blob: b12e7a1124d5132d71799a403e63aea912de2662 [file] [log] [blame]
/* Function summary pass.
Copyright (C) 2003-2022 Free Software Foundation, Inc.
Contributed by Jan Hubicka
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
/* Analysis of function bodies used by inter-procedural passes
We estimate for each function
- function body size and size after specializing into given context
- average function execution time in a given context
- function frame size
For each call
- call statement size, time and how often the parameters change
ipa_fn_summary data structures store above information locally (i.e.
parameters of the function itself) and globally (i.e. parameters of
the function created by applying all the inline decisions already
present in the callgraph).
We provide access to the ipa_fn_summary data structure and
basic logic updating the parameters when inlining is performed.
The summaries are context sensitive. Context means
1) partial assignment of known constant values of operands
2) whether function is inlined into the call or not.
It is easy to add more variants. To represent function size and time
that depends on context (i.e. it is known to be optimized away when
context is known either by inlining or from IP-CP and cloning),
we use predicates.
estimate_edge_size_and_time can be used to query
function size/time in the given context. ipa_merge_fn_summary_after_inlining merges
properties of caller and callee after inlining.
Finally pass_inline_parameters is exported. This is used to drive
computation of function parameters used by the early inliner. IPA
inlined performs analysis via its analyze_function method. */
#include "config.h"
#define INCLUDE_VECTOR
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "target.h"
#include "tree.h"
#include "gimple.h"
#include "alloc-pool.h"
#include "tree-pass.h"
#include "ssa.h"
#include "tree-streamer.h"
#include "cgraph.h"
#include "diagnostic.h"
#include "fold-const.h"
#include "print-tree.h"
#include "tree-inline.h"
#include "gimple-pretty-print.h"
#include "cfganal.h"
#include "gimple-iterator.h"
#include "tree-cfg.h"
#include "tree-ssa-loop-niter.h"
#include "tree-ssa-loop.h"
#include "symbol-summary.h"
#include "ipa-prop.h"
#include "ipa-fnsummary.h"
#include "cfgloop.h"
#include "tree-scalar-evolution.h"
#include "ipa-utils.h"
#include "cfgexpand.h"
#include "gimplify.h"
#include "stringpool.h"
#include "attribs.h"
#include "tree-into-ssa.h"
#include "symtab-clones.h"
#include "gimple-range.h"
#include "tree-dfa.h"
/* Summaries. */
fast_function_summary <ipa_fn_summary *, va_gc> *ipa_fn_summaries;
fast_function_summary <ipa_size_summary *, va_heap> *ipa_size_summaries;
fast_call_summary <ipa_call_summary *, va_heap> *ipa_call_summaries;
/* Edge predicates goes here. */
static object_allocator<ipa_predicate> edge_predicate_pool ("edge predicates");
/* Dump IPA hints. */
void
ipa_dump_hints (FILE *f, ipa_hints hints)
{
if (!hints)
return;
fprintf (f, "IPA hints:");
if (hints & INLINE_HINT_indirect_call)
{
hints &= ~INLINE_HINT_indirect_call;
fprintf (f, " indirect_call");
}
if (hints & INLINE_HINT_loop_iterations)
{
hints &= ~INLINE_HINT_loop_iterations;
fprintf (f, " loop_iterations");
}
if (hints & INLINE_HINT_loop_stride)
{
hints &= ~INLINE_HINT_loop_stride;
fprintf (f, " loop_stride");
}
if (hints & INLINE_HINT_same_scc)
{
hints &= ~INLINE_HINT_same_scc;
fprintf (f, " same_scc");
}
if (hints & INLINE_HINT_in_scc)
{
hints &= ~INLINE_HINT_in_scc;
fprintf (f, " in_scc");
}
if (hints & INLINE_HINT_cross_module)
{
hints &= ~INLINE_HINT_cross_module;
fprintf (f, " cross_module");
}
if (hints & INLINE_HINT_declared_inline)
{
hints &= ~INLINE_HINT_declared_inline;
fprintf (f, " declared_inline");
}
if (hints & INLINE_HINT_known_hot)
{
hints &= ~INLINE_HINT_known_hot;
fprintf (f, " known_hot");
}
if (hints & INLINE_HINT_builtin_constant_p)
{
hints &= ~INLINE_HINT_builtin_constant_p;
fprintf (f, " builtin_constant_p");
}
gcc_assert (!hints);
}
/* Record SIZE and TIME to SUMMARY.
The accounted code will be executed when EXEC_PRED is true.
When NONCONST_PRED is false the code will evaluate to constant and
will get optimized out in specialized clones of the function.
If CALL is true account to call_size_time_table rather than
size_time_table. */
void
ipa_fn_summary::account_size_time (int size, sreal time,
const ipa_predicate &exec_pred,
const ipa_predicate &nonconst_pred_in,
bool call)
{
size_time_entry *e;
bool found = false;
int i;
ipa_predicate nonconst_pred;
vec<size_time_entry> *table = call ? &call_size_time_table : &size_time_table;
if (exec_pred == false)
return;
nonconst_pred = nonconst_pred_in & exec_pred;
if (nonconst_pred == false)
return;
/* We need to create initial empty unconditional clause, but otherwise
we don't need to account empty times and sizes. */
if (!size && time == 0 && table->length ())
return;
/* Only for calls we are unaccounting what we previously recorded. */
gcc_checking_assert (time >= 0 || call);
for (i = 0; table->iterate (i, &e); i++)
if (e->exec_predicate == exec_pred
&& e->nonconst_predicate == nonconst_pred)
{
found = true;
break;
}
if (i == max_size_time_table_size)
{
i = 0;
found = true;
e = &(*table)[0];
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"\t\tReached limit on number of entries, "
"ignoring the predicate.");
}
if (dump_file && (dump_flags & TDF_DETAILS) && (time != 0 || size))
{
fprintf (dump_file,
"\t\tAccounting size:%3.2f, time:%3.2f on %spredicate exec:",
((double) size) / ipa_fn_summary::size_scale,
(time.to_double ()), found ? "" : "new ");
exec_pred.dump (dump_file, conds, 0);
if (exec_pred != nonconst_pred)
{
fprintf (dump_file, " nonconst:");
nonconst_pred.dump (dump_file, conds);
}
else
fprintf (dump_file, "\n");
}
if (!found)
{
class size_time_entry new_entry;
new_entry.size = size;
new_entry.time = time;
new_entry.exec_predicate = exec_pred;
new_entry.nonconst_predicate = nonconst_pred;
if (call)
call_size_time_table.safe_push (new_entry);
else
size_time_table.safe_push (new_entry);
}
else
{
e->size += size;
e->time += time;
/* FIXME: PR bootstrap/92653 gcc_checking_assert (e->time >= -1); */
/* Tolerate small roundoff issues. */
if (e->time < 0)
e->time = 0;
}
}
/* We proved E to be unreachable, redirect it to __builtin_unreachable. */
static struct cgraph_edge *
redirect_to_unreachable (struct cgraph_edge *e)
{
struct cgraph_node *callee = !e->inline_failed ? e->callee : NULL;
struct cgraph_node *target = cgraph_node::get_create
(builtin_decl_implicit (BUILT_IN_UNREACHABLE));
if (e->speculative)
e = cgraph_edge::resolve_speculation (e, target->decl);
else if (!e->callee)
e = cgraph_edge::make_direct (e, target);
else
e->redirect_callee (target);
class ipa_call_summary *es = ipa_call_summaries->get (e);
e->inline_failed = CIF_UNREACHABLE;
e->count = profile_count::zero ();
es->call_stmt_size = 0;
es->call_stmt_time = 0;
if (callee)
callee->remove_symbol_and_inline_clones ();
return e;
}
/* Set predicate for edge E. */
static void
edge_set_predicate (struct cgraph_edge *e, ipa_predicate *predicate)
{
/* If the edge is determined to be never executed, redirect it
to BUILTIN_UNREACHABLE to make it clear to IPA passes the call will
be optimized out. */
if (predicate && *predicate == false
/* When handling speculative edges, we need to do the redirection
just once. Do it always on the direct edge, so we do not
attempt to resolve speculation while duplicating the edge. */
&& (!e->speculative || e->callee))
e = redirect_to_unreachable (e);
class ipa_call_summary *es = ipa_call_summaries->get (e);
if (predicate && *predicate != true)
{
if (!es->predicate)
es->predicate = edge_predicate_pool.allocate ();
*es->predicate = *predicate;
}
else
{
if (es->predicate)
edge_predicate_pool.remove (es->predicate);
es->predicate = NULL;
}
}
/* Set predicate for hint *P. */
static void
set_hint_predicate (ipa_predicate **p, ipa_predicate new_predicate)
{
if (new_predicate == false || new_predicate == true)
{
if (*p)
edge_predicate_pool.remove (*p);
*p = NULL;
}
else
{
if (!*p)
*p = edge_predicate_pool.allocate ();
**p = new_predicate;
}
}
/* Find if NEW_PREDICATE is already in V and if so, increment its freq.
Otherwise add a new item to the vector with this predicate and frerq equal
to add_freq, unless the number of predicates would exceed MAX_NUM_PREDICATES
in which case the function does nothing. */
static void
add_freqcounting_predicate (vec<ipa_freqcounting_predicate, va_gc> **v,
const ipa_predicate &new_predicate, sreal add_freq,
unsigned max_num_predicates)
{
if (new_predicate == false || new_predicate == true)
return;
ipa_freqcounting_predicate *f;
for (int i = 0; vec_safe_iterate (*v, i, &f); i++)
if (new_predicate == f->predicate)
{
f->freq += add_freq;
return;
}
if (vec_safe_length (*v) >= max_num_predicates)
/* Too many different predicates to account for. */
return;
ipa_freqcounting_predicate fcp;
fcp.predicate = NULL;
set_hint_predicate (&fcp.predicate, new_predicate);
fcp.freq = add_freq;
vec_safe_push (*v, fcp);
return;
}
/* Compute what conditions may or may not hold given information about
parameters. RET_CLAUSE returns truths that may hold in a specialized copy,
while RET_NONSPEC_CLAUSE returns truths that may hold in an nonspecialized
copy when called in a given context. It is a bitmask of conditions. Bit
0 means that condition is known to be false, while bit 1 means that condition
may or may not be true. These differs - for example NOT_INLINED condition
is always false in the second and also builtin_constant_p tests cannot use
the fact that parameter is indeed a constant.
When INLINE_P is true, assume that we are inlining. AVAL contains known
information about argument values. The function does not modify its content
and so AVALs could also be of type ipa_call_arg_values but so far all
callers work with the auto version and so we avoid the conversion for
convenience.
