#include "tree-flow.h"
#include "ipa-prop.h"
#include "lto-streamer.h"
+#include "data-streamer.h"
+#include "tree-streamer.h"
#include "ipa-inline.h"
+#include "alloc-pool.h"
/* Estimate runtime of function can easilly run into huge numbers with many
- nested loops. Be sure we can compute time * INLINE_SIZE_SCALE in integer.
- For anything larger we use gcov_type. */
-#define MAX_TIME 1000000
+ nested loops. Be sure we can compute time * INLINE_SIZE_SCALE * 2 in an
+ integer. For anything larger we use gcov_type. */
+#define MAX_TIME 500000
/* Number of bits in integer, but we really want to be stable across different
hosts. */
/* Special condition code we use to represent test that operand is compile time
constant. */
#define IS_NOT_CONSTANT ERROR_MARK
+/* Special condition code we use to represent test that operand is not changed
+ across invocation of the function. When operand IS_NOT_CONSTANT it is always
+ CHANGED, however i.e. loop invariants can be NOT_CHANGED given percentage
+ of executions even when they are not compile time constants. */
+#define CHANGED IDENTIFIER_NODE
/* Holders of ipa cgraph hooks: */
static struct cgraph_node_hook_list *function_insertion_hook_holder;
static struct cgraph_node_hook_list *node_removal_hook_holder;
static struct cgraph_2node_hook_list *node_duplication_hook_holder;
+static struct cgraph_2edge_hook_list *edge_duplication_hook_holder;
static struct cgraph_edge_hook_list *edge_removal_hook_holder;
static void inline_node_removal_hook (struct cgraph_node *, void *);
static void inline_node_duplication_hook (struct cgraph_node *,
struct cgraph_node *, void *);
+static void inline_edge_removal_hook (struct cgraph_edge *, void *);
+static void inline_edge_duplication_hook (struct cgraph_edge *,
+ struct cgraph_edge *,
+ void *);
/* VECtor holding inline summaries.
In GGC memory because conditions might point to constant trees. */
VEC(inline_summary_t,gc) *inline_summary_vec;
+VEC(inline_edge_summary_t,heap) *inline_edge_summary_vec;
/* Cached node/edge growths. */
VEC(int,heap) *node_growth_cache;
VEC(edge_growth_cache_entry,heap) *edge_growth_cache;
+/* Edge predicates goes here. */
+static alloc_pool edge_predicate_pool;
/* Return true predicate (tautology).
We represent it by empty list of clauses. */
}
+/* Return true if P is (false). */
+
+static inline bool
+true_predicate_p (struct predicate *p)
+{
+ return !p->clause[0];
+}
+
+
+/* Return true if P is (false). */
+
+static inline bool
+false_predicate_p (struct predicate *p)
+{
+ if (p->clause[0] == (1 << predicate_false_condition))
+ {
+ gcc_checking_assert (!p->clause[1]
+ && p->clause[0] == 1 << predicate_false_condition);
+ return true;
+ }
+ return false;
+}
+
+
/* Return predicate that is set true when function is not inlined. */
static inline struct predicate
not_inlined_predicate (void)
}
-/* Add clause CLAUSE into the predicate. */
+/* Add clause CLAUSE into the predicate P. */
static inline void
-add_clause (struct predicate *p, clause_t clause)
+add_clause (conditions conditions, struct predicate *p, clause_t clause)
{
int i;
- int insert_here = 0;
+ int i2;
+ int insert_here = -1;
+ int c1, c2;
+
/* True clause. */
if (!clause)
return;
- /* Flase clause makes the whole predicate false. Kill the other variants. */
- if (clause & (1 << predicate_false_condition))
+ /* False clause makes the whole predicate false. Kill the other variants. */
+ if (clause == (1 << predicate_false_condition))
{
p->clause[0] = (1 << predicate_false_condition);
p->clause[1] = 0;
+ return;
}
- for (i = 0; i < MAX_CLAUSES; i++)
+ if (false_predicate_p (p))
+ return;
+
+ /* No one should be sily enough to add false into nontrivial clauses. */
+ gcc_checking_assert (!(clause & (1 << predicate_false_condition)));
+
+ /* Look where to insert the clause. At the same time prune out
+ clauses of P that are implied by the new clause and thus
+ redundant. */
+ for (i = 0, i2 = 0; i <= MAX_CLAUSES; i++)
{
- if (p->clause[i] == clause)
- return;
+ p->clause[i2] = p->clause[i];
+
if (!p->clause[i])
break;
- if (p->clause[i] < clause && !insert_here)
- insert_here = i;
+
+ /* If p->clause[i] implies clause, there is nothing to add. */
+ if ((p->clause[i] & clause) == p->clause[i])
+ {
+ /* We had nothing to add, none of clauses should've become
+ redundant. */
+ gcc_checking_assert (i == i2);
+ return;
+ }
+
+ if (p->clause[i] < clause && insert_here < 0)
+ insert_here = i2;
+
+ /* If clause implies p->clause[i], then p->clause[i] becomes redundant.
+ Otherwise the p->clause[i] has to stay. */
+ if ((p->clause[i] & clause) != clause)
+ i2++;
}
- /* We run out of variants. Be conservative in positive direciton. */
- if (i == MAX_CLAUSES)
+
+ /* Look for clauses that are obviously true. I.e.
+ op0 == 5 || op0 != 5. */
+ for (c1 = predicate_first_dynamic_condition; c1 < NUM_CONDITIONS; c1++)
+ {
+ condition *cc1;
+ if (!(clause & (1 << c1)))
+ continue;
+ cc1 = VEC_index (condition,
+ conditions,
+ c1 - predicate_first_dynamic_condition);
+ /* We have no way to represent !CHANGED and !IS_NOT_CONSTANT
+ and thus there is no point for looking for them. */
+ if (cc1->code == CHANGED
+ || cc1->code == IS_NOT_CONSTANT)
+ continue;
+ for (c2 = c1 + 1; c2 <= NUM_CONDITIONS; c2++)
+ if (clause & (1 << c2))
+ {
+ condition *cc1 = VEC_index (condition,
+ conditions,
+ c1 - predicate_first_dynamic_condition);
+ condition *cc2 = VEC_index (condition,
+ conditions,
+ c2 - predicate_first_dynamic_condition);
+ if (cc1->operand_num == cc2->operand_num
+ && cc1->val == cc2->val
+ && cc2->code != IS_NOT_CONSTANT
+ && cc2->code != CHANGED
+ && cc1->code == invert_tree_comparison
+ (cc2->code,
+ HONOR_NANS (TYPE_MODE (TREE_TYPE (cc1->val)))))
+ return;
+ }
+ }
+
+
+ /* We run out of variants. Be conservative in positive direction. */
+ if (i2 == MAX_CLAUSES)
return;
- /* Keep clauses ordered by index, so equivalence testing is easy. */
- p->clause[i + 1] = 0;
- for (;i > insert_here; i--)
- p->clause[i] = p->clause[i - 1];
+ /* Keep clauses in decreasing order. This makes equivalence testing easy. */
+ p->clause[i2 + 1] = 0;
+ if (insert_here >= 0)
+ for (;i2 > insert_here; i2--)
+ p->clause[i2] = p->clause[i2 - 1];
+ else
+ insert_here = i2;
p->clause[insert_here] = clause;
}
/* Return P & P2. */
static struct predicate
-and_predicates (struct predicate *p, struct predicate *p2)
+and_predicates (conditions conditions,
+ struct predicate *p, struct predicate *p2)
{
struct predicate out = *p;
int i;
- for (i = 0; p2->clause[i]; i++)
- add_clause (&out, p2->clause[i]);
- return out;
-}
+ /* Avoid busy work. */
+ if (false_predicate_p (p2) || true_predicate_p (p))
+ return *p2;
