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 (DECL_BUILT_IN (edge
->callee
->decl
)
642 && DECL_BUILT_IN_CLASS (edge
->callee
->decl
) == BUILT_IN_NORMAL
)
644 if (DECL_FUNCTION_CODE (edge
->callee
->decl
) == BUILT_IN_VA_ARG_PACK
)
645 reason
= "external function which calls va_arg_pack";
646 if (DECL_FUNCTION_CODE (edge
->callee
->decl
)
647 == BUILT_IN_VA_ARG_PACK_LEN
)
648 reason
= "external function which calls va_arg_pack_len";
651 if (reason
&& dump_file
&& !node
->alias
&& !node
->thunk
.thunk_p
)
652 fprintf (dump_file
, "Function %s is not versionable, reason: %s.\n",
653 node
->dump_name (), reason
);
655 info
->versionable
= (reason
== NULL
);
658 /* Return true if it is at all technically possible to create clones of a
662 ipcp_versionable_function_p (struct cgraph_node
*node
)
664 return IPA_NODE_REF (node
)->versionable
;
667 /* Structure holding accumulated information about callers of a node. */
669 struct caller_statistics
671 profile_count count_sum
;
672 int n_calls
, n_hot_calls
, freq_sum
;
675 /* Initialize fields of STAT to zeroes. */
678 init_caller_stats (struct caller_statistics
*stats
)
680 stats
->count_sum
= profile_count::zero ();
682 stats
->n_hot_calls
= 0;
686 /* Worker callback of cgraph_for_node_and_aliases accumulating statistics of
687 non-thunk incoming edges to NODE. */
690 gather_caller_stats (struct cgraph_node
*node
, void *data
)
692 struct caller_statistics
*stats
= (struct caller_statistics
*) data
;
693 struct cgraph_edge
*cs
;
695 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
696 if (!cs
->caller
->thunk
.thunk_p
)
698 if (cs
->count
.ipa ().initialized_p ())
699 stats
->count_sum
+= cs
->count
.ipa ();
700 stats
->freq_sum
+= cs
->frequency ();
702 if (cs
->maybe_hot_p ())
703 stats
->n_hot_calls
++;
709 /* Return true if this NODE is viable candidate for cloning. */
712 ipcp_cloning_candidate_p (struct cgraph_node
*node
)
714 struct caller_statistics stats
;
716 gcc_checking_assert (node
->has_gimple_body_p ());
718 if (!opt_for_fn (node
->decl
, flag_ipa_cp_clone
))
721 fprintf (dump_file
, "Not considering %s for cloning; "
722 "-fipa-cp-clone disabled.\n",
727 if (node
->optimize_for_size_p ())
730 fprintf (dump_file
, "Not considering %s for cloning; "
731 "optimizing it for size.\n",
736 init_caller_stats (&stats
);
737 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
, false);
739 if (ipa_fn_summaries
->get (node
)->self_size
< stats
.n_calls
)
742 fprintf (dump_file
, "Considering %s for cloning; code might shrink.\n",
747 /* When profile is available and function is hot, propagate into it even if
748 calls seems cold; constant propagation can improve function's speed
750 if (max_count
> profile_count::zero ())
752 if (stats
.count_sum
> node
->count
.ipa ().apply_scale (90, 100))
755 fprintf (dump_file
, "Considering %s for cloning; "
756 "usually called directly.\n",
761 if (!stats
.n_hot_calls
)
764 fprintf (dump_file
, "Not considering %s for cloning; no hot calls.\n",
769 fprintf (dump_file
, "Considering %s for cloning.\n",
774 template <typename valtype
>
775 class value_topo_info
778 /* Head of the linked list of topologically sorted values. */
779 ipcp_value
<valtype
> *values_topo
;
780 /* Stack for creating SCCs, represented by a linked list too. */
781 ipcp_value
<valtype
> *stack
;
782 /* Counter driving the algorithm in add_val_to_toposort. */
785 value_topo_info () : values_topo (NULL
), stack (NULL
), dfs_counter (0)
787 void add_val (ipcp_value
<valtype
> *cur_val
);
788 void propagate_effects ();
791 /* Arrays representing a topological ordering of call graph nodes and a stack
792 of nodes used during constant propagation and also data required to perform
793 topological sort of values and propagation of benefits in the determined
799 /* Array with obtained topological order of cgraph nodes. */
800 struct cgraph_node
**order
;
801 /* Stack of cgraph nodes used during propagation within SCC until all values
802 in the SCC stabilize. */
803 struct cgraph_node
**stack
;
804 int nnodes
, stack_top
;
806 value_topo_info
<tree
> constants
;
807 value_topo_info
<ipa_polymorphic_call_context
> contexts
;
809 ipa_topo_info () : order(NULL
), stack(NULL
), nnodes(0), stack_top(0),
814 /* Allocate the arrays in TOPO and topologically sort the nodes into order. */
817 build_toporder_info (struct ipa_topo_info
*topo
)
819 topo
->order
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
820 topo
->stack
= XCNEWVEC (struct cgraph_node
*, symtab
->cgraph_count
);
822 gcc_checking_assert (topo
->stack_top
== 0);
823 topo
->nnodes
= ipa_reduced_postorder (topo
->order
, true, NULL
);
826 /* Free information about strongly connected components and the arrays in
830 free_toporder_info (struct ipa_topo_info
*topo
)
832 ipa_free_postorder_info ();
837 /* Add NODE to the stack in TOPO, unless it is already there. */
840 push_node_to_stack (struct ipa_topo_info
*topo
, struct cgraph_node
*node
)
842 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
843 if (info
->node_enqueued
)
845 info
->node_enqueued
= 1;
846 topo
->stack
[topo
->stack_top
++] = node
;
849 /* Pop a node from the stack in TOPO and return it or return NULL if the stack
852 static struct cgraph_node
*
853 pop_node_from_stack (struct ipa_topo_info
*topo
)
857 struct cgraph_node
*node
;
859 node
= topo
->stack
[topo
->stack_top
];
860 IPA_NODE_REF (node
)->node_enqueued
= 0;
867 /* Set lattice LAT to bottom and return true if it previously was not set as
870 template <typename valtype
>
872 ipcp_lattice
<valtype
>::set_to_bottom ()
879 /* Mark lattice as containing an unknown value and return true if it previously
880 was not marked as such. */
882 template <typename valtype
>
884 ipcp_lattice
<valtype
>::set_contains_variable ()
886 bool ret
= !contains_variable
;
887 contains_variable
= true;
891 /* Set all aggegate lattices in PLATS to bottom and return true if they were
892 not previously set as such. */
895 set_agg_lats_to_bottom (struct ipcp_param_lattices
*plats
)
897 bool ret
= !plats
->aggs_bottom
;
898 plats
->aggs_bottom
= true;
902 /* Mark all aggegate lattices in PLATS as containing an unknown value and
903 return true if they were not previously marked as such. */
906 set_agg_lats_contain_variable (struct ipcp_param_lattices
*plats
)
908 bool ret
= !plats
->aggs_contain_variable
;
909 plats
->aggs_contain_variable
= true;
914 ipcp_vr_lattice::meet_with (const ipcp_vr_lattice
&other
)
916 return meet_with_1 (&other
.m_vr
);
919 /* Meet the current value of the lattice with value ranfge described by VR
923 ipcp_vr_lattice::meet_with (const value_range
*p_vr
)
925 return meet_with_1 (p_vr
);
928 /* Meet the current value of the lattice with value ranfge described by
932 ipcp_vr_lattice::meet_with_1 (const value_range
*other_vr
)
934 tree min
= m_vr
.min
, max
= m_vr
.max
;
935 value_range_type type
= m_vr
.type
;
940 if (other_vr
->type
== VR_VARYING
)
941 return set_to_bottom ();
943 vrp_meet (&m_vr
, other_vr
);
944 if (type
!= m_vr
.type
952 /* Return true if value range information in the lattice is yet unknown. */
955 ipcp_vr_lattice::top_p () const
957 return m_vr
.type
== VR_UNDEFINED
;
960 /* Return true if value range information in the lattice is known to be
964 ipcp_vr_lattice::bottom_p () const
966 return m_vr
.type
== VR_VARYING
;
969 /* Set value range information in the lattice to bottom. Return true if it
970 previously was in a different state. */
973 ipcp_vr_lattice::set_to_bottom ()
975 if (m_vr
.type
== VR_VARYING
)
977 m_vr
.type
= VR_VARYING
;
981 /* Set lattice value to bottom, if it already isn't the case. */
984 ipcp_bits_lattice::set_to_bottom ()
988 m_lattice_val
= IPA_BITS_VARYING
;
994 /* Set to constant if it isn't already. Only meant to be called
995 when switching state from TOP. */
998 ipcp_bits_lattice::set_to_constant (widest_int value
, widest_int mask
)
1000 gcc_assert (top_p ());
1001 m_lattice_val
= IPA_BITS_CONSTANT
;
1007 /* Convert operand to value, mask form. */
1010 ipcp_bits_lattice::get_value_and_mask (tree operand
, widest_int
*valuep
, widest_int
*maskp
)
1012 wide_int
get_nonzero_bits (const_tree
);
1014 if (TREE_CODE (operand
) == INTEGER_CST
)
1016 *valuep
= wi::to_widest (operand
);
1026 /* Meet operation, similar to ccp_lattice_meet, we xor values
1027 if this->value, value have different values at same bit positions, we want
1028 to drop that bit to varying. Return true if mask is changed.
1029 This function assumes that the lattice value is in CONSTANT state */
1032 ipcp_bits_lattice::meet_with_1 (widest_int value
, widest_int mask
,
1035 gcc_assert (constant_p ());
1037 widest_int old_mask
= m_mask
;
1038 m_mask
= (m_mask
| mask
) | (m_value
^ value
);
1040 if (wi::sext (m_mask
, precision
) == -1)
1041 return set_to_bottom ();
1043 return m_mask
!= old_mask
;
1046 /* Meet the bits lattice with operand
1047 described by <value, mask, sgn, precision. */
1050 ipcp_bits_lattice::meet_with (widest_int value
, widest_int mask
,
1058 if (wi::sext (mask
, precision
) == -1)
1059 return set_to_bottom ();
1060 return set_to_constant (value
, mask
);
1063 return meet_with_1 (value
, mask
, precision
);
1066 /* Meet bits lattice with the result of bit_value_binop (other, operand)
1067 if code is binary operation or bit_value_unop (other) if code is unary op.
1068 In the case when code is nop_expr, no adjustment is required. */
1071 ipcp_bits_lattice::meet_with (ipcp_bits_lattice
& other
, unsigned precision
,
1072 signop sgn
, enum tree_code code
, tree operand
)
1074 if (other
.bottom_p ())
1075 return set_to_bottom ();
1077 if (bottom_p () || other
.top_p ())
1080 widest_int adjusted_value
, adjusted_mask
;
1082 if (TREE_CODE_CLASS (code
) == tcc_binary
)
1084 tree type
= TREE_TYPE (operand
);
1085 gcc_assert (INTEGRAL_TYPE_P (type
));
1086 widest_int o_value
, o_mask
;
1087 get_value_and_mask (operand
, &o_value
, &o_mask
);
1089 bit_value_binop (code
, sgn
, precision
, &adjusted_value
, &adjusted_mask
,
1090 sgn
, precision
, other
.get_value (), other
.get_mask (),
1091 TYPE_SIGN (type
), TYPE_PRECISION (type
), o_value
, o_mask
);
1093 if (wi::sext (adjusted_mask
, precision
) == -1)
1094 return set_to_bottom ();
1097 else if (TREE_CODE_CLASS (code
) == tcc_unary
)
1099 bit_value_unop (code
, sgn
, precision
, &adjusted_value
,
1100 &adjusted_mask
, sgn
, precision
, other
.get_value (),
1103 if (wi::sext (adjusted_mask
, precision
) == -1)
1104 return set_to_bottom ();
1108 return set_to_bottom ();
1112 if (wi::sext (adjusted_mask
, precision
) == -1)
1113 return set_to_bottom ();
1114 return set_to_constant (adjusted_value
, adjusted_mask
);
1117 return meet_with_1 (adjusted_value
, adjusted_mask
, precision
);
1120 /* Mark bot aggregate and scalar lattices as containing an unknown variable,
1121 return true is any of them has not been marked as such so far. */
1124 set_all_contains_variable (struct ipcp_param_lattices
*plats
)
1127 ret
= plats
->itself
.set_contains_variable ();
1128 ret
|= plats
->ctxlat
.set_contains_variable ();
1129 ret
|= set_agg_lats_contain_variable (plats
);
1130 ret
|= plats
->bits_lattice
.set_to_bottom ();
1131 ret
|= plats
->m_value_range
.set_to_bottom ();
1135 /* Worker of call_for_symbol_thunks_and_aliases, increment the integer DATA
1136 points to by the number of callers to NODE. */
1139 count_callers (cgraph_node
*node
, void *data
)
1141 int *caller_count
= (int *) data
;
1143 for (cgraph_edge
*cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
1144 /* Local thunks can be handled transparently, but if the thunk can not
1145 be optimized out, count it as a real use. */
1146 if (!cs
->caller
->thunk
.thunk_p
|| !cs
->caller
->local
.local
)
1151 /* Worker of call_for_symbol_thunks_and_aliases, it is supposed to be called on
1152 the one caller of some other node. Set the caller's corresponding flag. */
1155 set_single_call_flag (cgraph_node
*node
, void *)
1157 cgraph_edge
*cs
= node
->callers
;
1158 /* Local thunks can be handled transparently, skip them. */
1159 while (cs
&& cs
->caller
->thunk
.thunk_p
&& cs
->caller
->local
.local
)
1160 cs
= cs
->next_caller
;
1163 IPA_NODE_REF (cs
->caller
)->node_calling_single_call
= true;
1169 /* Initialize ipcp_lattices. */
1172 initialize_node_lattices (struct cgraph_node
*node
)
1174 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1175 struct cgraph_edge
*ie
;
1176 bool disable
= false, variable
= false;
1179 gcc_checking_assert (node
->has_gimple_body_p ());
1180 if (cgraph_local_p (node
))
1182 int caller_count
= 0;
1183 node
->call_for_symbol_thunks_and_aliases (count_callers
, &caller_count
,
1185 gcc_checking_assert (caller_count
> 0);
1186 if (caller_count
== 1)
1187 node
->call_for_symbol_thunks_and_aliases (set_single_call_flag
,
1192 /* When cloning is allowed, we can assume that externally visible
1193 functions are not called. We will compensate this by cloning
1195 if (ipcp_versionable_function_p (node
)
1196 && ipcp_cloning_candidate_p (node
))
1202 for (i
= 0; i
< ipa_get_param_count (info
); i
++)
1204 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1205 plats
->m_value_range
.init ();
1208 if (disable
|| variable
)
1210 for (i
= 0; i
< ipa_get_param_count (info
); i
++)
1212 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
1215 plats
->itself
.set_to_bottom ();
1216 plats
->ctxlat
.set_to_bottom ();
1217 set_agg_lats_to_bottom (plats
);
1218 plats
->bits_lattice
.set_to_bottom ();
1219 plats
->m_value_range
.set_to_bottom ();
1222 set_all_contains_variable (plats
);
1224 if (dump_file
&& (dump_flags
& TDF_DETAILS
)
1225 && !node
->alias
&& !node
->thunk
.thunk_p
)
1226 fprintf (dump_file
, "Marking all lattices of %s as %s\n",
1227 node
->dump_name (), disable
? "BOTTOM" : "VARIABLE");
1230 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
1231 if (ie
->indirect_info
->polymorphic
1232 && ie
->indirect_info
->param_index
>= 0)
1234 gcc_checking_assert (ie
->indirect_info
->param_index
>= 0);
1235 ipa_get_parm_lattices (info
,
1236 ie
->indirect_info
->param_index
)->virt_call
= 1;
1240 /* Return the result of a (possibly arithmetic) pass through jump function
1241 JFUNC on the constant value INPUT. RES_TYPE is the type of the parameter
1242 to which the result is passed. Return NULL_TREE if that cannot be
1243 determined or be considered an interprocedural invariant. */
1246 ipa_get_jf_pass_through_result (struct ipa_jump_func
*jfunc
, tree input
,
1251 if (ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
1253 if (!is_gimple_ip_invariant (input
))
1256 tree_code opcode
= ipa_get_jf_pass_through_operation (jfunc
);
1259 if (TREE_CODE_CLASS (opcode
) == tcc_comparison
)
1260 res_type
= boolean_type_node
;
1261 else if (expr_type_first_operand_type_p (opcode
))
1262 res_type
= TREE_TYPE (input
);
1267 if (TREE_CODE_CLASS (opcode
) == tcc_unary
)
1268 res
= fold_unary (opcode
, res_type
, input
);
1270 res
= fold_binary (opcode
, res_type
, input
,
1271 ipa_get_jf_pass_through_operand (jfunc
));
1273 if (res
&& !is_gimple_ip_invariant (res
))
1279 /* Return the result of an ancestor jump function JFUNC on the constant value
1280 INPUT. Return NULL_TREE if that cannot be determined. */
1283 ipa_get_jf_ancestor_result (struct ipa_jump_func
*jfunc
, tree input
)
1285 gcc_checking_assert (TREE_CODE (input
) != TREE_BINFO
);
1286 if (TREE_CODE (input
) == ADDR_EXPR
)
1288 tree t
= TREE_OPERAND (input
, 0);
1289 t
= build_ref_for_offset (EXPR_LOCATION (t
), t
,
1290 ipa_get_jf_ancestor_offset (jfunc
), false,
1291 ptr_type_node
, NULL
, false);
1292 return build_fold_addr_expr (t
);
1298 /* Determine whether JFUNC evaluates to a single known constant value and if