ERROR_MARK value of an argument means compile time invariant. */
static void
evaluate_conditions_for_known_args (struct cgraph_node *node,
bool inline_p,
ipa_auto_call_arg_values *avals,
clause_t *ret_clause,
clause_t *ret_nonspec_clause)
{
clause_t clause = inline_p ? 0 : 1 << ipa_predicate::not_inlined_condition;
clause_t nonspec_clause = 1 << ipa_predicate::not_inlined_condition;
class ipa_fn_summary *info = ipa_fn_summaries->get (node);
int i;
struct condition *c;
for (i = 0; vec_safe_iterate (info->conds, i, &c); i++)
{
tree val = NULL;
tree res;
int j;
struct expr_eval_op *op;
/* We allow call stmt to have fewer arguments than the callee function
(especially for K&R style programs). So bound check here (we assume
m_known_aggs vector is either empty or has the same length as
m_known_vals). */
gcc_checking_assert (!avals->m_known_aggs.length ()
|| !avals->m_known_vals.length ()
|| (avals->m_known_vals.length ()
== avals->m_known_aggs.length ()));
if (c->agg_contents)
{
if (c->code == ipa_predicate::changed
&& !c->by_ref
&& (avals->safe_sval_at(c->operand_num) == error_mark_node))
continue;
if (ipa_agg_value_set *agg = avals->safe_aggval_at (c->operand_num))
{
tree sval = avals->safe_sval_at (c->operand_num);
val = ipa_find_agg_cst_for_param (agg, sval, c->offset,
c->by_ref);
}
else
val = NULL_TREE;
}
else
{
val = avals->safe_sval_at (c->operand_num);
if (val && val == error_mark_node
&& c->code != ipa_predicate::changed)
val = NULL_TREE;
}
if (!val
&& (c->code == ipa_predicate::changed
|| c->code == ipa_predicate::is_not_constant))
{
clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
continue;
}
if (c->code == ipa_predicate::changed)
{
nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
continue;
}
if (c->code == ipa_predicate::is_not_constant)
{
nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
continue;
}
if (val && TYPE_SIZE (c->type) == TYPE_SIZE (TREE_TYPE (val)))
{
if (c->type != TREE_TYPE (val))
val = fold_unary (VIEW_CONVERT_EXPR, c->type, val);
for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
{
if (!val)
break;
if (!op->val[0])
val = fold_unary (op->code, op->type, val);
else if (!op->val[1])
val = fold_binary (op->code, op->type,
op->index ? op->val[0] : val,
op->index ? val : op->val[0]);
else if (op->index == 0)
val = fold_ternary (op->code, op->type,
val, op->val[0], op->val[1]);
else if (op->index == 1)
val = fold_ternary (op->code, op->type,
op->val[0], val, op->val[1]);
else if (op->index == 2)
val = fold_ternary (op->code, op->type,
op->val[0], op->val[1], val);
else
val = NULL_TREE;
}
res = val
? fold_binary_to_constant (c->code, boolean_type_node, val, c->val)
: NULL;
if (res && integer_zerop (res))
continue;
if (res && integer_onep (res))
{
clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
nonspec_clause
|= 1 << (i + ipa_predicate::first_dynamic_condition);
continue;
}
}
if (c->operand_num < (int) avals->m_known_value_ranges.length ()
&& !c->agg_contents
&& (!val || TREE_CODE (val) != INTEGER_CST))
{
value_range vr = avals->m_known_value_ranges[c->operand_num];
if (!vr.undefined_p ()
&& !vr.varying_p ()
&& (TYPE_SIZE (c->type) == TYPE_SIZE (vr.type ())))
{
if (!useless_type_conversion_p (c->type, vr.type ()))
{
value_range res;
range_fold_unary_expr (&res, NOP_EXPR,
c->type, &vr, vr.type ());
vr = res;
}
tree type = c->type;
for (j = 0; vec_safe_iterate (c->param_ops, j, &op); j++)
{
if (vr.varying_p () || vr.undefined_p ())
break;
value_range res;
if (!op->val[0])
range_fold_unary_expr (&res, op->code, op->type, &vr, type);
else if (!op->val[1])
{
value_range op0 (op->val[0], op->val[0]);
range_fold_binary_expr (&res, op->code, op->type,
op->index ? &op0 : &vr,
op->index ? &vr : &op0);
}
else
res.set_varying (op->type);
type = op->type;
vr = res;
}
if (!vr.varying_p () && !vr.undefined_p ())
{
value_range res;
value_range val_vr (c->val, c->val);
range_fold_binary_expr (&res, c->code, boolean_type_node,
&vr,
&val_vr);
if (res.zero_p ())
continue;
}
}
}
clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
nonspec_clause |= 1 << (i + ipa_predicate::first_dynamic_condition);
}
*ret_clause = clause;
if (ret_nonspec_clause)
*ret_nonspec_clause = nonspec_clause;
}
/* Return true if VRP will be exectued on the function.
We do not want to anticipate optimizations that will not happen.
FIXME: This can be confused with -fdisable and debug counters and thus
it should not be used for correctness (only to make heuristics work).
This means that inliner should do its own optimizations of expressions
that it predicts to be constant so wrong code can not be triggered by
builtin_constant_p. */
static bool
vrp_will_run_p (struct cgraph_node *node)
{
return (opt_for_fn (node->decl, optimize)
&& !opt_for_fn (node->decl, optimize_debug)
&& opt_for_fn (node->decl, flag_tree_vrp));
}
/* Similarly about FRE. */
static bool
fre_will_run_p (struct cgraph_node *node)
{
return (opt_for_fn (node->decl, optimize)
&& !opt_for_fn (node->decl, optimize_debug)
&& opt_for_fn (node->decl, flag_tree_fre));
}
/* Work out what conditions might be true at invocation of E.
Compute costs for inlined edge if INLINE_P is true.
Return in CLAUSE_PTR the evaluated conditions and in NONSPEC_CLAUSE_PTR
(if non-NULL) conditions evaluated for nonspecialized clone called
in a given context.
Vectors in AVALS will be populated with useful known information about
argument values - information not known to have any uses will be omitted -
except for m_known_contexts which will only be calculated if
COMPUTE_CONTEXTS is true. */
void
evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
clause_t *clause_ptr,
clause_t *nonspec_clause_ptr,
ipa_auto_call_arg_values *avals,
bool compute_contexts)
{
struct cgraph_node *callee = e->callee->ultimate_alias_target ();
class ipa_fn_summary *info = ipa_fn_summaries->get (callee);
class ipa_edge_args *args;
if (clause_ptr)
*clause_ptr = inline_p ? 0 : 1 << ipa_predicate::not_inlined_condition;
if (ipa_node_params_sum
&& !e->call_stmt_cannot_inline_p
&& (info->conds || compute_contexts)
&& (args = ipa_edge_args_sum->get (e)) != NULL)
{
struct cgraph_node *caller;
class ipa_node_params *caller_parms_info, *callee_pi = NULL;
class ipa_call_summary *es = ipa_call_summaries->get (e);
int i, count = ipa_get_cs_argument_count (args);
if (count)
{
if (e->caller->inlined_to)
caller = e->caller->inlined_to;
else
caller = e->caller;
caller_parms_info = ipa_node_params_sum->get (caller);
callee_pi = ipa_node_params_sum->get (callee);
/* Watch for thunks. */
if (callee_pi)
/* Watch for variadic functions. */
count = MIN (count, ipa_get_param_count (callee_pi));
}
if (callee_pi)
for (i = 0; i < count; i++)
{
struct ipa_jump_func *jf = ipa_get_ith_jump_func (args, i);
if (ipa_is_param_used_by_indirect_call (callee_pi, i)
|| ipa_is_param_used_by_ipa_predicates (callee_pi, i))
{
/* Determine if we know constant value of the parameter. */
tree cst = ipa_value_from_jfunc (caller_parms_info, jf,
ipa_get_type (callee_pi, i));
if (!cst && e->call_stmt
&& i < (int)gimple_call_num_args (e->call_stmt))
{
cst = gimple_call_arg (e->call_stmt, i);
if (!is_gimple_min_invariant (cst))
cst = NULL;
}
if (cst)
{
gcc_checking_assert (TREE_CODE (cst) != TREE_BINFO);
if (!avals->m_known_vals.length ())
avals->m_known_vals.safe_grow_cleared (count, true);
avals->m_known_vals[i] = cst;
}
else if (inline_p && !es->param[i].change_prob)
{
if (!avals->m_known_vals.length ())
avals->m_known_vals.safe_grow_cleared (count, true);
avals->m_known_vals[i] = error_mark_node;
}
/* If we failed to get simple constant, try value range. */
if ((!cst || TREE_CODE (cst) != INTEGER_CST)
&& vrp_will_run_p (caller)
&& ipa_is_param_used_by_ipa_predicates (callee_pi, i))
{
value_range vr
= ipa_value_range_from_jfunc (caller_parms_info, e, jf,
ipa_get_type (callee_pi,
i));
if (!vr.undefined_p () && !vr.varying_p ())
{
if (!avals->m_known_value_ranges.length ())
{
avals->m_known_value_ranges.safe_grow (count, true);
for (int i = 0; i < count; ++i)
new (&avals->m_known_value_ranges[i])
value_range ();
}
avals->m_known_value_ranges[i] = vr;
}
}
/* Determine known aggregate values. */
if (fre_will_run_p (caller))
{
ipa_agg_value_set agg
= ipa_agg_value_set_from_jfunc (caller_parms_info,
caller, &jf->agg);
if (agg.items.length ())
{
if (!avals->m_known_aggs.length ())
avals->m_known_aggs.safe_grow_cleared (count, true);
avals->m_known_aggs[i] = agg;
}
}
}
/* For calls used in polymorphic calls we further determine
polymorphic call context. */
if (compute_contexts
&& ipa_is_param_used_by_polymorphic_call (callee_pi, i))
{
ipa_polymorphic_call_context
ctx = ipa_context_from_jfunc (caller_parms_info, e, i, jf);
if (!ctx.useless_p ())
{
if (!avals->m_known_contexts.length ())
avals->m_known_contexts.safe_grow_cleared (count, true);
avals->m_known_contexts[i]
= ipa_context_from_jfunc (caller_parms_info, e, i, jf);
}
}
}
else
gcc_assert (!count || callee->thunk);
}
else if (e->call_stmt && !e->call_stmt_cannot_inline_p && info->conds)
{
int i, count = (int)gimple_call_num_args (e->call_stmt);
for (i = 0; i < count; i++)
{
tree cst = gimple_call_arg (e->call_stmt, i);
if (!is_gimple_min_invariant (cst))
cst = NULL;
if (cst)
{
if (!avals->m_known_vals.length ())
avals->m_known_vals.safe_grow_cleared (count, true);
avals->m_known_vals[i] = cst;
}
}
}
evaluate_conditions_for_known_args (callee, inline_p, avals, clause_ptr,
nonspec_clause_ptr);
}
/* Allocate the function summary. */
static void
ipa_fn_summary_alloc (void)
{
gcc_checking_assert (!ipa_fn_summaries);
ipa_size_summaries = new ipa_size_summary_t (symtab);
ipa_fn_summaries = ipa_fn_summary_t::create_ggc (symtab);
ipa_call_summaries = new ipa_call_summary_t (symtab);
}
ipa_call_summary::~ipa_call_summary ()
{
if (predicate)
edge_predicate_pool.remove (predicate);
param.release ();
}
ipa_fn_summary::~ipa_fn_summary ()
{
unsigned len = vec_safe_length (loop_iterations);
for (unsigned i = 0; i < len; i++)
edge_predicate_pool.remove ((*loop_iterations)[i].predicate);
len = vec_safe_length (loop_strides);
for (unsigned i = 0; i < len; i++)
edge_predicate_pool.remove ((*loop_strides)[i].predicate);
vec_free (conds);
call_size_time_table.release ();
vec_free (loop_iterations);
vec_free (loop_strides);
builtin_constant_p_parms.release ();
}
void
ipa_fn_summary_t::remove_callees (cgraph_node *node)
{
cgraph_edge *e;
for (e = node->callees; e; e = e->next_callee)
ipa_call_summaries->remove (e);
for (e = node->indirect_calls; e; e = e->next_callee)
ipa_call_summaries->remove (e);
}
/* Duplicate predicates in loop hint vector, allocating memory for them and
remove and deallocate any uninteresting (true or false) ones. Return the
result. */
static vec<ipa_freqcounting_predicate, va_gc> *
remap_freqcounting_preds_after_dup (vec<ipa_freqcounting_predicate, va_gc> *v,
clause_t possible_truths)
{
if (vec_safe_length (v) == 0)
return NULL;
vec<ipa_freqcounting_predicate, va_gc> *res = v->copy ();
int len = res->length();
for (int i = len - 1; i >= 0; i--)
{
ipa_predicate new_predicate
= (*res)[i].predicate->remap_after_duplication (possible_truths);
/* We do not want to free previous predicate; it is used by node
origin. */
(*res)[i].predicate = NULL;
set_hint_predicate (&(*res)[i].predicate, new_predicate);
if (!(*res)[i].predicate)
res->unordered_remove (i);
}
return res;
}
/* Hook that is called by cgraph.cc when a node is duplicated. */
void
ipa_fn_summary_t::duplicate (cgraph_node *src,
cgraph_node *dst,
ipa_fn_summary *src_info,
ipa_fn_summary *info)
{
new (info) ipa_fn_summary (*src_info);
/* TODO: as an optimization, we may avoid copying conditions
that are known to be false or true. */
info->conds = vec_safe_copy (info->conds);
clone_info *cinfo = clone_info::get (dst);
/* When there are any replacements in the function body, see if we can figure
out that something was optimized out. */
if (ipa_node_params_sum && cinfo && cinfo->tree_map)
{
/* Use SRC parm info since it may not be copied yet. */
ipa_node_params *parms_info = ipa_node_params_sum->get (src);
ipa_auto_call_arg_values avals;
int count = ipa_get_param_count (parms_info);
int i, j;
clause_t possible_truths;
ipa_predicate true_pred = true;
size_time_entry *e;
int optimized_out_size = 0;
bool inlined_to_p = false;
struct cgraph_edge *edge, *next;
info->size_time_table.release ();
avals.m_known_vals.safe_grow_cleared (count, true);
for (i = 0; i < count; i++)
{
struct ipa_replace_map *r;
for (j = 0; vec_safe_iterate (cinfo->tree_map, j, &r); j++)
{
if (r->parm_num == i)
{
avals.m_known_vals[i] = r->new_tree;
break;
}
}
}
evaluate_conditions_for_known_args (dst, false,
&avals,
&possible_truths,
/* We are going to specialize,
so ignore nonspec truths. */
NULL);
info->account_size_time (0, 0, true_pred, true_pred);
/* Remap size_time vectors.