+ if (false_predicate_p (p) || true_predicate_p (p2))
+ return *p;
-/* Return P | P2. */
-
-static struct predicate
-or_predicates (struct predicate *p, struct predicate *p2)
-{
- struct predicate out = true_predicate ();
- int i,j;
- /* If one of conditions is false, return the other. */
- if (p2->clause[0] == 1 << predicate_false_condition)
+ /* See how far predicates match. */
+ for (i = 0; p->clause[i] && p->clause[i] == p2->clause[i]; i++)
{
- gcc_checking_assert (!p2->clause[1]);
- return *p;
+ gcc_checking_assert (i < MAX_CLAUSES);
}
- if (p->clause[0] == 1 << predicate_false_condition)
+
+ /* Combine the predicates rest. */
+ for (; p2->clause[i]; i++)
{
- gcc_checking_assert (!p->clause[1]);
- return *p2;
+ gcc_checking_assert (i < MAX_CLAUSES);
+ add_clause (conditions, &out, p2->clause[i]);
}
- for (i = 0; p->clause[i]; i++)
- for (j = 0; p2->clause[j]; j++)
- add_clause (&out, p->clause[i] | p2->clause[j]);
return out;
}
{
int i;
for (i = 0; p->clause[i]; i++)
- if (p->clause[i] != p2->clause[i])
- return false;
+ {
+ gcc_checking_assert (i < MAX_CLAUSES);
+ gcc_checking_assert (p->clause [i] > p->clause[i + 1]);
+ gcc_checking_assert (!p2->clause[i]
+ || p2->clause [i] > p2->clause[i + 1]);
+ if (p->clause[i] != p2->clause[i])
+ return false;
+ }
return !p2->clause[i];
}
-/* Having partial truth assignment in POSSIBLE_TRUTHS, return false if predicate P
- to be false. */
+/* Return P | P2. */
+
+static struct predicate
+or_predicates (conditions conditions, struct predicate *p, struct predicate *p2)
+{
+ struct predicate out = true_predicate ();
+ int i,j;
+
+ /* Avoid busy work. */
+ if (false_predicate_p (p2) || true_predicate_p (p))
+ return *p;
+ if (false_predicate_p (p) || true_predicate_p (p2))
+ return *p2;
+ if (predicates_equal_p (p, p2))
+ return *p;
+
+ /* OK, combine the predicates. */
+ for (i = 0; p->clause[i]; i++)
+ for (j = 0; p2->clause[j]; j++)
+ {
+ gcc_checking_assert (i < MAX_CLAUSES && j < MAX_CLAUSES);
+ add_clause (conditions, &out, p->clause[i] | p2->clause[j]);
+ }
+ return out;
+}
+
+
+/* Having partial truth assignment in POSSIBLE_TRUTHS, return false
+ if predicate P is known to be false. */
static bool
-evaulate_predicate (struct predicate *p, clause_t possible_truths)
+evaluate_predicate (struct predicate *p, clause_t possible_truths)
{
int i;
/* True remains true. */
- if (!p->clause[0])
+ if (true_predicate_p (p))
return true;
+ gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
+
/* See if we can find clause we can disprove. */
for (i = 0; p->clause[i]; i++)
- if (!(p->clause[i] & possible_truths))
- return false;
+ {
+ gcc_checking_assert (i < MAX_CLAUSES);
+ if (!(p->clause[i] & possible_truths))
+ return false;
+ }
return true;
}
+/* Return the probability in range 0...REG_BR_PROB_BASE that the predicated
+ instruction will be recomputed per invocation of the inlined call. */
+
+static int
+predicate_probability (conditions conds,
+ struct predicate *p, clause_t possible_truths,
+ VEC (inline_param_summary_t, heap) *inline_param_summary)
+{
+ int i;
+ int combined_prob = REG_BR_PROB_BASE;
+
+ /* True remains true. */
+ if (true_predicate_p (p))
+ return REG_BR_PROB_BASE;
+
+ if (false_predicate_p (p))
+ return 0;
+
+ gcc_assert (!(possible_truths & (1 << predicate_false_condition)));
+
+ /* See if we can find clause we can disprove. */
+ for (i = 0; p->clause[i]; i++)
+ {
+ gcc_checking_assert (i < MAX_CLAUSES);
+ if (!(p->clause[i] & possible_truths))
+ return 0;
+ else
+ {
+ int this_prob = 0;
+ int i2;
+ if (!inline_param_summary)
+ return REG_BR_PROB_BASE;
+ for (i2 = 0; i2 < NUM_CONDITIONS; i2++)
+ if ((p->clause[i] & possible_truths) & (1 << i2))
+ {
+ if (i2 >= predicate_first_dynamic_condition)
+ {
+ condition *c = VEC_index
+ (condition, conds,
+ i2 - predicate_first_dynamic_condition);
+ if (c->code == CHANGED
+ && (c->operand_num
+ < (int) VEC_length (inline_param_summary_t,
+ inline_param_summary)))
+ {
+ int iprob = VEC_index (inline_param_summary_t,
+ inline_param_summary,
+ c->operand_num)->change_prob;
+ this_prob = MAX (this_prob, iprob);
+ }
+ else
+ this_prob = REG_BR_PROB_BASE;
+ }
+ else
+ this_prob = REG_BR_PROB_BASE;
+ }
+ combined_prob = MIN (this_prob, combined_prob);
+ if (!combined_prob)
+ return 0;
+ }
+ }
+ return combined_prob;
+}
+
/* Dump conditional COND. */
fprintf (f, "not inlined");
else
{
- c = VEC_index (condition, conditions, cond - predicate_first_dynamic_condition);
+ c = VEC_index (condition, conditions,
+ cond - predicate_first_dynamic_condition);
fprintf (f, "op%i", c->operand_num);
if (c->code == IS_NOT_CONSTANT)
{
fprintf (f, " not constant");
return;
}
+ if (c->code == CHANGED)
+ {
+ fprintf (f, " changed");
+ return;
+ }
fprintf (f, " %s ", op_symbol_code (c->code));
print_generic_expr (f, c->val, 1);
}
dump_predicate (FILE *f, conditions conds, struct predicate *pred)
{
int i;
- if (!pred->clause[0])
+ if (true_predicate_p (pred))
dump_clause (f, conds, 0);
else
for (i = 0; pred->clause[i]; i++)
/* Record SIZE and TIME under condition PRED into the inline summary. */
static void
-account_size_time (struct inline_summary *summary, int size, int time, struct predicate *pred)
+account_size_time (struct inline_summary *summary, int size, int time,
+ struct predicate *pred)
{
size_time_entry *e;
bool found = false;
int i;
- if (pred->clause[0] == (1 << predicate_false_condition))
+ if (false_predicate_p (pred))
return;
/* We need to create initial empty unconitional clause, but otherwie
we don't need to account empty times and sizes. */
- if (!size && !time && summary->conds)
+ if (!size && !time && summary->entry)
return;
/* Watch overflow that might result from insane profiles. */
if (dump_file && (dump_flags & TDF_DETAILS) && (time || size))
{
fprintf (dump_file, "\t\tAccounting size:%3.2f, time:%3.2f on %spredicate:",
- ((double)size) / INLINE_SIZE_SCALE, ((double)time) / INLINE_TIME_SCALE,
+ ((double)size) / INLINE_SIZE_SCALE,
+ ((double)time) / INLINE_TIME_SCALE,
found ? "" : "new ");
dump_predicate (dump_file, summary->conds, pred);
}
}
}
+/* Set predicate for edge E. */
+
+static void
+edge_set_predicate (struct cgraph_edge *e, struct predicate *predicate)
+{
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ if (predicate && !true_predicate_p (predicate))
+ {
+ if (!es->predicate)
+ es->predicate = (struct predicate *)pool_alloc (edge_predicate_pool);
+ *es->predicate = *predicate;
+ }
+ else
+ {
+ if (es->predicate)
+ pool_free (edge_predicate_pool, es->predicate);
+ es->predicate = NULL;
+ }
+}
-/* Work out what conditions might be true at invocation of E. */
+
+/* KNOWN_VALS is partial mapping of parameters of NODE to constant values.
