_for_param (agg
, known_csts
[param_index
],
2424 ie
->indirect_info
->offset
, true);
2427 /* If we found the virtual table pointer, lookup the target. */
2431 unsigned HOST_WIDE_INT offset
;
2432 if (vtable_pointer_value_to_vtable (t
, &vtable
, &offset
))
2435 target
= gimple_get_virt_method_for_vtable (ie
->indirect_info
->otr_token
,
2436 vtable
, offset
, &can_refer
);
2440 || (TREE_CODE (TREE_TYPE (target
)) == FUNCTION_TYPE
2441 && DECL_FUNCTION_CODE (target
) == BUILT_IN_UNREACHABLE
)
2442 || !possible_polymorphic_call_target_p
2443 (ie
, cgraph_node::get (target
)))
2445 /* Do not speculate builtin_unreachable, it is stupid! */
2446 if (ie
->indirect_info
->vptr_changed
)
2448 target
= ipa_impossible_devirt_target (ie
, target
);
2450 *speculative
= ie
->indirect_info
->vptr_changed
;
2457 /* Do we know the constant value of pointer? */
2459 t
= known_csts
[param_index
];
2461 gcc_checking_assert (!t
|| TREE_CODE (t
) != TREE_BINFO
);
2463 ipa_polymorphic_call_context context
;
2464 if (known_contexts
.length () > (unsigned int) param_index
)
2466 context
= known_contexts
[param_index
];
2467 context
.offset_by (anc_offset
);
2468 if (ie
->indirect_info
->vptr_changed
)
2469 context
.possible_dynamic_type_change (ie
->in_polymorphic_cdtor
,
2470 ie
->indirect_info
->otr_type
);
2473 ipa_polymorphic_call_context ctx2
= ipa_polymorphic_call_context
2474 (t
, ie
->indirect_info
->otr_type
, anc_offset
);
2475 if (!ctx2
.useless_p ())
2476 context
.combine_with (ctx2
, ie
->indirect_info
->otr_type
);
2481 context
= ipa_polymorphic_call_context (t
, ie
->indirect_info
->otr_type
,
2483 if (ie
->indirect_info
->vptr_changed
)
2484 context
.possible_dynamic_type_change (ie
->in_polymorphic_cdtor
,
2485 ie
->indirect_info
->otr_type
);
2490 vec
<cgraph_node
*>targets
;
2493 targets
= possible_polymorphic_call_targets
2494 (ie
->indirect_info
->otr_type
,
2495 ie
->indirect_info
->otr_token
,
2497 if (!final
|| targets
.length () > 1)
2499 struct cgraph_node
*node
;
2502 if (!opt_for_fn (ie
->caller
->decl
, flag_devirtualize_speculatively
)
2503 || ie
->speculative
|| !ie
->maybe_hot_p ())
2505 node
= try_speculative_devirtualization (ie
->indirect_info
->otr_type
,
2506 ie
->indirect_info
->otr_token
,
2510 *speculative
= true;
2511 target
= node
->decl
;
2518 *speculative
= false;
2519 if (targets
.length () == 1)
2520 target
= targets
[0]->decl
;
2522 target
= ipa_impossible_devirt_target (ie
, NULL_TREE
);
2525 if (target
&& !possible_polymorphic_call_target_p (ie
,
2526 cgraph_node::get (target
)))
2530 target
= ipa_impossible_devirt_target (ie
, target
);
2537 /* If an indirect edge IE can be turned into a direct one based on KNOWN_CSTS,
2538 KNOWN_CONTEXTS (which can be vNULL) or KNOWN_AGGS (which also can be vNULL)
2539 return the destination. */
2542 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
2543 vec
<tree
> known_csts
,
2544 vec
<ipa_polymorphic_call_context
> known_contexts
,
2545 vec
<ipa_agg_jump_function_p
> known_aggs
,
2548 return ipa_get_indirect_edge_target_1 (ie
, known_csts
, known_contexts
,
2549 known_aggs
, NULL
, speculative
);
2552 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
2553 and KNOWN_CONTEXTS. */
2556 devirtualization_time_bonus (struct cgraph_node
*node
,
2557 vec
<tree
> known_csts
,
2558 vec
<ipa_polymorphic_call_context
> known_contexts
,
2559 vec
<ipa_agg_jump_function_p
> known_aggs
)
2561 struct cgraph_edge
*ie
;
2564 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
2566 struct cgraph_node
*callee
;
2567 struct ipa_fn_summary
*isummary
;
2568 enum availability avail
;
2572 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_contexts
,
2573 known_aggs
, &speculative
);
2577 /* Only bare minimum benefit for clearly un-inlineable targets. */
2579 callee
= cgraph_node::get (target
);
2580 if (!callee
|| !callee
->definition
)
2582 callee
= callee
->function_symbol (&avail
);
2583 if (avail
< AVAIL_AVAILABLE
)
2585 isummary
= ipa_fn_summaries
->get (callee
);
2586 if (!isummary
->inlinable
)
2589 /* FIXME: The values below need re-considering and perhaps also
2590 integrating into the cost metrics, at lest in some very basic way. */
2591 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
2592 res
+= 31 / ((int)speculative
+ 1);
2593 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
2594 res
+= 15 / ((int)speculative
+ 1);
2595 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
2596 || DECL_DECLARED_INLINE_P (callee
->decl
))
2597 res
+= 7 / ((int)speculative
+ 1);
2603 /* Return time bonus incurred because of HINTS. */
2606 hint_time_bonus (ipa_hints hints
)
2609 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
2610 result
+= PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
2611 if (hints
& INLINE_HINT_array_index
)
2612 result
+= PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS
);
2616 /* If there is a reason to penalize the function described by INFO in the
2617 cloning goodness evaluation, do so. */
2619 static inline int64_t
2620 incorporate_penalties (ipa_node_params
*info
, int64_t evaluation
)
2622 if (info
->node_within_scc
)
2623 evaluation
= (evaluation
2624 * (100 - PARAM_VALUE (PARAM_IPA_CP_RECURSION_PENALTY
))) / 100;
2626 if (info
->node_calling_single_call
)
2627 evaluation
= (evaluation
2628 * (100 - PARAM_VALUE (PARAM_IPA_CP_SINGLE_CALL_PENALTY
)))
2634 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
2635 and SIZE_COST and with the sum of frequencies of incoming edges to the
2636 potential new clone in FREQUENCIES. */
2639 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
2640 int freq_sum
, profile_count count_sum
, int size_cost
)
2642 if (time_benefit
== 0
2643 || !opt_for_fn (node
->decl
, flag_ipa_cp_clone
)
2644 || node
->optimize_for_size_p ())
2647 gcc_assert (size_cost
> 0);
2649 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2650 if (max_count
> profile_count::zero ())
2652 int factor
= RDIV (count_sum
.probability_in
2653 (max_count
).to_reg_br_prob_base ()
2654 * 1000, REG_BR_PROB_BASE
);
2655 int64_t evaluation
= (((int64_t) time_benefit
* factor
)
2657 evaluation
= incorporate_penalties (info
, evaluation
);
2659 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2661 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
2662 "size: %i, count_sum: ", time_benefit
, size_cost
);
2663 count_sum
.dump (dump_file
);
2664 fprintf (dump_file
, "%s%s) -> evaluation: " "%" PRId64
2665 ", threshold: %i\n",
2666 info
->node_within_scc
? ", scc" : "",
2667 info
->node_calling_single_call
? ", single_call" : "",
2668 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
2671 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
2675 int64_t evaluation
= (((int64_t) time_benefit
* freq_sum
)
2677 evaluation
= incorporate_penalties (info
, evaluation
);
2679 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2680 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
2681 "size: %i, freq_sum: %i%s%s) -> evaluation: "
2682 "%" PRId64
", threshold: %i\n",
2683 time_benefit
, size_cost
, freq_sum
,
2684 info
->node_within_scc
? ", scc" : "",
2685 info
->node_calling_single_call
? ", single_call" : "",
2686 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
2688 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
2692 /* Return all context independent values from aggregate lattices in PLATS in a
2693 vector. Return NULL if there are none. */
2695 static vec
<ipa_agg_jf_item
, va_gc
> *
2696 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
2698 vec
<ipa_agg_jf_item
, va_gc
> *res
= NULL
;
2700 if (plats
->aggs_bottom
2701 || plats
->aggs_contain_variable
2702 || plats
->aggs_count
== 0)
2705 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
2707 aglat
= aglat
->next
)
2708 if (aglat
->is_single_const ())
2710 struct ipa_agg_jf_item item
;
2711 item
.offset
= aglat
->offset
;
2712 item
.value
= aglat
->values
->value
;
2713 vec_safe_push (res
, item
);
2718 /* Allocate KNOWN_CSTS, KNOWN_CONTEXTS and, if non-NULL, KNOWN_AGGS and
2719 populate them with values of parameters that are known independent of the
2720 context. INFO describes the function. If REMOVABLE_PARAMS_COST is
2721 non-NULL, the movement cost of all removable parameters will be stored in
2725 gather_context_independent_values (struct ipa_node_params
*info
,
2726 vec
<tree
> *known_csts
,
2727 vec
<ipa_polymorphic_call_context
>
2729 vec
<ipa_agg_jump_function
> *known_aggs
,
2730 int *removable_params_cost
)
2732 int i
, count
= ipa_get_param_count (info
);
2735 known_csts
->create (0);
2736 known_contexts
->create (0);
2737 known_csts
->safe_grow_cleared (count
);
2738 known_contexts
->safe_grow_cleared (count
);
2741 known_aggs
->create (0);
2742 known_aggs
->safe_grow_cleared (count
);
2745 if (removable_params_cost
)
2746 *removable_params_cost
= 0;
2748 for (i
= 0; i
< count
; i
++)
2750 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2751 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
2753 if (lat
->is_single_const ())
2755 ipcp_value
<tree
> *val
= lat
->values
;
2756 gcc_checking_assert (TREE_CODE (val
->value
) != TREE_BINFO
);
2757 (*known_csts
)[i
] = val
->value
;
2758 if (removable_params_cost
)
2759 *removable_params_cost
2760 += estimate_move_cost (TREE_TYPE (val
->value
), false);
2763 else if (removable_params_cost
2764 && !ipa_is_param_used (info
, i
))
2765 *removable_params_cost
2766 += ipa_get_param_move_cost (info
, i
);
2768 if (!ipa_is_param_used (info
, i
))
2771 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
2772 /* Do not account known context as reason for cloning. We can see
2773 if it permits devirtualization. */
2774 if (ctxlat
->is_single_const ())
2775 (*known_contexts
)[i
] = ctxlat
->values
->value
;
2779 vec
<ipa_agg_jf_item
, va_gc
> *agg_items
;
2780 struct ipa_agg_jump_function
*ajf
;
2782 agg_items
= context_independent_aggregate_values (plats
);
2783 ajf
= &(*known_aggs
)[i
];
2784 ajf
->items
= agg_items
;
2785 ajf
->by_ref
= plats
->aggs_by_ref
;
2786 ret
|= agg_items
!= NULL
;
2793 /* The current interface in ipa-inline-analysis requires a pointer vector.