1299 so, return it. Otherwise return NULL. INFO describes the caller node or
1300 the one it is inlined to, so that pass-through jump functions can be
1301 evaluated. PARM_TYPE is the type of the parameter to which the result is
1305 ipa_value_from_jfunc (struct ipa_node_params
*info
, struct ipa_jump_func
*jfunc
,
1308 if (jfunc
->type
== IPA_JF_CONST
)
1309 return ipa_get_jf_constant (jfunc
);
1310 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1311 || jfunc
->type
== IPA_JF_ANCESTOR
)
1316 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1317 idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1319 idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1321 if (info
->ipcp_orig_node
)
1322 input
= info
->known_csts
[idx
];
1325 ipcp_lattice
<tree
> *lat
;
1328 || idx
>= ipa_get_param_count (info
))
1330 lat
= ipa_get_scalar_lat (info
, idx
);
1331 if (!lat
->is_single_const ())
1333 input
= lat
->values
->value
;
1339 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1340 return ipa_get_jf_pass_through_result (jfunc
, input
, parm_type
);
1342 return ipa_get_jf_ancestor_result (jfunc
, input
);
1348 /* Determie whether JFUNC evaluates to single known polymorphic context, given
1349 that INFO describes the caller node or the one it is inlined to, CS is the
1350 call graph edge corresponding to JFUNC and CSIDX index of the described
1353 ipa_polymorphic_call_context
1354 ipa_context_from_jfunc (ipa_node_params
*info
, cgraph_edge
*cs
, int csidx
,
1355 ipa_jump_func
*jfunc
)
1357 ipa_edge_args
*args
= IPA_EDGE_REF (cs
);
1358 ipa_polymorphic_call_context ctx
;
1359 ipa_polymorphic_call_context
*edge_ctx
1360 = cs
? ipa_get_ith_polymorhic_call_context (args
, csidx
) : NULL
;
1362 if (edge_ctx
&& !edge_ctx
->useless_p ())
1365 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1366 || jfunc
->type
== IPA_JF_ANCESTOR
)
1368 ipa_polymorphic_call_context srcctx
;
1370 bool type_preserved
= true;
1371 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1373 if (ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1375 type_preserved
= ipa_get_jf_pass_through_type_preserved (jfunc
);
1376 srcidx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1380 type_preserved
= ipa_get_jf_ancestor_type_preserved (jfunc
);
1381 srcidx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1383 if (info
->ipcp_orig_node
)
1385 if (info
->known_contexts
.exists ())
1386 srcctx
= info
->known_contexts
[srcidx
];
1391 || srcidx
>= ipa_get_param_count (info
))
1393 ipcp_lattice
<ipa_polymorphic_call_context
> *lat
;
1394 lat
= ipa_get_poly_ctx_lat (info
, srcidx
);
1395 if (!lat
->is_single_const ())
1397 srcctx
= lat
->values
->value
;
1399 if (srcctx
.useless_p ())
1401 if (jfunc
->type
== IPA_JF_ANCESTOR
)
1402 srcctx
.offset_by (ipa_get_jf_ancestor_offset (jfunc
));
1403 if (!type_preserved
)
1404 srcctx
.possible_dynamic_type_change (cs
->in_polymorphic_cdtor
);
1405 srcctx
.combine_with (ctx
);
1412 /* If checking is enabled, verify that no lattice is in the TOP state, i.e. not
1413 bottom, not containing a variable component and without any known value at
1417 ipcp_verify_propagated_values (void)
1419 struct cgraph_node
*node
;
1421 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
1423 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
1424 int i
, count
= ipa_get_param_count (info
);
1426 for (i
= 0; i
< count
; i
++)
1428 ipcp_lattice
<tree
> *lat
= ipa_get_scalar_lat (info
, i
);
1431 && !lat
->contains_variable
1432 && lat
->values_count
== 0)
1436 symtab
->dump (dump_file
);
1437 fprintf (dump_file
, "\nIPA lattices after constant "
1438 "propagation, before gcc_unreachable:\n");
1439 print_all_lattices (dump_file
, true, false);
1448 /* Return true iff X and Y should be considered equal values by IPA-CP. */
1451 values_equal_for_ipcp_p (tree x
, tree y
)
1453 gcc_checking_assert (x
!= NULL_TREE
&& y
!= NULL_TREE
);
1458 if (TREE_CODE (x
) == ADDR_EXPR
1459 && TREE_CODE (y
) == ADDR_EXPR
1460 && TREE_CODE (TREE_OPERAND (x
, 0)) == CONST_DECL
1461 && TREE_CODE (TREE_OPERAND (y
, 0)) == CONST_DECL
)
1462 return operand_equal_p (DECL_INITIAL (TREE_OPERAND (x
, 0)),
1463 DECL_INITIAL (TREE_OPERAND (y
, 0)), 0);
1465 return operand_equal_p (x
, y
, 0);
1468 /* Return true iff X and Y should be considered equal contexts by IPA-CP. */
1471 values_equal_for_ipcp_p (ipa_polymorphic_call_context x
,
1472 ipa_polymorphic_call_context y
)
1474 return x
.equal_to (y
);
1478 /* Add a new value source to the value represented by THIS, marking that a
1479 value comes from edge CS and (if the underlying jump function is a
1480 pass-through or an ancestor one) from a caller value SRC_VAL of a caller
1481 parameter described by SRC_INDEX. OFFSET is negative if the source was the
1482 scalar value of the parameter itself or the offset within an aggregate. */
1484 template <typename valtype
>
1486 ipcp_value
<valtype
>::add_source (cgraph_edge
*cs
, ipcp_value
*src_val
,
1487 int src_idx
, HOST_WIDE_INT offset
)
1489 ipcp_value_source
<valtype
> *src
;
1491 src
= new (ipcp_sources_pool
.allocate ()) ipcp_value_source
<valtype
>;
1492 src
->offset
= offset
;
1495 src
->index
= src_idx
;
1497 src
->next
= sources
;
1501 /* Allocate a new ipcp_value holding a tree constant, initialize its value to
1502 SOURCE and clear all other fields. */
1504 static ipcp_value
<tree
> *
1505 allocate_and_init_ipcp_value (tree source
)
1507 ipcp_value
<tree
> *val
;
1509 val
= new (ipcp_cst_values_pool
.allocate ()) ipcp_value
<tree
>();
1510 val
->value
= source
;
1514 /* Allocate a new ipcp_value holding a polymorphic context, initialize its
1515 value to SOURCE and clear all other fields. */
1517 static ipcp_value
<ipa_polymorphic_call_context
> *
1518 allocate_and_init_ipcp_value (ipa_polymorphic_call_context source
)
1520 ipcp_value
<ipa_polymorphic_call_context
> *val
;
1523 val
= new (ipcp_poly_ctx_values_pool
.allocate ())
1524 ipcp_value
<ipa_polymorphic_call_context
>();
1525 val
->value
= source
;
1529 /* Try to add NEWVAL to LAT, potentially creating a new ipcp_value for it. CS,
1530 SRC_VAL SRC_INDEX and OFFSET are meant for add_source and have the same
1531 meaning. OFFSET -1 means the source is scalar and not a part of an
1534 template <typename valtype
>
1536 ipcp_lattice
<valtype
>::add_value (valtype newval
, cgraph_edge
*cs
,
1537 ipcp_value
<valtype
> *src_val
,
1538 int src_idx
, HOST_WIDE_INT offset
)
1540 ipcp_value
<valtype
> *val
;
1545 for (val
= values
; val
; val
= val
->next
)
1546 if (values_equal_for_ipcp_p (val
->value
, newval
))
1548 if (ipa_edge_within_scc (cs
))
1550 ipcp_value_source
<valtype
> *s
;
1551 for (s
= val
->sources
; s
; s
= s
->next
)
1558 val
->add_source (cs
, src_val
, src_idx
, offset
);
1562 if (values_count
== PARAM_VALUE (PARAM_IPA_CP_VALUE_LIST_SIZE
))
1564 /* We can only free sources, not the values themselves, because sources
1565 of other values in this SCC might point to them. */
1566 for (val
= values
; val
; val
= val
->next
)
1568 while (val
->sources
)
1570 ipcp_value_source
<valtype
> *src
= val
->sources
;
1571 val
->sources
= src
->next
;
1572 ipcp_sources_pool
.remove ((ipcp_value_source
<tree
>*)src
);
1577 return set_to_bottom ();
1581 val
= allocate_and_init_ipcp_value (newval
);
1582 val
->add_source (cs
, src_val
, src_idx
, offset
);
1588 /* Propagate values through a pass-through jump function JFUNC associated with
1589 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1590 is the index of the source parameter. PARM_TYPE is the type of the
1591 parameter to which the result is passed. */
1594 propagate_vals_across_pass_through (cgraph_edge
*cs
, ipa_jump_func
*jfunc
,
1595 ipcp_lattice
<tree
> *src_lat
,
1596 ipcp_lattice
<tree
> *dest_lat
, int src_idx
,
1599 ipcp_value
<tree
> *src_val
;
1602 /* Do not create new values when propagating within an SCC because if there
1603 are arithmetic functions with circular dependencies, there is infinite
1604 number of them and we would just make lattices bottom. If this condition
1605 is ever relaxed we have to detect self-feeding recursive calls in
1606 cgraph_edge_brings_value_p in a smarter way. */
1607 if ((ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1608 && ipa_edge_within_scc (cs
))
1609 ret
= dest_lat
->set_contains_variable ();
1611 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1613 tree cstval
= ipa_get_jf_pass_through_result (jfunc
, src_val
->value
,
1617 ret
|= dest_lat
->add_value (cstval
, cs
, src_val
, src_idx
);
1619 ret
|= dest_lat
->set_contains_variable ();
1625 /* Propagate values through an ancestor jump function JFUNC associated with
1626 edge CS, taking values from SRC_LAT and putting them into DEST_LAT. SRC_IDX
1627 is the index of the source parameter. */
1630 propagate_vals_across_ancestor (struct cgraph_edge
*cs
,
1631 struct ipa_jump_func
*jfunc
,
1632 ipcp_lattice
<tree
> *src_lat
,
1633 ipcp_lattice
<tree
> *dest_lat
, int src_idx
)
1635 ipcp_value
<tree
> *src_val
;
1638 if (ipa_edge_within_scc (cs
))
1639 return dest_lat
->set_contains_variable ();
1641 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1643 tree t
= ipa_get_jf_ancestor_result (jfunc
, src_val
->value
);
1646 ret
|= dest_lat
->add_value (t
, cs
, src_val
, src_idx
);
1648 ret
|= dest_lat
->set_contains_variable ();
1654 /* Propagate scalar values across jump function JFUNC that is associated with
1655 edge CS and put the values into DEST_LAT. PARM_TYPE is the type of the
1656 parameter to which the result is passed. */
1659 propagate_scalar_across_jump_function (struct cgraph_edge
*cs
,
1660 struct ipa_jump_func
*jfunc
,
1661 ipcp_lattice
<tree
> *dest_lat
,
1664 if (dest_lat
->bottom
)
1667 if (jfunc
->type
== IPA_JF_CONST
)
1669 tree val
= ipa_get_jf_constant (jfunc
);
1670 return dest_lat
->add_value (val
, cs
, NULL
, 0);
1672 else if (jfunc
->type
== IPA_JF_PASS_THROUGH
1673 || jfunc
->type
== IPA_JF_ANCESTOR
)
1675 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1676 ipcp_lattice
<tree
> *src_lat
;
1680 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1681 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1683 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1685 src_lat
= ipa_get_scalar_lat (caller_info
, src_idx
);
1686 if (src_lat
->bottom
)
1687 return dest_lat
->set_contains_variable ();
1689 /* If we would need to clone the caller and cannot, do not propagate. */
1690 if (!ipcp_versionable_function_p (cs
->caller
)
1691 && (src_lat
->contains_variable
1692 || (src_lat
->values_count
> 1)))
1693 return dest_lat
->set_contains_variable ();
1695 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1696 ret
= propagate_vals_across_pass_through (cs
, jfunc
, src_lat
,
1697 dest_lat
, src_idx
, param_type
);
1699 ret
= propagate_vals_across_ancestor (cs
, jfunc
, src_lat
, dest_lat
,
1702 if (src_lat
->contains_variable
)
1703 ret
|= dest_lat
->set_contains_variable ();
1708 /* TODO: We currently do not handle member method pointers in IPA-CP (we only
1709 use it for indirect inlining), we should propagate them too. */
1710 return dest_lat
->set_contains_variable ();
1713 /* Propagate scalar values across jump function JFUNC that is associated with
1714 edge CS and describes argument IDX and put the values into DEST_LAT. */
1717 propagate_context_across_jump_function (cgraph_edge
*cs
,
1718 ipa_jump_func
*jfunc
, int idx
,
1719 ipcp_lattice
<ipa_polymorphic_call_context
> *dest_lat
)
1721 ipa_edge_args
*args
= IPA_EDGE_REF (cs
);
1722 if (dest_lat
->bottom
)
1725 bool added_sth
= false;
1726 bool type_preserved
= true;
1728 ipa_polymorphic_call_context edge_ctx
, *edge_ctx_ptr
1729 = ipa_get_ith_polymorhic_call_context (args
, idx
);
1732 edge_ctx
= *edge_ctx_ptr
;
1734 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1735 || jfunc
->type
== IPA_JF_ANCESTOR
)
1737 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1739 ipcp_lattice
<ipa_polymorphic_call_context
> *src_lat
;
1741 /* TODO: Once we figure out how to propagate speculations, it will
1742 probably be a good idea to switch to speculation if type_preserved is
1743 not set instead of punting. */
1744 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1746 if (ipa_get_jf_pass_through_operation (jfunc
) != NOP_EXPR
)
1748 type_preserved
= ipa_get_jf_pass_through_type_preserved (jfunc
);
1749 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1753 type_preserved
= ipa_get_jf_ancestor_type_preserved (jfunc
);
1754 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1757 src_lat
= ipa_get_poly_ctx_lat (caller_info
, src_idx
);
1758 /* If we would need to clone the caller and cannot, do not propagate. */
1759 if (!ipcp_versionable_function_p (cs
->caller
)
1760 && (src_lat
->contains_variable
1761 || (src_lat
->values_count
> 1)))
1764 ipcp_value
<ipa_polymorphic_call_context
> *src_val
;
1765 for (src_val
= src_lat
->values
; src_val
; src_val
= src_val
->next
)
1767 ipa_polymorphic_call_context cur
= src_val
->value
;
1769 if (!type_preserved
)
1770 cur
.possible_dynamic_type_change (cs
->in_polymorphic_cdtor
);
1771 if (jfunc
->type
== IPA_JF_ANCESTOR
)
1772 cur
.offset_by (ipa_get_jf_ancestor_offset (jfunc
));
1773 /* TODO: In cases we know how the context is going to be used,
1774 we can improve the result by passing proper OTR_TYPE. */
1775 cur
.combine_with (edge_ctx
);
1776 if (!cur
.useless_p ())
1778 if (src_lat
->contains_variable
1779 && !edge_ctx
.equal_to (cur
))
1780 ret
|= dest_lat
->set_contains_variable ();
1781 ret
|= dest_lat
->add_value (cur
, cs
, src_val
, src_idx
);
1791 if (!edge_ctx
.useless_p ())
1792 ret
|= dest_lat
->add_value (edge_ctx
, cs
);
1794 ret
|= dest_lat
->set_contains_variable ();
1800 /* Propagate bits across jfunc that is associated with
1801 edge cs and update dest_lattice accordingly. */
1804 propagate_bits_across_jump_function (cgraph_edge
*cs
, int idx
,
1805 ipa_jump_func
*jfunc
,
1806 ipcp_bits_lattice
*dest_lattice
)
1808 if (dest_lattice
->bottom_p ())
1811 enum availability availability
;
1812 cgraph_node
*callee
= cs
->callee
->function_symbol (&availability
);
1813 struct ipa_node_params
*callee_info
= IPA_NODE_REF (callee
);
1814 tree parm_type
= ipa_get_type (callee_info
, idx
);
1816 /* For K&R C programs, ipa_get_type() could return NULL_TREE. Avoid the
1817 transform for these cases. Similarly, we can have bad type mismatches
1818 with LTO, avoid doing anything with those too. */
1820 || (!INTEGRAL_TYPE_P (parm_type
) && !POINTER_TYPE_P (parm_type
)))
1822 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1823 fprintf (dump_file
, "Setting dest_lattice to bottom, because type of "
1824 "param %i of %s is NULL or unsuitable for bits propagation\n",
1825 idx
, cs
->callee
->name ());
1827 return dest_lattice
->set_to_bottom ();
1830 unsigned precision
= TYPE_PRECISION (parm_type
);
1831 signop sgn
= TYPE_SIGN (parm_type
);
1833 if (jfunc
->type
== IPA_JF_PASS_THROUGH
1834 || jfunc
->type
== IPA_JF_ANCESTOR
)
1836 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1837 tree operand
= NULL_TREE
;
1838 enum tree_code code
;
1841 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1843 code
= ipa_get_jf_pass_through_operation (jfunc
);
1844 src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1845 if (code
!= NOP_EXPR
)
1846 operand
= ipa_get_jf_pass_through_operand (jfunc
);
1850 code
= POINTER_PLUS_EXPR
;
1851 src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
1852 unsigned HOST_WIDE_INT offset
= ipa_get_jf_ancestor_offset (jfunc
) / BITS_PER_UNIT
;
1853 operand
= build_int_cstu (size_type_node
, offset
);
1856 struct ipcp_param_lattices
*src_lats
1857 = ipa_get_parm_lattices (caller_info
, src_idx
);
1859 /* Try to propagate bits if src_lattice is bottom, but jfunc is known.