+ Return clause of possible truths. When INLINE_P is true, assume that
+ we are inlining.
+
+ ERROR_MARK means compile time invariant. */
static clause_t
-evaulate_conditions_for_edge (struct cgraph_edge *e, bool inline_p)
+evaluate_conditions_for_known_args (struct cgraph_node *node,
+ bool inline_p,
+ VEC (tree, heap) *known_vals)
{
clause_t clause = inline_p ? 0 : 1 << predicate_not_inlined_condition;
- struct inline_summary *info = inline_summary (e->callee);
+ struct inline_summary *info = inline_summary (node);
int i;
+ struct condition *c;
+
+ for (i = 0; VEC_iterate (condition, info->conds, i, c); i++)
+ {
+ tree val;
+ tree res;
+
+ /* We allow call stmt to have fewer arguments than the callee
+ function (especially for K&R style programs). So bound
+ check here. */
+ if (c->operand_num < (int)VEC_length (tree, known_vals))
+ val = VEC_index (tree, known_vals, c->operand_num);
+ else
+ val = NULL;
+
+ if (val == error_mark_node && c->code != CHANGED)
+ val = NULL;
- if (ipa_node_params_vector && info->conds
- /* FIXME: it seems that we forget to get argument count in some cases,
- probaby for previously indirect edges or so. */
- && ipa_get_cs_argument_count (IPA_EDGE_REF (e)))
+ if (!val)
+ {
+ clause |= 1 << (i + predicate_first_dynamic_condition);
+ continue;
+ }
+ if (c->code == IS_NOT_CONSTANT || c->code == CHANGED)
+ continue;
+ res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val);
+ if (res
+ && integer_zerop (res))
+ continue;
+ clause |= 1 << (i + predicate_first_dynamic_condition);
+ }
+ return clause;
+}
+
+
+/* Work out what conditions might be true at invocation of E. */
+
+static void
+evaluate_properties_for_edge (struct cgraph_edge *e, bool inline_p,
+ clause_t *clause_ptr,
+ VEC (tree, heap) **known_vals_ptr,
+ VEC (tree, heap) **known_binfos_ptr)
+{
+ struct cgraph_node *callee = cgraph_function_or_thunk_node (e->callee, NULL);
+ struct inline_summary *info = inline_summary (callee);
+ VEC (tree, heap) *known_vals = NULL;
+
+ if (clause_ptr)
+ *clause_ptr = inline_p ? 0 : 1 << predicate_not_inlined_condition;
+ if (known_vals_ptr)
+ *known_vals_ptr = NULL;
+ if (known_binfos_ptr)
+ *known_binfos_ptr = NULL;
+
+ if (ipa_node_params_vector
+ && !e->call_stmt_cannot_inline_p
+ && ((clause_ptr && info->conds) || known_vals_ptr || known_binfos_ptr))
{
struct ipa_node_params *parms_info;
struct ipa_edge_args *args = IPA_EDGE_REF (e);
+ struct inline_edge_summary *es = inline_edge_summary (e);
int i, count = ipa_get_cs_argument_count (args);
- struct ipcp_lattice lat;
- struct condition *c;
- VEC (tree, heap) *known_vals = NULL;
if (e->caller->global.inlined_to)
parms_info = IPA_NODE_REF (e->caller->global.inlined_to);
else
parms_info = IPA_NODE_REF (e->caller);
- VEC_safe_grow_cleared (tree, heap, known_vals, count);
+ if (count && (info->conds || known_vals_ptr))
+ VEC_safe_grow_cleared (tree, heap, known_vals, count);
+ if (count && known_binfos_ptr)
+ VEC_safe_grow_cleared (tree, heap, *known_binfos_ptr, count);
+
for (i = 0; i < count; i++)
{
- ipa_lattice_from_jfunc (parms_info, &lat, ipa_get_ith_jump_func (args, i));
- if (lat.type == IPA_CONST_VALUE)
- VEC_replace (tree, known_vals, i, lat.constant);
- }
- for (i = 0; VEC_iterate (condition, info->conds, i, c); i++)
- {
- tree val = VEC_index (tree, known_vals, c->operand_num);
- tree res;
-
- if (!val)
+ tree cst = ipa_value_from_jfunc (parms_info,
+ ipa_get_ith_jump_func (args, i));
+ if (cst)
{
- clause |= 1 << (i + predicate_first_dynamic_condition);
- continue;
+ if (known_vals && TREE_CODE (cst) != TREE_BINFO)
+ VEC_replace (tree, known_vals, i, cst);
+ else if (known_binfos_ptr != NULL && TREE_CODE (cst) == TREE_BINFO)
+ VEC_replace (tree, *known_binfos_ptr, i, cst);
}
- if (c->code == IS_NOT_CONSTANT)
- continue;
- res = fold_binary_to_constant (c->code, boolean_type_node, val, c->val);
- if (res
- && integer_zerop (res))
- continue;
- clause |= 1 << (i + predicate_first_dynamic_condition);
+ else if (inline_p
+ && !VEC_index (inline_param_summary_t,
+ es->param,
+ i)->change_prob)
+ VEC_replace (tree, known_vals, i, error_mark_node);
}
- VEC_free (tree, heap, known_vals);
}
- else
- for (i = 0; i < (int)VEC_length (condition, info->conds); i++)
- clause |= 1 << (i + predicate_first_dynamic_condition);
- return clause;
+ if (clause_ptr)
+ *clause_ptr = evaluate_conditions_for_known_args (callee, inline_p,
+ known_vals);
+
+ if (known_vals_ptr)
+ *known_vals_ptr = known_vals;
+ else
+ VEC_free (tree, heap, known_vals);
}
if (!node_removal_hook_holder)
node_removal_hook_holder =
cgraph_add_node_removal_hook (&inline_node_removal_hook, NULL);
+ if (!edge_removal_hook_holder)
+ edge_removal_hook_holder =
+ cgraph_add_edge_removal_hook (&inline_edge_removal_hook, NULL);
if (!node_duplication_hook_holder)
node_duplication_hook_holder =
cgraph_add_node_duplication_hook (&inline_node_duplication_hook, NULL);
+ if (!edge_duplication_hook_holder)
+ edge_duplication_hook_holder =
+ cgraph_add_edge_duplication_hook (&inline_edge_duplication_hook, NULL);
if (VEC_length (inline_summary_t, inline_summary_vec)
<= (unsigned) cgraph_max_uid)
VEC_safe_grow_cleared (inline_summary_t, gc,
inline_summary_vec, cgraph_max_uid + 1);
+ if (VEC_length (inline_edge_summary_t, inline_edge_summary_vec)
+ <= (unsigned) cgraph_edge_max_uid)
+ VEC_safe_grow_cleared (inline_edge_summary_t, heap,
+ inline_edge_summary_vec, cgraph_edge_max_uid + 1);
+ if (!edge_predicate_pool)
+ edge_predicate_pool = create_alloc_pool ("edge predicates",
+ sizeof (struct predicate),
+ 10);
+}
+
+/* We are called multiple time for given function; clear
+ data from previous run so they are not cumulated. */
+
+static void
+reset_inline_edge_summary (struct cgraph_edge *e)
+{
+ if (e->uid
+ < (int)VEC_length (inline_edge_summary_t, inline_edge_summary_vec))
+ {
+ struct inline_edge_summary *es = inline_edge_summary (e);
+
+ es->call_stmt_size = es->call_stmt_time =0;
+ if (es->predicate)
+ pool_free (edge_predicate_pool, es->predicate);
+ es->predicate = NULL;
+ VEC_free (inline_param_summary_t, heap, es->param);
+ }
+}
+
+/* We are called multiple time for given function; clear
+ data from previous run so they are not cumulated. */
+
+static void
+reset_inline_summary (struct cgraph_node *node)
+{
+ struct inline_summary *info = inline_summary (node);
+ struct cgraph_edge *e;
+
+ info->self_size = info->self_time = 0;
+ info->estimated_stack_size = 0;
+ info->estimated_self_stack_size = 0;
+ info->stack_frame_offset = 0;
+ info->size = 0;
+ info->time = 0;
+ VEC_free (condition, gc, info->conds);
+ VEC_free (size_time_entry,gc, info->entry);
+ for (e = node->callees; e; e = e->next_callee)
+ reset_inline_edge_summary (e);
+ for (e = node->indirect_calls; e; e = e->next_callee)
+ reset_inline_edge_summary (e);
}