Simplify the predicate by pruning out alternatives that are known
to be false.
TODO: as on optimization, we can also eliminate conditions known
to be true. */
for (i = 0; src_info->size_time_table.iterate (i, &e); i++)
{
ipa_predicate new_exec_pred;
ipa_predicate new_nonconst_pred;
new_exec_pred = e->exec_predicate.remap_after_duplication
(possible_truths);
new_nonconst_pred = e->nonconst_predicate.remap_after_duplication
(possible_truths);
if (new_exec_pred == false || new_nonconst_pred == false)
optimized_out_size += e->size;
else
info->account_size_time (e->size, e->time, new_exec_pred,
new_nonconst_pred);
}
/* Remap edge predicates with the same simplification as above.
Also copy constantness arrays. */
for (edge = dst->callees; edge; edge = next)
{
ipa_predicate new_predicate;
class ipa_call_summary *es = ipa_call_summaries->get (edge);
next = edge->next_callee;
if (!edge->inline_failed)
inlined_to_p = true;
if (!es->predicate)
continue;
new_predicate = es->predicate->remap_after_duplication
(possible_truths);
if (new_predicate == false && *es->predicate != false)
optimized_out_size += es->call_stmt_size * ipa_fn_summary::size_scale;
edge_set_predicate (edge, &new_predicate);
}
/* Remap indirect edge predicates with the same simplification as above.
Also copy constantness arrays. */
for (edge = dst->indirect_calls; edge; edge = next)
{
ipa_predicate new_predicate;
class ipa_call_summary *es = ipa_call_summaries->get (edge);
next = edge->next_callee;
gcc_checking_assert (edge->inline_failed);
if (!es->predicate)
continue;
new_predicate = es->predicate->remap_after_duplication
(possible_truths);
if (new_predicate == false && *es->predicate != false)
optimized_out_size
+= es->call_stmt_size * ipa_fn_summary::size_scale;
edge_set_predicate (edge, &new_predicate);
}
info->loop_iterations
= remap_freqcounting_preds_after_dup (info->loop_iterations,
possible_truths);
info->loop_strides
= remap_freqcounting_preds_after_dup (info->loop_strides,
possible_truths);
if (info->builtin_constant_p_parms.length())
{
vec <int, va_heap, vl_ptr> parms = info->builtin_constant_p_parms;
int ip;
info->builtin_constant_p_parms = vNULL;
for (i = 0; parms.iterate (i, &ip); i++)
if (!avals.m_known_vals[ip])
info->builtin_constant_p_parms.safe_push (ip);
}
/* If inliner or someone after inliner will ever start producing
non-trivial clones, we will get trouble with lack of information
about updating self sizes, because size vectors already contains
sizes of the callees. */
gcc_assert (!inlined_to_p || !optimized_out_size);
}
else
{
info->size_time_table = src_info->size_time_table.copy ();
info->loop_iterations = vec_safe_copy (src_info->loop_iterations);
info->loop_strides = vec_safe_copy (info->loop_strides);
info->builtin_constant_p_parms
= info->builtin_constant_p_parms.copy ();
ipa_freqcounting_predicate *f;
for (int i = 0; vec_safe_iterate (info->loop_iterations, i, &f); i++)
{
ipa_predicate p = *f->predicate;
f->predicate = NULL;
set_hint_predicate (&f->predicate, p);
}
for (int i = 0; vec_safe_iterate (info->loop_strides, i, &f); i++)
{
ipa_predicate p = *f->predicate;
f->predicate = NULL;
set_hint_predicate (&f->predicate, p);
}
}
if (!dst->inlined_to)
ipa_update_overall_fn_summary (dst);
}
/* Hook that is called by cgraph.cc when a node is duplicated. */
void
ipa_call_summary_t::duplicate (struct cgraph_edge *src,
struct cgraph_edge *dst,
class ipa_call_summary *srcinfo,
class ipa_call_summary *info)
{
new (info) ipa_call_summary (*srcinfo);
info->predicate = NULL;
edge_set_predicate (dst, srcinfo->predicate);
info->param = srcinfo->param.copy ();
if (!dst->indirect_unknown_callee && src->indirect_unknown_callee)
{
info->call_stmt_size -= (eni_size_weights.indirect_call_cost
- eni_size_weights.call_cost);
info->call_stmt_time -= (eni_time_weights.indirect_call_cost
- eni_time_weights.call_cost);
}
}
/* Dump edge summaries associated to NODE and recursively to all clones.
Indent by INDENT. */
static void
dump_ipa_call_summary (FILE *f, int indent, struct cgraph_node *node,
class ipa_fn_summary *info)
{
struct cgraph_edge *edge;
for (edge = node->callees; edge; edge = edge->next_callee)
{
class ipa_call_summary *es = ipa_call_summaries->get (edge);
struct cgraph_node *callee = edge->callee->ultimate_alias_target ();
int i;
fprintf (f,
"%*s%s %s\n%*s freq:%4.2f",
indent, "", callee->dump_name (),
!edge->inline_failed
? "inlined" : cgraph_inline_failed_string (edge-> inline_failed),
indent, "", edge->sreal_frequency ().to_double ());
if (cross_module_call_p (edge))
fprintf (f, " cross module");
if (es)
fprintf (f, " loop depth:%2i size:%2i time: %2i",
es->loop_depth, es->call_stmt_size, es->call_stmt_time);
ipa_fn_summary *s = ipa_fn_summaries->get (callee);
ipa_size_summary *ss = ipa_size_summaries->get (callee);
if (s != NULL)
fprintf (f, " callee size:%2i stack:%2i",
(int) (ss->size / ipa_fn_summary::size_scale),
(int) s->estimated_stack_size);
if (es && es->predicate)
{
fprintf (f, " predicate: ");
es->predicate->dump (f, info->conds);
}
else
fprintf (f, "\n");
if (es && es->param.exists ())
for (i = 0; i < (int) es->param.length (); i++)
{
int prob = es->param[i].change_prob;
if (!prob)
fprintf (f, "%*s op%i is compile time invariant\n",
indent + 2, "", i);
else if (prob != REG_BR_PROB_BASE)
fprintf (f, "%*s op%i change %f%% of time\n", indent + 2, "", i,
prob * 100.0 / REG_BR_PROB_BASE);
if (es->param[i].points_to_local_or_readonly_memory)
fprintf (f, "%*s op%i points to local or readonly memory\n",
indent + 2, "", i);
}
if (!edge->inline_failed)
{
ipa_size_summary *ss = ipa_size_summaries->get (callee);
fprintf (f, "%*sStack frame offset %i, callee self size %i\n",
indent + 2, "",
(int) ipa_get_stack_frame_offset (callee),
(int) ss->estimated_self_stack_size);
dump_ipa_call_summary (f, indent + 2, callee, info);
}
}
for (edge = node->indirect_calls; edge; edge = edge->next_callee)
{
class ipa_call_summary *es = ipa_call_summaries->get (edge);
fprintf (f, "%*sindirect call loop depth:%2i freq:%4.2f size:%2i"
" time: %2i",
indent, "",
es->loop_depth,
edge->sreal_frequency ().to_double (), es->call_stmt_size,
es->call_stmt_time);
if (es->predicate)
{
fprintf (f, "predicate: ");
es->predicate->dump (f, info->conds);
}
else
fprintf (f, "\n");
}
}
void
ipa_dump_fn_summary (FILE *f, struct cgraph_node *node)
{
if (node->definition)
{
class ipa_fn_summary *s = ipa_fn_summaries->get (node);
class ipa_size_summary *ss = ipa_size_summaries->get (node);
if (s != NULL)
{
size_time_entry *e;
int i;
fprintf (f, "IPA function summary for %s", node->dump_name ());
if (DECL_DISREGARD_INLINE_LIMITS (node->decl))
fprintf (f, " always_inline");
if (s->inlinable)
fprintf (f, " inlinable");
if (s->fp_expressions)
fprintf (f, " fp_expression");
if (s->builtin_constant_p_parms.length ())
{
fprintf (f, " builtin_constant_p_parms");
for (unsigned int i = 0;
i < s->builtin_constant_p_parms.length (); i++)
fprintf (f, " %i", s->builtin_constant_p_parms[i]);
}
fprintf (f, "\n global time: %f\n", s->time.to_double ());
fprintf (f, " self size: %i\n", ss->self_size);
fprintf (f, " global size: %i\n", ss->size);
fprintf (f, " min size: %i\n", s->min_size);
fprintf (f, " self stack: %i\n",
(int) ss->estimated_self_stack_size);
fprintf (f, " global stack: %i\n", (int) s->estimated_stack_size);
if (s->growth)
fprintf (f, " estimated growth:%i\n", (int) s->growth);
if (s->scc_no)
fprintf (f, " In SCC: %i\n", (int) s->scc_no);
for (i = 0; s->size_time_table.iterate (i, &e); i++)
{
fprintf (f, " size:%f, time:%f",
(double) e->size / ipa_fn_summary::size_scale,
e->time.to_double ());
if (e->exec_predicate != true)
{
fprintf (f, ", executed if:");
e->exec_predicate.dump (f, s->conds, 0);
}
if (e->exec_predicate != e->nonconst_predicate)
{
fprintf (f, ", nonconst if:");
e->nonconst_predicate.dump (f, s->conds, 0);
}
fprintf (f, "\n");
}
ipa_freqcounting_predicate *fcp;