2796 FIXME: That interface should be re-worked, this is slightly silly. Still,
2797 I'd like to discuss how to change it first and this demonstrates the
2800 static vec
<ipa_agg_jump_function_p
>
2801 agg_jmp_p_vec_for_t_vec (vec
<ipa_agg_jump_function
> known_aggs
)
2803 vec
<ipa_agg_jump_function_p
> ret
;
2804 struct ipa_agg_jump_function
*ajf
;
2807 ret
.create (known_aggs
.length ());
2808 FOR_EACH_VEC_ELT (known_aggs
, i
, ajf
)
2809 ret
.quick_push (ajf
);
2813 /* Perform time and size measurement of NODE with the context given in
2814 KNOWN_CSTS, KNOWN_CONTEXTS and KNOWN_AGGS, calculate the benefit and cost
2815 given BASE_TIME of the node without specialization, REMOVABLE_PARAMS_COST of
2816 all context-independent removable parameters and EST_MOVE_COST of estimated
2817 movement of the considered parameter and store it into VAL. */
2820 perform_estimation_of_a_value (cgraph_node
*node
, vec
<tree
> known_csts
,
2821 vec
<ipa_polymorphic_call_context
> known_contexts
,
2822 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
,
2823 int removable_params_cost
,
2824 int est_move_cost
, ipcp_value_base
*val
)
2826 int size
, time_benefit
;
2827 sreal time
, base_time
;
2830 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_contexts
,
2831 known_aggs_ptrs
, &size
, &time
,
2832 &base_time
, &hints
);
2834 if (base_time
> 65535)
2836 time_benefit
= base_time
.to_int ()
2837 + devirtualization_time_bonus (node
, known_csts
, known_contexts
,
2839 + hint_time_bonus (hints
)
2840 + removable_params_cost
+ est_move_cost
;
2842 gcc_checking_assert (size
>=0);
2843 /* The inliner-heuristics based estimates may think that in certain
2844 contexts some functions do not have any size at all but we want
2845 all specializations to have at least a tiny cost, not least not to
2850 val
->local_time_benefit
= time_benefit
;
2851 val
->local_size_cost
= size
;
2854 /* Iterate over known values of parameters of NODE and estimate the local
2855 effects in terms of time and size they have. */
2858 estimate_local_effects (struct cgraph_node
*node
)
2860 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2861 int i
, count
= ipa_get_param_count (info
);
2862 vec
<tree
> known_csts
;
2863 vec
<ipa_polymorphic_call_context
> known_contexts
;
2864 vec
<ipa_agg_jump_function
> known_aggs
;
2865 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
;
2867 int removable_params_cost
;
2869 if (!count
|| !ipcp_versionable_function_p (node
))
2872 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2873 fprintf (dump_file
, "\nEstimating effects for %s.\n", node
->dump_name ());
2875 always_const
= gather_context_independent_values (info
, &known_csts
,
2876 &known_contexts
, &known_aggs
,
2877 &removable_params_cost
);
2878 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
2879 int devirt_bonus
= devirtualization_time_bonus (node
, known_csts
,
2880 known_contexts
, known_aggs_ptrs
);
2881 if (always_const
|| devirt_bonus
2882 || (removable_params_cost
&& node
->local
.can_change_signature
))
2884 struct caller_statistics stats
;
2886 sreal time
, base_time
;
2889 init_caller_stats (&stats
);
2890 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
2892 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_contexts
,
2893 known_aggs_ptrs
, &size
, &time
,
2894 &base_time
, &hints
);
2895 time
-= devirt_bonus
;
2896 time
-= hint_time_bonus (hints
);
2897 time
-= removable_params_cost
;
2898 size
-= stats
.n_calls
* removable_params_cost
;
2901 fprintf (dump_file
, " - context independent values, size: %i, "
2902 "time_benefit: %f\n", size
, (base_time
- time
).to_double ());
2904 if (size
<= 0 || node
->local
.local
)
2906 info
->do_clone_for_all_contexts
= true;
2909 fprintf (dump_file
, " Decided to specialize for all "
2910 "known contexts, code not going to grow.\n");
2912 else if (good_cloning_opportunity_p (node
,
2913 MIN ((base_time
- time
).to_int (),
2915 stats
.freq_sum
, stats
.count_sum
,
2918 if (size
+ overall_size
<= max_new_size
)
2920 info
->do_clone_for_all_contexts
= true;
2921 overall_size
+= size
;
2924 fprintf (dump_file
, " Decided to specialize for all "
2925 "known contexts, growth deemed beneficial.\n");
2927 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2928 fprintf (dump_file
, " Not cloning for all contexts because "
2929 "max_new_size would be reached with %li.\n",
2930 size
+ overall_size
);
2932 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2933 fprintf (dump_file
, " Not cloning for all contexts because "
2934 "!good_cloning_opportunity_p.\n");
2938 for (i
= 0; i
< count
; i
++)
2940 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2941 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
2942 ipcp_value
<tree
> *val
;
2949 for (val
= lat
->values
; val
; val
= val
->next
)
2951 gcc_checking_assert (TREE_CODE (val
->value
) != TREE_BINFO
);
2952 known_csts
[i
] = val
->value
;
2954 int emc
= estimate_move_cost (TREE_TYPE (val
->value
), true);
2955 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
2957 removable_params_cost
, emc
, val
);
2959 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2961 fprintf (dump_file
, " - estimates for value ");
2962 print_ipcp_constant_value (dump_file
, val
->value
);
2963 fprintf (dump_file
, " for ");
2964 ipa_dump_param (dump_file
, info
, i
);
2965 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
2966 val
->local_time_benefit
, val
->local_size_cost
);
2969 known_csts
[i
] = NULL_TREE
;
2972 for (i
= 0; i
< count
; i
++)
2974 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2976 if (!plats
->virt_call
)
2979 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
2980 ipcp_value
<ipa_polymorphic_call_context
> *val
;
2984 || !known_contexts
[i
].useless_p ())
2987 for (val
= ctxlat
->values
; val
; val
= val
->next
)
2989 known_contexts
[i
] = val
->value
;
2990 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
2992 removable_params_cost
, 0, val
);
2994 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2996 fprintf (dump_file
, " - estimates for polymorphic context ");
2997 print_ipcp_constant_value (dump_file
, val
->value
);
2998 fprintf (dump_file
, " for ");
2999 ipa_dump_param (dump_file
, info
, i
);
3000 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
3001 val
->local_time_benefit
, val
->local_size_cost
);
3004 known_contexts
[i
] = ipa_polymorphic_call_context ();
3007 for (i
= 0; i
< count
; i
++)
3009 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3010 struct ipa_agg_jump_function
*ajf
;
3011 struct ipcp_agg_lattice
*aglat
;
3013 if (plats
->aggs_bottom
|| !plats
->aggs
)
3016 ajf
= &known_aggs
[i
];
3017 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3019 ipcp_value
<tree
> *val
;
3020 if (aglat
->bottom
|| !aglat
->values
3021 /* If the following is true, the one value is in known_aggs. */
3022 || (!plats
->aggs_contain_variable
3023 && aglat
->is_single_const ()))
3026 for (val
= aglat
->values
; val
; val
= val
->next
)
3028 struct ipa_agg_jf_item item
;
3030 item
.offset
= aglat
->offset
;
3031 item
.value
= val
->value
;
3032 vec_safe_push (ajf
->items
, item
);
3034 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
3036 removable_params_cost
, 0, val
);
3038 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3040 fprintf (dump_file
, " - estimates for value ");
3041 print_ipcp_constant_value (dump_file
, val
->value
);
3042 fprintf (dump_file
, " for ");
3043 ipa_dump_param (dump_file
, info
, i
);
3044 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
3045 "]: time_benefit: %i, size: %i\n",
3046 plats
->aggs_by_ref
? "ref " : "",
3048 val
->local_time_benefit
, val
->local_size_cost
);
3056 for (i
= 0; i
< count
; i
++)
3057 vec_free (known_aggs
[i
].items
);
3059 known_csts
.release ();
3060 known_contexts
.release ();
3061 known_aggs
.release ();
3062 known_aggs_ptrs
.release ();
3066 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
3067 topological sort of values. */
3069 template <typename valtype
>
3071 value_topo_info
<valtype
>::add_val (ipcp_value
<valtype
> *cur_val
)
3073 ipcp_value_source
<valtype
> *src
;
3079 cur_val
->dfs
= dfs_counter
;
3080 cur_val
->low_link
= dfs_counter
;
3082 cur_val
->topo_next
= stack
;
3084 cur_val
->on_stack
= true;
3086 for (src
= cur_val
->sources
; src
; src
= src
->next
)
3089 if (src
->val
->dfs
== 0)
3092 if (src
->val
->low_link
< cur_val
->low_link
)
3093 cur_val
->low_link
= src
->val
->low_link
;
3095 else if (src
->val
->on_stack
3096 && src
->val
->dfs
< cur_val
->low_link
)
3097 cur_val
->low_link
= src
->val
->dfs
;
3100 if (cur_val
->dfs
== cur_val
->low_link
)
3102 ipcp_value
<valtype
> *v
, *scc_list
= NULL
;
3107 stack
= v
->topo_next
;
3108 v
->on_stack
= false;
3110 v
->scc_next
= scc_list
;
3113 while (v
!= cur_val
);
3115 cur_val