/* Hook that is called by cgraph.c when a node is removed. */
<= (unsigned)node->uid)
return;
info = inline_summary (node);
- reset_node_growth_cache (node);
- VEC_free (condition, gc, info->conds);
- VEC_free (size_time_entry, gc, info->entry);
- info->conds = NULL;
- info->entry = NULL;
+ reset_inline_summary (node);
memset (info, 0, sizeof (inline_summary_t));
}
+
/* Hook that is called by cgraph.c when a node is duplicated. */
static void
info = inline_summary (dst);
memcpy (info, inline_summary (src),
sizeof (struct inline_summary));
+ /* TODO: as an optimization, we may avoid copying conditions
+ that are known to be false or true. */
info->conds = VEC_copy (condition, gc, info->conds);
- info->entry = VEC_copy (size_time_entry, gc, info->entry);
+
+ /* When there are any replacements in the function body, see if we can figure
+ out that something was optimized out. */
+ if (ipa_node_params_vector && dst->clone.tree_map)
+ {
+ VEC(size_time_entry,gc) *entry = info->entry;
+ /* Use SRC parm info since it may not be copied yet. */
+ struct ipa_node_params *parms_info = IPA_NODE_REF (src);
+ VEC (tree, heap) *known_vals = NULL;
+ int count = ipa_get_param_count (parms_info);
+ int i,j;
+ clause_t possible_truths;
+ struct predicate true_pred = true_predicate ();
+ size_time_entry *e;
+ int optimized_out_size = 0;
+ gcov_type optimized_out_time = 0;
+ bool inlined_to_p = false;
+ struct cgraph_edge *edge;
+
+ info->entry = 0;
+ VEC_safe_grow_cleared (tree, heap, known_vals, count);
+ for (i = 0; i < count; i++)
+ {
+ tree t = ipa_get_param (parms_info, i);
+ struct ipa_replace_map *r;
+
+ for (j = 0;
+ VEC_iterate (ipa_replace_map_p, dst->clone.tree_map, j, r);
+ j++)
+ {
+ if (r->old_tree == t
+ && r->replace_p
+ && !r->ref_p)
+ {
+ VEC_replace (tree, known_vals, i, r->new_tree);
+ break;
+ }
+ }
+ }
+ possible_truths = evaluate_conditions_for_known_args (dst,
+ false, known_vals);
+ VEC_free (tree, heap, known_vals);
+
+ account_size_time (info, 0, 0, &true_pred);
+
+ /* Remap size_time vectors.
+ Simplify the predicate by prunning out alternatives that are known
+ to be false.
+ TODO: as on optimization, we can also eliminate conditions known
+ to be true. */
+ for (i = 0; VEC_iterate (size_time_entry, entry, i, e); i++)
+ {
+ struct predicate new_predicate = true_predicate ();
+ for (j = 0; e->predicate.clause[j]; j++)
+ if (!(possible_truths & e->predicate.clause[j]))
+ {
+ new_predicate = false_predicate ();
+ break;
+ }
+ else
+ add_clause (info->conds, &new_predicate,
+ possible_truths & e->predicate.clause[j]);
+ if (false_predicate_p (&new_predicate))
+ {
+ optimized_out_size += e->size;
+ optimized_out_time += e->time;
+ }
+ else
+ account_size_time (info, e->size, e->time, &new_predicate);
+ }
+
+ /* Remap edge predicates with the same simplification as above.
+ Also copy constantness arrays. */
+ for (edge = dst->callees; edge; edge = edge->next_callee)
+ {
+ struct predicate new_predicate = true_predicate ();
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+
+ if (!edge->inline_failed)
+ inlined_to_p = true;
+ if (!es->predicate)
+ continue;
+ for (j = 0; es->predicate->clause[j]; j++)
+ if (!(possible_truths & es->predicate->clause[j]))
+ {
+ new_predicate = false_predicate ();
+ break;
+ }
+ else
+ add_clause (info->conds, &new_predicate,
+ possible_truths & es->predicate->clause[j]);
+ if (false_predicate_p (&new_predicate)
+ && !false_predicate_p (es->predicate))
+ {
+ optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
+ optimized_out_time += (es->call_stmt_time
+ * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE)
+ * edge->frequency);
+ edge->frequency = 0;
+ }
+ *es->predicate = new_predicate;
+ }
+
+ /* Remap indirect edge predicates with the same simplificaiton as above.
+ Also copy constantness arrays. */
+ for (edge = dst->indirect_calls; edge; edge = edge->next_callee)
+ {
+ struct predicate new_predicate = true_predicate ();
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+
+ if (!edge->inline_failed)
+ inlined_to_p = true;
+ if (!es->predicate)
+ continue;
+ for (j = 0; es->predicate->clause[j]; j++)
+ if (!(possible_truths & es->predicate->clause[j]))
+ {
+ new_predicate = false_predicate ();
+ break;
+ }
+ else
+ add_clause (info->conds, &new_predicate,
+ possible_truths & es->predicate->clause[j]);
+ if (false_predicate_p (&new_predicate)
+ && !false_predicate_p (es->predicate))
+ {
+ optimized_out_size += es->call_stmt_size * INLINE_SIZE_SCALE;
+ optimized_out_time += (es->call_stmt_time
+ * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE)
+ * edge->frequency);
+ edge->frequency = 0;
+ }
+ *es->predicate = new_predicate;
+ }
+
+ /* 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 calees. */
+ gcc_assert (!inlined_to_p
+ || (!optimized_out_size && !optimized_out_time));
+
+ info->size -= optimized_out_size / INLINE_SIZE_SCALE;
+ info->self_size -= optimized_out_size / INLINE_SIZE_SCALE;
+ gcc_assert (info->size > 0);
+ gcc_assert (info->self_size > 0);
+
+ optimized_out_time /= INLINE_TIME_SCALE;
+ if (optimized_out_time > MAX_TIME)
+ optimized_out_time = MAX_TIME;
+ info->time -= optimized_out_time;
+ info->self_time -= optimized_out_time;
+ if (info->time < 0)
+ info->time = 0;
+ if (info->self_time < 0)
+ info->self_time = 0;
+ }
+ else
+ info->entry = VEC_copy (size_time_entry, gc, info->entry);
+}
+
+
+/* Hook that is called by cgraph.c when a node is duplicated. */
+
+static void
+inline_edge_duplication_hook (struct cgraph_edge *src, struct cgraph_edge *dst,
+ ATTRIBUTE_UNUSED void *data)
+{
+ struct inline_edge_summary *info;
+ struct inline_edge_summary *srcinfo;
+ inline_summary_alloc ();
+ info = inline_edge_summary (dst);
+ srcinfo = inline_edge_summary (src);
+ memcpy (info, srcinfo,
+ sizeof (struct inline_edge_summary));
+ info->predicate = NULL;
+ edge_set_predicate (dst, srcinfo->predicate);
+ info->param = VEC_copy (inline_param_summary_t, heap, srcinfo->param);
}
static void
inline_edge_removal_hook (struct cgraph_edge *edge, void *data ATTRIBUTE_UNUSED)
{
- reset_edge_growth_cache (edge);
+ if (edge_growth_cache)
+ reset_edge_growth_cache (edge);
+ reset_inline_edge_summary (edge);
}
void
initialize_growth_caches (void)
{
- if (!edge_removal_hook_holder)
- edge_removal_hook_holder =
- cgraph_add_edge_removal_hook (&inline_edge_removal_hook, NULL);
if (cgraph_edge_max_uid)
VEC_safe_grow_cleared (edge_growth_cache_entry, heap, edge_growth_cache,
cgraph_edge_max_uid);
void
free_growth_caches (void)
{
- if (edge_removal_hook_holder)
- cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
VEC_free (edge_growth_cache_entry, heap, edge_growth_cache);
edge_growth_cache = 0;
VEC_free (int, heap, node_growth_cache);
}
+/* Dump edge summaries associated to NODE and recursively to all clones.