1865 Assume lattice for x is bottom, however we can still propagate
1866 result of x & 0xff == 0xff, which gets computed during ccp1 pass
1867 and we store it in jump function during analysis stage. */
1869 if (src_lats
->bits_lattice
.bottom_p ()
1871 return dest_lattice
->meet_with (jfunc
->bits
->value
, jfunc
->bits
->mask
,
1874 return dest_lattice
->meet_with (src_lats
->bits_lattice
, precision
, sgn
,
1878 else if (jfunc
->type
== IPA_JF_ANCESTOR
)
1879 return dest_lattice
->set_to_bottom ();
1880 else if (jfunc
->bits
)
1881 return dest_lattice
->meet_with (jfunc
->bits
->value
, jfunc
->bits
->mask
,
1884 return dest_lattice
->set_to_bottom ();
1887 /* Emulate effects of unary OPERATION and/or conversion from SRC_TYPE to
1888 DST_TYPE on value range in SRC_VR and store it to DST_VR. Return true if
1889 the result is a range or an anti-range. */
1892 ipa_vr_operation_and_type_effects (value_range
*dst_vr
, value_range
*src_vr
,
1893 enum tree_code operation
,
1894 tree dst_type
, tree src_type
)
1896 memset (dst_vr
, 0, sizeof (*dst_vr
));
1897 extract_range_from_unary_expr (dst_vr
, operation
, dst_type
, src_vr
, src_type
);
1898 if (dst_vr
->type
== VR_RANGE
|| dst_vr
->type
== VR_ANTI_RANGE
)
1904 /* Propagate value range across jump function JFUNC that is associated with
1905 edge CS with param of callee of PARAM_TYPE and update DEST_PLATS
1909 propagate_vr_across_jump_function (cgraph_edge
*cs
, ipa_jump_func
*jfunc
,
1910 struct ipcp_param_lattices
*dest_plats
,
1913 ipcp_vr_lattice
*dest_lat
= &dest_plats
->m_value_range
;
1915 if (dest_lat
->bottom_p ())
1919 || (!INTEGRAL_TYPE_P (param_type
)
1920 && !POINTER_TYPE_P (param_type
)))
1921 return dest_lat
->set_to_bottom ();
1923 if (jfunc
->type
== IPA_JF_PASS_THROUGH
)
1925 enum tree_code operation
= ipa_get_jf_pass_through_operation (jfunc
);
1927 if (TREE_CODE_CLASS (operation
) == tcc_unary
)
1929 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
1930 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
1931 tree operand_type
= ipa_get_type (caller_info
, src_idx
);
1932 struct ipcp_param_lattices
*src_lats
1933 = ipa_get_parm_lattices (caller_info
, src_idx
);
1935 if (src_lats
->m_value_range
.bottom_p ())
1936 return dest_lat
->set_to_bottom ();
1938 if (ipa_vr_operation_and_type_effects (&vr
,
1939 &src_lats
->m_value_range
.m_vr
,
1940 operation
, param_type
,
1942 return dest_lat
->meet_with (&vr
);
1945 else if (jfunc
->type
== IPA_JF_CONST
)
1947 tree val
= ipa_get_jf_constant (jfunc
);
1948 if (TREE_CODE (val
) == INTEGER_CST
)
1950 val
= fold_convert (param_type
, val
);
1951 if (TREE_OVERFLOW_P (val
))
1952 val
= drop_tree_overflow (val
);
1955 memset (&tmpvr
, 0, sizeof (tmpvr
));
1956 tmpvr
.type
= VR_RANGE
;
1959 return dest_lat
->meet_with (&tmpvr
);
1965 && ipa_vr_operation_and_type_effects (&vr
, jfunc
->m_vr
, NOP_EXPR
,
1967 TREE_TYPE (jfunc
->m_vr
->min
)))
1968 return dest_lat
->meet_with (&vr
);
1970 return dest_lat
->set_to_bottom ();
1973 /* If DEST_PLATS already has aggregate items, check that aggs_by_ref matches
1974 NEW_AGGS_BY_REF and if not, mark all aggs as bottoms and return true (in all
1975 other cases, return false). If there are no aggregate items, set
1976 aggs_by_ref to NEW_AGGS_BY_REF. */
1979 set_check_aggs_by_ref (struct ipcp_param_lattices
*dest_plats
,
1980 bool new_aggs_by_ref
)
1982 if (dest_plats
->aggs
)
1984 if (dest_plats
->aggs_by_ref
!= new_aggs_by_ref
)
1986 set_agg_lats_to_bottom (dest_plats
);
1991 dest_plats
->aggs_by_ref
= new_aggs_by_ref
;
1995 /* Walk aggregate lattices in DEST_PLATS from ***AGLAT on, until ***aglat is an
1996 already existing lattice for the given OFFSET and SIZE, marking all skipped
1997 lattices as containing variable and checking for overlaps. If there is no
1998 already existing lattice for the OFFSET and VAL_SIZE, create one, initialize
1999 it with offset, size and contains_variable to PRE_EXISTING, and return true,
2000 unless there are too many already. If there are two many, return false. If
2001 there are overlaps turn whole DEST_PLATS to bottom and return false. If any
2002 skipped lattices were newly marked as containing variable, set *CHANGE to
2006 merge_agg_lats_step (struct ipcp_param_lattices
*dest_plats
,
2007 HOST_WIDE_INT offset
, HOST_WIDE_INT val_size
,
2008 struct ipcp_agg_lattice
***aglat
,
2009 bool pre_existing
, bool *change
)
2011 gcc_checking_assert (offset
>= 0);
2013 while (**aglat
&& (**aglat
)->offset
< offset
)
2015 if ((**aglat
)->offset
+ (**aglat
)->size
> offset
)
2017 set_agg_lats_to_bottom (dest_plats
);
2020 *change
|= (**aglat
)->set_contains_variable ();
2021 *aglat
= &(**aglat
)->next
;
2024 if (**aglat
&& (**aglat
)->offset
== offset
)
2026 if ((**aglat
)->size
!= val_size
2028 && (**aglat
)->next
->offset
< offset
+ val_size
))
2030 set_agg_lats_to_bottom (dest_plats
);
2033 gcc_checking_assert (!(**aglat
)->next
2034 || (**aglat
)->next
->offset
>= offset
+ val_size
);
2039 struct ipcp_agg_lattice
*new_al
;
2041 if (**aglat
&& (**aglat
)->offset
< offset
+ val_size
)
2043 set_agg_lats_to_bottom (dest_plats
);
2046 if (dest_plats
->aggs_count
== PARAM_VALUE (PARAM_IPA_MAX_AGG_ITEMS
))
2048 dest_plats
->aggs_count
++;
2049 new_al
= ipcp_agg_lattice_pool
.allocate ();
2050 memset (new_al
, 0, sizeof (*new_al
));
2052 new_al
->offset
= offset
;
2053 new_al
->size
= val_size
;
2054 new_al
->contains_variable
= pre_existing
;
2056 new_al
->next
= **aglat
;
2062 /* Set all AGLAT and all other aggregate lattices reachable by next pointers as
2063 containing an unknown value. */
2066 set_chain_of_aglats_contains_variable (struct ipcp_agg_lattice
*aglat
)
2071 ret
|= aglat
->set_contains_variable ();
2072 aglat
= aglat
->next
;
2077 /* Merge existing aggregate lattices in SRC_PLATS to DEST_PLATS, subtracting
2078 DELTA_OFFSET. CS is the call graph edge and SRC_IDX the index of the source
2079 parameter used for lattice value sources. Return true if DEST_PLATS changed
2083 merge_aggregate_lattices (struct cgraph_edge
*cs
,
2084 struct ipcp_param_lattices
*dest_plats
,
2085 struct ipcp_param_lattices
*src_plats
,
2086 int src_idx
, HOST_WIDE_INT offset_delta
)
2088 bool pre_existing
= dest_plats
->aggs
!= NULL
;
2089 struct ipcp_agg_lattice
**dst_aglat
;
2092 if (set_check_aggs_by_ref (dest_plats
, src_plats
->aggs_by_ref
))
2094 if (src_plats
->aggs_bottom
)
2095 return set_agg_lats_contain_variable (dest_plats
);
2096 if (src_plats
->aggs_contain_variable
)
2097 ret
|= set_agg_lats_contain_variable (dest_plats
);
2098 dst_aglat
= &dest_plats
->aggs
;
2100 for (struct ipcp_agg_lattice
*src_aglat
= src_plats
->aggs
;
2102 src_aglat
= src_aglat
->next
)
2104 HOST_WIDE_INT new_offset
= src_aglat
->offset
- offset_delta
;
2108 if (merge_agg_lats_step (dest_plats
, new_offset
, src_aglat
->size
,
2109 &dst_aglat
, pre_existing
, &ret
))
2111 struct ipcp_agg_lattice
*new_al
= *dst_aglat
;
2113 dst_aglat
= &(*dst_aglat
)->next
;
2114 if (src_aglat
->bottom
)
2116 ret
|= new_al
->set_contains_variable ();
2119 if (src_aglat
->contains_variable
)
2120 ret
|= new_al
->set_contains_variable ();
2121 for (ipcp_value
<tree
> *val
= src_aglat
->values
;
2124 ret
|= new_al
->add_value (val
->value
, cs
, val
, src_idx
,
2127 else if (dest_plats
->aggs_bottom
)
2130 ret
|= set_chain_of_aglats_contains_variable (*dst_aglat
);
2134 /* Determine whether there is anything to propagate FROM SRC_PLATS through a
2135 pass-through JFUNC and if so, whether it has conform and conforms to the
2136 rules about propagating values passed by reference. */
2139 agg_pass_through_permissible_p (struct ipcp_param_lattices
*src_plats
,
2140 struct ipa_jump_func
*jfunc
)
2142 return src_plats
->aggs
2143 && (!src_plats
->aggs_by_ref
2144 || ipa_get_jf_pass_through_agg_preserved (jfunc
));
2147 /* Propagate scalar values across jump function JFUNC that is associated with
2148 edge CS and put the values into DEST_LAT. */
2151 propagate_aggs_across_jump_function (struct cgraph_edge
*cs
,
2152 struct ipa_jump_func
*jfunc
,
2153 struct ipcp_param_lattices
*dest_plats
)
2157 if (dest_plats
->aggs_bottom
)
2160 if (jfunc
->type
== IPA_JF_PASS_THROUGH
2161 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
2163 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2164 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
2165 struct ipcp_param_lattices
*src_plats
;
2167 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2168 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
2170 /* Currently we do not produce clobber aggregate jump
2171 functions, replace with merging when we do. */
2172 gcc_assert (!jfunc
->agg
.items
);
2173 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
,
2177 ret
|= set_agg_lats_contain_variable (dest_plats
);
2179 else if (jfunc
->type
== IPA_JF_ANCESTOR
2180 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
2182 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
2183 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
2184 struct ipcp_param_lattices
*src_plats
;
2186 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
2187 if (src_plats
->aggs
&& src_plats
->aggs_by_ref
)
2189 /* Currently we do not produce clobber aggregate jump
2190 functions, replace with merging when we do. */
2191 gcc_assert (!jfunc
->agg
.items
);
2192 ret
|= merge_aggregate_lattices (cs
, dest_plats
, src_plats
, src_idx
,
2193 ipa_get_jf_ancestor_offset (jfunc
));
2195 else if (!src_plats
->aggs_by_ref
)
2196 ret
|= set_agg_lats_to_bottom (dest_plats
);
2198 ret
|= set_agg_lats_contain_variable (dest_plats
);
2200 else if (jfunc
->agg
.items
)
2202 bool pre_existing
= dest_plats
->aggs
!= NULL
;
2203 struct ipcp_agg_lattice
**aglat
= &dest_plats
->aggs
;
2204 struct ipa_agg_jf_item
*item
;
2207 if (set_check_aggs_by_ref (dest_plats
, jfunc
->agg
.by_ref
))
2210 FOR_EACH_VEC_ELT (*jfunc
->agg
.items
, i
, item
)
2212 HOST_WIDE_INT val_size
;
2214 if (item
->offset
< 0)
2216 gcc_checking_assert (is_gimple_ip_invariant (item
->value
));
2217 val_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (item
->value
)));
2219 if (merge_agg_lats_step (dest_plats
, item
->offset
, val_size
,
2220 &aglat
, pre_existing
, &ret
))
2222 ret
|= (*aglat
)->add_value (item
->value
, cs
, NULL
, 0, 0);
2223 aglat
= &(*aglat
)->next
;
2225 else if (dest_plats
->aggs_bottom
)
2229 ret
|= set_chain_of_aglats_contains_variable (*aglat
);
2232 ret
|= set_agg_lats_contain_variable (dest_plats
);
2237 /* Return true if on the way cfrom CS->caller to the final (non-alias and
2238 non-thunk) destination, the call passes through a thunk. */
2241 call_passes_through_thunk_p (cgraph_edge
*cs
)
2243 cgraph_node
*alias_or_thunk
= cs
->callee
;
2244 while (alias_or_thunk
->alias
)
2245 alias_or_thunk
= alias_or_thunk
->get_alias_target ();
2246 return alias_or_thunk
->thunk
.thunk_p
;
2249 /* Propagate constants from the caller to the callee of CS. INFO describes the
2253 propagate_constants_across_call (struct cgraph_edge
*cs
)
2255 struct ipa_node_params
*callee_info
;
2256 enum availability availability
;
2257 cgraph_node
*callee
;
2258 struct ipa_edge_args
*args
;
2260 int i
, args_count
, parms_count
;
2262 callee
= cs
->callee
->function_symbol (&availability
);
2263 if (!callee
->definition
)
2265 gcc_checking_assert (callee
->has_gimple_body_p ());
2266 callee_info
= IPA_NODE_REF (callee
);
2268 args
= IPA_EDGE_REF (cs
);
2269 args_count
= ipa_get_cs_argument_count (args
);
2270 parms_count
= ipa_get_param_count (callee_info
);
2271 if (parms_count
== 0)
2274 /* No propagation through instrumentation thunks is available yet.
2275 It should be possible with proper mapping of call args and
2276 instrumented callee params in the propagation loop below. But
2277 this case mostly occurs when legacy code calls instrumented code
2278 and it is not a primary target for optimizations.
2279 We detect instrumentation thunks in aliases and thunks chain by
2280 checking instrumentation_clone flag for chain source and target.