bool first_fcp = true;
for (int i = 0; vec_safe_iterate (s->loop_iterations, i, &fcp); i++)
{
if (first_fcp)
{
fprintf (f, " loop iterations:");
first_fcp = false;
}
fprintf (f, " %3.2f for ", fcp->freq.to_double ());
fcp->predicate->dump (f, s->conds);
}
first_fcp = true;
for (int i = 0; vec_safe_iterate (s->loop_strides, i, &fcp); i++)
{
if (first_fcp)
{
fprintf (f, " loop strides:");
first_fcp = false;
}
fprintf (f, " %3.2f for :", fcp->freq.to_double ());
fcp->predicate->dump (f, s->conds);
}
fprintf (f, " calls:\n");
dump_ipa_call_summary (f, 4, node, s);
fprintf (f, "\n");
if (s->target_info)
fprintf (f, " target_info: %x\n", s->target_info);
}
else
fprintf (f, "IPA summary for %s is missing.\n", node->dump_name ());
}
}
DEBUG_FUNCTION void
ipa_debug_fn_summary (struct cgraph_node *node)
{
ipa_dump_fn_summary (stderr, node);
}
void
ipa_dump_fn_summaries (FILE *f)
{
struct cgraph_node *node;
FOR_EACH_DEFINED_FUNCTION (node)
if (!node->inlined_to)
ipa_dump_fn_summary (f, node);
}
/* Callback of walk_aliased_vdefs. Flags that it has been invoked to the
boolean variable pointed to by DATA. */
static bool
mark_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
void *data)
{
bool *b = (bool *) data;
*b = true;
return true;
}
/* If OP refers to value of function parameter, return the corresponding
parameter. If non-NULL, the size of the memory load (or the SSA_NAME of the
PARM_DECL) will be stored to *SIZE_P in that case too. */
static tree
unmodified_parm_1 (ipa_func_body_info *fbi, gimple *stmt, tree op,
poly_int64 *size_p)
{
/* SSA_NAME referring to parm default def? */
if (TREE_CODE (op) == SSA_NAME
&& SSA_NAME_IS_DEFAULT_DEF (op)
&& TREE_CODE (SSA_NAME_VAR (op)) == PARM_DECL)
{
if (size_p)
*size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
return SSA_NAME_VAR (op);
}
/* Non-SSA parm reference? */
if (TREE_CODE (op) == PARM_DECL
&& fbi->aa_walk_budget > 0)
{
bool modified = false;
ao_ref refd;
ao_ref_init (&refd, op);
int walked = walk_aliased_vdefs (&refd, gimple_vuse (stmt),
mark_modified, &modified, NULL, NULL,
fbi->aa_walk_budget);
if (walked < 0)
{
fbi->aa_walk_budget = 0;
return NULL_TREE;
}
fbi->aa_walk_budget -= walked;
if (!modified)
{
if (size_p)
*size_p = tree_to_poly_int64 (TYPE_SIZE (TREE_TYPE (op)));
return op;
}
}
return NULL_TREE;
}
/* If OP refers to value of function parameter, return the corresponding
parameter. Also traverse chains of SSA register assignments. If non-NULL,
the size of the memory load (or the SSA_NAME of the PARM_DECL) will be
stored to *SIZE_P in that case too. */
static tree
unmodified_parm (ipa_func_body_info *fbi, gimple *stmt, tree op,
poly_int64 *size_p)
{
tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
if (res)
return res;
if (TREE_CODE (op) == SSA_NAME
&& !SSA_NAME_IS_DEFAULT_DEF (op)
&& gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
return unmodified_parm (fbi, SSA_NAME_DEF_STMT (op),
gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)),
size_p);
return NULL_TREE;
}
/* If OP refers to a value of a function parameter or value loaded from an
aggregate passed to a parameter (either by value or reference), return TRUE
and store the number of the parameter to *INDEX_P, the access size into
*SIZE_P, and information whether and how it has been loaded from an
aggregate into *AGGPOS. INFO describes the function parameters, STMT is the
statement in which OP is used or loaded. */
static bool
unmodified_parm_or_parm_agg_item (struct ipa_func_body_info *fbi,
gimple *stmt, tree op, int *index_p,
poly_int64 *size_p,
struct agg_position_info *aggpos)
{
tree res = unmodified_parm_1 (fbi, stmt, op, size_p);
gcc_checking_assert (aggpos);
if (res)
{
*index_p = ipa_get_param_decl_index (fbi->info, res);
if (*index_p < 0)
return false;
aggpos->agg_contents = false;
aggpos->by_ref = false;
return true;
}
if (TREE_CODE (op) == SSA_NAME)
{
if (SSA_NAME_IS_DEFAULT_DEF (op)
|| !gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
return false;
stmt = SSA_NAME_DEF_STMT (op);
op = gimple_assign_rhs1 (stmt);
if (!REFERENCE_CLASS_P (op))
return unmodified_parm_or_parm_agg_item (fbi, stmt, op, index_p, size_p,
aggpos);
}
aggpos->agg_contents = true;
return ipa_load_from_parm_agg (fbi, fbi->info->descriptors,
stmt, op, index_p, &aggpos->offset,
size_p, &aggpos->by_ref);
}
/* See if statement might disappear after inlining.
0 - means not eliminated
1 - half of statements goes away
2 - for sure it is eliminated.
We are not terribly sophisticated, basically looking for simple abstraction
penalty wrappers. */
static int
eliminated_by_inlining_prob (ipa_func_body_info *fbi, gimple *stmt)
{
enum gimple_code code = gimple_code (stmt);
enum tree_code rhs_code;
if (!optimize)
return 0;
switch (code)
{
case GIMPLE_RETURN:
return 2;
case GIMPLE_ASSIGN:
if (gimple_num_ops (stmt) != 2)
return 0;
rhs_code = gimple_assign_rhs_code (stmt);
/* Casts of parameters, loads from parameters passed by reference
and stores to return value or parameters are often free after
inlining due to SRA and further combining.
Assume that half of statements goes away. */
if (CONVERT_EXPR_CODE_P (rhs_code)
|| rhs_code == VIEW_CONVERT_EXPR
|| rhs_code == ADDR_EXPR
|| gimple_assign_rhs_class (stmt) == GIMPLE_SINGLE_RHS)
{
tree rhs = gimple_assign_rhs1 (stmt);
tree lhs = gimple_assign_lhs (stmt);
tree inner_rhs = get_base_address (rhs);
tree inner_lhs = get_base_address (lhs);
bool rhs_free = false;
bool lhs_free = false;
if (!inner_rhs)
inner_rhs = rhs;
if (!inner_lhs)
inner_lhs = lhs;
/* Reads of parameter are expected to be free. */
if (unmodified_parm (fbi, stmt, inner_rhs, NULL))
rhs_free = true;
/* Match expressions of form &this->field. Those will most likely
combine with something upstream after inlining. */
else if (TREE_CODE (inner_rhs) == ADDR_EXPR)
{
tree op = get_base_address (TREE_OPERAND (inner_rhs, 0));
if (TREE_CODE (op) == PARM_DECL)
rhs_free = true;
else if (TREE_CODE (op) == MEM_REF
&& unmodified_parm (fbi, stmt, TREE_OPERAND (op, 0),
NULL))
rhs_free = true;
}
/* When parameter is not SSA register because its address is taken
and it is just copied into one, the statement will be completely
free after inlining (we will copy propagate backward). */
if (rhs_free && is_gimple_reg (lhs))
return 2;
/* Reads of parameters passed by reference
expected to be free (i.e. optimized out after inlining). */
if (TREE_CODE (inner_rhs) == MEM_REF
&& unmodified_parm (fbi, stmt, TREE_OPERAND (inner_rhs, 0), NULL))
rhs_free = true;
/* Copying parameter passed by reference into gimple register is
probably also going to copy propagate, but we can't be quite
sure. */
if (rhs_free && is_gimple_reg (lhs))
lhs_free = true;
/* Writes to parameters, parameters passed by value and return value
(either directly or passed via invisible reference) are free.
TODO: We ought to handle testcase like
struct a {int a,b;};
struct a
returnstruct (void)
{
struct a a ={1,2};
return a;
}
This translate into:
returnstruct ()
{
int a$b;
int a$a;
struct a a;
struct a D.2739;
<bb 2>:
D.2739.a = 1;
D.2739.b = 2;
return D.2739;
}
For that we either need to copy ipa-split logic detecting writes
to return value. */
if (TREE_CODE (inner_lhs) == PARM_DECL
|| TREE_CODE (inner_lhs) == RESULT_DECL
|| (TREE_CODE (inner_lhs) == MEM_REF
&& (unmodified_parm (fbi, stmt, TREE_OPERAND (inner_lhs, 0),
NULL)
|| (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
&& SSA_NAME_VAR (TREE_OPERAND (inner_lhs, 0))
&& TREE_CODE (SSA_NAME_VAR (TREE_OPERAND
(inner_lhs,
0))) == RESULT_DECL))))
lhs_free = true;
if (lhs_free
&& (is_gimple_reg (rhs) || is_gimple_min_invariant (rhs)))
rhs_free = true;
if (lhs_free && rhs_free)
return 1;
}
return 0;
default:
return 0;
}
}
/* Analyze EXPR if it represents a series of simple operations performed on
a function parameter and return true if so. FBI, STMT, EXPR, INDEX_P and
AGGPOS have the same meaning like in unmodified_parm_or_parm_agg_item.