+ Indent by INDENT. */
+
static void
+dump_inline_edge_summary (FILE * f, int indent, struct cgraph_node *node,
+ struct inline_summary *info)
+{
+ struct cgraph_edge *edge;
+ for (edge = node->callees; edge; edge = edge->next_callee)
+ {
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+ struct cgraph_node *callee = cgraph_function_or_thunk_node (edge->callee, NULL);
+ int i;
+
+ fprintf (f, "%*s%s/%i %s\n%*s loop depth:%2i freq:%4i size:%2i time: %2i callee size:%2i stack:%2i",
+ indent, "", cgraph_node_name (callee),
+ callee->uid,
+ !edge->inline_failed ? "inlined"
+ : cgraph_inline_failed_string (edge->inline_failed),
+ indent, "",
+ es->loop_depth,
+ edge->frequency,
+ es->call_stmt_size,
+ es->call_stmt_time,
+ (int)inline_summary (callee)->size / INLINE_SIZE_SCALE,
+ (int)inline_summary (callee)->estimated_stack_size);
+
+ if (es->predicate)
+ {
+ fprintf (f, " predicate: ");
+ dump_predicate (f, info->conds, es->predicate);
+ }
+ else
+ fprintf (f, "\n");
+ if (es->param)
+ for (i = 0; i < (int)VEC_length (inline_param_summary_t, es->param);
+ i++)
+ {
+ int prob = VEC_index (inline_param_summary_t,
+ 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 (!edge->inline_failed)
+ {
+ fprintf (f, "%*sStack frame offset %i, callee self size %i,"
+ " callee size %i\n",
+ indent+2, "",
+ (int)inline_summary (callee)->stack_frame_offset,
+ (int)inline_summary (callee)->estimated_self_stack_size,
+ (int)inline_summary (callee)->estimated_stack_size);
+ dump_inline_edge_summary (f, indent+2, callee, info);
+ }
+ }
+ for (edge = node->indirect_calls; edge; edge = edge->next_callee)
+ {
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+ fprintf (f, "%*sindirect call loop depth:%2i freq:%4i size:%2i"
+ " time: %2i",
+ indent, "",
+ es->loop_depth,
+ edge->frequency,
+ es->call_stmt_size,
+ es->call_stmt_time);
+ if (es->predicate)
+ {
+ fprintf (f, "predicate: ");
+ dump_predicate (f, info->conds, es->predicate);
+ }
+ else
+ fprintf (f, "\n");
+ }
+}
+
+
+void
dump_inline_summary (FILE * f, struct cgraph_node *node)
{
if (node->analyzed)
fprintf (f, " always_inline");
if (s->inlinable)
fprintf (f, " inlinable");
- if (s->versionable)
- fprintf (f, " versionable");
fprintf (f, "\n self time: %i\n",
s->self_time);
fprintf (f, " global time: %i\n", s->time);
(double) e->time / INLINE_TIME_SCALE);
dump_predicate (f, s->conds, &e->predicate);
}
+ fprintf (f, " calls:\n");
+ dump_inline_edge_summary (f, 4, node, s);
fprintf (f, "\n");
}
}
-void
+DEBUG_FUNCTION void
debug_inline_summary (struct cgraph_node *node)
{
dump_inline_summary (stderr, node);
struct cgraph_node *node;
for (node = cgraph_nodes; node; node = node->next)
- if (node->analyzed)
+ if (node->analyzed && !node->global.inlined_to)
dump_inline_summary (f, node);
}
e->inline_failed = CIF_BODY_NOT_AVAILABLE;
else if (callee->local.redefined_extern_inline)
e->inline_failed = CIF_REDEFINED_EXTERN_INLINE;
- else if (e->call_stmt && gimple_call_cannot_inline_p (e->call_stmt))
+ else if (e->call_stmt_cannot_inline_p)
e->inline_failed = CIF_MISMATCHED_ARGUMENTS;
else
e->inline_failed = CIF_FUNCTION_NOT_CONSIDERED;
}
+/* 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 reffers to value of function parameter, return
+ the corresponding parameter. */
+
+static tree
+unmodified_parm (gimple stmt, tree op)
+{
+ /* 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)
+ return SSA_NAME_VAR (op);
+ /* Non-SSA parm reference? */
+ if (TREE_CODE (op) == PARM_DECL)
+ {
+ bool modified = false;
+
+ ao_ref refd;
+ ao_ref_init (&refd, op);
+ walk_aliased_vdefs (&refd, gimple_vuse (stmt), mark_modified, &modified,
+ NULL);
+ if (!modified)
+ return op;
+ }
+ /* Assignment from a parameter? */
+ if (TREE_CODE (op) == SSA_NAME
+ && !SSA_NAME_IS_DEFAULT_DEF (op)
+ && gimple_assign_single_p (SSA_NAME_DEF_STMT (op)))
+ return unmodified_parm (SSA_NAME_DEF_STMT (op),
+ gimple_assign_rhs1 (SSA_NAME_DEF_STMT (op)));
+ return NULL;
+}
+
/* See if statement might disappear after inlining.
0 - means not eliminated
1 - half of statements goes away
eliminated_by_inlining_prob (gimple stmt)
{
enum gimple_code code = gimple_code (stmt);
+
+ if (!optimize)
+ return 0;
+
switch (code)
{
case GIMPLE_RETURN:
{
tree rhs = gimple_assign_rhs1 (stmt);
tree lhs = gimple_assign_lhs (stmt);
- tree inner_rhs = rhs;
- tree inner_lhs = lhs;
+ tree inner_rhs = get_base_address (rhs);
+ tree inner_lhs = get_base_address (lhs);
bool rhs_free = false;
bool lhs_free = false;
- while (handled_component_p (inner_lhs)
- || TREE_CODE (inner_lhs) == MEM_REF)
- inner_lhs = TREE_OPERAND (inner_lhs, 0);
- while (handled_component_p (inner_rhs)
- || TREE_CODE (inner_rhs) == ADDR_EXPR
- || TREE_CODE (inner_rhs) == MEM_REF)
- inner_rhs = TREE_OPERAND (inner_rhs, 0);
+ if (!inner_rhs)
+ inner_rhs = rhs;
+ if (!inner_lhs)
+ inner_lhs = lhs;
+
+ /* Reads of parameter are expected to be free. */
+ if (unmodified_parm (stmt, inner_rhs))
+ 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;
- if (TREE_CODE (inner_rhs) == PARM_DECL
- || (TREE_CODE (inner_rhs) == SSA_NAME
- && SSA_NAME_IS_DEFAULT_DEF (inner_rhs)
- && TREE_CODE (SSA_NAME_VAR (inner_rhs)) == PARM_DECL))
+ /* 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 (stmt, TREE_OPERAND (inner_rhs, 0)))
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;
- if (((TREE_CODE (inner_lhs) == PARM_DECL
- || (TREE_CODE (inner_lhs) == SSA_NAME
- && SSA_NAME_IS_DEFAULT_DEF (inner_lhs)
- && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == PARM_DECL))
- && inner_lhs != lhs)
- || TREE_CODE (inner_lhs) == RESULT_DECL
- || (TREE_CODE (inner_lhs) == SSA_NAME
- && TREE_CODE (SSA_NAME_VAR (inner_lhs)) == RESULT_DECL))
+
+ /* Writes to parameters, parameters passed by value and return value
+ (either dirrectly or passed via invisible reference) are free.