2281 Going through instrumentation thunks we always have it changed
2282 from 0 to 1 and all other nodes do not change it. */
2283 if (!cs
->callee
->instrumentation_clone
2284 && callee
->instrumentation_clone
)
2286 for (i
= 0; i
< parms_count
; i
++)
2287 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
2292 /* If this call goes through a thunk we must not propagate to the first (0th)
2293 parameter. However, we might need to uncover a thunk from below a series
2294 of aliases first. */
2295 if (call_passes_through_thunk_p (cs
))
2297 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
,
2304 for (; (i
< args_count
) && (i
< parms_count
); i
++)
2306 struct ipa_jump_func
*jump_func
= ipa_get_ith_jump_func (args
, i
);
2307 struct ipcp_param_lattices
*dest_plats
;
2308 tree param_type
= ipa_get_type (callee_info
, i
);
2310 dest_plats
= ipa_get_parm_lattices (callee_info
, i
);
2311 if (availability
== AVAIL_INTERPOSABLE
)
2312 ret
|= set_all_contains_variable (dest_plats
);
2315 ret
|= propagate_scalar_across_jump_function (cs
, jump_func
,
2316 &dest_plats
->itself
,
2318 ret
|= propagate_context_across_jump_function (cs
, jump_func
, i
,
2319 &dest_plats
->ctxlat
);
2321 |= propagate_bits_across_jump_function (cs
, i
, jump_func
,
2322 &dest_plats
->bits_lattice
);
2323 ret
|= propagate_aggs_across_jump_function (cs
, jump_func
,
2325 if (opt_for_fn (callee
->decl
, flag_ipa_vrp
))
2326 ret
|= propagate_vr_across_jump_function (cs
, jump_func
,
2327 dest_plats
, param_type
);
2329 ret
|= dest_plats
->m_value_range
.set_to_bottom ();
2332 for (; i
< parms_count
; i
++)
2333 ret
|= set_all_contains_variable (ipa_get_parm_lattices (callee_info
, i
));
2338 /* If an indirect edge IE can be turned into a direct one based on KNOWN_VALS
2339 KNOWN_CONTEXTS, KNOWN_AGGS or AGG_REPS return the destination. The latter
2340 three can be NULL. If AGG_REPS is not NULL, KNOWN_AGGS is ignored. */
2343 ipa_get_indirect_edge_target_1 (struct cgraph_edge
*ie
,
2344 vec
<tree
> known_csts
,
2345 vec
<ipa_polymorphic_call_context
> known_contexts
,
2346 vec
<ipa_agg_jump_function_p
> known_aggs
,
2347 struct ipa_agg_replacement_value
*agg_reps
,
2350 int param_index
= ie
->indirect_info
->param_index
;
2351 HOST_WIDE_INT anc_offset
;
2355 *speculative
= false;
2357 if (param_index
== -1
2358 || known_csts
.length () <= (unsigned int) param_index
)
2361 if (!ie
->indirect_info
->polymorphic
)
2365 if (ie
->indirect_info
->agg_contents
)
2368 if (agg_reps
&& ie
->indirect_info
->guaranteed_unmodified
)
2372 if (agg_reps
->index
== param_index
2373 && agg_reps
->offset
== ie
->indirect_info
->offset
2374 && agg_reps
->by_ref
== ie
->indirect_info
->by_ref
)
2376 t
= agg_reps
->value
;
2379 agg_reps
= agg_reps
->next
;
2384 struct ipa_agg_jump_function
*agg
;
2385 if (known_aggs
.length () > (unsigned int) param_index
)
2386 agg
= known_aggs
[param_index
];
2389 bool from_global_constant
;
2390 t
= ipa_find_agg_cst_for_param (agg
, known_csts
[param_index
],
2391 ie
->indirect_info
->offset
,
2392 ie
->indirect_info
->by_ref
,
2393 &from_global_constant
);
2395 && !from_global_constant
2396 && !ie
->indirect_info
->guaranteed_unmodified
)
2401 t
= known_csts
[param_index
];
2404 && TREE_CODE (t
) == ADDR_EXPR
2405 && TREE_CODE (TREE_OPERAND (t
, 0)) == FUNCTION_DECL
)
2406 return TREE_OPERAND (t
, 0);
2411 if (!opt_for_fn (ie
->caller
->decl
, flag_devirtualize
))
2414 gcc_assert (!ie
->indirect_info
->agg_contents
);
2415 anc_offset
= ie
->indirect_info
->offset
;
2419 /* Try to work out value of virtual table pointer value in replacemnets. */
2420 if (!t
&& agg_reps
&& !ie
->indirect_info
->by_ref
)
2424 if (agg_reps
->index
== param_index
2425 && agg_reps
->offset
== ie
->indirect_info
->offset
2426 && agg_reps
->by_ref
)
2428 t
= agg_reps
->value
;
2431 agg_reps
= agg_reps
->next
;
2435 /* Try to work out value of virtual table pointer value in known
2436 aggregate values. */
2437 if (!t
&& known_aggs
.length () > (unsigned int) param_index
2438 && !ie
->indirect_info
->by_ref
)
2440 struct ipa_agg_jump_function
*agg
;
2441 agg
= known_aggs
[param_index
];
2442 t
= ipa_find_agg_cst_for_param (agg
, known_csts
[param_index
],
2443 ie
->indirect_info
->offset
, true);
2446 /* If we found the virtual table pointer, lookup the target. */
2450 unsigned HOST_WIDE_INT offset
;
2451 if (vtable_pointer_value_to_vtable (t
, &vtable
, &offset
))
2454 target
= gimple_get_virt_method_for_vtable (ie
->indirect_info
->otr_token
,
2455 vtable
, offset
, &can_refer
);
2459 || (TREE_CODE (TREE_TYPE (target
)) == FUNCTION_TYPE
2460 && DECL_FUNCTION_CODE (target
) == BUILT_IN_UNREACHABLE
)
2461 || !possible_polymorphic_call_target_p
2462 (ie
, cgraph_node::get (target
)))
2464 /* Do not speculate builtin_unreachable, it is stupid! */
2465 if (ie
->indirect_info
->vptr_changed
)
2467 target
= ipa_impossible_devirt_target (ie
, target
);
2469 *speculative
= ie
->indirect_info
->vptr_changed
;
2476 /* Do we know the constant value of pointer? */
2478 t
= known_csts
[param_index
];
2480 gcc_checking_assert (!t
|| TREE_CODE (t
) != TREE_BINFO
);
2482 ipa_polymorphic_call_context context
;
2483 if (known_contexts
.length () > (unsigned int) param_index
)
2485 context
= known_contexts
[param_index
];
2486 context
.offset_by (anc_offset
);
2487 if (ie
->indirect_info
->vptr_changed
)
2488 context
.possible_dynamic_type_change (ie
->in_polymorphic_cdtor
,
2489 ie
->indirect_info
->otr_type
);
2492 ipa_polymorphic_call_context ctx2
= ipa_polymorphic_call_context
2493 (t
, ie
->indirect_info
->otr_type
, anc_offset
);
2494 if (!ctx2
.useless_p ())
2495 context
.combine_with (ctx2
, ie
->indirect_info
->otr_type
);
2500 context
= ipa_polymorphic_call_context (t
, ie
->indirect_info
->otr_type
,
2502 if (ie
->indirect_info
->vptr_changed
)
2503 context
.possible_dynamic_type_change (ie
->in_polymorphic_cdtor
,
2504 ie
->indirect_info
->otr_type
);
2509 vec
<cgraph_node
*>targets
;
2512 targets
= possible_polymorphic_call_targets
2513 (ie
->indirect_info
->otr_type
,
2514 ie
->indirect_info
->otr_token
,
2516 if (!final
|| targets
.length () > 1)
2518 struct cgraph_node
*node
;
2521 if (!opt_for_fn (ie
->caller
->decl
, flag_devirtualize_speculatively
)
2522 || ie
->speculative
|| !ie
->maybe_hot_p ())
2524 node
= try_speculative_devirtualization (ie
->indirect_info
->otr_type
,
2525 ie
->indirect_info
->otr_token
,
2529 *speculative
= true;
2530 target
= node
->decl
;
2537 *speculative
= false;
2538 if (targets
.length () == 1)
2539 target
= targets
[0]->decl
;
2541 target
= ipa_impossible_devirt_target (ie
, NULL_TREE
);
2544 if (target
&& !possible_polymorphic_call_target_p (ie
,
2545 cgraph_node::get (target
)))
2549 target
= ipa_impossible_devirt_target (ie
, target
);
2556 /* If an indirect edge IE can be turned into a direct one based on KNOWN_CSTS,
2557 KNOWN_CONTEXTS (which can be vNULL) or KNOWN_AGGS (which also can be vNULL)
2558 return the destination. */
2561 ipa_get_indirect_edge_target (struct cgraph_edge
*ie
,
2562 vec
<tree
> known_csts
,
2563 vec
<ipa_polymorphic_call_context
> known_contexts
,
2564 vec
<ipa_agg_jump_function_p
> known_aggs
,
2567 return ipa_get_indirect_edge_target_1 (ie
, known_csts
, known_contexts
,
2568 known_aggs
, NULL
, speculative
);
2571 /* Calculate devirtualization time bonus for NODE, assuming we know KNOWN_CSTS
2572 and KNOWN_CONTEXTS. */
2575 devirtualization_time_bonus (struct cgraph_node
*node
,
2576 vec
<tree
> known_csts
,
2577 vec
<ipa_polymorphic_call_context
> known_contexts
,
2578 vec
<ipa_agg_jump_function_p
> known_aggs
)
2580 struct cgraph_edge
*ie
;
2583 for (ie
= node
->indirect_calls
; ie
; ie
= ie
->next_callee
)
2585 struct cgraph_node
*callee
;
2586 struct ipa_fn_summary
*isummary
;
2587 enum availability avail
;
2591 target
= ipa_get_indirect_edge_target (ie
, known_csts
, known_contexts
,
2592 known_aggs
, &speculative
);
2596 /* Only bare minimum benefit for clearly un-inlineable targets. */
2598 callee
= cgraph_node::get (target
);
2599 if (!callee
|| !callee
->definition
)
2601 callee
= callee
->function_symbol (&avail
);
2602 if (avail
< AVAIL_AVAILABLE
)
2604 isummary
= ipa_fn_summaries
->get (callee
);
2605 if (!isummary
->inlinable
)
2608 /* FIXME: The values below need re-considering and perhaps also
2609 integrating into the cost metrics, at lest in some very basic way. */
2610 if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 4)
2611 res
+= 31 / ((int)speculative
+ 1);
2612 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
/ 2)
2613 res
+= 15 / ((int)speculative
+ 1);
2614 else if (isummary
->size
<= MAX_INLINE_INSNS_AUTO
2615 || DECL_DECLARED_INLINE_P (callee
->decl
))
2616 res
+= 7 / ((int)speculative
+ 1);
2622 /* Return time bonus incurred because of HINTS. */
2625 hint_time_bonus (ipa_hints hints
)
2628 if (hints
& (INLINE_HINT_loop_iterations
| INLINE_HINT_loop_stride
))
2629 result
+= PARAM_VALUE (PARAM_IPA_CP_LOOP_HINT_BONUS
);
2630 if (hints
& INLINE_HINT_array_index
)
2631 result
+= PARAM_VALUE (PARAM_IPA_CP_ARRAY_INDEX_HINT_BONUS
);
2635 /* If there is a reason to penalize the function described by INFO in the
2636 cloning goodness evaluation, do so. */
2638 static inline int64_t
2639 incorporate_penalties (ipa_node_params
*info
, int64_t evaluation
)
2641 if (info
->node_within_scc
)
2642 evaluation
= (evaluation
2643 * (100 - PARAM_VALUE (PARAM_IPA_CP_RECURSION_PENALTY
))) / 100;
2645 if (info
->node_calling_single_call
)
2646 evaluation
= (evaluation
2647 * (100 - PARAM_VALUE (PARAM_IPA_CP_SINGLE_CALL_PENALTY
)))
2653 /* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
2654 and SIZE_COST and with the sum of frequencies of incoming edges to the
2655 potential new clone in FREQUENCIES. */
2658 good_cloning_opportunity_p (struct cgraph_node
*node
, int time_benefit
,
2659 int freq_sum
, profile_count count_sum
, int size_cost
)
2661 if (time_benefit
== 0
2662 || !opt_for_fn (node
->decl
, flag_ipa_cp_clone
)
2663 || node
->optimize_for_size_p ())
2666 gcc_assert (size_cost
> 0);
2668 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2669 if (max_count
> profile_count::zero ())
2671 int factor
= RDIV (count_sum
.probability_in
2672 (max_count
).to_reg_br_prob_base ()
2673 * 1000, REG_BR_PROB_BASE
);
2674 int64_t evaluation
= (((int64_t) time_benefit
* factor
)
2676 evaluation
= incorporate_penalties (info
, evaluation
);
2678 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2680 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
2681 "size: %i, count_sum: ", time_benefit
, size_cost
);
2682 count_sum
.dump (dump_file
);
2683 fprintf (dump_file
, "%s%s) -> evaluation: " "%" PRId64
2684 ", threshold: %i\n",
2685 info
->node_within_scc
? ", scc" : "",
2686 info
->node_calling_single_call
? ", single_call" : "",
2687 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
2690 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
2694 int64_t evaluation
= (((int64_t) time_benefit
* freq_sum
)
2696 evaluation
= incorporate_penalties (info
, evaluation
);
2698 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2699 fprintf (dump_file
, " good_cloning_opportunity_p (time: %i, "
2700 "size: %i, freq_sum: %i%s%s) -> evaluation: "
2701 "%" PRId64
", threshold: %i\n",
2702 time_benefit
, size_cost
, freq_sum
,
2703 info
->node_within_scc
? ", scc" : "",
2704 info
->node_calling_single_call
? ", single_call" : "",
2705 evaluation
, PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
));
2707 return evaluation
>= PARAM_VALUE (PARAM_IPA_CP_EVAL_THRESHOLD
);
2711 /* Return all context independent values from aggregate lattices in PLATS in a
2712 vector. Return NULL if there are none. */
2714 static vec
<ipa_agg_jf_item
, va_gc
> *
2715 context_independent_aggregate_values (struct ipcp_param_lattices
*plats
)
2717 vec
<ipa_agg_jf_item
, va_gc
> *res
= NULL
;
2719 if (plats
->aggs_bottom
2720 || plats
->aggs_contain_variable
2721 || plats
->aggs_count
== 0)
2724 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
;
2726 aglat
= aglat
->next
)
2727 if (aglat
->is_single_const ())
2729 struct ipa_agg_jf_item item
;
2730 item
.offset
= aglat
->offset
;
2731 item
.value
= aglat
->values
->value
;
2732 vec_safe_push (res
, item
);
2737 /* Allocate KNOWN_CSTS, KNOWN_CONTEXTS and, if non-NULL, KNOWN_AGGS and
2738 populate them with values of parameters that are known independent of the
2739 context. INFO describes the function. If REMOVABLE_PARAMS_COST is
2740 non-NULL, the movement cost of all removable parameters will be stored in
2744 gather_context_independent_values (struct ipa_node_params
*info
,
2745 vec
<tree
> *known_csts
,
2746 vec
<ipa_polymorphic_call_context
>
2748 vec
<ipa_agg_jump_function
> *known_aggs
,
2749 int *removable_params_cost
)
2751 int i
, count
= ipa_get_param_count (info
);
2754 known_csts
->create (0);
2755 known_contexts
->create (0);
2756 known_csts
->safe_grow_cleared (count
);
2757 known_contexts
->safe_grow_cleared (count
);
2760 known_aggs
->create (0);
2761 known_aggs
->safe_grow_cleared (count
);
2764 if (removable_params_cost
)
2765 *removable_params_cost
= 0;
2767 for (i
= 0; i
< count
; i
++)
2769 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2770 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
2772 if (lat
->is_single_const ())
2774 ipcp_value
<tree
> *val
= lat
->values
;
2775 gcc_checking_assert (TREE_CODE (val
->value
) != TREE_BINFO
);
2776 (*known_csts
)[i
] = val
->value
;
2777 if (removable_params_cost
)
2778 *removable_params_cost
2779 += estimate_move_cost (TREE_TYPE (val
->value
), false);
2782 else if (removable_params_cost
2783 && !ipa_is_param_used (info
, i
))
2784 *removable_params_cost
2785 += ipa_get_param_move_cost (info
, i
);
2787 if (!ipa_is_param_used (info
, i
))
2790 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
2791 /* Do not account known context as reason for cloning. We can see
2792 if it permits devirtualization. */
2793 if (ctxlat
->is_single_const ())
2794 (*known_contexts
)[i
] = ctxlat
->values
->value
;
2798 vec
<ipa_agg_jf_item
, va_gc
> *agg_items
;
2799 struct ipa_agg_jump_function
*ajf
;
2801 agg_items
= context_independent_aggregate_values (plats
);
2802 ajf
= &(*known_aggs
)[i
];
2803 ajf
->items
= agg_items
;
2804 ajf
->by_ref
= plats
->aggs_by_ref
;
2805 ret
|= agg_items
!= NULL
;
2812 /* The current interface in ipa-inline-analysis requires a pointer vector.