Type of the parameter or load from an aggregate via the parameter is
stored in *TYPE_P. Operations on the parameter are recorded to
PARAM_OPS_P if it is not NULL. */
static bool
decompose_param_expr (struct ipa_func_body_info *fbi,
gimple *stmt, tree expr,
int *index_p, tree *type_p,
struct agg_position_info *aggpos,
expr_eval_ops *param_ops_p = NULL)
{
int op_limit = opt_for_fn (fbi->node->decl, param_ipa_max_param_expr_ops);
int op_count = 0;
if (param_ops_p)
*param_ops_p = NULL;
while (true)
{
expr_eval_op eval_op;
unsigned rhs_count;
unsigned cst_count = 0;
if (unmodified_parm_or_parm_agg_item (fbi, stmt, expr, index_p, NULL,
aggpos))
{
tree type = TREE_TYPE (expr);
if (aggpos->agg_contents)
{
/* Stop if containing bit-field. */
if (TREE_CODE (expr) == BIT_FIELD_REF
|| contains_bitfld_component_ref_p (expr))
break;
}
*type_p = type;
return true;
}
if (TREE_CODE (expr) != SSA_NAME || SSA_NAME_IS_DEFAULT_DEF (expr))
break;
if (!is_gimple_assign (stmt = SSA_NAME_DEF_STMT (expr)))
break;
switch (gimple_assign_rhs_class (stmt))
{
case GIMPLE_SINGLE_RHS:
expr = gimple_assign_rhs1 (stmt);
continue;
case GIMPLE_UNARY_RHS:
rhs_count = 1;
break;
case GIMPLE_BINARY_RHS:
rhs_count = 2;
break;
case GIMPLE_TERNARY_RHS:
rhs_count = 3;
break;
default:
goto fail;
}
/* Stop if expression is too complex. */
if (op_count++ == op_limit)
break;
if (param_ops_p)
{
eval_op.code = gimple_assign_rhs_code (stmt);
eval_op.type = TREE_TYPE (gimple_assign_lhs (stmt));
eval_op.val[0] = NULL_TREE;
eval_op.val[1] = NULL_TREE;
}
expr = NULL_TREE;
for (unsigned i = 0; i < rhs_count; i++)
{
tree op = gimple_op (stmt, i + 1);
gcc_assert (op && !TYPE_P (op));
if (is_gimple_ip_invariant (op))
{
if (++cst_count == rhs_count)
goto fail;
eval_op.val[cst_count - 1] = op;
}
else if (!expr)
{
/* Found a non-constant operand, and record its index in rhs
operands. */
eval_op.index = i;
expr = op;
}
else
{
/* Found more than one non-constant operands. */
goto fail;
}
}
if (param_ops_p)
vec_safe_insert (*param_ops_p, 0, eval_op);
}
/* Failed to decompose, free resource and return. */
fail:
if (param_ops_p)
vec_free (*param_ops_p);
return false;
}
/* Record to SUMMARY that PARM is used by builtin_constant_p. */
static void
add_builtin_constant_p_parm (class ipa_fn_summary *summary, int parm)
{
int ip;
/* Avoid duplicates. */
for (unsigned int i = 0;
summary->builtin_constant_p_parms.iterate (i, &ip); i++)
if (ip == parm)
return;
summary->builtin_constant_p_parms.safe_push (parm);
}
/* If BB ends by a conditional we can turn into predicates, attach corresponding
predicates to the CFG edges. */
static void
set_cond_stmt_execution_predicate (struct ipa_func_body_info *fbi,
class ipa_fn_summary *summary,
class ipa_node_params *params_summary,
basic_block bb)
{
gimple *last;
tree op, op2;
int index;
struct agg_position_info aggpos;
enum tree_code code, inverted_code;
edge e;
edge_iterator ei;
gimple *set_stmt;
tree param_type;
expr_eval_ops param_ops;
last = last_stmt (bb);
if (!last || gimple_code (last) != GIMPLE_COND)
return;
if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
return;
op = gimple_cond_lhs (last);
if (decompose_param_expr (fbi, last, op, &index, &param_type, &aggpos,
&param_ops))
{
code = gimple_cond_code (last);
inverted_code = invert_tree_comparison (code, HONOR_NANS (op));
FOR_EACH_EDGE (e, ei, bb->succs)
{
enum tree_code this_code = (e->flags & EDGE_TRUE_VALUE
? code : inverted_code);
/* invert_tree_comparison will return ERROR_MARK on FP
comparisons that are not EQ/NE instead of returning proper
unordered one. Be sure it is not confused with NON_CONSTANT.
And if the edge's target is the final block of diamond CFG graph
of this conditional statement, we do not need to compute
predicate for the edge because the final block's predicate must
be at least as that of the first block of the statement. */
if (this_code != ERROR_MARK
&& !dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
{
ipa_predicate p
= add_condition (summary, params_summary, index,
param_type, &aggpos,
this_code, gimple_cond_rhs (last), param_ops);
e->aux = edge_predicate_pool.allocate ();
*(ipa_predicate *) e->aux = p;
}
}
vec_free (param_ops);
}
if (TREE_CODE (op) != SSA_NAME)
return;
/* Special case
if (builtin_constant_p (op))
constant_code
else
nonconstant_code.
Here we can predicate nonconstant_code. We can't
really handle constant_code since we have no predicate
for this and also the constant code is not known to be
optimized away when inliner doesn't see operand is constant.
Other optimizers might think otherwise. */
if (gimple_cond_code (last) != NE_EXPR
|| !integer_zerop (gimple_cond_rhs (last)))
return;
set_stmt = SSA_NAME_DEF_STMT (op);
if (!gimple_call_builtin_p (set_stmt, BUILT_IN_CONSTANT_P)
|| gimple_call_num_args (set_stmt) != 1)
return;
op2 = gimple_call_arg (set_stmt, 0);
if (!decompose_param_expr (fbi, set_stmt, op2, &index, &param_type, &aggpos))
return;
if (!aggpos.by_ref)
add_builtin_constant_p_parm (summary, index);
FOR_EACH_EDGE (e, ei, bb->succs) if (e->flags & EDGE_FALSE_VALUE)
{
ipa_predicate p = add_condition (summary, params_summary, index,
param_type, &aggpos,
ipa_predicate::is_not_constant, NULL_TREE);
e->aux = edge_predicate_pool.allocate ();
*(ipa_predicate *) e->aux = p;
}
}
/* If BB ends by a switch we can turn into predicates, attach corresponding
predicates to the CFG edges. */
static void
set_switch_stmt_execution_predicate (struct ipa_func_body_info *fbi,
class ipa_fn_summary *summary,
class ipa_node_params *params_summary,
basic_block bb)
{
gimple *lastg;
tree op;
int index;
struct agg_position_info aggpos;
edge e;
edge_iterator ei;
size_t n;
size_t case_idx;
tree param_type;
expr_eval_ops param_ops;
lastg = last_stmt (bb);
if (!lastg || gimple_code (lastg) != GIMPLE_SWITCH)
return;
gswitch *last = as_a <gswitch *> (lastg);
op = gimple_switch_index (last);
if (!decompose_param_expr (fbi, last, op, &index, &param_type, &aggpos,
&param_ops))
return;
auto_vec<std::pair<tree, tree> > ranges;
tree type = TREE_TYPE (op);
int bound_limit = opt_for_fn (fbi->node->decl,
param_ipa_max_switch_predicate_bounds);
int bound_count = 0;
value_range vr;
get_range_query (cfun)->range_of_expr (vr, op);
if (vr.undefined_p ())
vr.set_varying (TREE_TYPE (op));
value_range_kind vr_type = vr.kind ();
wide_int vr_wmin = wi::to_wide (vr.min ());
wide_int vr_wmax = wi::to_wide (vr.max ());
FOR_EACH_EDGE (e, ei, bb->succs)
{
e->aux = edge_predicate_pool.allocate ();
*(ipa_predicate *) e->aux = false;
}
e = gimple_switch_edge (cfun, last, 0);
/* Set BOUND_COUNT to maximum count to bypass computing predicate for
default case if its target basic block is in convergence point of all
switch cases, which can be determined by checking whether it
post-dominates the switch statement. */
if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
bound_count = INT_MAX;
n = gimple_switch_num_labels (last);
for (case_idx = 1; case_idx < n; ++case_idx)
{
tree cl = gimple_switch_label (last, case_idx);
tree min = CASE_LOW (cl);
tree max = CASE_HIGH (cl);
ipa_predicate p;
e = gimple_switch_edge (cfun, last, case_idx);
/* The case value might not have same type as switch expression,
extend the value based on the expression type. */
if (TREE_TYPE (min) != type)
min = wide_int_to_tree (type, wi::to_wide (min));
if (!max)
max = min;
else if (TREE_TYPE (max) != type)
max = wide_int_to_tree (type, wi::to_wide (max));
/* The case's target basic block is in convergence point of all switch
cases, its predicate should be at least as that of the switch
statement. */
if (dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest))
p = true;
else if (min == max)
p = add_condition (summary, params_summary, index, param_type,
&aggpos, EQ_EXPR, min, param_ops);
else
{
ipa_predicate p1, p2;
p1 = add_condition (summary, params_summary, index, param_type,
&aggpos, GE_EXPR, min, param_ops);
p2 = add_condition (summary, params_summary,index, param_type,
&aggpos, LE_EXPR, max, param_ops);
p = p1 & p2;
}
*(ipa_predicate *) e->aux
= p.or_with (summary->conds, *(ipa_predicate *) e->aux);
/* If there are too many disjoint case ranges, predicate for default
case might become too complicated. So add a limit here. */
if (bound_count > bound_limit)
continue;
bool new_range = true;
if (!ranges.is_empty ())
{
wide_int curr_wmin = wi::to_wide (min);
wide_int last_wmax = wi::to_wide (ranges.last ().second);
/* Merge case ranges if they are continuous. */
if (curr_wmin == last_wmax + 1)
new_range = false;
else if (vr_type == VR_ANTI_RANGE)
{
/* If two disjoint case ranges can be connected by anti-range
of switch index, combine them to one range. */
if (wi::lt_p (vr_wmax, curr_wmin - 1, TYPE_SIGN (type)))
vr_type = VR_UNDEFINED;
else if (wi::le_p (vr_wmin, last_wmax + 1, TYPE_SIGN (type)))
new_range = false;
}
}
/* Create/extend a case range. And we count endpoints of range set,
this number nearly equals to number of conditions that we will create
for predicate of default case. */
if (new_range)
{
bound_count += (min == max) ? 1 : 2;
ranges.safe_push (std::make_pair (min, max));
}
else
{
bound_count += (ranges.last ().first == ranges.last ().second);
ranges.last ().second = max;
}
}
e = gimple_switch_edge (cfun, last, 0);
if (bound_count > bound_limit)
{
*(ipa_predicate *) e->aux = true;
vec_free (param_ops);
return;
}
ipa_predicate p_seg = true;
ipa_predicate p_all = false;
if (vr_type != VR_RANGE)
{
vr_wmin = wi::to_wide (TYPE_MIN_VALUE (type));
vr_wmax = wi::to_wide (TYPE_MAX_VALUE (type));
}
/* Construct predicate to represent default range set that is negation of
all case ranges. Case range is classified as containing single/non-single
values. Suppose a piece of case ranges in the following.
[D1...D2] [S1] ... [Sn] [D3...D4]
To represent default case's range sets between two non-single value
case ranges (From D2 to D3), we construct predicate as:
D2 < x < D3 && x != S1 && ... && x != Sn
*/
for (size_t i = 0; i < ranges.length (); i++)
{
tree min = ranges[i].first;
tree max = ranges[i].second;
if (min == max)
p_seg &= add_condition (summary, params_summary, index,
param_type, &aggpos, NE_EXPR,
min, param_ops);
else
{
/* Do not create sub-predicate for range that is beyond low bound
of switch index. */
if (wi::lt_p (vr_wmin, wi::to_wide (min), TYPE_SIGN (type)))
{
p_seg &= add_condition (summary, params_summary, index,
param_type, &aggpos,
LT_EXPR, min, param_ops);
p_all = p_all.or_with (summary->conds, p_seg);
}
/* Do not create sub-predicate for range that is beyond up bound
of switch index. */
if (wi::le_p (vr_wmax, wi::to_wide (max), TYPE_SIGN (type)))
{
p_seg = false;
break;
}
p_seg = add_condition (summary, params_summary, index,
param_type, &aggpos, GT_EXPR,
max, param_ops);
}
}
p_all = p_all.or_with (summary->conds, p_seg);
*(ipa_predicate *) e->aux
= p_all.or_with (summary->conds, *(ipa_predicate *) e->aux);
vec_free (param_ops);
}
/* For each BB in NODE attach to its AUX pointer predicate under
which it is executable. */
static void
compute_bb_predicates (struct ipa_func_body_info *fbi,
struct cgraph_node *node,
class ipa_fn_summary *summary,
class ipa_node_params *params_summary)
{
struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
bool done = false;
basic_block bb;
FOR_EACH_BB_FN (bb, my_function)
{
set_cond_stmt_execution_predicate (fbi, summary, params_summary, bb);
set_switch_stmt_execution_predicate (fbi, summary, params_summary, bb);
}
/* Entry block is always executable. */
ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux
= edge_predicate_pool.allocate ();
*(ipa_predicate *) ENTRY_BLOCK_PTR_FOR_FN (my_function)->aux = true;
/* A simple dataflow propagation of predicates forward in the CFG.