+
+ TODO: We ought to handle testcase like
+ struct a {int a,b;};
+ struct a
+ retrurnsturct (void)
+ {
+ struct a a ={1,2};
+ return a;
+ }
+
+ This translate into:
+
+ retrurnsturct ()
+ {
+ 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 (stmt, TREE_OPERAND (inner_lhs, 0))
+ || (TREE_CODE (TREE_OPERAND (inner_lhs, 0)) == SSA_NAME
+ && 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)))
}
-/* Return predicate that must be true when is E executable. */
+/* If BB ends by a conditional we can turn into predicates, attach corresponding
+ predicates to the CFG edges. */
-static struct predicate
-edge_execution_predicate (struct ipa_node_params *info,
- struct inline_summary *summary,
- edge e)
+static void
+set_cond_stmt_execution_predicate (struct ipa_node_params *info,
+ struct inline_summary *summary,
+ basic_block bb)
{
- struct predicate p = true_predicate ();
gimple last;
tree op;
int index;
- enum tree_code code;
-
- if (e->src == ENTRY_BLOCK_PTR)
- return p;
-
- last = last_stmt (e->src);
- /* TODO: handle switch. */
+ enum tree_code code, inverted_code;
+ edge e;
+ edge_iterator ei;
+ gimple set_stmt;
+ tree op2;
+ tree parm;
+ tree base;
+
+ last = last_stmt (bb);
if (!last
|| gimple_code (last) != GIMPLE_COND)
- return p;
+ return;
if (!is_gimple_ip_invariant (gimple_cond_rhs (last)))
- return p;
+ return;
op = gimple_cond_lhs (last);
/* TODO: handle conditionals like
var = op0 < 4;
- if (var != 0)
- and __bulitin_constant_p. */
- if (TREE_CODE (op) != SSA_NAME
- || !SSA_NAME_IS_DEFAULT_DEF (op))
- return p;
- index = ipa_get_param_decl_index (info, SSA_NAME_VAR (op));
+ if (var != 0). */
+ parm = unmodified_parm (last, op);
+ if (parm)
+ {
+ index = ipa_get_param_decl_index (info, parm);
+ if (index == -1)
+ return;
+ code = gimple_cond_code (last);
+ inverted_code
+ = invert_tree_comparison (code,
+ HONOR_NANS (TYPE_MODE (TREE_TYPE (op))));
+
+ FOR_EACH_EDGE (e, ei, bb->succs)
+ {
+ struct predicate p = add_condition (summary,
+ index,
+ e->flags & EDGE_TRUE_VALUE
+ ? code : inverted_code,
+ gimple_cond_rhs (last));
+ e->aux = pool_alloc (edge_predicate_pool);
+ *(struct predicate *)e->aux = p;
+ }
+ }
+
+ 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 doen't see operand is constant.
+ Other optimizers might think otherwise. */
+ 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);
+ base = get_base_address (op2);
+ parm = unmodified_parm (set_stmt, base ? base : op2);
+ if (!parm)
+ return;
+ index = ipa_get_param_decl_index (info, parm);
if (index == -1)
- return p;
- code = gimple_cond_code (last);
+ return;
+ if (gimple_cond_code (last) != NE_EXPR
+ || !integer_zerop (gimple_cond_rhs (last)))
+ return;
+ FOR_EACH_EDGE (e, ei, bb->succs)
+ if (e->flags & EDGE_FALSE_VALUE)
+ {
+ struct predicate p = add_condition (summary,
+ index,
+ IS_NOT_CONSTANT,
+ NULL);
+ e->aux = pool_alloc (edge_predicate_pool);
+ *(struct predicate *)e->aux = p;
+ }
+}
- if (EDGE_TRUE_VALUE)
- code = invert_tree_comparison (code,
- HONOR_NANS (TYPE_MODE (TREE_TYPE (op))));
- return add_condition (summary,
- index,
- gimple_cond_code (last),
- gimple_cond_rhs (last));
+/* 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_node_params *info,
+ struct inline_summary *summary,
+ basic_block bb)
+{
+ gimple last;
+ tree op;
+ int index;
+ edge e;
+ edge_iterator ei;
+ size_t n;
+ size_t case_idx;
+ tree parm;
+
+ last = last_stmt (bb);
+ if (!last
+ || gimple_code (last) != GIMPLE_SWITCH)
+ return;
+ op = gimple_switch_index (last);
+ parm = unmodified_parm (last, op);
+ if (!parm)
+ return;
+ index = ipa_get_param_decl_index (info, parm);
+ if (index == -1)
+ return;
+
+ FOR_EACH_EDGE (e, ei, bb->succs)
+ {
+ e->aux = pool_alloc (edge_predicate_pool);
+ *(struct predicate *)e->aux = false_predicate ();
+ }
+ n = gimple_switch_num_labels(last);
+ for (case_idx = 0; case_idx < n; ++case_idx)
+ {
+ tree cl = gimple_switch_label (last, case_idx);
+ tree min, max;
+ struct predicate p;
+
+ e = find_edge (bb, label_to_block (CASE_LABEL (cl)));
+ min = CASE_LOW (cl);
+ max = CASE_HIGH (cl);
+
+ /* For default we might want to construct predicate that none
+ of cases is met, but it is bit hard to do not having negations
+ of conditionals handy. */
+ if (!min && !max)
+ p = true_predicate ();
+ else if (!max)
+ p = add_condition (summary, index,
+ EQ_EXPR,
+ min);
+ else
+ {
+ struct predicate p1, p2;
+ p1 = add_condition (summary, index,
+ GE_EXPR,
+ min);
+ p2 = add_condition (summary, index,
+ LE_EXPR,
+ max);
+ p = and_predicates (summary->conds, &p1, &p2);
+ }
+ *(struct predicate *)e->aux
+ = or_predicates (summary->conds, &p, (struct predicate *)e->aux);
+ }
}
+
+/* For each BB in NODE attach to its AUX pointer predicate under
+ which it is executable. */
+
+static void
+compute_bb_predicates (struct cgraph_node *node,
+ struct ipa_node_params *parms_info,
+ struct inline_summary *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 (parms_info, summary, bb);
+ set_switch_stmt_execution_predicate (parms_info, summary, bb);
+ }
+
+ /* Entry block is always executable. */
+ ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function)->aux
+ = pool_alloc (edge_predicate_pool);
+ *(struct predicate *)ENTRY_BLOCK_PTR_FOR_FUNCTION (my_function)->aux
+ = true_predicate ();
+
+ /* 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)
+ {
+ struct predicate p = false_predicate ();
+ edge e;
+ edge_iterator ei;
+ FOR_EACH_EDGE (e, ei, bb->preds)
+ {
+ if (e->src->aux)
+ {
+ struct predicate this_bb_predicate
+ = *(struct predicate *)e->src->aux;
+ if (e->aux)
+ this_bb_predicate
+ = and_predicates (summary->conds, &this_bb_predicate,
+ (struct predicate *)e->aux);
+ p = or_predicates (summary->conds, &p, &this_bb_predicate);
+ if (true_predicate_p (&p))
+ break;
+ }
+ }
+ if (false_predicate_p (&p))
+ gcc_assert (!bb->aux);
+ else
+ {
+ if (!bb->aux)
+ {
+ done = false;
+ bb->aux = pool_alloc (edge_predicate_pool);
+ *((struct predicate *)bb->aux) = p;
+ }
+ else if (!predicates_equal_p (&p, (struct predicate *)bb->aux))
+ {
+ done = false;
+ *((struct predicate *)bb->aux) = p;
+ }
+ }
+ }
+ }
+}
+
+
+/* We keep info about constantness of SSA names. */
+
+typedef struct predicate predicate_t;
+DEF_VEC_O (predicate_t);
+DEF_VEC_ALLOC_O (predicate_t, heap);
+
+
+/* Return predicate specifying when the STMT might have result that is not
+ a compile time constant. */
+
static struct predicate
will_be_nonconstant_predicate (struct ipa_node_params *info,
struct inline_summary *summary,
- gimple stmt)
+ gimple stmt,
+ VEC (predicate_t, heap) *nonconstant_names)
+
{
struct predicate p = true_predicate ();
ssa_op_iter iter;
tree use;
struct predicate op_non_const;
+ bool is_load;
+
+ /* What statments might be optimized away
+ when their arguments are constant
+ TODO: also trivial builtins.