2815 FIXME: That interface should be re-worked, this is slightly silly. Still,
2816 I'd like to discuss how to change it first and this demonstrates the
2819 static vec
<ipa_agg_jump_function_p
>
2820 agg_jmp_p_vec_for_t_vec (vec
<ipa_agg_jump_function
> known_aggs
)
2822 vec
<ipa_agg_jump_function_p
> ret
;
2823 struct ipa_agg_jump_function
*ajf
;
2826 ret
.create (known_aggs
.length ());
2827 FOR_EACH_VEC_ELT (known_aggs
, i
, ajf
)
2828 ret
.quick_push (ajf
);
2832 /* Perform time and size measurement of NODE with the context given in
2833 KNOWN_CSTS, KNOWN_CONTEXTS and KNOWN_AGGS, calculate the benefit and cost
2834 given BASE_TIME of the node without specialization, REMOVABLE_PARAMS_COST of
2835 all context-independent removable parameters and EST_MOVE_COST of estimated
2836 movement of the considered parameter and store it into VAL. */
2839 perform_estimation_of_a_value (cgraph_node
*node
, vec
<tree
> known_csts
,
2840 vec
<ipa_polymorphic_call_context
> known_contexts
,
2841 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
,
2842 int removable_params_cost
,
2843 int est_move_cost
, ipcp_value_base
*val
)
2845 int size
, time_benefit
;
2846 sreal time
, base_time
;
2849 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_contexts
,
2850 known_aggs_ptrs
, &size
, &time
,
2851 &base_time
, &hints
);
2853 if (base_time
> 65535)
2856 /* Extern inline functions have no cloning local time benefits because they
2857 will be inlined anyway. The only reason to clone them is if it enables
2858 optimization in any of the functions they call. */
2859 if (DECL_EXTERNAL (node
->decl
) && DECL_DECLARED_INLINE_P (node
->decl
))
2862 time_benefit
= base_time
.to_int ()
2863 + devirtualization_time_bonus (node
, known_csts
, known_contexts
,
2865 + hint_time_bonus (hints
)
2866 + removable_params_cost
+ est_move_cost
;
2868 gcc_checking_assert (size
>=0);
2869 /* The inliner-heuristics based estimates may think that in certain
2870 contexts some functions do not have any size at all but we want
2871 all specializations to have at least a tiny cost, not least not to
2876 val
->local_time_benefit
= time_benefit
;
2877 val
->local_size_cost
= size
;
2880 /* Iterate over known values of parameters of NODE and estimate the local
2881 effects in terms of time and size they have. */
2884 estimate_local_effects (struct cgraph_node
*node
)
2886 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
2887 int i
, count
= ipa_get_param_count (info
);
2888 vec
<tree
> known_csts
;
2889 vec
<ipa_polymorphic_call_context
> known_contexts
;
2890 vec
<ipa_agg_jump_function
> known_aggs
;
2891 vec
<ipa_agg_jump_function_p
> known_aggs_ptrs
;
2893 int removable_params_cost
;
2895 if (!count
|| !ipcp_versionable_function_p (node
))
2898 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2899 fprintf (dump_file
, "\nEstimating effects for %s.\n", node
->dump_name ());
2901 always_const
= gather_context_independent_values (info
, &known_csts
,
2902 &known_contexts
, &known_aggs
,
2903 &removable_params_cost
);
2904 known_aggs_ptrs
= agg_jmp_p_vec_for_t_vec (known_aggs
);
2905 int devirt_bonus
= devirtualization_time_bonus (node
, known_csts
,
2906 known_contexts
, known_aggs_ptrs
);
2907 if (always_const
|| devirt_bonus
2908 || (removable_params_cost
&& node
->local
.can_change_signature
))
2910 struct caller_statistics stats
;
2912 sreal time
, base_time
;
2915 init_caller_stats (&stats
);
2916 node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
2918 estimate_ipcp_clone_size_and_time (node
, known_csts
, known_contexts
,
2919 known_aggs_ptrs
, &size
, &time
,
2920 &base_time
, &hints
);
2921 time
-= devirt_bonus
;
2922 time
-= hint_time_bonus (hints
);
2923 time
-= removable_params_cost
;
2924 size
-= stats
.n_calls
* removable_params_cost
;
2927 fprintf (dump_file
, " - context independent values, size: %i, "
2928 "time_benefit: %f\n", size
, (base_time
- time
).to_double ());
2930 if (size
<= 0 || node
->local
.local
)
2932 info
->do_clone_for_all_contexts
= true;
2935 fprintf (dump_file
, " Decided to specialize for all "
2936 "known contexts, code not going to grow.\n");
2938 else if (good_cloning_opportunity_p (node
,
2939 MAX ((base_time
- time
).to_int (),
2941 stats
.freq_sum
, stats
.count_sum
,
2944 if (size
+ overall_size
<= max_new_size
)
2946 info
->do_clone_for_all_contexts
= true;
2947 overall_size
+= size
;
2950 fprintf (dump_file
, " Decided to specialize for all "
2951 "known contexts, growth deemed beneficial.\n");
2953 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2954 fprintf (dump_file
, " Not cloning for all contexts because "
2955 "max_new_size would be reached with %li.\n",
2956 size
+ overall_size
);
2958 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2959 fprintf (dump_file
, " Not cloning for all contexts because "
2960 "!good_cloning_opportunity_p.\n");
2964 for (i
= 0; i
< count
; i
++)
2966 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
2967 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
2968 ipcp_value
<tree
> *val
;
2975 for (val
= lat
->values
; val
; val
= val
->next
)
2977 gcc_checking_assert (TREE_CODE (val
->value
) != TREE_BINFO
);
2978 known_csts
[i
] = val
->value
;
2980 int emc
= estimate_move_cost (TREE_TYPE (val
->value
), true);
2981 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
2983 removable_params_cost
, emc
, val
);
2985 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2987 fprintf (dump_file
, " - estimates for value ");
2988 print_ipcp_constant_value (dump_file
, val
->value
);
2989 fprintf (dump_file
, " for ");
2990 ipa_dump_param (dump_file
, info
, i
);
2991 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
2992 val
->local_time_benefit
, val
->local_size_cost
);
2995 known_csts
[i
] = NULL_TREE
;
2998 for (i
= 0; i
< count
; i
++)
3000 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3002 if (!plats
->virt_call
)
3005 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
3006 ipcp_value
<ipa_polymorphic_call_context
> *val
;
3010 || !known_contexts
[i
].useless_p ())
3013 for (val
= ctxlat
->values
; val
; val
= val
->next
)
3015 known_contexts
[i
] = val
->value
;
3016 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
3018 removable_params_cost
, 0, val
);
3020 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3022 fprintf (dump_file
, " - estimates for polymorphic context ");
3023 print_ipcp_constant_value (dump_file
, val
->value
);
3024 fprintf (dump_file
, " for ");
3025 ipa_dump_param (dump_file
, info
, i
);
3026 fprintf (dump_file
, ": time_benefit: %i, size: %i\n",
3027 val
->local_time_benefit
, val
->local_size_cost
);
3030 known_contexts
[i
] = ipa_polymorphic_call_context ();
3033 for (i
= 0; i
< count
; i
++)
3035 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3036 struct ipa_agg_jump_function
*ajf
;
3037 struct ipcp_agg_lattice
*aglat
;
3039 if (plats
->aggs_bottom
|| !plats
->aggs
)
3042 ajf
= &known_aggs
[i
];
3043 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3045 ipcp_value
<tree
> *val
;
3046 if (aglat
->bottom
|| !aglat
->values
3047 /* If the following is true, the one value is in known_aggs. */
3048 || (!plats
->aggs_contain_variable
3049 && aglat
->is_single_const ()))
3052 for (val
= aglat
->values
; val
; val
= val
->next
)
3054 struct ipa_agg_jf_item item
;
3056 item
.offset
= aglat
->offset
;
3057 item
.value
= val
->value
;
3058 vec_safe_push (ajf
->items
, item
);
3060 perform_estimation_of_a_value (node
, known_csts
, known_contexts
,
3062 removable_params_cost
, 0, val
);
3064 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3066 fprintf (dump_file
, " - estimates for value ");
3067 print_ipcp_constant_value (dump_file
, val
->value
);
3068 fprintf (dump_file
, " for ");
3069 ipa_dump_param (dump_file
, info
, i
);
3070 fprintf (dump_file
, "[%soffset: " HOST_WIDE_INT_PRINT_DEC
3071 "]: time_benefit: %i, size: %i\n",
3072 plats
->aggs_by_ref
? "ref " : "",
3074 val
->local_time_benefit
, val
->local_size_cost
);
3082 for (i
= 0; i
< count
; i
++)
3083 vec_free (known_aggs
[i
].items
);
3085 known_csts
.release ();
3086 known_contexts
.release ();
3087 known_aggs
.release ();
3088 known_aggs_ptrs
.release ();
3092 /* Add value CUR_VAL and all yet-unsorted values it is dependent on to the
3093 topological sort of values. */
3095 template <typename valtype
>
3097 value_topo_info
<valtype
>::add_val (ipcp_value
<valtype
> *cur_val
)
3099 ipcp_value_source
<valtype
> *src
;
3105 cur_val
->dfs
= dfs_counter
;
3106 cur_val
->low_link
= dfs_counter
;
3108 cur_val
->topo_next
= stack
;
3110 cur_val
->on_stack
= true;
3112 for (src
= cur_val
->sources
; src
; src
= src
->next
)
3115 if (src
->val
->dfs
== 0)
3118 if (src
->val
->low_link
< cur_val
->low_link
)
3119 cur_val
->low_link
= src
->val
->low_link
;
3121 else if (src
->val
->on_stack
3122 && src
->val
->dfs
< cur_val
->low_link
)
3123 cur_val
->low_link
= src
->val
->dfs
;
3126 if (cur_val
->dfs
== cur_val
->low_link
)
3128 ipcp_value
<valtype
> *v
, *scc_list
= NULL
;
3133 stack
= v
->topo_next
;
3134 v
->on_stack
= false;
3136 v
->scc_next
= scc_list
;
3139 while (v
!= cur_val
);
3141 cur_val
->topo_next
= values_topo
;
3142 values_topo
= cur_val
;
3146 /* Add all values in lattices associated with NODE to the topological sort if
3147 they are not there yet. */
3150 add_all_node_vals_to_toposort (cgraph_node
*node
, ipa_topo_info
*topo
)
3152 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3153 int i
, count
= ipa_get_param_count (info
);
3155 for (i
= 0; i
< count
; i
++)
3157 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
3158 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
3159 struct ipcp_agg_lattice
*aglat
;
3163 ipcp_value
<tree
> *val
;
3164 for (val
= lat
->values
; val
; val
= val
->next
)
3165 topo
->constants
.add_val (val
);
3168 if (!plats
->aggs_bottom
)
3169 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3172 ipcp_value
<tree
> *val
;
3173 for (val
= aglat
->values
; val
; val
= val
->next
)
3174 topo
->constants
.add_val (val
);
3177 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
3178 if (!ctxlat
->bottom
)
3180 ipcp_value
<ipa_polymorphic_call_context
> *ctxval
;
3181 for (ctxval
= ctxlat
->values
; ctxval
; ctxval
= ctxval
->next
)
3182 topo
->contexts
.add_val (ctxval
);
3187 /* One pass of constants propagation along the call graph edges, from callers
3188 to callees (requires topological ordering in TOPO), iterate over strongly
3189 connected components. */
3192 propagate_constants_topo (struct ipa_topo_info
*topo
)
3196 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
3199 struct cgraph_node
*v
, *node
= topo
->order
[i
];
3200 vec
<cgraph_node
*> cycle_nodes
= ipa_get_nodes_in_cycle (node
);
3202 /* First, iteratively propagate within the strongly connected component
3203 until all lattices stabilize. */
3204 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
3205 if (v
->has_gimple_body_p ())
3206 push_node_to_stack (topo
, v
);
3208 v
= pop_node_from_stack (topo
);
3211 struct cgraph_edge
*cs
;
3213 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
3214 if (ipa_edge_within_scc (cs
))
3216 IPA_NODE_REF (v
)->node_within_scc
= true;
3217 if (propagate_constants_across_call (cs
))
3218 push_node_to_stack (topo
, cs
->callee
->function_symbol ());
3220 v
= pop_node_from_stack (topo
);
3223 /* Afterwards, propagate along edges leading out of the SCC, calculates
3224 the local effects of the discovered constants and all valid values to
3225 their topological sort. */
3226 FOR_EACH_VEC_ELT (cycle_nodes
, j
, v
)
3227 if (v
->has_gimple_body_p ())
3229 struct cgraph_edge
*cs
;
3231 estimate_local_effects (v
);
3232 add_all_node_vals_to_toposort (v
, topo
);
3233 for (cs
= v
->callees
; cs
; cs
= cs
->next_callee
)
3234 if (!ipa_edge_within_scc (cs
))
3235 propagate_constants_across_call (cs
);
3237 cycle_nodes
.release ();
3242 /* Return the sum of A and B if none of them is bigger than INT_MAX/2, return
3243 the bigger one if otherwise. */
3246 safe_add (int a
, int b
)
3248 if (a
> INT_MAX
/2 || b
> INT_MAX
/2)
3249 return a
> b
? a
: b
;
3255 /* Propagate the estimated effects of individual values along the topological
3256 from the dependent values to those they depend on. */
3258 template <typename valtype
>
3260 value_topo_info
<valtype
>::propagate_effects ()
3262 ipcp_value
<valtype
> *base
;
3264 for (base
= values_topo
; base
; base
= base
->topo_next
)
3266 ipcp_value_source
<valtype
> *src
;
3267 ipcp_value
<valtype
> *val
;
3268 int time
= 0, size
= 0;
3270 for (val
= base
; val
; val
= val
->scc_next
)
3272 time
= safe_add (time
,
3273 val
->local_time_benefit
+ val
->prop_time_benefit
);
3274 size
= safe_add (size
, val
->local_size_cost
+ val
->prop_size_cost
);
3277 for (val
= base
; val
; val
= val
->scc_next
)
3278 for (src
= val
->sources
; src
; src
= src
->next
)
3280 && src
->cs
->maybe_hot_p ())
3282 src
->val
->prop_time_benefit
= safe_add (time
,
3283 src
->val
->prop_time_benefit
);
3284 src
->val
->prop_size_cost
= safe_add (size
,
3285 src
->val
->prop_size_cost
);
3291 /* Propagate constants, polymorphic contexts and their effects from the
3292 summaries interprocedurally. */
3295 ipcp_propagate_stage (struct ipa_topo_info
*topo
)
3297 struct cgraph_node
*node
;
3300 fprintf (dump_file
, "\n Propagating constants:\n\n");
3302 max_count
= profile_count::uninitialized ();
3304 FOR_EACH_DEFINED_FUNCTION (node
)
3306 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3308 determine_versionability (node
, info
);
3309 if (node
->has_gimple_body_p ())
3311 info
->lattices
= XCNEWVEC (struct ipcp_param_lattices
,
3312 ipa_get_param_count (info
));
3313 initialize_node_lattices (node
);
3315 if (node
->definition
&& !node
->alias
)
3316 overall_size
+= ipa_fn_summaries
->get (node
)->self_size
;
3317 max_count
= max_count
.max (node
->count
.ipa ());
3320 max_new_size
= overall_size
;
3321 if (max_new_size
< PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
))
3322 max_new_size
= PARAM_VALUE (PARAM_LARGE_UNIT_INSNS
);
3323 max_new_size
+= max_new_size
* PARAM_VALUE (PARAM_IPCP_UNIT_GROWTH
) / 100 + 1;
3326 fprintf (dump_file
, "\noverall_size: %li, max_new_size: %li\n",
3327 overall_size
, max_new_size
);
3329 propagate_constants_topo (topo
);
3331 ipcp_verify_propagated_values ();
3332 topo
->constants
.propagate_effects ();
3333 topo
->contexts
.propagate_effects ();
3337 fprintf (dump_file
, "\nIPA lattices after all propagation:\n");
3338 print_all_lattices (dump_file
, (dump_flags
& TDF_DETAILS
), true);
3342 /* Discover newly direct outgoing edges from NODE which is a new clone with
3343 known KNOWN_CSTS and make them direct. */
3346 ipcp_discover_new_direct_edges (struct cgraph_node
*node
,
3347 vec
<tree
> known_csts
,
3348 vec
<ipa_polymorphic_call_context
>
3350 struct ipa_agg_replacement_value
*aggvals
)
3352 struct cgraph_edge
*ie
, *next_ie
;
3355 for (ie
= node
->indirect_calls
; ie
; ie
= next_ie
)
3360 next_ie
= ie
->next_callee
;
3361 target
= ipa_get_indirect_edge_target_1 (ie
, known_csts
, known_contexts
,
3362 vNULL
, aggvals
, &speculative
);
3365 bool agg_contents
= ie
->indirect_info
->agg_contents
;
3366 bool polymorphic
= ie
->indirect_info
->polymorphic
;
3367 int param_index
= ie
->indirect_info
->param_index
;
3368 struct cgraph_edge
*cs
= ipa_make_edge_direct_to_target (ie
, target
,
3372 if (cs
&& !agg_contents
&& !polymorphic
)
3374 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3375 int c
= ipa_get_controlled_uses (info
, param_index
);
3376 if (c
!= IPA_UNDESCRIBED_USE
)
3378 struct ipa_ref
*to_del
;
3381 ipa_set_controlled_uses (info
, param_index
, c
);
3382 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3383 fprintf (dump_file
, " controlled uses count of param "
3384 "%i bumped down to %i\n", param_index
, c
);
3386 && (to_del
= node
->find_reference (cs
->callee
, NULL
, 0)))
3388 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3389 fprintf (dump_file
, " and even removing its "
3390 "cloning-created reference\n");
3391 to_del
->remove_reference ();
3397 /* Turning calls to direct calls will improve overall summary. */
3399 ipa_update_overall_fn_summary (node
);
3402 /* Vector of pointers which for linked lists of clones of an original crgaph
3405 static vec
<cgraph_edge
*> next_edge_clone
;
3406 static vec
<cgraph_edge
*> prev_edge_clone
;
3409 grow_edge_clone_vectors (void)
3411 if (next_edge_clone
.length ()
3412 <= (unsigned) symtab
->edges_max_uid
)
3413 next_edge_clone
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
3414 if (prev_edge_clone
.length ()
3415 <= (unsigned) symtab
->edges_max_uid
)
3416 prev_edge_clone
.safe_grow_cleared (symtab
->edges_max_uid
+ 1);
3419 /* Edge duplication hook to grow the appropriate linked list in
3423 ipcp_edge_duplication_hook (struct cgraph_edge
*src
, struct cgraph_edge
*dst
,
3426 grow_edge_clone_vectors ();
3428 struct cgraph_edge
*old_next
= next_edge_clone
[src
->uid
];
3430 prev_edge_clone
[old_next
->uid
] = dst
;
3431 prev_edge_clone
[dst
->uid
] = src
;
3433 next_edge_clone
[dst
->uid
] = old_next
;
3434 next_edge_clone
[src
->uid
] = dst
;
3437 /* Hook that is called by cgraph.c when an edge is removed. */
3440 ipcp_edge_removal_hook (struct cgraph_edge
*cs
, void *)
3442 grow_edge_clone_vectors ();
3444 struct cgraph_edge
*prev
= prev_edge_clone
[cs
->uid
];
3445 struct cgraph_edge
*next
= next_edge_clone
[cs
->uid
];
3447 next_edge_clone
[prev
->uid
] = next
;
3449 prev_edge_clone
[next
->uid
] = prev
;
3452 /* See if NODE is a clone with a known aggregate value at a given OFFSET of a
3453 parameter with the given INDEX. */
3456 get_clone_agg_value (struct cgraph_node
*node
, HOST_WIDE_INT offset
,
3459 struct ipa_agg_replacement_value
*aggval
;
3461 aggval
= ipa_get_agg_replacements_for_node (node
);
3464 if (aggval
->offset
== offset
3465 && aggval
->index
== index
)
3466 return aggval
->value
;
3467 aggval
= aggval
->next
;
3472 /* Return true is NODE is DEST or its clone for all contexts. */
3475 same_node_or_its_all_contexts_clone_p (cgraph_node
*node
, cgraph_node
*dest
)
3480 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3481 return info
->is_all_contexts_clone
&& info
->ipcp_orig_node
== dest
;
3484 /* Return true if edge CS does bring about the value described by SRC to
3485 DEST_VAL of node DEST or its clone for all contexts. */
3488 cgraph_edge_brings_value_p (cgraph_edge
*cs
, ipcp_value_source
<tree
> *src
,
3489 cgraph_node
*dest
, ipcp_value
<tree
> *dest_val
)
3491 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3492 enum availability availability
;
3493 cgraph_node
*real_dest
= cs
->callee
->function_symbol (&availability
);
3495 if (!same_node_or_its_all_contexts_clone_p (real_dest
, dest
)
3496 || availability
<= AVAIL_INTERPOSABLE
3497 || caller_info
->node_dead
)
3503 if (caller_info
->ipcp_orig_node
)
3506 if (src
->offset
== -1)
3507 t
= caller_info
->known_csts
[src
->index
];
3509 t
= get_clone_agg_value (cs
->caller
, src
->offset
, src
->index
);
3510 return (t
!= NULL_TREE
3511 && values_equal_for_ipcp_p (src
->val
->value
, t
));
3515 /* At the moment we do not propagate over arithmetic jump functions in
3516 SCCs, so it is safe to detect self-feeding recursive calls in this
3518 if (src
->val
== dest_val
)
3521 struct ipcp_agg_lattice
*aglat
;
3522 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
3524 if (src
->offset
== -1)
3525 return (plats
->itself
.is_single_const ()
3526 && values_equal_for_ipcp_p (src
->val
->value
,
3527 plats
->itself
.values
->value
));
3530 if (plats
->aggs_bottom
|| plats
->aggs_contain_variable
)
3532 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
3533 if (aglat
->offset
== src
->offset
)
3534 return (aglat
->is_single_const ()
3535 && values_equal_for_ipcp_p (src
->val
->value
,
3536 aglat
->values
->value
));
3542 /* Return true if edge CS does bring about the value described by SRC to
3543 DST_VAL of node DEST or its clone for all contexts. */
3546 cgraph_edge_brings_value_p (cgraph_edge
*cs
,
3547 ipcp_value_source
<ipa_polymorphic_call_context
> *src
,
3549 ipcp_value
<ipa_polymorphic_call_context
> *)
3551 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
3552 cgraph_node
*real_dest
= cs
->callee
->function_symbol ();
3554 if (!same_node_or_its_all_contexts_clone_p (real_dest
, dest
)
3555 || caller_info
->node_dead
)
3560 if (caller_info
->ipcp_orig_node
)
3561 return (caller_info
->known_contexts
.length () > (unsigned) src
->index
)
3562 && values_equal_for_ipcp_p (src
->val
->value
,
3563 caller_info
->known_contexts
[src
->index
]);
3565 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (caller_info
,
3567 return plats
->ctxlat
.is_single_const ()
3568 && values_equal_for_ipcp_p (src
->val
->value
,
3569 plats
->ctxlat
.values
->value
);
3572 /* Get the next clone in the linked list of clones of an edge. */
3574 static inline struct cgraph_edge
*
3575 get_next_cgraph_edge_clone (struct cgraph_edge
*cs
)
3577 return next_edge_clone
[cs
->uid
];
3580 /* Given VAL that is intended for DEST, iterate over all its sources and if any
3581 of them is viable and hot, return true. In that case, for those that still
3582 hold, add their edge frequency and their number into *FREQUENCY and
3583 *CALLER_COUNT respectively. */
3585 template <typename valtype
>
3587 get_info_about_necessary_edges (ipcp_value
<valtype
> *val
, cgraph_node
*dest
,
3589 profile_count
*count_sum
, int *caller_count
)
3591 ipcp_value_source
<valtype
> *src
;
3592 int freq
= 0, count
= 0;
3593 profile_count cnt
= profile_count::zero ();
3595 bool non_self_recursive
= false;
3597 for (src
= val
->sources
; src
; src
= src
->next
)
3599 struct cgraph_edge
*cs
= src
->cs
;
3602 if (cgraph_edge_brings_value_p (cs
, src
, dest
, val
))
3605 freq
+= cs
->frequency ();
3606 if (cs
->count
.ipa ().initialized_p ())
3607 cnt
+= cs
->count
.ipa ();
3608 hot
|= cs
->maybe_hot_p ();
3609 if (cs
->caller
!= dest
)
3610 non_self_recursive
= true;
3612 cs
= get_next_cgraph_edge_clone (cs
);
3616 /* If the only edges bringing a value are self-recursive ones, do not bother
3618 if (!non_self_recursive
)
3623 *caller_count
= count
;
3627 /* Return a vector of incoming edges that do bring value VAL to node DEST. It
3628 is assumed their number is known and equal to CALLER_COUNT. */
3630 template <typename valtype
>
3631 static vec
<cgraph_edge
*>
3632 gather_edges_for_value (ipcp_value
<valtype
> *val
, cgraph_node
*dest
,
3635 ipcp_value_source
<valtype
> *src
;
3636 vec
<cgraph_edge
*> ret
;
3638 ret
.create (caller_count
);
3639 for (src
= val
->sources
; src
; src
= src
->next
)
3641 struct cgraph_edge
*cs
= src
->cs
;
3644 if (cgraph_edge_brings_value_p (cs
, src
, dest
, val
))
3645 ret
.quick_push (cs
);
3646 cs
= get_next_cgraph_edge_clone (cs
);
3653 /* Construct a replacement map for a know VALUE for a formal parameter PARAM.