TODO: work in reverse postorder. */
while (!done)
{
done = true;
FOR_EACH_BB_FN (bb, my_function)
{
ipa_predicate p = false;
edge e;
edge_iterator ei;
FOR_EACH_EDGE (e, ei, bb->preds)
{
if (e->src->aux)
{
ipa_predicate this_bb_predicate
= *(ipa_predicate *) e->src->aux;
if (e->aux)
this_bb_predicate &= (*(ipa_predicate *) e->aux);
p = p.or_with (summary->conds, this_bb_predicate);
if (p == true)
break;
}
}
if (p != false)
{
basic_block pdom_bb;
if (!bb->aux)
{
done = false;
bb->aux = edge_predicate_pool.allocate ();
*((ipa_predicate *) bb->aux) = p;
}
else if (p != *(ipa_predicate *) bb->aux)
{
/* This OR operation is needed to ensure monotonous data flow
in the case we hit the limit on number of clauses and the
and/or operations above give approximate answers. */
p = p.or_with (summary->conds, *(ipa_predicate *)bb->aux);
if (p != *(ipa_predicate *)bb->aux)
{
done = false;
*((ipa_predicate *)bb->aux) = p;
}
}
/* For switch/if statement, we can OR-combine predicates of all
its cases/branches to get predicate for basic block in their
convergence point, but sometimes this will generate very
complicated predicate. Actually, we can get simplified
predicate in another way by using the fact that predicate
for a basic block must also hold true for its post dominators.
To be specific, basic block in convergence point of
conditional statement should include predicate of the
statement. */
pdom_bb = get_immediate_dominator (CDI_POST_DOMINATORS, bb);
if (pdom_bb == EXIT_BLOCK_PTR_FOR_FN (my_function) || !pdom_bb)
;
else if (!pdom_bb->aux)
{
done = false;
pdom_bb->aux = edge_predicate_pool.allocate ();
*((ipa_predicate *)pdom_bb->aux) = p;
}
else if (p != *(ipa_predicate *)pdom_bb->aux)
{
p = p.or_with (summary->conds,
*(ipa_predicate *)pdom_bb->aux);
if (p != *(ipa_predicate *)pdom_bb->aux)
{
done = false;
*((ipa_predicate *)pdom_bb->aux) = p;
}
}
}
}
}
}
/* Return predicate specifying when the STMT might have result that is not
a compile time constant. */
static ipa_predicate
will_be_nonconstant_expr_predicate (ipa_func_body_info *fbi,
class ipa_fn_summary *summary,
class ipa_node_params *params_summary,
tree expr,
vec<ipa_predicate> nonconstant_names)
{
tree parm;
int index;
while (UNARY_CLASS_P (expr))
expr = TREE_OPERAND (expr, 0);
parm = unmodified_parm (fbi, NULL, expr, NULL);
if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
return add_condition (summary, params_summary, index, TREE_TYPE (parm), NULL,
ipa_predicate::changed, NULL_TREE);
if (is_gimple_min_invariant (expr))
return false;
if (TREE_CODE (expr) == SSA_NAME)
return nonconstant_names[SSA_NAME_VERSION (expr)];
if (BINARY_CLASS_P (expr) || COMPARISON_CLASS_P (expr))
{
ipa_predicate p1
= will_be_nonconstant_expr_predicate (fbi, summary,
params_summary,
TREE_OPERAND (expr, 0),
nonconstant_names);
if (p1 == true)
return p1;
ipa_predicate p2
= will_be_nonconstant_expr_predicate (fbi, summary,
params_summary,
TREE_OPERAND (expr, 1),
nonconstant_names);
return p1.or_with (summary->conds, p2);
}
else if (TREE_CODE (expr) == COND_EXPR)
{
ipa_predicate p1
= will_be_nonconstant_expr_predicate (fbi, summary,
params_summary,
TREE_OPERAND (expr, 0),
nonconstant_names);
if (p1 == true)
return p1;
ipa_predicate p2
= will_be_nonconstant_expr_predicate (fbi, summary,
params_summary,
TREE_OPERAND (expr, 1),
nonconstant_names);
if (p2 == true)
return p2;
p1 = p1.or_with (summary->conds, p2);
p2 = will_be_nonconstant_expr_predicate (fbi, summary,
params_summary,
TREE_OPERAND (expr, 2),
nonconstant_names);
return p2.or_with (summary->conds, p1);
}
else if (TREE_CODE (expr) == CALL_EXPR)
return true;
else
{
debug_tree (expr);
gcc_unreachable ();
}
}
/* Return predicate specifying when the STMT might have result that is not
a compile time constant. */
static ipa_predicate
will_be_nonconstant_predicate (struct ipa_func_body_info *fbi,
class ipa_fn_summary *summary,
class ipa_node_params *params_summary,
gimple *stmt,
vec<ipa_predicate> nonconstant_names)
{
ipa_predicate p = true;
ssa_op_iter iter;
tree use;
tree param_type = NULL_TREE;
ipa_predicate op_non_const;
bool is_load;
int base_index;
struct agg_position_info aggpos;
/* What statements might be optimized away
when their arguments are constant. */
if (gimple_code (stmt) != GIMPLE_ASSIGN
&& gimple_code (stmt) != GIMPLE_COND
&& gimple_code (stmt) != GIMPLE_SWITCH
&& (gimple_code (stmt) != GIMPLE_CALL
|| !(gimple_call_flags (stmt) & ECF_CONST)))
return p;
/* Stores will stay anyway. */
if (gimple_store_p (stmt))
return p;
is_load = gimple_assign_load_p (stmt);
/* Loads can be optimized when the value is known. */
if (is_load)
{
tree op = gimple_assign_rhs1 (stmt);
if (!decompose_param_expr (fbi, stmt, op, &base_index, &param_type,
&aggpos))
return p;
}
else
base_index = -1;
/* See if we understand all operands before we start
adding conditionals. */
FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
{
tree parm = unmodified_parm (fbi, stmt, use, NULL);
/* For arguments we can build a condition. */
if (parm && ipa_get_param_decl_index (fbi->info, parm) >= 0)
continue;
if (TREE_CODE (use) != SSA_NAME)
return p;
/* If we know when operand is constant,
we still can say something useful. */
if (nonconstant_names[SSA_NAME_VERSION (use)] != true)
continue;
return p;
}
if (is_load)
op_non_const =
add_condition (summary, params_summary,
base_index, param_type, &aggpos,
ipa_predicate::changed, NULL_TREE);
else
op_non_const = false;
FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
{
tree parm = unmodified_parm (fbi, stmt, use, NULL);
int index;
if (parm && (index = ipa_get_param_decl_index (fbi->info, parm)) >= 0)
{
if (index != base_index)
p = add_condition (summary, params_summary, index,
TREE_TYPE (parm), NULL,
ipa_predicate::changed, NULL_TREE);
else
continue;
}
else
p = nonconstant_names[SSA_NAME_VERSION (use)];
op_non_const = p.or_with (summary->conds, op_non_const);
}
if ((gimple_code (stmt) == GIMPLE_ASSIGN || gimple_code (stmt) == GIMPLE_CALL)
&& gimple_op (stmt, 0)
&& TREE_CODE (gimple_op (stmt, 0)) == SSA_NAME)
nonconstant_names[SSA_NAME_VERSION (gimple_op (stmt, 0))]
= op_non_const;
return op_non_const;
}
struct record_modified_bb_info
{
tree op;
bitmap bb_set;
gimple *stmt;
};
/* Value is initialized in INIT_BB and used in USE_BB. We want to compute
probability how often it changes between USE_BB.
INIT_BB->count/USE_BB->count is an estimate, but if INIT_BB
is in different loop nest, we can do better.
This is all just estimate. In theory we look for minimal cut separating
INIT_BB and USE_BB, but we only want to anticipate loop invariant motion
anyway. */
static basic_block
get_minimal_bb (basic_block init_bb, basic_block use_bb)
{
class loop *l = find_common_loop (init_bb->loop_father, use_bb->loop_father);
if (l && l->header->count < init_bb->count)
return l->header;
return init_bb;
}
/* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
set except for info->stmt. */
static bool
record_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef, void *data)
{
struct record_modified_bb_info *info =
(struct record_modified_bb_info *) data;
if (SSA_NAME_DEF_STMT (vdef) == info->stmt)
return false;
if (gimple_clobber_p (SSA_NAME_DEF_STMT (vdef)))
return false;
bitmap_set_bit (info->bb_set,
SSA_NAME_IS_DEFAULT_DEF (vdef)
? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
: get_minimal_bb
(gimple_bb (SSA_NAME_DEF_STMT (vdef)),
gimple_bb (info->stmt))->index);
if (dump_file)
{
fprintf (dump_file, " Param ");
print_generic_expr (dump_file, info->op, TDF_SLIM);
fprintf (dump_file, " changed at bb %i, minimal: %i stmt: ",
gimple_bb (SSA_NAME_DEF_STMT (vdef))->index,
get_minimal_bb
(gimple_bb (SSA_NAME_DEF_STMT (vdef)),
gimple_bb (info->stmt))->index);
print_gimple_stmt (dump_file, SSA_NAME_DEF_STMT (vdef), 0);
}
return false;
}
/* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
will change since last invocation of STMT.
Value 0 is reserved for compile time invariants.
For common parameters it is REG_BR_PROB_BASE. For loop invariants it
ought to be REG_BR_PROB_BASE / estimated_iters. */
static int
param_change_prob (ipa_func_body_info *fbi, gimple *stmt, int i)
{
tree op = gimple_call_arg (stmt, i);
basic_block bb = gimple_bb (stmt);
if (TREE_CODE (op) == WITH_SIZE_EXPR)
op = TREE_OPERAND (op, 0);
tree base = get_base_address (op);
/* Global invariants never change. */
if (is_gimple_min_invariant (base))
return 0;
/* We would have to do non-trivial analysis to really work out what
is the probability of value to change (i.e. when init statement
is in a sibling loop of the call).