->topo_next
= values_topo
;
3116 values_topo
= cur_val
;
3120 /* Add all values in lattices associated with NODE to the topological sort if
3121 they are not there yet. */
3124 add_all_node_vals_to_toposort (cgraph_node
*node
, ipa_topo_info
*topo
)
3126 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3127 int i
, count
= ipa_get_param_count (info
);
3129 for (i
= 0; i
< count
; i
++)
3131 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3132 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
3133 struct ipcp_agg_lattice
*aglat
;
3137 ipcp_value
<tree
> *val
;
3138 for (val
= lat
->values
; val
; val
= val
->next
)
3139 topo
->constants
.add_val (val
);
3142 if (!plats
->aggs_bottom
)
3143 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3146 ipcp_value
<tree
> *val
;
3147 for (val
= aglat
->values
; val
; val
= val
->next
)
3148 topo
->constants
.add_val (val
);
3151 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
3152 if (!ctxlat
->bottom
)
3154 ipcp_value
<ipa_polymorphic_call_context
> *ctxval
;
3155 for (ctxval
= ctxlat
->values
; ctxval
; ctxval
= ctxval
->next
)
3156 topo
->contexts
.add_val (ctxval
);
3161 /* One pass of constants propagation along the call graph edges, from callers
3162 to callees (requires topological ordering in TOPO), iterate over strongly
3163 connected components. */
3166 propagate_constants_topo (struct ipa_topo_info
*topo
)
3170 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3173 struct cgraph_node
*v
, *node
= topo
->order
[i
];
3174 vec
<cgraph_node
*> cycle_nodes
= ipa_get_nodes_in_cycle (node
);
3176 /* First, iteratively propagate within the strongly connected component
3177 until all lattices stabilize. */
3178 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
3179 if (v
->has_gimple_body_p ())
3180 push_node_to_stack (topo
, v
);
3182 v
= pop_node_from_stack (topo
);
3185 struct cgraph_edge
*cs
;
3187 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
3188 if (ipa_edge_within_scc (cs
))
3190 IPA_NODE_REF (v
)->node_within_scc
= true;
3191 if (propagate_constants_across_call (cs
))
3192 push_node_to_stack (topo
, cs
->callee
->function_symbol ());
3194 v
= pop_node_from_stack (topo
);
3197 /* Afterwards, propagate along edges leading out of the SCC, calculates
3198 the local effects of the discovered constants and all valid values to
3199 their topological sort. */
3200 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
3201 if (v
->has_gimple_body_p ())
3203 struct cgraph_edge
*cs
;
3205 estimate_local_effects (v
);
3206 add_all_node_vals_to_toposort (v
, topo
);
3207 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
3208 if (!ipa_edge_within_scc (cs
))
3209 propagate_constants_across_call (cs
);
3211 cycle_nodes
.release ();
3216 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
3217 the bigger one if otherwise. */
3220 safe_add (int a
, int b
)
3222 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
3223 return a
> b
? a
: b
;
3229 /* Propagate the estimated effects of individual values along the topological
3230 from the dependent values to those they depend on. */
3232 template <typename valtype
>
3234 value_topo_info
<valtype
>::propagate_effects ()
3236 ipcp_value
<valtype
> *base
;
3238 for (base
= values_topo
; base
; base
= base
->topo_next
)
3240 ipcp_value_source
<valtype
> *src
;
3241 ipcp_value
<valtype
> *val
;
3242 int time
= 0, size
= 0;
3244 for (val
= base
; val
; val
= val
->scc_next
)
3246 time
= safe_add (time
,
3247 val
->local_time_benefit
+ val
->prop_time_benefit
);
3248 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
3251 for (val
= base
; val
; val
= val
->scc_next
)
3252 for (src
= val
->sources
; src
; src
= src
->next
)
3254 && src
->cs
->maybe_hot_p ())
3256 src
->val
->prop_time_benefit
= safe_add (time
,
3257 src
->val
->prop_time_benefit
);
3258 src
->val
->prop_size_cost
= safe_add (size
,
3259 src
->val
->prop_size_cost
);
3265 /* Propagate constants, polymorphic contexts and their effects from the
3266 summaries interprocedurally. */
3269 ipcp_propagate_stage (struct ipa_topo_info
*topo
)
3271 struct cgraph_node
*node
;
3274 fprintf (dump_file
, "\n Propagating constants:\n\n");
3276 max_count
= profile_count::uninitialized ();
3278 FOR_EACH_DEFINED_FUNCTION (node
)
3280 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3282 determine_versionability (node
, info
);
3283 if (node
->has_gimple_body_p ())
3285 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
3286 ipa_get_param_count (info
));
3287 initialize_node_lattices (node
);
3289 ipa_fn_summary
*s
= ipa_fn_summaries
->get (node
);
3290 if (node
->definition
&& !node
->alias
&& s
!= NULL
)
3291 overall_size
+= s
->self_size
;
3292 max_count
= max_count
.max (node
->count
.ipa ());
3295 max_new_size
= overall_size
;
3296 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
3297 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
3298 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
3301 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
3302 overall_size
, max_new_size
);
3304 propagate_constants_topo (topo
);
3306 ipcp_verify_propagated_values ();
3307 topo
->constants
.propagate_effects ();
3308 topo
->contexts
.propagate_effects ();
3312 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
3313 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
3317 /* Discover newly direct outgoing edges from NODE which is a new clone with
3318 known KNOWN_CSTS and make them direct. */
3321 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
3322 vec
<tree
> known_csts
,
3323 vec
<ipa_polymorphic_call_context
>
3325 struct ipa_agg_replacement_value
*aggvals
)
3327 struct cgraph_edge
*ie
, *next_ie
;
3330 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
3335 next_ie
= ie
->next_callee
;
3336 target
= ipa_get_indirect_edge_target_1 (ie
, known_csts
, known_contexts
,
3337 vNULL
, aggvals
, &speculative
);
3340 bool agg_contents
= ie
->indirect_info
->agg_contents
;
3341 bool polymorphic
= ie
->indirect_info
->polymorphic
;
3342 int param_index
= ie
->indirect_info
->param_index
;
3343 struct cgraph_edge
*cs
= ipa_make_edge_direct_to_target (ie
, target
,
3347 if (cs
&& !agg_contents
&& !polymorphic
)
3349 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3350 int c
= ipa_get_controlled_uses (info
, param_index
);
3351 if (c
!= IPA_UNDESCRIBED_USE
)
3353 struct ipa_ref
*to_del
;
3356 ipa_set_controlled_uses (info
, param_index
, c
);
3357 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3358 fprintf (dump_file
, " controlled uses count of param "
3359 "%i bumped down to %i\n", param_index
, c
);
3361 && (to_del
= node
->find_reference (cs
->callee
, NULL
, 0)))
3363 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3364 fprintf (dump_file
, " and even removing its "
3365 "cloning-created reference\n");
3366 to_del
->remove_reference ();
3372 /* Turning calls to direct calls will improve overall summary. */
3374 ipa_update_overall_fn_summary (node
);
3377 class edge_clone_summary
;
3378 static call_summary
<edge_clone_summary
*> *edge_clone_summaries
= NULL
;
3380 /* Edge clone summary. */
3382 struct edge_clone_summary
3384 /* Default constructor. */
3385 edge_clone_summary (): prev_clone (NULL
), next_clone (NULL
) {}
3387 /* Default destructor. */
3388 ~edge_clone_summary ()
3391 edge_clone_summaries
->get (prev_clone
)->next_clone
= next_clone
;
3393 edge_clone_summaries
->get (next_clone
)->prev_clone
= prev_clone
;
3396 cgraph_edge
*prev_clone
;
3397 cgraph_edge
*next_clone
;
3400 class edge_clone_summary_t
:
3401 public call_summary
<edge_clone_summary
*>
3404 edge_clone_summary_t (symbol_table
*symtab
):
3405 call_summary
<edge_clone_summary
*> (symtab
)
3407 m_initialize_when_cloning
= true;
3410 virtual void duplicate (cgraph_edge
*src_edge
, cgraph_edge
*dst_edge
,
3411 edge_clone_summary
*src_data
,
3412 edge_clone_summary
*dst_data
);
3415 /* Edge duplication hook. */
3418 edge_clone_summary_t::duplicate (cgraph_edge
*src_edge
, cgraph_edge
*dst_edge
,
3419 edge_clone_summary
*src_data
,
3420 edge_clone_summary
*dst_data
)
3422 if (src_data
->next_clone
)
3423 edge_clone_summaries
->get (src_data
->next_clone
)->prev_clone
= dst_edge
;
3424 dst_data
->prev_clone
= src_edge
;
3425 dst_data
->next_clone
= src_data
->next_clone
;
3426 src_data
->next_clone
= dst_edge
;
3429 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
3430 parameter with the given INDEX. */
3433 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDE_INT offset
,
3436 struct ipa_agg_replacement_value
*aggval
;
3438 aggval
= ipa_get_agg_replacements_for_node (node
);
3441 if (aggval
->offset
== offset
3442 && aggval
->index
== index
)
3443 return aggval
->value
;
3444 aggval
= aggval
->next
;
3449 /* Return true is NODE is DEST or its clone for all contexts. */