+ builtin_constant_p is already handled later. */
+ if (gimple_code (stmt) != GIMPLE_ASSIGN
+ && gimple_code (stmt) != GIMPLE_COND
+ && gimple_code (stmt) != GIMPLE_SWITCH)
+ return p;
+
+ /* Stores will stay anyway. */
+ if (gimple_vdef (stmt))
+ return p;
+
+ is_load = gimple_vuse (stmt) != NULL;
+
+ /* Loads can be optimized when the value is known. */
+ if (is_load)
+ {
+ tree op = gimple_assign_rhs1 (stmt);
+ tree base = get_base_address (op);
+ tree parm;
+
+ gcc_assert (gimple_assign_single_p (stmt));
+ if (!base)
+ return p;
+ parm = unmodified_parm (stmt, base);
+ if (!parm )
+ return p;
+ if (ipa_get_param_decl_index (info, parm) < 0)
+ return p;
+ }
+
+ /* 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 (stmt, use);
+ /* For arguments we can build a condition. */
+ if (parm && ipa_get_param_decl_index (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 (!true_predicate_p (VEC_index (predicate_t, nonconstant_names,
+ SSA_NAME_VERSION (use))))
+ continue;
+ return p;
+ }
+ op_non_const = false_predicate ();
+ if (is_load)
+ {
+ tree parm = unmodified_parm
+ (stmt, get_base_address (gimple_assign_rhs1 (stmt)));
+ p = add_condition (summary,
+ ipa_get_param_decl_index (info, parm),
+ CHANGED, NULL);
+ op_non_const = or_predicates (summary->conds, &p, &op_non_const);
+ }
+ FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
+ {
+ tree parm = unmodified_parm (stmt, use);
+ if (parm && ipa_get_param_decl_index (info, parm) >= 0)
+ p = add_condition (summary,
+ ipa_get_param_decl_index (info, parm),
+ CHANGED, NULL);
+ else
+ p = *VEC_index (predicate_t, nonconstant_names,
+ SSA_NAME_VERSION (use));
+ op_non_const = or_predicates (summary->conds, &p, &op_non_const);
+ }
+ if (gimple_code (stmt) == GIMPLE_ASSIGN
+ && TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME)
+ VEC_replace (predicate_t, nonconstant_names,
+ SSA_NAME_VERSION (gimple_assign_lhs (stmt)), &op_non_const);
+ return op_non_const;
+}
+
+struct record_modified_bb_info
+{
+ bitmap bb_set;
+ gimple stmt;
+};
+
+/* 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;
+ bitmap_set_bit (info->bb_set,
+ SSA_NAME_IS_DEFAULT_DEF (vdef)
+ ? ENTRY_BLOCK_PTR->index : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
+ return false;
+}
- /* What statments might be optimized away
- when their arguments are constant
- TODO: also trivial builtins, especially builtin_constant_p. */
- if (gimple_code (stmt) != GIMPLE_ASSIGN
- && gimple_code (stmt) != GIMPLE_COND
- && gimple_code (stmt) != GIMPLE_SWITCH)
- return p;
+/* Return probability (based on REG_BR_PROB_BASE) that I-th parameter of STMT
+ will change since last invocation of STMT.
- /* Stores and loads will stay anyway. */
- if (gimple_vuse (stmt))
- return p;
+ 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. */
- /* See if we understand all operands before we start
- adding conditionals. */
- FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
+static int
+param_change_prob (gimple stmt, int i)
+{
+ tree op = gimple_call_arg (stmt, i);
+ basic_block bb = gimple_bb (stmt);
+ tree base;
+
+ if (is_gimple_min_invariant (op))
+ 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 (op) == SSA_NAME)
{
- /* TODO: handle nested expressions and constant
- array accesses. */
- if (TREE_CODE (use) != SSA_NAME
- || !SSA_NAME_IS_DEFAULT_DEF (use)
- || ipa_get_param_decl_index (info, SSA_NAME_VAR (use)) < 0)
- return p;
+ int init_freq;
+
+ if (!bb->frequency)
+ return REG_BR_PROB_BASE;
+
+ if (SSA_NAME_IS_DEFAULT_DEF (op))
+ init_freq = ENTRY_BLOCK_PTR->frequency;
+ else
+ init_freq = gimple_bb (SSA_NAME_DEF_STMT (op))->frequency;
+
+ if (!init_freq)
+ init_freq = 1;
+ if (init_freq < bb->frequency)
+ return MAX ((init_freq * REG_BR_PROB_BASE +
+ bb->frequency / 2) / bb->frequency, 1);
+ else
+ return REG_BR_PROB_BASE;
}
- op_non_const = false_predicate ();
- FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
+
+ base = get_base_address (op);
+ if (base)
{
- p = add_condition (summary,
- ipa_get_param_decl_index (info, SSA_NAME_VAR (use)),
- IS_NOT_CONSTANT, NULL);
- op_non_const = or_predicates (&p, &op_non_const);
+ ao_ref refd;
+ int max;
+ struct record_modified_bb_info info;
+ bitmap_iterator bi;
+ unsigned index;
+
+ if (const_value_known_p (base))
+ return 0;
+ if (!bb->frequency)
+ return REG_BR_PROB_BASE;
+ ao_ref_init (&refd, op);
+ info.stmt = stmt;
+ info.bb_set = BITMAP_ALLOC (NULL);
+ walk_aliased_vdefs (&refd, gimple_vuse (stmt), record_modified, &info,
+ NULL);
+ if (bitmap_bit_p (info.bb_set, bb->index))
+ {
+ BITMAP_FREE (info.bb_set);
+ return REG_BR_PROB_BASE;
+ }
+
+ /* Assume that every memory is initialized at entry.