We do an conservative estimate: when call is executed N times more often
than the statement defining value, we take the frequency 1/N. */
if (TREE_CODE (base) == SSA_NAME)
{
profile_count init_count;
if (!bb->count.nonzero_p ())
return REG_BR_PROB_BASE;
if (SSA_NAME_IS_DEFAULT_DEF (base))
init_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
else
init_count = get_minimal_bb
(gimple_bb (SSA_NAME_DEF_STMT (base)),
gimple_bb (stmt))->count;
if (init_count < bb->count)
return MAX ((init_count.to_sreal_scale (bb->count)
* REG_BR_PROB_BASE).to_int (), 1);
return REG_BR_PROB_BASE;
}
else
{
ao_ref refd;
profile_count max = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
struct record_modified_bb_info info;
tree init = ctor_for_folding (base);
if (init != error_mark_node)
return 0;
if (!bb->count.nonzero_p () || fbi->aa_walk_budget == 0)
return REG_BR_PROB_BASE;
if (dump_file)
{
fprintf (dump_file, " Analyzing param change probability of ");
print_generic_expr (dump_file, op, TDF_SLIM);
fprintf (dump_file, "\n");
}
ao_ref_init (&refd, op);
info.op = op;
info.stmt = stmt;
info.bb_set = BITMAP_ALLOC (NULL);
int walked
= walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
NULL, NULL, fbi->aa_walk_budget);
if (walked > 0)
fbi->aa_walk_budget -= walked;
if (walked < 0 || bitmap_bit_p (info.bb_set, bb->index))
{
if (walked < 0)
fbi->aa_walk_budget = 0;
if (dump_file)
{
if (walked < 0)
fprintf (dump_file, " Ran out of AA walking budget.\n");
else
fprintf (dump_file, " Set in same BB as used.\n");
}
BITMAP_FREE (info.bb_set);
return REG_BR_PROB_BASE;
}
bitmap_iterator bi;
unsigned index;
/* Lookup the most frequent update of the value and believe that
it dominates all the other; precise analysis here is difficult. */
EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
max = max.max (BASIC_BLOCK_FOR_FN (cfun, index)->count);
if (dump_file)
{
fprintf (dump_file, " Set with count ");
max.dump (dump_file);
fprintf (dump_file, " and used with count ");
bb->count.dump (dump_file);
fprintf (dump_file, " freq %f\n",
max.to_sreal_scale (bb->count).to_double ());
}
BITMAP_FREE (info.bb_set);
if (max < bb->count)
return MAX ((max.to_sreal_scale (bb->count)
* REG_BR_PROB_BASE).to_int (), 1);
return REG_BR_PROB_BASE;
}
}
/* Find whether a basic block BB is the final block of a (half) diamond CFG
sub-graph and if the predicate the condition depends on is known. If so,
return true and store the pointer the predicate in *P. */
static bool
phi_result_unknown_predicate (ipa_func_body_info *fbi,
ipa_fn_summary *summary,
class ipa_node_params *params_summary,
basic_block bb,
ipa_predicate *p,
vec<ipa_predicate> nonconstant_names)
{
edge e;
edge_iterator ei;
basic_block first_bb = NULL;
gimple *stmt;
if (single_pred_p (bb))
{
*p = false;
return true;
}
FOR_EACH_EDGE (e, ei, bb->preds)
{
if (single_succ_p (e->src))
{
if (!single_pred_p (e->src))
return false;
if (!first_bb)
first_bb = single_pred (e->src);
else if (single_pred (e->src) != first_bb)
return false;
}
else
{
if (!first_bb)
first_bb = e->src;
else if (e->src != first_bb)
return false;
}
}
if (!first_bb)
return false;
stmt = last_stmt (first_bb);
if (!stmt
|| gimple_code (stmt) != GIMPLE_COND
|| !is_gimple_ip_invariant (gimple_cond_rhs (stmt)))
return false;
*p = will_be_nonconstant_expr_predicate (fbi, summary, params_summary,
gimple_cond_lhs (stmt),
nonconstant_names);
if (*p == true)
return false;
else
return true;
}
/* Given a PHI statement in a function described by inline properties SUMMARY
and *P being the predicate describing whether the selected PHI argument is
known, store a predicate for the result of the PHI statement into
NONCONSTANT_NAMES, if possible. */
static void
predicate_for_phi_result (class ipa_fn_summary *summary, gphi *phi,
ipa_predicate *p,
vec<ipa_predicate> nonconstant_names)
{
unsigned i;
for (i = 0; i < gimple_phi_num_args (phi); i++)
{
tree arg = gimple_phi_arg (phi, i)->def;
if (!is_gimple_min_invariant (arg))
{
gcc_assert (TREE_CODE (arg) == SSA_NAME);
*p = p->or_with (summary->conds,
nonconstant_names[SSA_NAME_VERSION (arg)]);
if (*p == true)
return;
}
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "\t\tphi predicate: ");
p->dump (dump_file, summary->conds);
}
nonconstant_names[SSA_NAME_VERSION (gimple_phi_result (phi))] = *p;
}
/* For a typical usage of __builtin_expect (a<b, 1), we
may introduce an extra relation stmt:
With the builtin, we have
t1 = a <= b;
t2 = (long int) t1;
t3 = __builtin_expect (t2, 1);
if (t3 != 0)
goto ...
Without the builtin, we have
if (a<=b)
goto...
This affects the size/time estimation and may have
an impact on the earlier inlining.
Here find this pattern and fix it up later. */
static gimple *
find_foldable_builtin_expect (basic_block bb)
{
gimple_stmt_iterator bsi;
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
{
gimple *stmt = gsi_stmt (bsi);
if (gimple_call_builtin_p (stmt, BUILT_IN_EXPECT)
|| gimple_call_builtin_p (stmt, BUILT_IN_EXPECT_WITH_PROBABILITY)
|| gimple_call_internal_p (stmt, IFN_BUILTIN_EXPECT))
{
tree var = gimple_call_lhs (stmt);
tree arg = gimple_call_arg (stmt, 0);
use_operand_p use_p;
gimple *use_stmt;
bool match = false;
bool done = false;
if (!var || !arg)
continue;
gcc_assert (TREE_CODE (var) == SSA_NAME);
while (TREE_CODE (arg) == SSA_NAME)
{
gimple *stmt_tmp = SSA_NAME_DEF_STMT (arg);
if (!is_gimple_assign (stmt_tmp))
break;
switch (gimple_assign_rhs_code (stmt_tmp))
{
case LT_EXPR:
case LE_EXPR:
case GT_EXPR:
case GE_EXPR:
case EQ_EXPR:
case NE_EXPR:
match = true;
done = true;
break;
CASE_CONVERT:
break;
default:
done = true;
break;
}
if (done)
break;
arg = gimple_assign_rhs1 (stmt_tmp);
}
if (match && single_imm_use (var, &use_p, &use_stmt)
&& gimple_code (use_stmt) == GIMPLE_COND)
return use_stmt;
}
}
return NULL;
}
/* Return true when the basic blocks contains only clobbers followed by RESX.
Such BBs are kept around to make removal of dead stores possible with
presence of EH and will be optimized out by optimize_clobbers later in the
game.
NEED_EH is used to recurse in case the clobber has non-EH predecessors
that can be clobber only, too.. When it is false, the RESX is not necessary
on the end of basic block. */
static bool
clobber_only_eh_bb_p (basic_block bb, bool need_eh = true)
{
gimple_stmt_iterator gsi = gsi_last_bb (bb);
edge_iterator ei;
edge e;
if (need_eh)
{
if (gsi_end_p (gsi))
return false;
if (gimple_code (gsi_stmt (gsi)) != GIMPLE_RESX)
return false;
gsi_prev (&gsi);
}
else if (!single_succ_p (bb))
return false;
for (; !gsi_end_p (gsi); gsi_prev (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (is_gimple_debug (stmt))
continue;
if (gimple_clobber_p (stmt))
continue;
if (gimple_code (stmt) == GIMPLE_LABEL)
break;
return false;
}
/* See if all predecessors are either throws or clobber only BBs. */
FOR_EACH_EDGE (e, ei, bb->preds)
if (!(e->flags & EDGE_EH)
&& !clobber_only_eh_bb_p (e->src, false))
return false;
return true;
}
/* Return true if STMT compute a floating point expression that may be affected
by -ffast-math and similar flags. */
static bool
fp_expression_p (gimple *stmt)
{
ssa_op_iter i;
tree op;
FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_DEF|SSA_OP_USE)
if (FLOAT_TYPE_P (TREE_TYPE (op)))
return true;
return false;
}
/* Return true if T references memory location that is local
for the function (that means, dead after return) or read-only. */
bool
refs_local_or_readonly_memory_p (tree t)
{
/* Non-escaping memory is fine. */
t = get_base_address (t);
if ((TREE_CODE (t) == MEM_REF
|| TREE_CODE (t) == TARGET_MEM_REF))
return points_to_local_or_readonly_memory_p (TREE_OPERAND (t, 0));
/* Automatic variables are fine. */
if (DECL_P (t)
&& auto_var_in_fn_p (t, current_function_decl))
return true;
/* Read-only variables are fine. */
if (DECL_P (t) && TREE_READONLY (t))
return true;
return false;
}
/* Return true if T is a pointer pointing to memory location that is local
for the function (that means, dead after return) or read-only. */
bool
points_to_local_or_readonly_memory_p (tree t)
{
/* See if memory location is clearly invalid. */
if (integer_zerop (t))
return flag_delete_null_pointer_checks;
if (TREE_CODE (t) == SSA_NAME)
{
/* For IPA passes we can consinder accesses to return slot local
even if it is not local in the sense that memory is dead by
the end of founction.
The outer function will see a store in the call assignment
and thus this will do right thing for all uses of this
function in the current IPA passes (modref, pure/const discovery
and inlining heuristics). */
if (DECL_RESULT (current_function_decl)
&& DECL_BY_REFERENCE (DECL_RESULT (current_function_decl))
&& t == ssa_default_def (cfun, DECL_RESULT (current_function_decl)))
return true;
return !ptr_deref_may_alias_global_p (t, false);
}
if (TREE_CODE (t) == ADDR_EXPR)
return refs_local_or_readonly_memory_p (TREE_OPERAND (t, 0));
return false;
}
/* Analyze function body for NODE.
EARLY indicates run from early optimization pipeline. */
static void
analyze_function_body (struct cgraph_node *node, bool early)
{
sreal time = opt_for_fn (node->decl, param_uninlined_function_time);
/* Estimate static overhead for function prologue/epilogue and alignment. */
int size = opt_for_fn (node->decl, param_uninlined_function_insns);
/* Benefits are scaled by probability of elimination that is in range
<0,2>. */
basic_block bb;
struct function *my_function = DECL_STRUCT_FUNCTION (node->decl);
sreal freq;
class ipa_fn_summary *info = ipa_fn_summaries->get_create (node);
ipa_node_params *params_summary
= early ? NULL : ipa_node_params_sum->get (node);
ipa_predicate bb_predicate;
struct ipa_func_body_info fbi;
vec<ipa_predicate> nonconstant_names = vNULL;
int nblocks, n;
int *order;
gimple *fix_builtin_expect_stmt;
gcc_assert (my_function && my_function->cfg);
gcc_assert (cfun == my_function);
memset(&fbi, 0, sizeof(fbi));
vec_free (info->conds);
info->conds = NULL;
info->size_time_table.release ();
info->call_size_time_table.release ();
/* When optimizing and analyzing for IPA inliner, initialize loop optimizer
so we can produce proper inline hints.