3452 same_node_or_its_all_contexts_clone_p (cgraph_node
*node
, cgraph_node
*dest
)
3457 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3458 return info
->is_all_contexts_clone
&& info
->ipcp_orig_node
== dest
;
3461 /* Return true if edge CS does bring about the value described by SRC to
3462 DEST_VAL of node DEST or its clone for all contexts. */
3465 cgraph_edge_brings_value_p (cgraph_edge
*cs
, ipcp_value_source
<tree
> *src
,
3466 cgraph_node
*dest
, ipcp_value
<tree
> *dest_val
)
3468 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3469 enum availability availability
;
3470 cgraph_node
*real_dest
= cs
->callee
->function_symbol (&availability
);
3472 if (!same_node_or_its_all_contexts_clone_p (real_dest
, dest
)
3473 || availability
<= AVAIL_INTERPOSABLE
3474 || caller_info
->node_dead
)
3480 if (caller_info
->ipcp_orig_node
)
3483 if (src
->offset
== -1)
3484 t
= caller_info
->known_csts
[src
->index
];
3486 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
3487 return (t
!= NULL_TREE
3488 && values_equal_for_ipcp_p (src
->val
->value
, t
));
3492 /* At the moment we do not propagate over arithmetic jump functions in
3493 SCCs, so it is safe to detect self-feeding recursive calls in this
3495 if (src
->val
== dest_val
)
3498 struct ipcp_agg_lattice
*aglat
;
3499 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
3501 if (src
->offset
== -1)
3502 return (plats
->itself
.is_single_const ()
3503 && values_equal_for_ipcp_p (src
->val
->value
,
3504 plats
->itself
.values
->value
));
3507 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
3509 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3510 if (aglat
->offset
== src
->offset
)
3511 return (aglat
->is_single_const ()
3512 && values_equal_for_ipcp_p (src
->val
->value
,
3513 aglat
->values
->value
));
3519 /* Return true if edge CS does bring about the value described by SRC to
3520 DST_VAL of node DEST or its clone for all contexts. */
3523 cgraph_edge_brings_value_p (cgraph_edge
*cs
,
3524 ipcp_value_source
<ipa_polymorphic_call_context
> *src
,
3526 ipcp_value
<ipa_polymorphic_call_context
> *)
3528 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3529 cgraph_node
*real_dest
= cs
->callee
->function_symbol ();
3531 if (!same_node_or_its_all_contexts_clone_p (real_dest
, dest
)
3532 || caller_info
->node_dead
)
3537 if (caller_info
->ipcp_orig_node
)
3538 return (caller_info
->known_contexts
.length () > (unsigned) src
->index
)
3539 && values_equal_for_ipcp_p (src
->val
->value
,
3540 caller_info
->known_contexts
[src
->index
]);
3542 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
3544 return plats
->ctxlat
.is_single_const ()
3545 && values_equal_for_ipcp_p (src
->val
->value
,
3546 plats
->ctxlat
.values
->value
);
3549 /* Get the next clone in the linked list of clones of an edge. */
3551 static inline struct cgraph_edge
*
3552 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
3554 edge_clone_summary
*s
= edge_clone_summaries
->get (cs
);
3555 return s
!= NULL
? s
->next_clone
: NULL
;
3558 /* Given VAL that is intended for DEST, iterate over all its sources and if any
3559 of them is viable and hot, return true. In that case, for those that still
3560 hold, add their edge frequency and their number into *FREQUENCY and
3561 *CALLER_COUNT respectively. */
3563 template <typename valtype
>
3565 get_info_about_necessary_edges (ipcp_value
<valtype
> *val
, cgraph_node
*dest
,
3567 profile_count
*count_sum
, int *caller_count
)
3569 ipcp_value_source
<valtype
> *src
;
3570 int freq
= 0, count
= 0;
3571 profile_count cnt
= profile_count::zero ();
3573 bool non_self_recursive
= false;
3575 for (src
= val
->sources
; src
; src
= src
->next
)
3577 struct cgraph_edge
*cs
= src
->cs
;
3580 if (cgraph_edge_brings_value_p (cs
, src
, dest
, val
))
3583 freq
+= cs
->frequency ();
3584 if (cs
->count
.ipa ().initialized_p ())
3585 cnt
+= cs
->count
.ipa ();
3586 hot
|= cs
->maybe_hot_p ();
3587 if (cs
->caller
!= dest
)
3588 non_self_recursive
= true;
3590 cs
= get_next_cgraph_edge_clone (cs
);
3594 /* If the only edges bringing a value are self-recursive ones, do not bother
3596 if (!non_self_recursive
)
3601 *caller_count
= count
;
3605 /* Return a vector of incoming edges that do bring value VAL to node DEST. It
3606 is assumed their number is known and equal to CALLER_COUNT. */
3608 template <typename valtype
>
3609 static vec
<cgraph_edge
*>
3610 gather_edges_for_value (ipcp_value
<valtype
> *val
, cgraph_node
*dest
,
3613 ipcp_value_source
<valtype
> *src
;
3614 vec
<cgraph_edge
*> ret
;
3616 ret
.create (caller_count
);
3617 for (src
= val
->sources
; src
; src
= src
->next
)
3619 struct cgraph_edge
*cs
= src
->cs
;
3622 if (cgraph_edge_brings_value_p (cs
, src
, dest
, val
))
3623 ret
.quick_push (cs
);
3624 cs
= get_next_cgraph_edge_clone (cs
);
3631 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
3632 Return it or NULL if for some reason it cannot be created. */
3634 static struct ipa_replace_map
*
3635 get_replacement_map (struct ipa_node_params
*info
, tree value
, int parm_num
)
3637 struct ipa_replace_map
*replace_map
;
3640 replace_map
= ggc_alloc
<ipa_replace_map
> ();
3643 fprintf (dump_file
, " replacing ");
3644 ipa_dump_param (dump_file
, info
, parm_num
);
3646 fprintf (dump_file
, " with const ");
3647 print_generic_expr (dump_file
, value
);
3648 fprintf (dump_file
, "\n");
3650 replace_map
->old_tree
= NULL
;
3651 replace_map
->parm_num
= parm_num
;
3652 replace_map
->new_tree
= value
;
3653 replace_map
->replace_p
= true;
3654 replace_map
->ref_p
= false;
3659 /* Dump new profiling counts */
3662 dump_profile_updates (struct cgraph_node
*orig_node
,
3663 struct cgraph_node
*new_node
)
3665 struct cgraph_edge
*cs
;
3667 fprintf (dump_file
, " setting count of the specialized node to ");
3668 new_node
->count
.dump (dump_file
);
3669 fprintf (dump_file
, "\n");
3670 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3672 fprintf (dump_file
, " edge to %s has count ",
3673 cs
->callee
->name ());
3674 cs
->count
.dump (dump_file
);
3675 fprintf (dump_file
, "\n");
3678 fprintf (dump_file
, " setting count of the original node to ");
3679 orig_node
->count
.dump (dump_file
);
3680 fprintf (dump_file
, "\n");
3681 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3683 fprintf (dump_file
, " edge to %s is left with ",
3684 cs
->callee
->name ());
3685 cs
->count
.dump (dump_file
);
3686 fprintf (dump_file
, "\n");
3690 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
3691 their profile information to reflect this. */
3694 update_profiling_info (struct cgraph_node
*orig_node
,
3695 struct cgraph_node
*new_node
)
3697 struct cgraph_edge
*cs
;
3698 struct caller_statistics stats
;
3699 profile_count new_sum
, orig_sum
;
3700 profile_count remainder
, orig_node_count
= orig_node
->count
;
3702 if (!(orig_node_count
.ipa () > profile_count::zero ()))
3705 init_caller_stats (&stats
);
3706 orig_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
3708 orig_sum
= stats
.count_sum
;
3709 init_caller_stats (&stats
);
3710 new_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
3712 new_sum
= stats
.count_sum
;
3714 if (orig_node_count
< orig_sum
+ new_sum
)
3718 fprintf (dump_file
, " Problem: node %s has too low count ",
3719 orig_node
->dump_name ());
3720 orig_node_count
.dump (dump_file
);
3721 fprintf (dump_file
, "while the sum of incoming count is ");
3722 (orig_sum
+ new_sum
).dump (dump_file
);
3723 fprintf (dump_file
, "\n");
3726 orig_node_count
= (orig_sum
+ new_sum
).apply_scale (12, 10);
3729 fprintf (dump_file
, " proceeding by pretending it was ");
3730 orig_node_count
.dump (dump_file
);
3731 fprintf (dump_file
, "\n");
3735 remainder
= orig_node_count
.combine_with_ipa_count (orig_node_count
.ipa ()
3737 new_sum
= orig_node_count
.combine_with_ipa_count (new_sum
);
3738 orig_node
->count
= remainder
;
3740 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3741 cs
->count
= cs
->count
.apply_scale (new_sum
, orig_node_count
);
3743 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3744 cs
->count
= cs
->count
.apply_scale (remainder
, orig_node_count
);
3747 dump_profile_updates (orig_node
, new_node
);
3750 /* Update the respective profile of specialized NEW_NODE and the original
3751 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
3752 have been redirected to the specialized version. */
3755 update_specialized_profile (struct cgraph_node
*new_node
,
3756 struct cgraph_node
*orig_node
,
3757 profile_count redirected_sum
)
3759 struct cgraph_edge
*cs
;
3760 profile_count new_node_count
, orig_node_count
= orig_node
->count
;
3764 fprintf (dump_file
, " the sum of counts of redirected edges is ");
3765 redirected_sum
.dump (dump_file
);
3766 fprintf (dump_file
, "\n");
3768 if (!(orig_node_count