3654 Return it or NULL if for some reason it cannot be created. */
3656 static struct ipa_replace_map
*
3657 get_replacement_map (struct ipa_node_params
*info
, tree value
, int parm_num
)
3659 struct ipa_replace_map
*replace_map
;
3662 replace_map
= ggc_alloc
<ipa_replace_map
> ();
3665 fprintf (dump_file
, " replacing ");
3666 ipa_dump_param (dump_file
, info
, parm_num
);
3668 fprintf (dump_file
, " with const ");
3669 print_generic_expr (dump_file
, value
);
3670 fprintf (dump_file
, "\n");
3672 replace_map
->old_tree
= NULL
;
3673 replace_map
->parm_num
= parm_num
;
3674 replace_map
->new_tree
= value
;
3675 replace_map
->replace_p
= true;
3676 replace_map
->ref_p
= false;
3681 /* Dump new profiling counts */
3684 dump_profile_updates (struct cgraph_node
*orig_node
,
3685 struct cgraph_node
*new_node
)
3687 struct cgraph_edge
*cs
;
3689 fprintf (dump_file
, " setting count of the specialized node to ");
3690 new_node
->count
.dump (dump_file
);
3691 fprintf (dump_file
, "\n");
3692 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3694 fprintf (dump_file
, " edge to %s has count ",
3695 cs
->callee
->name ());
3696 cs
->count
.dump (dump_file
);
3697 fprintf (dump_file
, "\n");
3700 fprintf (dump_file
, " setting count of the original node to ");
3701 orig_node
->count
.dump (dump_file
);
3702 fprintf (dump_file
, "\n");
3703 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3705 fprintf (dump_file
, " edge to %s is left with ",
3706 cs
->callee
->name ());
3707 cs
->count
.dump (dump_file
);
3708 fprintf (dump_file
, "\n");
3712 /* After a specialized NEW_NODE version of ORIG_NODE has been created, update
3713 their profile information to reflect this. */
3716 update_profiling_info (struct cgraph_node
*orig_node
,
3717 struct cgraph_node
*new_node
)
3719 struct cgraph_edge
*cs
;
3720 struct caller_statistics stats
;
3721 profile_count new_sum
, orig_sum
;
3722 profile_count remainder
, orig_node_count
= orig_node
->count
;
3724 if (!(orig_node_count
.ipa () > profile_count::zero ()))
3727 init_caller_stats (&stats
);
3728 orig_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
3730 orig_sum
= stats
.count_sum
;
3731 init_caller_stats (&stats
);
3732 new_node
->call_for_symbol_thunks_and_aliases (gather_caller_stats
, &stats
,
3734 new_sum
= stats
.count_sum
;
3736 if (orig_node_count
< orig_sum
+ new_sum
)
3740 fprintf (dump_file
, " Problem: node %s has too low count ",
3741 orig_node
->dump_name ());
3742 orig_node_count
.dump (dump_file
);
3743 fprintf (dump_file
, "while the sum of incoming count is ");
3744 (orig_sum
+ new_sum
).dump (dump_file
);
3745 fprintf (dump_file
, "\n");
3748 orig_node_count
= (orig_sum
+ new_sum
).apply_scale (12, 10);
3751 fprintf (dump_file
, " proceeding by pretending it was ");
3752 orig_node_count
.dump (dump_file
);
3753 fprintf (dump_file
, "\n");
3757 remainder
= orig_node_count
.combine_with_ipa_count (orig_node_count
.ipa ()
3759 new_sum
= orig_node_count
.combine_with_ipa_count (new_sum
);
3760 orig_node
->count
= remainder
;
3762 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3763 cs
->count
= cs
->count
.apply_scale (new_sum
, orig_node_count
);
3765 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3766 cs
->count
= cs
->count
.apply_scale (remainder
, orig_node_count
);
3769 dump_profile_updates (orig_node
, new_node
);
3772 /* Update the respective profile of specialized NEW_NODE and the original
3773 ORIG_NODE after additional edges with cumulative count sum REDIRECTED_SUM
3774 have been redirected to the specialized version. */
3777 update_specialized_profile (struct cgraph_node
*new_node
,
3778 struct cgraph_node
*orig_node
,
3779 profile_count redirected_sum
)
3781 struct cgraph_edge
*cs
;
3782 profile_count new_node_count
, orig_node_count
= orig_node
->count
;
3786 fprintf (dump_file
, " the sum of counts of redirected edges is ");
3787 redirected_sum
.dump (dump_file
);
3788 fprintf (dump_file
, "\n");
3790 if (!(orig_node_count
> profile_count::zero ()))
3793 gcc_assert (orig_node_count
>= redirected_sum
);
3795 new_node_count
= new_node
->count
;
3796 new_node
->count
+= redirected_sum
;
3797 orig_node
->count
-= redirected_sum
;
3799 for (cs
= new_node
->callees
; cs
; cs
= cs
->next_callee
)
3800 cs
->count
+= cs
->count
.apply_scale (redirected_sum
, new_node_count
);
3802 for (cs
= orig_node
->callees
; cs
; cs
= cs
->next_callee
)
3804 profile_count dec
= cs
->count
.apply_scale (redirected_sum
,
3810 dump_profile_updates (orig_node
, new_node
);
3813 /* Create a specialized version of NODE with known constants in KNOWN_CSTS,
3814 known contexts in KNOWN_CONTEXTS and known aggregate values in AGGVALS and
3815 redirect all edges in CALLERS to it. */
3817 static struct cgraph_node
*
3818 create_specialized_node (struct cgraph_node
*node
,
3819 vec
<tree
> known_csts
,
3820 vec
<ipa_polymorphic_call_context
> known_contexts
,
3821 struct ipa_agg_replacement_value
*aggvals
,
3822 vec
<cgraph_edge
*> callers
)
3824 struct ipa_node_params
*new_info
, *info
= IPA_NODE_REF (node
);
3825 vec
<ipa_replace_map
*, va_gc
> *replace_trees
= NULL
;
3826 struct ipa_agg_replacement_value
*av
;
3827 struct cgraph_node
*new_node
;
3828 int i
, count
= ipa_get_param_count (info
);
3829 bitmap args_to_skip
;
3831 gcc_assert (!info
->ipcp_orig_node
);
3833 if (node
->local
.can_change_signature
)
3835 args_to_skip
= BITMAP_GGC_ALLOC ();
3836 for (i
= 0; i
< count
; i
++)
3838 tree t
= known_csts
[i
];
3840 if (t
|| !ipa_is_param_used (info
, i
))
3841 bitmap_set_bit (args_to_skip
, i
);
3846 args_to_skip
= NULL
;
3847 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3848 fprintf (dump_file
, " cannot change function signature\n");
3851 for (i
= 0; i
< count
; i
++)
3853 tree t
= known_csts
[i
];
3856 struct ipa_replace_map
*replace_map
;
3858 gcc_checking_assert (TREE_CODE (t
) != TREE_BINFO
);
3859 replace_map
= get_replacement_map (info
, t
, i
);
3861 vec_safe_push (replace_trees
, replace_map
);
3864 auto_vec
<cgraph_edge
*, 2> self_recursive_calls
;
3865 for (i
= callers
.length () - 1; i
>= 0; i
--)
3867 cgraph_edge
*cs
= callers
[i
];
3868 if (cs
->caller
== node
)
3870 self_recursive_calls
.safe_push (cs
);
3871 callers
.unordered_remove (i
);
3875 new_node
= node
->create_virtual_clone (callers
, replace_trees
,
3876 args_to_skip
, "constprop");
3878 bool have_self_recursive_calls
= !self_recursive_calls
.is_empty ();
3879 for (unsigned j
= 0; j
< self_recursive_calls
.length (); j
++)
3881 cgraph_edge
*cs
= next_edge_clone
[self_recursive_calls
[j
]->uid
];
3882 /* Cloned edges can disappear during cloning as speculation can be
3883 resolved, check that we have one and that it comes from the last
3885 if (cs
&& cs
->caller
== new_node
)
3886 cs
->redirect_callee_duplicating_thunks (new_node
);
3887 /* Any future code that would make more than one clone of an outgoing
3888 edge would confuse this mechanism, so let's check that does not
3890 gcc_checking_assert (!cs
3891 || !next_edge_clone
[cs
->uid
]
3892 || next_edge_clone
[cs
->uid
]->caller
!= new_node
);
3894 if (have_self_recursive_calls
)
3895 new_node
->expand_all_artificial_thunks ();
3897 ipa_set_node_agg_value_chain (new_node
, aggvals
);
3898 for (av
= aggvals
; av
; av
= av
->next
)
3899 new_node
->maybe_create_reference (av
->value
, NULL
);
3901 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3903 fprintf (dump_file
, " the new node is %s.\n", new_node
->dump_name ());
3904 if (known_contexts
.exists ())
3906 for (i
= 0; i
< count
; i
++)
3907 if (!known_contexts
[i
].useless_p ())
3909 fprintf (dump_file
, " known ctx %i is ", i
);
3910 known_contexts
[i
].dump (dump_file
);
3914 ipa_dump_agg_replacement_values (dump_file
, aggvals
);
3916 ipa_check_create_node_params ();
3917 update_profiling_info (node
, new_node
);
3918 new_info
= IPA_NODE_REF (new_node
);
3919 new_info
->ipcp_orig_node
= node
;
3920 new_info
->known_csts
= known_csts
;
3921 new_info
->known_contexts
= known_contexts
;
3923 ipcp_discover_new_direct_edges (new_node
, known_csts
, known_contexts
, aggvals
);
3929 /* Return true, if JFUNC, which describes a i-th parameter of call CS, is a
3930 simple no-operation pass-through function to itself. */
3933 self_recursive_pass_through_p (cgraph_edge
*cs
, ipa_jump_func
*jfunc
, int i
)
3935 enum availability availability
;
3936 if (cs
->caller
== cs
->callee
->function_symbol (&availability
)
3937 && availability
> AVAIL_INTERPOSABLE
3938 && jfunc
->type
== IPA_JF_PASS_THROUGH
3939 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
3940 && ipa_get_jf_pass_through_formal_id (jfunc
) == i
)
3945 /* Given a NODE, and a subset of its CALLERS, try to populate blanks slots in
3946 KNOWN_CSTS with constants that are also known for all of the CALLERS. */
3949 find_more_scalar_values_for_callers_subset (struct cgraph_node
*node
,
3950 vec
<tree
> known_csts
,
3951 vec
<cgraph_edge
*> callers
)
3953 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
3954 int i
, count
= ipa_get_param_count (info
);
3956 for (i
= 0; i
< count
; i
++)
3958 struct cgraph_edge
*cs
;
3959 tree newval
= NULL_TREE
;
3962 tree type
= ipa_get_type (info
, i
);
3964 if (ipa_get_scalar_lat (info
, i
)->bottom
|| known_csts
[i
])
3967 FOR_EACH_VEC_ELT (callers
, j
, cs
)
3969 struct ipa_jump_func
*jump_func
;
3972 if (IPA_NODE_REF (cs
->caller
)->node_dead
)
3975 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
))
3977 && call_passes_through_thunk_p (cs
))
3978 || (!cs
->callee
->instrumentation_clone
3979 && cs
->callee
->function_symbol ()->instrumentation_clone
))
3984 jump_func
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
3985 if (self_recursive_pass_through_p (cs
, jump_func
, i
))
3988 t
= ipa_value_from_jfunc (IPA_NODE_REF (cs
->caller
), jump_func
, type
);
3991 && !values_equal_for_ipcp_p (t
, newval
))
3992 || (!first
&& !newval
))
4004 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4006 fprintf (dump_file
, " adding an extra known scalar value ");
4007 print_ipcp_constant_value (dump_file
, newval
);
4008 fprintf (dump_file
, " for ");
4009 ipa_dump_param (dump_file
, info
, i
);
4010 fprintf (dump_file
, "\n");
4013 known_csts
[i
] = newval
;
4018 /* Given a NODE and a subset of its CALLERS, try to populate plank slots in
4019 KNOWN_CONTEXTS with polymorphic contexts that are also known for all of the
4023 find_more_contexts_for_caller_subset (cgraph_node
*node
,
4024 vec
<ipa_polymorphic_call_context
>
4026 vec
<cgraph_edge
*> callers
)
4028 ipa_node_params
*info
= IPA_NODE_REF (node
);
4029 int i
, count
= ipa_get_param_count (info
);
4031 for (i
= 0; i
< count
; i
++)
4035 if (ipa_get_poly_ctx_lat (info
, i
)->bottom
4036 || (known_contexts
->exists ()
4037 && !(*known_contexts
)[i
].useless_p ()))
4040 ipa_polymorphic_call_context newval
;
4044 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4046 if (i
>= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
)))
4048 ipa_jump_func
*jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
),
4050 ipa_polymorphic_call_context ctx