When optimizing and analyzing for early inliner, initialize node params
so we can produce correct BB predicates. */
if (opt_for_fn (node->decl, optimize))
{
calculate_dominance_info (CDI_DOMINATORS);
calculate_dominance_info (CDI_POST_DOMINATORS);
if (!early)
loop_optimizer_init (LOOPS_NORMAL | LOOPS_HAVE_RECORDED_EXITS);
else
{
ipa_check_create_node_params ();
ipa_initialize_node_params (node);
}
if (ipa_node_params_sum)
{
fbi.node = node;
fbi.info = ipa_node_params_sum->get (node);
fbi.bb_infos = vNULL;
fbi.bb_infos.safe_grow_cleared (last_basic_block_for_fn (cfun), true);
fbi.param_count = count_formal_params (node->decl);
fbi.aa_walk_budget = opt_for_fn (node->decl, param_ipa_max_aa_steps);
nonconstant_names.safe_grow_cleared
(SSANAMES (my_function)->length (), true);
}
}
if (dump_file)
fprintf (dump_file, "\nAnalyzing function body size: %s\n",
node->dump_name ());
/* When we run into maximal number of entries, we assign everything to the
constant truth case. Be sure to have it in list. */
bb_predicate = true;
info->account_size_time (0, 0, bb_predicate, bb_predicate);
bb_predicate = ipa_predicate::not_inlined ();
info->account_size_time (opt_for_fn (node->decl,
param_uninlined_function_insns)
* ipa_fn_summary::size_scale,
opt_for_fn (node->decl,
param_uninlined_function_time),
bb_predicate,
bb_predicate);
/* Only look for target information for inlinable functions. */
bool scan_for_target_info =
info->inlinable
&& targetm.target_option.need_ipa_fn_target_info (node->decl,
info->target_info);
if (fbi.info)
compute_bb_predicates (&fbi, node, info, params_summary);
const profile_count entry_count = ENTRY_BLOCK_PTR_FOR_FN (cfun)->count;
order = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
nblocks = pre_and_rev_post_order_compute (NULL, order, false);
for (n = 0; n < nblocks; n++)
{
bb = BASIC_BLOCK_FOR_FN (cfun, order[n]);
freq = bb->count.to_sreal_scale (entry_count);
if (clobber_only_eh_bb_p (bb))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "\n Ignoring BB %i;"
" it will be optimized away by cleanup_clobbers\n",
bb->index);
continue;
}
/* TODO: Obviously predicates can be propagated down across CFG. */
if (fbi.info)
{
if (bb->aux)
bb_predicate = *(ipa_predicate *)bb->aux;
else
bb_predicate = false;
}
else
bb_predicate = true;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "\n BB %i predicate:", bb->index);
bb_predicate.dump (dump_file, info->conds);
}
if (fbi.info && nonconstant_names.exists ())
{
ipa_predicate phi_predicate;
bool first_phi = true;
for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);
gsi_next (&bsi))
{
if (first_phi
&& !phi_result_unknown_predicate (&fbi, info,
params_summary,
bb,
&phi_predicate,
nonconstant_names))
break;
first_phi = false;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " ");
print_gimple_stmt (dump_file, gsi_stmt (bsi), 0);
}
predicate_for_phi_result (info, bsi.phi (), &phi_predicate,
nonconstant_names);
}
}
fix_builtin_expect_stmt = find_foldable_builtin_expect (bb);
for (gimple_stmt_iterator bsi = gsi_start_nondebug_bb (bb);
!gsi_end_p (bsi); gsi_next_nondebug (&bsi))
{
gimple *stmt = gsi_stmt (bsi);
int this_size = estimate_num_insns (stmt, &eni_size_weights);
int this_time = estimate_num_insns (stmt, &eni_time_weights);
int prob;
ipa_predicate will_be_nonconstant;
/* This relation stmt should be folded after we remove
__builtin_expect call. Adjust the cost here. */
if (stmt == fix_builtin_expect_stmt)
{
this_size--;
this_time--;
}
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " ");
print_gimple_stmt (dump_file, stmt, 0);
fprintf (dump_file, "\t\tfreq:%3.2f size:%3i time:%3i\n",
freq.to_double (), this_size,
this_time);
}
if (is_gimple_call (stmt)
&& !gimple_call_internal_p (stmt))
{
struct cgraph_edge *edge = node->get_edge (stmt);
ipa_call_summary *es = ipa_call_summaries->get_create (edge);
/* Special case: results of BUILT_IN_CONSTANT_P will be always
resolved as constant. We however don't want to optimize
out the cgraph edges. */
if (nonconstant_names.exists ()
&& gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
&& gimple_call_lhs (stmt)
&& TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
{
ipa_predicate false_p = false;
nonconstant_names[SSA_NAME_VERSION (gimple_call_lhs (stmt))]
= false_p;
}
if (ipa_node_params_sum)
{
int count = gimple_call_num_args (stmt);
int i;
if (count)
es->param.safe_grow_cleared (count, true);
for (i = 0; i < count; i++)
{
int prob = param_change_prob (&fbi, stmt, i);
gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
es->param[i].change_prob = prob;
es->param[i].points_to_local_or_readonly_memory
= points_to_local_or_readonly_memory_p
(gimple_call_arg (stmt, i));
}
}
/* We cannot setup VLA parameters during inlining. */
for (unsigned int i = 0; i < gimple_call_num_args (stmt); ++i)
if (TREE_CODE (gimple_call_arg (stmt, i)) == WITH_SIZE_EXPR)
{
edge->inline_failed = CIF_FUNCTION_NOT_INLINABLE;
break;
}
es->call_stmt_size = this_size;
es->call_stmt_time = this_time;
es->loop_depth = bb_loop_depth (bb);
edge_set_predicate (edge, &bb_predicate);
if (edge->speculative)
{
cgraph_edge *indirect
= edge->speculative_call_indirect_edge ();
ipa_call_summary *es2
= ipa_call_summaries->get_create (indirect);
ipa_call_summaries->duplicate (edge, indirect,
es, es2);
/* Edge is the first direct call.
create and duplicate call summaries for multiple
speculative call targets. */
for (cgraph_edge *direct
= edge->next_speculative_call_target ();
direct;
direct = direct->next_speculative_call_target ())
{
ipa_call_summary *es3
= ipa_call_summaries->get_create (direct);
ipa_call_summaries->duplicate (edge, direct,
es, es3);
}
}
}
/* TODO: When conditional jump or switch is known to be constant, but
we did not translate it into the predicates, we really can account
just maximum of the possible paths. */
if (fbi.info)
will_be_nonconstant
= will_be_nonconstant_predicate (&fbi, info, params_summary,
stmt, nonconstant_names);
else
will_be_nonconstant = true;
if (this_time || this_size)
{
sreal final_time = (sreal)this_time * freq;
prob = eliminated_by_inlining_prob (&fbi, stmt);
if (prob == 1 && dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file,
"\t\t50%% will be eliminated by inlining\n");
if (prob == 2 && dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
ipa_predicate p = bb_predicate & will_be_nonconstant;
/* We can ignore statement when we proved it is never going
to happen, but we cannot do that for call statements
because edges are accounted specially. */
if (*(is_gimple_call (stmt) ? &bb_predicate : &p) != false)
{
time += final_time;
size += this_size;
}
/* We account everything but the calls. Calls have their own
size/time info attached to cgraph edges. This is necessary
in order to make the cost disappear after inlining. */
if (!is_gimple_call (stmt))
{
if (prob)
{
ipa_predicate ip
= bb_predicate & ipa_predicate::not_inlined ();
info->account_size_time (this_size * prob,
(final_time * prob) / 2, ip,
p);
}
if (prob != 2)
info->account_size_time (this_size * (2 - prob),
(final_time * (2 - prob) / 2),
bb_predicate,
p);
}
if (!info->fp_expressions && fp_expression_p (stmt))
{
info->fp_expressions = true;
if (dump_file)
fprintf (dump_file, " fp_expression set\n");
}
}
/* For target specific information, we want to scan all statements
rather than those statements with non-zero weights, to avoid
missing to scan something interesting for target information,
such as: internal function calls. */
if (scan_for_target_info)
scan_for_target_info =
targetm.target_option.update_ipa_fn_target_info
(info->target_info, stmt);
/* Account cost of address calculations in the statements. */
for (unsigned int i = 0; i < gimple_num_ops (stmt); i++)
{
for (tree op = gimple_op (stmt, i);
op && handled_component_p (op);
op = TREE_OPERAND (op, 0))
if ((TREE_CODE (op) == ARRAY_REF
|| TREE_CODE (op) == ARRAY_RANGE_REF)
&& TREE_CODE (TREE_OPERAND (op, 1)) == SSA_NAME)
{
ipa_predicate p = bb_predicate;
if (fbi.info)
p = p & will_be_nonconstant_expr_predicate
(&fbi, info, params_summary,
TREE_OPERAND (op, 1),
nonconstant_names);
if (p != false)
{
time += freq;
size += 1;
if (dump_file)
fprintf (dump_file,
"\t\tAccounting address calculation.\n");
info->account_size_time (ipa_fn_summary::size_scale,
freq,
bb_predicate,
p);
}
}
}
}
}
free (order);
if (nonconstant_names.exists () && !early)
{
ipa_fn_summary *s = ipa_fn_summaries->get (node);
unsigned max_loop_predicates = opt_for_fn (node->decl,
param_ipa_max_loop_predicates);
if (dump_file && (dump_flags & TDF_DETAILS))
flow_loops_dump (dump_file, NULL, 0);
scev_initialize ();
for (auto loop : loops_list (cfun, 0))
{
ipa_predicate loop_iterations = true;
sreal header_freq;
edge ex;
unsigned int j;
class tree_niter_desc niter_desc;
if (!loop->header->aux)
continue;
profile_count phdr_count = loop_preheader_edge (loop)->count ();
sreal phdr_freq = phdr_count.to_sreal_scale (entry_count);
bb_predicate = *(ipa_predicate *)loop->header->aux;
auto_vec<edge> exits = get_loop_exit_edges (loop);
FOR_EACH_VEC_ELT (exits, j, ex)
if (number_of_iterations_exit (loop, ex, &niter_desc, false)
&& !is_gimple_min_invariant (niter_desc.niter))
{
ipa_predicate will_be_nonconstant
= will_be_nonconstant_expr_predicate (&fbi, info,
params_summary,
niter_desc.niter,
nonconstant_names);
if (will_be_nonconstant != true)
will_be_nonconstant = bb_predicate & will_be_nonconstant;
if (will_be_nonconstant != true
&& will_be_nonconstant != false)
loop_iterations &= will_be_nonconstant;
}
add_freqcounting_predicate (&s->loop_iterations, loop_iterations,
phdr_freq, max_loop_predicates);
}
/* To avoid quadratic behavior we analyze stride predicates only
with respect to the containing loop. Thus we simply iterate
over all defs in the outermost loop body. */
for (class loop *loop = loops_for_fn (cfun)->tree_root->inner;
loop != NULL; loop = loop->next)
{
ipa_predicate loop_stride = true;
basic_block *body = get_loop_body (loop);
profile_count phdr_count = loop_preheader_edge (loop)->count ();
sreal phdr_freq = phdr_count.to_sreal_scale (entry_count);
for (unsigned i = 0; i < loop->num_nodes; i++)
{
gimple_stmt_iterator gsi;
if (!body[i]->aux)
continue;
bb_predicate = *(ipa_predicate *)body[i]->aux;
for (gsi = gsi_start_bb (body[i]); !gsi_end_p (gsi);
gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
if (!is_gimple_assign (stmt))
continue;
tree def = gimple_assign_lhs (stmt);
if (TREE_CODE (def) != SSA_NAME)
continue;
affine_iv iv;
if (!simple_iv (loop_containing_stmt (stmt),
loop_containing_stmt (stmt),
def, &iv, true)
|| is_gimple_min_invariant (iv.step))
continue;
ipa_predicate will_be_nonconstant
= will_be_nonconstant_expr_predicate (&fbi, info,
params_summary,
iv.step,
nonconstant_names);
if (will_be_nonconstant != true)
will_be_nonconstant = bb_predicate & will_be_nonconstant;
if (will_be_nonconstant != true
&& will_be_nonconstant != false)
loop_stride = loop_stride & will_be_nonconstant;
}
}
add_freqcounting_predicate (&s->loop_strides, loop_stride,
phdr_freq, max_loop_predicates);