> profile_count::zero ()))
3771 gcc_assert (orig_node_count
>= redirected_sum
);
3773 new_node_count
= new_node
->count
;
3774 new_node
->count
+= redirected_sum
;
3775 orig_node
->count
-= redirected_sum
;
3777 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3778 cs
->count
+= cs
->count
.apply_scale (redirected_sum
, new_node_count
);
3780 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3782 profile_count dec
= cs
->count
.apply_scale (redirected_sum
,
3788 dump_profile_updates (orig_node
, new_node
);
3791 /* Create a specialized version of NODE with known constants in KNOWN_CSTS,
3792 known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and
3793 redirect all edges in CALLERS to it. */
3795 static struct cgraph_node
*
3796 create_specialized_node (struct cgraph_node
*node
,
3797 vec
<tree
> known_csts
,
3798 vec
<ipa_polymorphic_call_context
> known_contexts
,
3799 struct ipa_agg_replacement_value
*aggvals
,
3800 vec
<cgraph_edge
*> callers
)
3802 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
3803 vec
<ipa_replace_map
*, va_gc
> *replace_trees
= NULL
;
3804 struct ipa_agg_replacement_value
*av
;
3805 struct cgraph_node
*new_node
;
3806 int i
, count
= ipa_get_param_count (info
);
3807 bitmap args_to_skip
;
3809 gcc_assert (!info
->ipcp_orig_node
);
3811 if (node
->local
.can_change_signature
)
3813 args_to_skip
= BITMAP_GGC_ALLOC ();
3814 for (i
= 0; i
< count
; i
++)
3816 tree t
= known_csts
[i
];
3818 if (t
|| !ipa_is_param_used (info
, i
))
3819 bitmap_set_bit (args_to_skip
, i
);
3824 args_to_skip
= NULL
;
3825 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3826 fprintf (dump_file
, " cannot change function signature\n");
3829 for (i
= 0; i
< count
; i
++)
3831 tree t
= known_csts
[i
];
3834 struct ipa_replace_map
*replace_map
;
3836 gcc_checking_assert (TREE_CODE (t
) != TREE_BINFO
);
3837 replace_map
= get_replacement_map (info
, t
, i
);
3839 vec_safe_push (replace_trees
, replace_map
);
3842 auto_vec
<cgraph_edge
*, 2> self_recursive_calls
;
3843 for (i
= callers
.length () - 1; i
>= 0; i
--)
3845 cgraph_edge
*cs
= callers
[i
];
3846 if (cs
->caller
== node
)
3848 self_recursive_calls
.safe_push (cs
);
3849 callers
.unordered_remove (i
);
3853 new_node
= node
->create_virtual_clone (callers
, replace_trees
,
3854 args_to_skip
, "constprop");
3856 bool have_self_recursive_calls
= !self_recursive_calls
.is_empty ();
3857 for (unsigned j
= 0; j
< self_recursive_calls
.length (); j
++)
3859 cgraph_edge
*cs
= get_next_cgraph_edge_clone (self_recursive_calls
[j
]);
3860 /* Cloned edges can disappear during cloning as speculation can be
3861 resolved, check that we have one and that it comes from the last
3863 if (cs
&& cs
->caller
== new_node
)
3864 cs
->redirect_callee_duplicating_thunks (new_node
);
3865 /* Any future code that would make more than one clone of an outgoing
3866 edge would confuse this mechanism, so let's check that does not
3868 gcc_checking_assert (!cs
3869 || !get_next_cgraph_edge_clone (cs
)
3870 || get_next_cgraph_edge_clone (cs
)->caller
!= new_node
);
3872 if (have_self_recursive_calls
)
3873 new_node
->expand_all_artificial_thunks ();
3875 ipa_set_node_agg_value_chain (new_node
, aggvals
);
3876 for (av
= aggvals
; av
; av
= av
->next
)
3877 new_node
->maybe_create_reference (av
->value
, NULL
);
3879 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3881 fprintf (dump_file
, " the new node is %s.\n", new_node
->dump_name ());
3882 if (known_contexts
.exists ())
3884 for (i
= 0; i
< count
; i
++)
3885 if (!known_contexts
[i
].useless_p ())
3887 fprintf (dump_file
, " known ctx %i is ", i
);
3888 known_contexts
[i
].dump (dump_file
);
3892 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
3894 ipa_check_create_node_params ();
3895 update_profiling_info (node
, new_node
);
3896 new_info
= IPA_NODE_REF (new_node
);
3897 new_info
->ipcp_orig_node
= node
;
3898 new_info
->known_csts
= known_csts
;
3899 new_info
->known_contexts
= known_contexts
;
3901 ipcp_discover_new_direct_edges (new_node
, known_csts
, known_contexts
, aggvals
);
3907 /* Return true, if JFUNC, which describes a i-th parameter of call CS, is a
3908 simple no-operation pass-through function to itself. */
3911 self_recursive_pass_through_p (cgraph_edge
*cs
, ipa_jump_func
*jfunc
, int i
)
3913 enum availability availability
;
3914 if (cs
->caller
== cs
->callee
->function_symbol (&availability
)
3915 && availability
> AVAIL_INTERPOSABLE
3916 && jfunc
->type
== IPA_JF_PASS_THROUGH
3917 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
3918 && ipa_get_jf_pass_through_formal_id (jfunc
) == i
)
3923 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
3924 KNOWN_CSTS with constants that are also known for all of the CALLERS. */
3927 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
3928 vec
<tree
> known_csts
,
3929 vec
<cgraph_edge
*> callers
)
3931 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3932 int i
, count
= ipa_get_param_count (info
);
3934 for (i
= 0; i
< count
; i
++)
3936 struct cgraph_edge
*cs
;
3937 tree newval
= NULL_TREE
;
3940 tree type
= ipa_get_type (info
, i
);
3942 if (ipa_get_scalar_lat (info
, i
)->bottom
|| known_csts
[i
])
3945 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3947 struct ipa_jump_func
*jump_func
;
3950 if (IPA_NODE_REF (cs
->caller
)->node_dead
)
3953 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
))
3955 && call_passes_through_thunk_p (cs
)))
3960 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
3961 if (self_recursive_pass_through_p (cs
, jump_func
, i
))
3964 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
, type
);
3967 && !values_equal_for_ipcp_p (t
, newval
))
3968 || (!first
&& !newval
))
3980 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3982 fprintf (dump_file
, " adding an extra known scalar value ");
3983 print_ipcp_constant_value (dump_file
, newval
);
3984 fprintf (dump_file
, " for ");
3985 ipa_dump_param (dump_file
, info
, i
);
3986 fprintf (dump_file
, "\n");
3989 known_csts
[i
] = newval
;
3994 /* Given a NODE and a subset of its CALLERS, try to populate plank slots in
3995 KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the
3999 find_more_contexts_for_caller_subset (cgraph_node
*node
,
4000 vec
<ipa_polymorphic_call_context
>
4002 vec
<cgraph_edge
*> callers
)
4004 ipa_node_params
*info
= IPA_NODE_REF (node
);
4005 int i
, count
= ipa_get_param_count (info
);
4007 for (i
= 0; i
< count
; i
++)
4011 if (ipa_get_poly_ctx_lat (info
, i
)->bottom
4012 || (known_contexts
->exists ()
4013 && !(*known_contexts
)[i
].useless_p ()))
4016 ipa_polymorphic_call_context newval
;
4020 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4022 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
4024 ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
),
4026 ipa_polymorphic_call_context ctx
;
4027 ctx
= ipa_context_from_jfunc (IPA_NODE_REF (cs
->caller
), cs
, i
,
4035 newval
.meet_with (ctx
);
4036 if (newval
.useless_p ())
4040 if (!newval
.useless_p ())
4042 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4044 fprintf (dump_file
, " adding an extra known polymorphic "
4046 print_ipcp_constant_value (dump_file
, newval
);
4047 fprintf (dump_file
, " for ");
4048 ipa_dump_param (dump_file
, info
, i
);
4049 fprintf (dump_file
, "\n");
4052 if (!known_contexts
->exists ())
4053 known_contexts
->safe_grow_cleared (ipa_get_param_count (info
));
4054 (*known_contexts
)[i
] = newval
;
4060 /* Go through PLATS and create a vector of values consisting of values and
4061 offsets (minus OFFSET) of lattices that contain only a single value. */
4063 static vec
<ipa_agg_jf_item
>
4064 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
4066 vec
<ipa_agg_jf_item
> res
= vNULL
;
4068 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
4071 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
4072 if (aglat
->is_single_const ())
4074 struct ipa_agg_jf_item ti
;
4075 ti
.offset
= aglat
->offset
- offset
;
4076 ti
.value
= aglat
->values
->value
;
4082 /* Intersect all values in INTER with single value lattices in PLATS (while
4083 subtracting OFFSET). */
4086 intersect_with_plats (struct ipcp_param_lattices
*plats
,
4087 vec
<ipa_agg_jf_item
> *inter
,
4088 HOST_WIDE_INT offset
)
4090 struct ipcp_agg_lattice
*aglat
;
4091 struct ipa_agg_jf_item
*item
;
4094 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
4100 aglat
= plats
->aggs
;
4101 FOR_EACH_VEC_ELT (*inter
, k
, item
)
4108 if (aglat
->offset
- offset
> item
->offset
)
4110 if (aglat
->offset
- offset
== item
->offset
)
4112 gcc_checking_assert (item
->value
);
4113 if (aglat
->is_single_const ()
4114 && values_equal_for_ipcp_p (item
->value
,
4115 aglat
->values
->value
))
4119 aglat
= aglat
->next
;
4122 item
->value
= NULL_TREE
;