;
4051 ctx
= ipa_context_from_jfunc (IPA_NODE_REF (cs
->caller
), cs
, i
,
4059 newval
.meet_with (ctx
);
4060 if (newval
.useless_p ())
4064 if (!newval
.useless_p ())
4066 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4068 fprintf (dump_file
, " adding an extra known polymorphic "
4070 print_ipcp_constant_value (dump_file
, newval
);
4071 fprintf (dump_file
, " for ");
4072 ipa_dump_param (dump_file
, info
, i
);
4073 fprintf (dump_file
, "\n");
4076 if (!known_contexts
->exists ())
4077 known_contexts
->safe_grow_cleared (ipa_get_param_count (info
));
4078 (*known_contexts
)[i
] = newval
;
4084 /* Go through PLATS and create a vector of values consisting of values and
4085 offsets (minus OFFSET) of lattices that contain only a single value. */
4087 static vec
<ipa_agg_jf_item
>
4088 copy_plats_to_inter (struct ipcp_param_lattices
*plats
, HOST_WIDE_INT offset
)
4090 vec
<ipa_agg_jf_item
> res
= vNULL
;
4092 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
4095 for (struct ipcp_agg_lattice
*aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
4096 if (aglat
->is_single_const ())
4098 struct ipa_agg_jf_item ti
;
4099 ti
.offset
= aglat
->offset
- offset
;
4100 ti
.value
= aglat
->values
->value
;
4106 /* Intersect all values in INTER with single value lattices in PLATS (while
4107 subtracting OFFSET). */
4110 intersect_with_plats (struct ipcp_param_lattices
*plats
,
4111 vec
<ipa_agg_jf_item
> *inter
,
4112 HOST_WIDE_INT offset
)
4114 struct ipcp_agg_lattice
*aglat
;
4115 struct ipa_agg_jf_item
*item
;
4118 if (!plats
->aggs
|| plats
->aggs_contain_variable
|| plats
->aggs_bottom
)
4124 aglat
= plats
->aggs
;
4125 FOR_EACH_VEC_ELT (*inter
, k
, item
)
4132 if (aglat
->offset
- offset
> item
->offset
)
4134 if (aglat
->offset
- offset
== item
->offset
)
4136 gcc_checking_assert (item
->value
);
4137 if (aglat
->is_single_const ()
4138 && values_equal_for_ipcp_p (item
->value
,
4139 aglat
->values
->value
))
4143 aglat
= aglat
->next
;
4146 item
->value
= NULL_TREE
;
4150 /* Copy aggregate replacement values of NODE (which is an IPA-CP clone) to the
4151 vector result while subtracting OFFSET from the individual value offsets. */
4153 static vec
<ipa_agg_jf_item
>
4154 agg_replacements_to_vector (struct cgraph_node
*node
, int index
,
4155 HOST_WIDE_INT offset
)
4157 struct ipa_agg_replacement_value
*av
;
4158 vec
<ipa_agg_jf_item
> res
= vNULL
;
4160 for (av
= ipa_get_agg_replacements_for_node (node
); av
; av
= av
->next
)
4161 if (av
->index
== index
4162 && (av
->offset
- offset
) >= 0)
4164 struct ipa_agg_jf_item item
;
4165 gcc_checking_assert (av
->value
);
4166 item
.offset
= av
->offset
- offset
;
4167 item
.value
= av
->value
;
4168 res
.safe_push (item
);
4174 /* Intersect all values in INTER with those that we have already scheduled to
4175 be replaced in parameter number INDEX of NODE, which is an IPA-CP clone
4176 (while subtracting OFFSET). */
4179 intersect_with_agg_replacements (struct cgraph_node
*node
, int index
,
4180 vec
<ipa_agg_jf_item
> *inter
,
4181 HOST_WIDE_INT offset
)
4183 struct ipa_agg_replacement_value
*srcvals
;
4184 struct ipa_agg_jf_item
*item
;
4187 srcvals
= ipa_get_agg_replacements_for_node (node
);
4194 FOR_EACH_VEC_ELT (*inter
, i
, item
)
4196 struct ipa_agg_replacement_value
*av
;
4200 for (av
= srcvals
; av
; av
= av
->next
)
4202 gcc_checking_assert (av
->value
);
4203 if (av
->index
== index
4204 && av
->offset
- offset
== item
->offset
)
4206 if (values_equal_for_ipcp_p (item
->value
, av
->value
))
4212 item
->value
= NULL_TREE
;
4216 /* Intersect values in INTER with aggregate values that come along edge CS to
4217 parameter number INDEX and return it. If INTER does not actually exist yet,
4218 copy all incoming values to it. If we determine we ended up with no values
4219 whatsoever, return a released vector. */
4221 static vec
<ipa_agg_jf_item
>
4222 intersect_aggregates_with_edge (struct cgraph_edge
*cs
, int index
,
4223 vec
<ipa_agg_jf_item
> inter
)
4225 struct ipa_jump_func
*jfunc
;
4226 jfunc
= ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), index
);
4227 if (jfunc
->type
== IPA_JF_PASS_THROUGH
4228 && ipa_get_jf_pass_through_operation (jfunc
) == NOP_EXPR
)
4230 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
4231 int src_idx
= ipa_get_jf_pass_through_formal_id (jfunc
);
4233 if (caller_info
->ipcp_orig_node
)
4235 struct cgraph_node
*orig_node
= caller_info
->ipcp_orig_node
;
4236 struct ipcp_param_lattices
*orig_plats
;
4237 orig_plats
= ipa_get_parm_lattices (IPA_NODE_REF (orig_node
),
4239 if (agg_pass_through_permissible_p (orig_plats
, jfunc
))
4241 if (!inter
.exists ())
4242 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, 0);
4244 intersect_with_agg_replacements (cs
->caller
, src_idx
,
4255 struct ipcp_param_lattices
*src_plats
;
4256 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
4257 if (agg_pass_through_permissible_p (src_plats
, jfunc
))
4259 /* Currently we do not produce clobber aggregate jump
4260 functions, adjust when we do. */
4261 gcc_checking_assert (!jfunc
->agg
.items
);
4262 if (!inter
.exists ())
4263 inter
= copy_plats_to_inter (src_plats
, 0);
4265 intersect_with_plats (src_plats
, &inter
, 0);
4274 else if (jfunc
->type
== IPA_JF_ANCESTOR
4275 && ipa_get_jf_ancestor_agg_preserved (jfunc
))
4277 struct ipa_node_params
*caller_info
= IPA_NODE_REF (cs
->caller
);
4278 int src_idx
= ipa_get_jf_ancestor_formal_id (jfunc
);
4279 struct ipcp_param_lattices
*src_plats
;
4280 HOST_WIDE_INT delta
= ipa_get_jf_ancestor_offset (jfunc
);
4282 if (caller_info
->ipcp_orig_node
)
4284 if (!inter
.exists ())
4285 inter
= agg_replacements_to_vector (cs
->caller
, src_idx
, delta
);
4287 intersect_with_agg_replacements (cs
->caller
, src_idx
, &inter
,
4292 src_plats
= ipa_get_parm_lattices (caller_info
, src_idx
);
4293 /* Currently we do not produce clobber aggregate jump
4294 functions, adjust when we do. */
4295 gcc_checking_assert (!src_plats
->aggs
|| !jfunc
->agg
.items
);
4296 if (!inter
.exists ())
4297 inter
= copy_plats_to_inter (src_plats
, delta
);
4299 intersect_with_plats (src_plats
, &inter
, delta
);
4302 else if (jfunc
->agg
.items
)
4304 struct ipa_agg_jf_item
*item
;
4307 if (!inter
.exists ())
4308 for (unsigned i
= 0; i
< jfunc
->agg
.items
->length (); i
++)
4309 inter
.safe_push ((*jfunc
->agg
.items
)[i
]);
4311 FOR_EACH_VEC_ELT (inter
, k
, item
)
4319 while ((unsigned) l
< jfunc
->agg
.items
->length ())
4321 struct ipa_agg_jf_item
*ti
;
4322 ti
= &(*jfunc
->agg
.items
)[l
];
4323 if (ti
->offset
> item
->offset
)
4325 if (ti
->offset
== item
->offset
)
4327 gcc_checking_assert (ti
->value
);
4328 if (values_equal_for_ipcp_p (item
->value
,
4342 return vec
<ipa_agg_jf_item
>();
4347 /* Look at edges in CALLERS and collect all known aggregate values that arrive
4348 from all of them. */
4350 static struct ipa_agg_replacement_value
*
4351 find_aggregate_values_for_callers_subset (struct cgraph_node
*node
,
4352 vec
<cgraph_edge
*> callers
)
4354 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
4355 struct ipa_agg_replacement_value
*res
;
4356 struct ipa_agg_replacement_value
**tail
= &res
;
4357 struct cgraph_edge
*cs
;
4358 int i
, j
, count
= ipa_get_param_count (dest_info
);
4360 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4362 int c
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
4367 for (i
= 0; i
< count
; i
++)
4369 struct cgraph_edge
*cs
;
4370 vec
<ipa_agg_jf_item
> inter
= vNULL
;
4371 struct ipa_agg_jf_item
*item
;
4372 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (dest_info
, i
);
4375 /* Among other things, the following check should deal with all by_ref
4377 if (plats
->aggs_bottom
)
4380 FOR_EACH_VEC_ELT (callers
, j
, cs
)
4382 struct ipa_jump_func
*jfunc
4383 = ipa_get_ith_jump_func (IPA_EDGE_REF (cs
), i
);
4384 if (self_recursive_pass_through_p (cs
, jfunc
, i
)
4385 && (!plats
->aggs_by_ref
4386 || ipa_get_jf_pass_through_agg_preserved (jfunc
)))
4388 inter
= intersect_aggregates_with_edge (cs
, i
, inter
);
4390 if (!inter
.exists ())
4394 FOR_EACH_VEC_ELT (inter
, j
, item
)
4396 struct ipa_agg_replacement_value
*v
;
4401 v
= ggc_alloc
<ipa_agg_replacement_value
> ();
4403 v
->offset
= item
->offset
;
4404 v
->value
= item
->value
;
4405 v
->by_ref
= plats
->aggs_by_ref
;
4411 if (inter
.exists ())
4418 /* Determine whether CS also brings all scalar values that the NODE is
4422 cgraph_edge_brings_all_scalars_for_node (struct cgraph_edge
*cs
,
4423 struct cgraph_node
*node
)
4425 struct ipa_node_params
*dest_info
= IPA_NODE_REF (node
);
4426 int count
= ipa_get_param_count (dest_info
);
4427 struct ipa_node_params
*caller_info
;
4428 struct ipa_edge_args
*args
;
4431 caller_info
= IPA_NODE_REF (cs
->caller
);
4432 args
= IPA_EDGE_REF (cs
);
4433 for (i
= 0; i
< count
; i
++)
4435 struct ipa_jump_func
*jump_func
;
4438 val
= dest_info
->known_csts
[i
];
4442 if (i
>= ipa_get_cs_argument_count (args
))
4444 jump_func
= ipa_get_ith_jump_func (args
, i
);
4445 t
= ipa_value_from_jfunc (caller_info
, jump_func
,
4446 ipa_get_type (dest_info
, i
));
4447 if (!t
|| !values_equal_for_ipcp_p (val
, t
))
4453 /* Determine whether CS also brings all aggregate values that NODE is
4456 cgraph_edge_brings_all_agg_vals_for_node (struct cgraph_edge
*cs
,
4457 struct cgraph_node
*node
)
4459 struct ipa_node_params
*orig_caller_info
= IPA_NODE_REF (cs
->caller
);
4460 struct ipa_node_params
*orig_node_info
;
4461 struct ipa_agg_replacement_value
*aggval
;
4464 aggval
= ipa_get_agg_replacements_for_node (node
);
4468 count
= ipa_get_param_count (IPA_NODE_REF (node
));
4469 ec
= ipa_get_cs_argument_count (IPA_EDGE_REF (cs
));
4471 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4472 if (aggval
->index
>= ec
)
4475 orig_node_info
= IPA_NODE_REF (IPA_NODE_REF (node
)->ipcp_orig_node
);
4476 if (orig_caller_info
->ipcp_orig_node
)
4477 orig_caller_info
= IPA_NODE_REF (orig_caller_info
->ipcp_orig_node
);
4479 for (i
= 0; i
< count
; i
++)
4481 static vec
<ipa_agg_jf_item
> values
= vec
<ipa_agg_jf_item
>();
4482 struct ipcp_param_lattices
*plats
;
4483 bool interesting
= false;
4484 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4485 if (aggval
->index
== i
)
4493 plats
= ipa_get_parm_lattices (orig_node_info
, aggval
->index
);
4494 if (plats
->aggs_bottom
)
4497 values
= intersect_aggregates_with_edge (cs
, i
, values
);
4498 if (!values
.exists ())
4501 for (struct ipa_agg_replacement_value
*av
= aggval
; av
; av
= av
->next
)
4502 if (aggval
->index
== i
)
4504 struct ipa_agg_jf_item
*item
;
4507 FOR_EACH_VEC_ELT (values
, j
, item
)
4509 && item
->offset
== av
->offset
4510 && values_equal_for_ipcp_p (item
->value
, av
->value
))
4525 /* Given an original NODE and a VAL for which we have already created a
4526 specialized clone, look whether there are incoming edges that still lead
4527 into the old node but now also bring the requested value and also conform to
4528 all other criteria such that they can be redirected the special node.