4126 /* Copy aggregate replacement values of NODE (which is an IPA-CP clone) to the
4127 vector result while subtracting OFFSET from the individual value offsets. */
4129 static vec
<ipa_agg_jf_item
>
4130 agg_replacements_to_vector (struct cgraph_node
*node
, int index
,
4131 HOST_WIDE_INT offset
)
4133 struct ipa_agg_replacement_value
*av
;
4134 vec
<ipa_agg_jf_item
> res
= vNULL
;
4136 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
4137 if (av
->index
== index
4138 && (av
->offset
- offset
) >= 0)
4140 struct ipa_agg_jf_item item
;
4141 gcc_checking_assert (av
->value
);
4142 item
.offset
= av
->offset
- offset
;
4143 item
.value
= av
->value
;
4144 res
.safe_push (item
);
4150 /* Intersect all values in INTER with those that we have already scheduled to
4151 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
4152 (while subtracting OFFSET). */
4155 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
4156 vec
<ipa_agg_jf_item
> *inter
,
4157 HOST_WIDE_INT offset
)
4159 struct ipa_agg_replacement_value
*srcvals
;
4160 struct ipa_agg_jf_item
*item
;
4163 srcvals
= ipa_get_agg_replacements_for_node (node
);
4170 FOR_EACH_VEC_ELT (*inter
, i
, item
)
4172 struct ipa_agg_replacement_value
*av
;
4176 for (av
= srcvals
; av
; av
= av
->next
)
4178 gcc_checking_assert (av
->value
);
4179 if (av
->index
== index
4180 && av
->offset
- offset
== item
->offset
)
4182 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
4188 item
->value
= NULL_TREE
;
4192 /* Intersect values in INTER with aggregate values that come along edge CS to
4193 parameter number INDEX and return it. If INTER does not actually exist yet,
4194 copy all incoming values to it. If we determine we ended up with no values
4195 whatsoever, return a released vector. */
4197 static vec
<ipa_agg_jf_item
>
4198 intersect_aggregates_with_edge (struct cgraph_edge
*cs
, int index
,
4199 vec
<ipa_agg_jf_item
> inter
)
4201 struct ipa_jump_func
*jfunc
;
4202 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), index
);
4203 if (jfunc
->type
== IPA_JF_PASS_THROUGH
4204 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
4206 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
4207 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
4209 if (caller_info
->ipcp_orig_node
)
4211 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
4212 struct ipcp_param_lattices
*orig_plats
;
4213 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
4215 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
4217 if (!inter
.exists ())
4218 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, 0);
4220 intersect_with_agg_replacements (cs
->caller
, src_idx
,
4231 struct ipcp_param_lattices
*src_plats
;
4232 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
4233 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
4235 /* Currently we do not produce clobber aggregate jump
4236 functions, adjust when we do. */
4237 gcc_checking_assert (!jfunc
->agg
.items
);
4238 if (!inter
.exists ())
4239 inter
= copy_plats_to_inter (src_plats
, 0);
4241 intersect_with_plats (src_plats
, &inter
, 0);
4250 else if (jfunc
->type
== IPA_JF_ANCESTOR
4251 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
4253 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
4254 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
4255 struct ipcp_param_lattices
*src_plats
;
4256 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
4258 if (caller_info
->ipcp_orig_node
)
4260 if (!inter
.exists ())
4261 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, delta
);
4263 intersect_with_agg_replacements (cs
->caller
, src_idx
, &inter
,
4268 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
4269 /* Currently we do not produce clobber aggregate jump
4270 functions, adjust when we do. */
4271 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
4272 if (!inter
.exists ())
4273 inter
= copy_plats_to_inter (src_plats
, delta
);
4275 intersect_with_plats (src_plats
, &inter
, delta
);
4278 else if (jfunc
->agg
.items
)
4280 struct ipa_agg_jf_item
*item
;
4283 if (!inter
.exists ())
4284 for (unsigned i
= 0; i
< jfunc
->agg
.items
->length (); i
++)
4285 inter
.safe_push ((*jfunc
->agg
.items
)[i
]);
4287 FOR_EACH_VEC_ELT (inter
, k
, item
)
4295 while ((unsigned) l
< jfunc
->agg
.items
->length ())
4297 struct ipa_agg_jf_item
*ti
;
4298 ti
= &(*jfunc
->agg
.items
)[l
];
4299 if (ti
->offset
> item
->offset
)
4301 if (ti
->offset
== item
->offset
)
4303 gcc_checking_assert (ti
->value
);
4304 if (values_equal_for_ipcp_p (item
->value
,
4318 return vec
<ipa_agg_jf_item
>();
4323 /* Look at edges in CALLERS and collect all known aggregate values that arrive
4324 from all of them. */
4326 static struct ipa_agg_replacement_value
*
4327 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
4328 vec
<cgraph_edge
*> callers
)
4330 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
4331 struct ipa_agg_replacement_value
*res
;
4332 struct ipa_agg_replacement_value
**tail
= &res
;
4333 struct cgraph_edge
*cs
;
4334 int i
, j
, count
= ipa_get_param_count (dest_info
);
4336 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4338 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
4343 for (i
= 0; i
< count
; i
++)
4345 struct cgraph_edge
*cs
;
4346 vec
<ipa_agg_jf_item
> inter
= vNULL
;
4347 struct ipa_agg_jf_item
*item
;
4348 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (dest_info
, i
);
4351 /* Among other things, the following check should deal with all by_ref
4353 if (plats
->aggs_bottom
)
4356 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4358 struct ipa_jump_func
*jfunc
4359 = ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
4360 if (self_recursive_pass_through_p (cs
, jfunc
, i
)
4361 && (!plats
->aggs_by_ref
4362 || ipa_get_jf_pass_through_agg_preserved (jfunc
)))
4364 inter
= intersect_aggregates_with_edge (cs
, i
, inter
);
4366 if (!inter
.exists ())
4370 FOR_EACH_VEC_ELT (inter
, j
, item
)
4372 struct ipa_agg_replacement_value
*v
;
4377 v
= ggc_alloc
<ipa_agg_replacement_value
> ();
4379 v
->offset
= item
->offset
;
4380 v
->value
= item
->value
;
4381 v
->by_ref
= plats
->aggs_by_ref
;
4387 if (inter
.exists ())
4394 /* Determine whether CS also brings all scalar values that the NODE is
4398 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
4399 struct cgraph_node
*node
)
4401 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
4402 int count
= ipa_get_param_count (dest_info
);
4403 struct ipa_node_params
*caller_info
;
4404 struct ipa_edge_args
*args
;
4407 caller_info
= IPA_NODE_REF (cs
->caller
);
4408 args
= IPA_EDGE_REF (cs
);
4409 for (i
= 0; i
< count
; i
++)
4411 struct ipa_jump_func
*jump_func
;
4414 val
= dest_info
->known_csts
[i
];
4418 if (i
>= ipa_get_cs_argument_count (args
))
4420 jump_func
= ipa_get_ith_jump_func (args
, i
);
4421 t
= ipa_value_from_jfunc (caller_info
, jump_func
,
4422 ipa_get_type (dest_info
, i
));
4423 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
4429 /* Determine whether CS also brings all aggregate values that NODE is
4432 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
4433 struct cgraph_node
*node
)
4435 struct ipa_node_params
*orig_caller_info
= IPA_NODE_REF (cs
->caller
);
4436 struct ipa_node_params
*orig_node_info
;
4437 struct ipa_agg_replacement_value
*aggval
;
4440 aggval
= ipa_get_agg_replacements_for_node (node
);
4444 count
= ipa_get_param_count (IPA_NODE_REF (node
));
4445 ec
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
4447 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4448 if (aggval
->index
>= ec
)
4451 orig_node_info
= IPA_NODE_REF (IPA_NODE_REF (node
)->ipcp_orig_node
);
4452 if (orig_caller_info
->ipcp_orig_node
)
4453 orig_caller_info
= IPA_NODE_REF (orig_caller_info
->ipcp_orig_node
);
4455 for (i
= 0; i
< count
; i
++)
4457 static vec
<ipa_agg_jf_item
> values
= vec
<ipa_agg_jf_item
>();
4458 struct ipcp_param_lattices
*plats
;
4459 bool interesting
= false;
4460 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4461 if (aggval
->index
== i
)
4469 plats
= ipa_get_parm_lattices (orig_node_info
, aggval
->index
);
4470 if (plats
->aggs_bottom
)
4473 values
= intersect_aggregates_with_edge (cs
, i
, values
);
4474 if (!values
.exists ())
4477 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4478 if (aggval
->index
== i
)
4480 struct ipa_agg_jf_item
*item
;
4483 FOR_EACH_VEC_ELT (values
, j
, item
)
4485 && item
->offset
== av
->offset
4486 && values_equal_for_ipcp_p (item
->value
, av
->value
))
4501 /* Given an original NODE and a VAL for which we have already created a
4502 specialized clone, look whether there are incoming edges that still lead
4503 into the old node but now also bring the requested value and also conform to
4504 all other criteria such that they can be redirected the special node.
4505 This function can therefore redirect the final edge in a SCC. */
4507 template <typename valtype
>
4509 perhaps_add_new_callers (cgraph_node
*node
, ipcp_value
<valtype
> *val
)
4511 ipcp_value_source
<valtype
> *src
;
4512 profile_count redirected_sum
= profile_count::zero ();
4514 for (src
= val
->sources
; src
; src
= src
->next
)
4516 struct cgraph_edge
*cs
= src
->cs
;
4519 if (cgraph_edge_brings_value_p (cs
, src
, node
, val
)
4520 && cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
4521 && cgraph_edge_brings_all_agg_vals_for_node (cs
, val
->spec_node
))
4524 fprintf (dump_file
, " - adding an extra caller %s of %s\n",
4525 cs
->caller
->dump_name (),
4526 val
->spec_node
->dump_name ());
4528 cs
->redirect_callee_duplicating_thunks (val
->spec_node
);
4529 val
->spec_node
->expand_all_artificial_thunks ();
4530 if (cs
->count
.ipa ().initialized_p ())
4531 redirected_sum
= redirected_sum
+ cs
->count
.ipa ();
4533 cs
= get_next_cgraph_edge_clone (cs
);
4537 if (redirected_sum
.nonzero_p ())
4538 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
4541 /* Return true if KNOWN_CONTEXTS contain at least one useful context. */
4544 known_contexts_useful_p (vec
<ipa_polymorphic_call_context
> known_contexts
)
4546 ipa_polymorphic_call_context
*ctx
;
4549 FOR_EACH_VEC_ELT (known_contexts
, i
, ctx
)
4550 if (!ctx
->useless_p ())
4555 /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */
4557 static vec
<ipa_polymorphic_call_context
>
4558 copy_useful_known_contexts (vec
<ipa_polymorphic_call_context
> known_contexts
)
4560 if (known_contexts_useful_p (known_contexts
))
4561 return known_contexts
.copy ();
4566 /* Copy KNOWN_CSTS and modify the copy according to VAL and INDEX. If
4567 non-empty, replace KNOWN_CONTEXTS with its copy too. */
4570 modify_known_vectors_with_val (vec
<tree
> *known_csts
,
4571 vec
<ipa_polymorphic_call_context
> *known_contexts
,
4572 ipcp_value
<tree
> *val
,
4575 *known_csts
= known_csts
->copy ();
4576 *known_contexts
= copy_useful_known_contexts (*known_contexts
);
4577 (*known_csts
)[index
] = val
->value
;
4580 /* Replace KNOWN_CSTS with its copy. Also copy KNOWN_CONTEXTS and modify the
4581 copy according to VAL and INDEX. */
4584 modify_known_vectors_with_val (vec
<tree
> *known_csts
,
4585 vec
<ipa_polymorphic_call_context
> *known_contexts
,
4586 ipcp_value
<ipa_polymorphic_call_context
> *val
,
4589 *known_csts
= known_csts
->copy ();
4590 *known_contexts
= known_contexts
->copy ();
4591 (*known_contexts
)[index
] = val
->value
;
4594 /* Return true if OFFSET indicates this was not an aggregate value or there is
4595 a replacement equivalent to VALUE, INDEX and OFFSET among those in the
4599 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value
*aggvals
,
4600 int index
, HOST_WIDE_INT offset
, tree value
)
4607 if (aggvals
->index
== index
4608 && aggvals
->offset
== offset
4609 && values_equal_for_ipcp_p (aggvals
->value
, value
))
4611 aggvals
= aggvals
->next
;
4616 /* Return true if offset is minus one because source of a polymorphic contect
4617 cannot be an aggregate value. */
4620 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value
*,
4621 int , HOST_WIDE_INT offset
,
4622 ipa_polymorphic_call_context
)
4624 return offset
== -1;
4627 /* Decide wheter to create a special version of NODE for value VAL of parameter
4628 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
4629 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
4630 KNOWN_CONTEXTS and KNOWN_AGGS describe the other already known values. */
4632 template <typename valtype
>
4634 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
4635 ipcp_value
<valtype
> *val
, vec
<tree
> known_csts
,
4636 vec
<ipa_polymorphic_call_context
> known_contexts
)
4638 struct ipa_agg_replacement_value
*aggvals
;
4639 int freq_sum
, caller_count
;
4640 profile_count count_sum
;
4641 vec
<cgraph_edge
*> callers
;
4645 perhaps_add_new_callers (node
, val
);
4648 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
4650 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4651 fprintf (dump_file
, " Ignoring candidate value because "
4652 "max_new_size would be reached with %li.\n",
4653 val
->local_size_cost
+ overall_size
);
4656 else if (!get_info_about_necessary_edges (val
, node
, &freq_sum
, &count_sum
,
4660 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4662 fprintf (dump_file
, " - considering value ");
4663 print_ipcp_constant_value (dump_file
, val
->value
);
4664 fprintf (dump_file
, " for ");
4665 ipa_dump_param (dump_file
, IPA_NODE_REF (node
), index
);
4667 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
4668 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
4671 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
4672 freq_sum
, count_sum
,
4673 val
->local_size_cost
)
4674 && !good_cloning_opportunity_p (node
,
4675 val
->local_time_benefit
4676 + val
->prop_time_benefit
,
4677 freq_sum
, count_sum
,
4678 val
->local_size_cost
4679 + val
->prop_size_cost
))
4683 fprintf (dump_file
, " Creating a specialized node of %s.\n",
4684 node
->dump_name ());
4686 callers
= gather_edges_for_value (val
, node
, caller_count
);
4688 modify_known_vectors_with_val (&known_csts
, &known_contexts
, val
, index
);
4691 known_csts
= known_csts
.copy ();
4692 known_contexts
= copy_useful_known_contexts (known_contexts
);
4694 find_more_scalar_values_for_callers_subset (node
, known_csts
, callers
);
4695 find_more_contexts_for_caller_subset (node
, &known_contexts
, callers
);
4696 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
4697 gcc_checking_assert (ipcp_val_agg_replacement_ok_p (aggvals
, index
,
4698 offset
, val
->value
));
4699 val
->spec_node
= create_specialized_node (node
, known_csts
, known_contexts
,
4701 overall_size
+= val
->local_size_cost
;
4703 /* TODO: If for some lattice there is only one other known value
4704 left, make a special node for it too. */
4709 /* Decide whether and what specialized clones of NODE should be created. */
4712 decide_whether_version_node (struct cgraph_node
*node
)
4714 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
4715 int i
, count
= ipa_get_param_count (info
);
4716 vec
<tree
> known_csts
;
4717 vec
<ipa_polymorphic_call_context
> known_contexts
;
4718 vec
<ipa_agg_jump_function
> known_aggs
= vNULL
;
4724 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4725 fprintf (dump_file
, "\nEvaluating opportunities for %s.\n",
4726 node
->dump_name ());
4728 gather_context_independent_values (info
, &known_csts
, &known_contexts
,
4729 info
->do_clone_for_all_contexts
? &known_aggs
4732 for (i
= 0; i
< count
;i
++)
4734 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4735 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
4736 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
4741 ipcp_value
<tree
> *val
;
4742 for (val
= lat
->values
; val
; val
= val
->next
)
4743 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
4747 if (!plats
->aggs_bottom
)
4749 struct ipcp_agg_lattice
*aglat
;
4750 ipcp_value
<tree
> *val
;
4751 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
4752 if (!aglat
->bottom
&& aglat
->values
4753 /* If the following is false, the one value is in
4755 && (plats
->aggs_contain_variable
4756 || !aglat
->is_single_const ()))
4757 for (val
= aglat
->values
; val
; val
= val
->next
)
4758 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
4759 known_csts
, known_contexts
);
4763 && known_contexts
[i
].useless_p ())
4765 ipcp_value
<ipa_polymorphic_call_context
> *val
;
4766 for (val
= ctxlat
->values
; val
; val
= val
->next
)
4767 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
4771 info
= IPA_NODE_REF (node
);
4774 if (info
->do_clone_for_all_contexts
)
4776 struct cgraph_node
*clone
;
4777 vec
<cgraph_edge
*> callers
;
4780 fprintf (dump_file
, " - Creating a specialized node of %s "
4781 "for all known contexts.\n", node
->dump_name ());
4783 callers
= node
->collect_callers ();
4784 find_more_scalar_values_for_callers_subset (node
, known_csts
, callers
);
4785 find_more_contexts_for_caller_subset (node
, &known_contexts
, callers
);
4786 ipa_agg_replacement_value
*aggvals
4787 = find_aggregate_values_for_callers_subset (node
, callers
);
4789 if (!known_contexts_useful_p (known_contexts
))
4791 known_contexts
.release ();
4792 known_contexts
= vNULL
;
4794 clone
= create_specialized_node (node
, known_csts
, known_contexts
,
4796 info
= IPA_NODE_REF (node
);
4797 info
->do_clone_for_all_contexts
= false;
4798 IPA_NODE_REF (clone
)->is_all_contexts_clone
= true;
4799 for (i
= 0; i
< count
; i
++)
4800 vec_free (known_aggs
[i
].items
);
4801 known_aggs
.release ();
4806 known_csts
.release ();
4807 known_contexts
.release ();
4813 /* Transitively mark all callees of NODE within the same SCC as not dead. */
4816 spread_undeadness (struct cgraph_node
*node
)
4818 struct cgraph_edge
*cs
;