4529 This function can therefore redirect the final edge in a SCC. */
4531 template <typename valtype
>
4533 perhaps_add_new_callers (cgraph_node
*node
, ipcp_value
<valtype
> *val
)
4535 ipcp_value_source
<valtype
> *src
;
4536 profile_count redirected_sum
= profile_count::zero ();
4538 for (src
= val
->sources
; src
; src
= src
->next
)
4540 struct cgraph_edge
*cs
= src
->cs
;
4543 if (cgraph_edge_brings_value_p (cs
, src
, node
, val
)
4544 && cgraph_edge_brings_all_scalars_for_node (cs
, val
->spec_node
)
4545 && cgraph_edge_brings_all_agg_vals_for_node (cs
, val
->spec_node
))
4548 fprintf (dump_file
, " - adding an extra caller %s of %s\n",
4549 cs
->caller
->dump_name (),
4550 val
->spec_node
->dump_name ());
4552 cs
->redirect_callee_duplicating_thunks (val
->spec_node
);
4553 val
->spec_node
->expand_all_artificial_thunks ();
4554 if (cs
->count
.ipa ().initialized_p ())
4555 redirected_sum
= redirected_sum
+ cs
->count
.ipa ();
4557 cs
= get_next_cgraph_edge_clone (cs
);
4561 if (redirected_sum
.nonzero_p ())
4562 update_specialized_profile (val
->spec_node
, node
, redirected_sum
);
4565 /* Return true if KNOWN_CONTEXTS contain at least one useful context. */
4568 known_contexts_useful_p (vec
<ipa_polymorphic_call_context
> known_contexts
)
4570 ipa_polymorphic_call_context
*ctx
;
4573 FOR_EACH_VEC_ELT (known_contexts
, i
, ctx
)
4574 if (!ctx
->useless_p ())
4579 /* Return a copy of KNOWN_CSTS if it is not empty, otherwise return vNULL. */
4581 static vec
<ipa_polymorphic_call_context
>
4582 copy_useful_known_contexts (vec
<ipa_polymorphic_call_context
> known_contexts
)
4584 if (known_contexts_useful_p (known_contexts
))
4585 return known_contexts
.copy ();
4590 /* Copy KNOWN_CSTS and modify the copy according to VAL and INDEX. If
4591 non-empty, replace KNOWN_CONTEXTS with its copy too. */
4594 modify_known_vectors_with_val (vec
<tree
> *known_csts
,
4595 vec
<ipa_polymorphic_call_context
> *known_contexts
,
4596 ipcp_value
<tree
> *val
,
4599 *known_csts
= known_csts
->copy ();
4600 *known_contexts
= copy_useful_known_contexts (*known_contexts
);
4601 (*known_csts
)[index
] = val
->value
;
4604 /* Replace KNOWN_CSTS with its copy. Also copy KNOWN_CONTEXTS and modify the
4605 copy according to VAL and INDEX. */
4608 modify_known_vectors_with_val (vec
<tree
> *known_csts
,
4609 vec
<ipa_polymorphic_call_context
> *known_contexts
,
4610 ipcp_value
<ipa_polymorphic_call_context
> *val
,
4613 *known_csts
= known_csts
->copy ();
4614 *known_contexts
= known_contexts
->copy ();
4615 (*known_contexts
)[index
] = val
->value
;
4618 /* Return true if OFFSET indicates this was not an aggregate value or there is
4619 a replacement equivalent to VALUE, INDEX and OFFSET among those in the
4623 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value
*aggvals
,
4624 int index
, HOST_WIDE_INT offset
, tree value
)
4631 if (aggvals
->index
== index
4632 && aggvals
->offset
== offset
4633 && values_equal_for_ipcp_p (aggvals
->value
, value
))
4635 aggvals
= aggvals
->next
;
4640 /* Return true if offset is minus one because source of a polymorphic contect
4641 cannot be an aggregate value. */
4644 ipcp_val_agg_replacement_ok_p (ipa_agg_replacement_value
*,
4645 int , HOST_WIDE_INT offset
,
4646 ipa_polymorphic_call_context
)
4648 return offset
== -1;
4651 /* Decide wheter to create a special version of NODE for value VAL of parameter
4652 at the given INDEX. If OFFSET is -1, the value is for the parameter itself,
4653 otherwise it is stored at the given OFFSET of the parameter. KNOWN_CSTS,
4654 KNOWN_CONTEXTS and KNOWN_AGGS describe the other already known values. */
4656 template <typename valtype
>
4658 decide_about_value (struct cgraph_node
*node
, int index
, HOST_WIDE_INT offset
,
4659 ipcp_value
<valtype
> *val
, vec
<tree
> known_csts
,
4660 vec
<ipa_polymorphic_call_context
> known_contexts
)
4662 struct ipa_agg_replacement_value
*aggvals
;
4663 int freq_sum
, caller_count
;
4664 profile_count count_sum
;
4665 vec
<cgraph_edge
*> callers
;
4669 perhaps_add_new_callers (node
, val
);
4672 else if (val
->local_size_cost
+ overall_size
> max_new_size
)
4674 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4675 fprintf (dump_file
, " Ignoring candidate value because "
4676 "max_new_size would be reached with %li.\n",
4677 val
->local_size_cost
+ overall_size
);
4680 else if (!get_info_about_necessary_edges (val
, node
, &freq_sum
, &count_sum
,
4684 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4686 fprintf (dump_file
, " - considering value ");
4687 print_ipcp_constant_value (dump_file
, val
->value
);
4688 fprintf (dump_file
, " for ");
4689 ipa_dump_param (dump_file
, IPA_NODE_REF (node
), index
);
4691 fprintf (dump_file
, ", offset: " HOST_WIDE_INT_PRINT_DEC
, offset
);
4692 fprintf (dump_file
, " (caller_count: %i)\n", caller_count
);
4695 if (!good_cloning_opportunity_p (node
, val
->local_time_benefit
,
4696 freq_sum
, count_sum
,
4697 val
->local_size_cost
)
4698 && !good_cloning_opportunity_p (node
,
4699 val
->local_time_benefit
4700 + val
->prop_time_benefit
,
4701 freq_sum
, count_sum
,
4702 val
->local_size_cost
4703 + val
->prop_size_cost
))
4707 fprintf (dump_file
, " Creating a specialized node of %s.\n",
4708 node
->dump_name ());
4710 callers
= gather_edges_for_value (val
, node
, caller_count
);
4712 modify_known_vectors_with_val (&known_csts
, &known_contexts
, val
, index
);
4715 known_csts
= known_csts
.copy ();
4716 known_contexts
= copy_useful_known_contexts (known_contexts
);
4718 find_more_scalar_values_for_callers_subset (node
, known_csts
, callers
);
4719 find_more_contexts_for_caller_subset (node
, &known_contexts
, callers
);
4720 aggvals
= find_aggregate_values_for_callers_subset (node
, callers
);
4721 gcc_checking_assert (ipcp_val_agg_replacement_ok_p (aggvals
, index
,
4722 offset
, val
->value
));
4723 val
->spec_node
= create_specialized_node (node
, known_csts
, known_contexts
,
4725 overall_size
+= val
->local_size_cost
;
4727 /* TODO: If for some lattice there is only one other known value
4728 left, make a special node for it too. */
4733 /* Decide whether and what specialized clones of NODE should be created. */
4736 decide_whether_version_node (struct cgraph_node
*node
)
4738 struct ipa_node_params
*info
= IPA_NODE_REF (node
);
4739 int i
, count
= ipa_get_param_count (info
);
4740 vec
<tree
> known_csts
;
4741 vec
<ipa_polymorphic_call_context
> known_contexts
;
4742 vec
<ipa_agg_jump_function
> known_aggs
= vNULL
;
4748 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4749 fprintf (dump_file
, "\nEvaluating opportunities for %s.\n",
4750 node
->dump_name ());
4752 gather_context_independent_values (info
, &known_csts
, &known_contexts
,
4753 info
->do_clone_for_all_contexts
? &known_aggs
4756 for (i
= 0; i
< count
;i
++)
4758 struct ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4759 ipcp_lattice
<tree
> *lat
= &plats
->itself
;
4760 ipcp_lattice
<ipa_polymorphic_call_context
> *ctxlat
= &plats
->ctxlat
;
4765 ipcp_value
<tree
> *val
;
4766 for (val
= lat
->values
; val
; val
= val
->next
)
4767 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
4771 if (!plats
->aggs_bottom
)
4773 struct ipcp_agg_lattice
*aglat
;
4774 ipcp_value
<tree
> *val
;
4775 for (aglat
= plats
->aggs
; aglat
; aglat
= aglat
->next
)
4776 if (!aglat
->bottom
&& aglat
->values
4777 /* If the following is false, the one value is in
4779 && (plats
->aggs_contain_variable
4780 || !aglat
->is_single_const ()))
4781 for (val
= aglat
->values
; val
; val
= val
->next
)
4782 ret
|= decide_about_value (node
, i
, aglat
->offset
, val
,
4783 known_csts
, known_contexts
);
4787 && known_contexts
[i
].useless_p ())
4789 ipcp_value
<ipa_polymorphic_call_context
> *val
;
4790 for (val
= ctxlat
->values
; val
; val
= val
->next
)
4791 ret
|= decide_about_value (node
, i
, -1, val
, known_csts
,
4795 info
= IPA_NODE_REF (node
);
4798 if (info
->do_clone_for_all_contexts
)
4800 struct cgraph_node
*clone
;
4801 vec
<cgraph_edge
*> callers
;
4804 fprintf (dump_file
, " - Creating a specialized node of %s "
4805 "for all known contexts.\n", node
->dump_name ());
4807 callers
= node
->collect_callers ();
4808 find_more_scalar_values_for_callers_subset (node
, known_csts
, callers
);
4809 find_more_contexts_for_caller_subset (node
, &known_contexts
, callers
);
4810 ipa_agg_replacement_value
*aggvals
4811 = find_aggregate_values_for_callers_subset (node
, callers
);
4813 if (!known_contexts_useful_p (known_contexts
))
4815 known_contexts
.release ();
4816 known_contexts
= vNULL
;
4818 clone
= create_specialized_node (node
, known_csts
, known_contexts
,
4820 info
= IPA_NODE_REF (node
);
4821 info
->do_clone_for_all_contexts
= false;
4822 IPA_NODE_REF (clone
)->is_all_contexts_clone
= true;
4823 for (i
= 0; i
< count
; i
++)
4824 vec_free (known_aggs
[i
].items
);
4825 known_aggs
.release ();
4830 known_csts
.release ();
4831 known_contexts
.release ();
4837 /* Transitively mark all callees of NODE within the same SCC as not dead. */
4840 spread_undeadness (struct cgraph_node
*node
)
4842 struct cgraph_edge
*cs
;
4844 for (cs
= node
->callees
; cs
; cs
= cs
->next_callee
)
4845 if (ipa_edge_within_scc (cs
))
4847 struct cgraph_node
*callee
;
4848 struct ipa_node_params
*info
;
4850 callee
= cs
->callee
->function_symbol (NULL
);
4851 info
= IPA_NODE_REF (callee
);
4853 if (info
->node_dead
)
4855 info
->node_dead
= 0;
4856 spread_undeadness (callee
);
4861 /* Return true if NODE has a caller from outside of its SCC that is not
4862 dead. Worker callback for cgraph_for_node_and_aliases. */
4865 has_undead_caller_from_outside_scc_p (struct cgraph_node
*node
,
4866 void *data ATTRIBUTE_UNUSED
)
4868 struct cgraph_edge
*cs
;
4870 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4871 if (cs
->caller
->thunk
.thunk_p
4872 && cs
->caller
->call_for_symbol_thunks_and_aliases
4873 (has_undead_caller_from_outside_scc_p
, NULL
, true))
4875 else if (!ipa_edge_within_scc (cs
)
4876 && !IPA_NODE_REF (cs
->caller
)->node_dead
)
4882 /* Identify nodes within the same SCC as NODE which are no longer needed
4883 because of new clones and will be removed as unreachable. */
4886 identify_dead_nodes (struct cgraph_node
*node
)
4888 struct cgraph_node
*v
;
4889 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4891 && !v
->call_for_symbol_thunks_and_aliases
4892 (has_undead_caller_from_outside_scc_p
, NULL
, true))
4893 IPA_NODE_REF (v
)->node_dead
= 1;
4895 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4896 if (!IPA_NODE_REF (v
)->node_dead
)
4897 spread_undeadness (v
);
4899 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4901 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4902 if (IPA_NODE_REF (v
)->node_dead
)
4903 fprintf (dump_file
, " Marking node as dead: %s.\n", v
->dump_name ());
4907 /* The decision stage. Iterate over the topological order of call graph nodes
4908 TOPO and make specialized clones if deemed beneficial. */
4911 ipcp_decision_stage (struct ipa_topo_info
*topo
)
4916 fprintf (dump_file
, "\nIPA decision stage:\n\n");
4918 for (i
= topo
->nnodes
- 1; i
>= 0; i
--)
4920 struct cgraph_node
*node
= topo
->order
[i
];
4921 bool change
= false, iterate
= true;
4925 struct cgraph_node
*v
;
4927 for (v
= node
; v
; v
= ((struct ipa_dfs_info
*) v
->aux
)->next_cycle
)
4928 if (v
->has_gimple_body_p ()
4929 && ipcp_versionable_function_p (v
))
4930 iterate
|= decide_whether_version_node (v
);
4935 identify_dead_nodes (node
);
4939 /* Look up all the bits information that we have discovered and copy it over
4940 to the transformation summary. */
4943 ipcp_store_bits_results (void)
4947 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
4949 ipa_node_params
*info
= IPA_NODE_REF (node
);
4950 bool dumped_sth
= false;
4951 bool found_useful_result
= false;
4953 if (!opt_for_fn (node
->decl
, flag_ipa_bit_cp
))
4956 fprintf (dump_file
, "Not considering %s for ipa bitwise propagation "
4957 "; -fipa-bit-cp: disabled.\n",
4962 if (info
->ipcp_orig_node
)
4963 info
= IPA_NODE_REF (info
->ipcp_orig_node
);
4965 unsigned count
= ipa_get_param_count (info
);
4966 for (unsigned i
= 0; i
< count
; i
++)
4968 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4969 if (plats
->bits_lattice
.constant_p ())
4971 found_useful_result
= true;
4976 if (!found_useful_result
)
4979 ipcp_grow_transformations_if_necessary ();
4980 ipcp_transformation_summary
*ts
= ipcp_get_transformation_summary (node
);
4981 vec_safe_reserve_exact (ts
->bits
, count
);
4983 for (unsigned i
= 0; i
< count
; i
++)
4985 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
4988 if (plats
->bits_lattice
.constant_p ())
4990 = ipa_get_ipa_bits_for_value (plats
->bits_lattice
.get_value (),
4991 plats
->bits_lattice
.get_mask ());
4995 ts
->bits
->quick_push (jfbits
);
4996 if (!dump_file
|| !jfbits
)
5000 fprintf (dump_file
, "Propagated bits info for function %s:\n",
5001 node
->dump_name ());
5004 fprintf (dump_file
, " param %i: value = ", i
);
5005 print_hex (jfbits
->value
, dump_file
);
5006 fprintf (dump_file
, ", mask = ");
5007 print_hex (jfbits
->mask
, dump_file
);
5008 fprintf (dump_file
, "\n");
5013 /* Look up all VR information that we have discovered and copy it over
5014 to the transformation summary. */
5017 ipcp_store_vr_results (void)
5021 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
5023 ipa_node_params
*info
= IPA_NODE_REF (node
);
5024 bool found_useful_result
= false;
5026 if (!opt_for_fn (node
->decl
, flag_ipa_vrp
))
5029 fprintf (dump_file
, "Not considering %s for VR discovery "
5030 "and propagate; -fipa-ipa-vrp: disabled.\n",
5035 if (info
->ipcp_orig_node
)
5036 info
= IPA_NODE_REF (info
->ipcp_orig_node
);
5038 unsigned count
= ipa_get_param_count (info
);
5039 for (unsigned i
= 0; i
< count
; i
++)
5041 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
5042 if (!plats
->m_value_range
.bottom_p ()
5043 && !plats
->m_value_range
.top_p ())
5045 found_useful_result
= true;
5049 if (!found_useful_result
)
5052 ipcp_grow_transformations_if_necessary ();
5053 ipcp_transformation_summary
*ts
= ipcp_get_transformation_summary (node
);
5054 vec_safe_reserve_exact (ts
->m_vr
, count
);
5056 for (unsigned i
= 0; i
< count
; i
++)
5058 ipcp_param_lattices
*plats
= ipa_get_parm_lattices (info
, i
);
5061 if (!plats
->m_value_range
.bottom_p ()
5062 && !plats
->m_value_range
.top_p ())
5065 vr
.type
= plats
->m_value_range
.m_vr
.type
;
5066 vr
.min
= wi::to_wide (plats
->m_value_range
.m_vr
.min
);
5067 vr
.max
= wi::to_wide (plats
->m_value_range
.m_vr
.max
);
5072 vr
.type
= VR_VARYING
;
5073 vr
.min
= vr
.max
= wi::zero (INT_TYPE_SIZE
);
5075 ts
->m_vr
->quick_push (vr
);
5080 /* The IPCP driver. */
5085 struct cgraph_2edge_hook_list
*edge_duplication_hook_holder
;
5086 struct cgraph_edge_hook_list
*edge_removal_hook_holder
;
5087 struct ipa_topo_info topo
;
5089 ipa_check_create_node_params ();
5090 ipa_check_create_edge_args ();
5091 grow_edge_clone_vectors ();
5092 edge_duplication_hook_holder
5093 = symtab
->add_edge_duplication_hook (&ipcp_edge_duplication_hook
, NULL
);
5094 edge_removal_hook_holder
5095 = symtab
->add_edge_removal_hook (&ipcp_edge_removal_hook
, NULL
);
5099 fprintf (dump_file
, "\nIPA structures before propagation:\n");
5100 if (dump_flags
& TDF_DETAILS
)
5101 ipa_print_all_params (dump_file
);
5102 ipa_print_all_jump_functions (dump_file
);
5105 /* Topological sort. */
5106 build_toporder_info (&topo
);
5107 /* Do the interprocedural propagation. */
5108 ipcp_propagate_stage (&topo
);
5109 /* Decide what constant propagation and cloning should be performed. */
5110 ipcp_decision_stage (&topo
);
5111 /* Store results of bits propagation. */
5112 ipcp_store_bits_results ();
5113 /* Store results of value range propagation. */
5114 ipcp_store_vr_results ();
5116 /* Free all IPCP structures. */
5117 free_toporder_info (&topo
);
5118 next_edge_clone
.release ();
5119 prev_edge_clone
.release ();
5120 symtab
->remove_edge_removal_hook (edge_removal_hook_holder
);
5121 symtab
->remove_edge_duplication_hook (edge_duplication_hook_holder
);
5122 ipa_free_all_structures_after_ipa_cp ();
5124 fprintf (dump_file
, "\nIPA constant propagation end\n");
5128 /* Initialization and computation of IPCP data structures. This is the initial
5129 intraprocedural analysis of functions, which gathers information to be
5130 propagated later on. */
5133 ipcp_generate_summary (void)
5135 struct cgraph_node
*node
;
5138 fprintf (dump_file
, "\nIPA constant propagation start:\n");
5139 ipa_register_cgraph_hooks ();
5141 FOR_EACH_FUNCTION_WITH_GIMPLE_BODY (node
)
5142 ipa_analyze_node (node
);
5145 /* Write ipcp summary for nodes in SET. */
5148 ipcp_write_summary (void)
5150 ipa_prop_write_jump_functions ();
5153 /* Read ipcp summary. */
5156 ipcp_read_summary (void)
5158 ipa_prop_read_jump_functions ();
5163 const pass_data pass_data_ipa_cp
=
5165 IPA_PASS
, /* type */
5167 OPTGROUP_NONE
, /* optinfo_flags */
5168 TV_IPA_CONSTANT_PROP
, /* tv_id */
5169 0, /* properties_required */
5170 0, /* properties_provided */
5171 0, /* properties_destroyed */
5172 0, /* todo_flags_start */
5173 ( TODO_dump_symtab
| TODO_remove_functions
), /* todo_flags_finish */
5176 class pass_ipa_cp
: public ipa_opt_pass_d
5179 pass_ipa_cp (gcc::context
*ctxt
)
5180 : ipa_opt_pass_d (pass_data_ipa_cp
, ctxt
,
5181 ipcp_generate_summary
, /* generate_summary */
5182 ipcp_write_summary
, /* write_summary */
5183 ipcp_read_summary
, /* read_summary */
5184 ipcp_write_transformation_summaries
, /*
5185 write_optimization_summary */
5186 ipcp_read_transformation_summaries
, /*
5187 read_optimization_summary */
5188 NULL
, /* stmt_fixup */
5189 0, /* function_transform_todo_flags_start */
5190 ipcp_transform_function
, /* function_transform */
5191 NULL
) /* variable_transform */
5194 /* opt_pass methods: */
5195 virtual bool gate (function
*)
5197 /* FIXME: We should remove the optimize check after we ensure we never run
5198 IPA passes when not optimizing. */
5199 return (flag_ipa_cp
&& optimize
) || in_lto_p
;
5202 virtual unsigned int execute (function
*) { return ipcp_driver (); }
5204 }; // class pass_ipa_cp
5209 make_pass_ipa_cp (gcc::context
*ctxt
)
5211 return new pass_ipa_cp (ctxt
);
5214 /* Reset all state within ipa-cp.c so that we can rerun the compiler
5215 within the same process. For use by toplev::finalize. */
5218 ipa_cp_c_finalize (void)
5220 max_count
= profile_count::uninitialized ();