4820 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
4821 if (ipa_edge_within_scc (cs
))
4823 struct cgraph_node
*callee
;
4824 struct ipa_node_params
*info
;
4826 callee
= cs
->callee
->function_symbol (NULL
);
4827 info
= IPA_NODE_REF (callee
);
4829 if (info
->node_dead
)
4831 info
->node_dead
= 0;
4832 spread_undeadness (callee
);
4837 /* Return true if NODE has a caller from outside of its SCC that is not
4838 dead. Worker callback for cgraph_for_node_and_aliases. */
4841 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
4842 void *data ATTRIBUTE_UNUSED
)
4844 struct cgraph_edge
*cs
;
4846 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4847 if (cs
->caller
->thunk
.thunk_p
4848 && cs
->caller
->call_for_symbol_thunks_and_aliases
4849 (has_undead_caller_from_outside_scc_p
, NULL
, true))
4851 else if (!ipa_edge_within_scc (cs
)
4852 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
4858 /* Identify nodes within the same SCC as NODE which are no longer needed
4859 because of new clones and will be removed as unreachable. */
4862 identify_dead_nodes (struct cgraph_node
*node
)
4864 struct cgraph_node
*v
;
4865 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4867 && !v
->call_for_symbol_thunks_and_aliases
4868 (has_undead_caller_from_outside_scc_p
, NULL
, true))
4869 IPA_NODE_REF (v
)->node_dead
= 1;
4871 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4872 if (!IPA_NODE_REF (v
)->node_dead
)
4873 spread_undeadness (v
);
4875 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4877 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4878 if (IPA_NODE_REF (v
)->node_dead
)
4879 fprintf (dump_file
, " Marking node as dead: %s.\n", v
->dump_name ());
4883 /* The decision stage. Iterate over the topological order of call graph nodes
4884 TOPO and make specialized clones if deemed beneficial. */
4887 ipcp_decision_stage (struct ipa_topo_info
*topo
)
4892 fprintf (dump_file
, "\nIPA decision stage:\n\n");
4894 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
4896 struct cgraph_node
*node
= topo
->order
[i
];
4897 bool change
= false, iterate
= true;
4901 struct cgraph_node
*v
;
4903 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4904 if (v
->has_gimple_body_p ()
4905 && ipcp_versionable_function_p (v
))
4906 iterate
|= decide_whether_version_node (v
);
4911 identify_dead_nodes (node
);
4915 /* Look up all the bits information that we have discovered and copy it over
4916 to the transformation summary. */
4919 ipcp_store_bits_results (void)
4923 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
4925 ipa_node_params
*info
= IPA_NODE_REF (node
);
4926 bool dumped_sth
= false;
4927 bool found_useful_result
= false;
4929 if (!opt_for_fn (node
->decl
, flag_ipa_bit_cp
))
4932 fprintf (dump_file
, "Not considering %s for ipa bitwise propagation "
4933 "; -fipa-bit-cp: disabled.\n",
4938 if (info
->ipcp_orig_node
)
4939 info
= IPA_NODE_REF (info
->ipcp_orig_node
);
4941 unsigned count
= ipa_get_param_count (info
);
4942 for (unsigned i
= 0; i
< count
; i
++)
4944 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4945 if (plats
->bits_lattice
.constant_p ())
4947 found_useful_result
= true;
4952 if (!found_useful_result
)
4955 ipcp_transformation_initialize ();
4956 ipcp_transformation
*ts
= ipcp_transformation_sum
->get_create (node
);
4957 vec_safe_reserve_exact (ts
->bits
, count
);
4959 for (unsigned i
= 0; i
< count
; i
++)
4961 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4964 if (plats
->bits_lattice
.constant_p ())
4966 = ipa_get_ipa_bits_for_value (plats
->bits_lattice
.get_value (),
4967 plats
->bits_lattice
.get_mask ());
4971 ts
->bits
->quick_push (jfbits
);
4972 if (!dump_file
|| !jfbits
)
4976 fprintf (dump_file
, "Propagated bits info for function %s:\n",
4977 node
->dump_name ());
4980 fprintf (dump_file
, " param %i: value = ", i
);
4981 print_hex (jfbits
->value
, dump_file
);
4982 fprintf (dump_file
, ", mask = ");
4983 print_hex (jfbits
->mask
, dump_file
);
4984 fprintf (dump_file
, "\n");
4989 /* Look up all VR information that we have discovered and copy it over
4990 to the transformation summary. */
4993 ipcp_store_vr_results (void)
4997 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
4999 ipa_node_params
*info
= IPA_NODE_REF (node
);
5000 bool found_useful_result
= false;
5002 if (!opt_for_fn (node
->decl
, flag_ipa_vrp
))
5005 fprintf (dump_file
, "Not considering %s for VR discovery "
5006 "and propagate; -fipa-ipa-vrp: disabled.\n",
5011 if (info
->ipcp_orig_node
)
5012 info
= IPA_NODE_REF (info
->ipcp_orig_node
);
5014 unsigned count
= ipa_get_param_count (info
);
5015 for (unsigned i
= 0; i
< count
; i
++)
5017 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
5018 if (!plats
->m_value_range
.bottom_p ()
5019 && !plats
->m_value_range
.top_p ())
5021 found_useful_result
= true;
5025 if (!found_useful_result
)
5028 ipcp_transformation_initialize ();
5029 ipcp_transformation
*ts
= ipcp_transformation_sum
->get_create (node
);
5030 vec_safe_reserve_exact (ts
->m_vr
, count
);
5032 for (unsigned i
= 0; i
< count
; i
++)
5034 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
5037 if (!plats
->m_value_range
.bottom_p ()
5038 && !plats
->m_value_range
.top_p ())
5041 vr
.type
= plats
->m_value_range
.m_vr
.type
;
5042 vr
.min
= wi::to_wide (plats
->m_value_range
.m_vr
.min
);
5043 vr
.max
= wi::to_wide (plats
->m_value_range
.m_vr
.max
);
5048 vr
.type
= VR_VARYING
;
5049 vr
.min
= vr
.max
= wi::zero (INT_TYPE_SIZE
);
5051 ts
->m_vr
->quick_push (vr
);
5056 /* The IPCP driver. */
5061 struct ipa_topo_info topo
;
5063 if (edge_clone_summaries
== NULL
)
5064 edge_clone_summaries
= new edge_clone_summary_t (symtab
);
5066 ipa_check_create_node_params ();
5067 ipa_check_create_edge_args ();
5071 fprintf (dump_file
, "\nIPA structures before propagation:\n");
5072 if (dump_flags
& TDF_DETAILS
)
5073 ipa_print_all_params (dump_file
);
5074 ipa_print_all_jump_functions (dump_file
);
5077 /* Topological sort. */
5078 build_toporder_info (&topo
);
5079 /* Do the interprocedural propagation. */
5080 ipcp_propagate_stage (&topo
);
5081 /* Decide what constant propagation and cloning should be performed. */
5082 ipcp_decision_stage (&topo
);
5083 /* Store results of bits propagation. */
5084 ipcp_store_bits_results ();
5085 /* Store results of value range propagation. */
5086 ipcp_store_vr_results ();
5088 /* Free all IPCP structures. */
5089 free_toporder_info (&topo
);
5090 delete edge_clone_summaries
;
5091 edge_clone_summaries
= NULL
;
5092 ipa_free_all_structures_after_ipa_cp ();
5094 fprintf (dump_file
, "\nIPA constant propagation end\n");
5098 /* Initialization and computation of IPCP data structures. This is the initial
5099 intraprocedural analysis of functions, which gathers information to be
5100 propagated later on. */
5103 ipcp_generate_summary (void)
5105 struct cgraph_node
*node
;
5108 fprintf (dump_file
, "\nIPA constant propagation start:\n");
5109 ipa_register_cgraph_hooks ();
5111 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
5112 ipa_analyze_node (node
);
5115 /* Write ipcp summary for nodes in SET. */
5118 ipcp_write_summary (void)
5120 ipa_prop_write_jump_functions ();
5123 /* Read ipcp summary. */
5126 ipcp_read_summary (void)
5128 ipa_prop_read_jump_functions ();
5133 const pass_data pass_data_ipa_cp
=
5135 IPA_PASS
, /* type */
5137 OPTGROUP_NONE
, /* optinfo_flags */
5138 TV_IPA_CONSTANT_PROP
, /* tv_id */
5139 0, /* properties_required */
5140 0, /* properties_provided */
5141 0, /* properties_destroyed */
5142 0, /* todo_flags_start */
5143 ( TODO_dump_symtab
| TODO_remove_functions
), /* todo_flags_finish */
5146 class pass_ipa_cp
: public ipa_opt_pass_d
5149 pass_ipa_cp (gcc::context
*ctxt
)
5150 : ipa_opt_pass_d (pass_data_ipa_cp
, ctxt
,
5151 ipcp_generate_summary
, /* generate_summary */
5152 ipcp_write_summary
, /* write_summary */
5153 ipcp_read_summary
, /* read_summary */
5154 ipcp_write_transformation_summaries
, /*
5155 write_optimization_summary */
5156 ipcp_read_transformation_summaries
, /*
5157 read_optimization_summary */
5158 NULL
, /* stmt_fixup */
5159 0, /* function_transform_todo_flags_start */
5160 ipcp_transform_function
, /* function_transform */
5161 NULL
) /* variable_transform */
5164 /* opt_pass methods: */
5165 virtual bool gate (function
*)
5167 /* FIXME: We should remove the optimize check after we ensure we never run
5168 IPA passes when not optimizing. */
5169 return (flag_ipa_cp
&& optimize
) || in_lto_p
;
5172 virtual unsigned int execute (function
*) { return ipcp_driver (); }
5174 }; // class pass_ipa_cp
5179 make_pass_ipa_cp (gcc::context
*ctxt
)
5181 return new pass_ipa_cp (ctxt
);
5184 /* Reset all state within ipa-cp.c so that we can rerun the compiler
5185 within the same process. For use by toplev::finalize. */
5188 ipa_cp_c_finalize (void)
5190 max_count
= profile_count::uninitialized ();