+ TODO: Can we easilly determine if value is always defined
+ and thus we may skip entry block? */
+ if (ENTRY_BLOCK_PTR->frequency)
+ max = ENTRY_BLOCK_PTR->frequency;
+ else
+ max = 1;
+
+ EXECUTE_IF_SET_IN_BITMAP (info.bb_set, 0, index, bi)
+ max = MIN (max, BASIC_BLOCK (index)->frequency);
+
+ BITMAP_FREE (info.bb_set);
+ if (max < bb->frequency)
+ return MAX ((max * REG_BR_PROB_BASE +
+ bb->frequency / 2) / bb->frequency, 1);
+ else
+ return REG_BR_PROB_BASE;
}
- return op_non_const;
+ return REG_BR_PROB_BASE;
}
int freq;
struct inline_summary *info = inline_summary (node);
struct predicate bb_predicate;
- struct ipa_node_params *parms_info;
+ struct ipa_node_params *parms_info = NULL;
+ VEC (predicate_t, heap) *nonconstant_names = NULL;
- parms_info = ipa_node_params_vector && !early ? IPA_NODE_REF (node) : NULL;
+ if (ipa_node_params_vector && !early && optimize)
+ {
+ parms_info = IPA_NODE_REF (node);
+ VEC_safe_grow_cleared (predicate_t, heap, nonconstant_names,
+ VEC_length (tree, SSANAMES (my_function)));
+ }
info->conds = 0;
info->entry = 0;
bb_predicate = not_inlined_predicate ();
account_size_time (info, 2 * INLINE_SIZE_SCALE, 0, &bb_predicate);
-
gcc_assert (my_function && my_function->cfg);
+ if (parms_info)
+ compute_bb_predicates (node, parms_info, info);
FOR_EACH_BB_FN (bb, my_function)
{
- edge e;
- edge_iterator ei;
-
freq = compute_call_stmt_bb_frequency (node->decl, bb);
/* TODO: Obviously predicates can be propagated down across CFG. */
if (parms_info)
{
- bb_predicate = false_predicate ();
- FOR_EACH_EDGE (e, ei, bb->preds)
- {
- struct predicate ep;
- ep = edge_execution_predicate (parms_info, info, e);
- bb_predicate = or_predicates (&ep, &bb_predicate);
- }
+ if (bb->aux)
+ bb_predicate = *(struct predicate *)bb->aux;
+ else
+ bb_predicate = false_predicate ();
}
else
bb_predicate = true_predicate ();
int this_size = estimate_num_insns (stmt, &eni_size_weights);
int this_time = estimate_num_insns (stmt, &eni_time_weights);
int prob;
+ struct predicate will_be_nonconstant;
if (dump_file && (dump_flags & TDF_DETAILS))
{
if (is_gimple_call (stmt))
{
struct cgraph_edge *edge = cgraph_edge (node, stmt);
- edge->call_stmt_size = this_size;
- edge->call_stmt_time = this_time;
-
- /* Do not inline calls where we cannot triviall work around
- mismatches in argument or return types. */
- if (edge->callee
- && !gimple_check_call_matching_types (stmt, edge->callee->decl))
+ struct inline_edge_summary *es = inline_edge_summary (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
+ && gimple_call_builtin_p (stmt, BUILT_IN_CONSTANT_P)
+ && gimple_call_lhs (stmt)
+ && TREE_CODE (gimple_call_lhs (stmt)) == SSA_NAME)
{
- edge->call_stmt_cannot_inline_p = true;
- gimple_call_set_cannot_inline (stmt, true);
+ struct predicate false_p = false_predicate ();
+ VEC_replace (predicate_t, nonconstant_names,
+ SSA_NAME_VERSION (gimple_call_lhs (stmt)),
+ &false_p);
}
- else
- gcc_assert (!gimple_call_cannot_inline_p (stmt));
+ if (ipa_node_params_vector)
+ {
+ int count = gimple_call_num_args (stmt);
+ int i;
+
+ if (count)
+ VEC_safe_grow_cleared (inline_param_summary_t, heap,
+ es->param, count);
+ for (i = 0; i < count; i++)
+ {
+ int prob = param_change_prob (stmt, i);
+ gcc_assert (prob >= 0 && prob <= REG_BR_PROB_BASE);
+ VEC_index (inline_param_summary_t,
+ es->param, i)->change_prob = prob;
+ }
+ }
+
+ es->call_stmt_size = this_size;
+ es->call_stmt_time = this_time;
+ es->loop_depth = bb->loop_depth;
+ edge_set_predicate (edge, &bb_predicate);
}
+ /* TODO: When conditional jump or swithc 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 (parms_info)
+ will_be_nonconstant
+ = will_be_nonconstant_predicate (parms_info, info,
+ stmt, nonconstant_names);
if (this_time || this_size)
{
- struct predicate will_be_nonconstant;
struct predicate p;
this_time *= freq;
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 eliminated by inlining\n");
+ fprintf (dump_file, "\t\tWill be eliminated by inlining\n");
if (parms_info)
- {
- will_be_nonconstant
- = will_be_nonconstant_predicate (parms_info, info, stmt);
- p = and_predicates (&bb_predicate, &will_be_nonconstant);
- }
+ p = and_predicates (info->conds, &bb_predicate,
+ &will_be_nonconstant);
else
p = true_predicate ();
if (prob)
{
struct predicate ip = not_inlined_predicate ();
- ip = and_predicates (&ip, &p);
+ ip = and_predicates (info->conds, &ip, &p);
account_size_time (info, this_size * prob,
this_time * prob, &ip);
}
}
}
}
+ FOR_ALL_BB_FN (bb, my_function)
+ {
+ edge e;
+ edge_iterator ei;
+
+ if (bb->aux)
+ pool_free (edge_predicate_pool, bb->aux);
+ bb->aux = NULL;
+ FOR_EACH_EDGE (e, ei, bb->succs)
+ {
+ if (e->aux)
+ pool_free (edge_predicate_pool, e->aux);
+ e->aux = NULL;
+ }
+ }
time = (time + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
if (time > MAX_TIME)
time = MAX_TIME;
inline_summary (node)->self_time = time;
inline_summary (node)->self_size = size;
+ VEC_free (predicate_t, heap, nonconstant_names);
if (dump_file)
{
fprintf (dump_file, "\n");
HOST_WIDE_INT self_stack_size;
struct cgraph_edge *e;
struct inline_summary *info;
+ tree old_decl = current_function_decl;
gcc_assert (!node->global.inlined_to);
inline_summary_alloc ();
info = inline_summary (node);
+ reset_inline_summary (node);
+
+ /* FIXME: Thunks are inlinable, but tree-inline don't know how to do that.
+ Once this happen, we will need to more curefully predict call
+ statement size. */
+ if (node->thunk.thunk_p)
+ {
+ struct inline_edge_summary *es = inline_edge_summary (node->callees);
+ struct predicate t = true_predicate ();
+
+ info->inlinable = 0;
+ node->callees->call_stmt_cannot_inline_p = true;
+ node->local.can_change_signature = false;
+ es->call_stmt_time = 1;
+ es->call_stmt_size = 1;
+ account_size_time (info, 0, 0, &t);
+ return;
+ }
+
+ /* Even is_gimple_min_invariant rely on current_function_decl. */
+ current_function_decl = node->decl;
+ push_cfun (DECL_STRUCT_FUNCTION (node->decl));
/* Estimate the stack size for the function if we're optimizing. */
self_stack_size = optimize ? estimated_stack_frame_size (node) : 0;
/* Can this function be inlined at all? */
info->inlinable = tree_inlinable_function_p (node->decl);
- /* Inlinable functions always can change signature. */
- if (info->inlinable)
- node->local.can_change_signature = true;
+ /* Type attributes can use parameter indices to describe them. */
+ if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
+ node->local.can_change_signature = false;
else
{
- /* Functions calling builtin_apply can not change signature. */
- for (e = node->callees; e; e = e->next_callee)
- if (DECL_BUILT_IN (e->callee->decl)
- && DECL_BUILT_IN_CLASS (e->callee->decl) == BUILT_IN_NORMAL
- && DECL_FUNCTION_CODE (e->callee->decl) == BUILT_IN_APPLY_ARGS)
- break;
- node->local.can_change_signature = !e;
+ /* Otherwise, inlinable functions always can change signature. */
+ if (info->inlinable)
+ node->local.can_change_signature = true;
+ else
+ {
+ /* Functions calling builtin_apply can not change signature. */
+ for (e = node->callees; e; e = e->next_callee)
+ {
+ tree cdecl = e->callee->decl;
+ if (DECL_BUILT_IN (cdecl)
+ && DECL_BUILT_IN_CLASS (cdecl) == BUILT_IN_NORMAL
+ && (DECL_FUNCTION_CODE (cdecl) == BUILT_IN_APPLY_ARGS
+ || DECL_FUNCTION_CODE (cdecl) == BUILT_IN_VA_START))
+ break;
+ }
+ node->local.can_change_signature = !e;
+ }
}
estimate_function_body_sizes (node, early);
info->size = info->self_size;
info->stack_frame_offset = 0;
info->estimated_stack_size = info->estimated_self_stack_size;
+ current_function_decl = old_decl;
+ pop_cfun ();
}
/* Increase SIZE and TIME for size and time needed to handle edge E. */
static void
-estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *time)
+estimate_edge_size_and_time (struct cgraph_edge *e, int *size, int *time,
+ int prob)
{
- *size += e->call_stmt_size * INLINE_SIZE_SCALE;
- *time += (e->call_stmt_time
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ *size += es->call_stmt_size * INLINE_SIZE_SCALE;
+ *time += (es->call_stmt_time * prob / REG_BR_PROB_BASE
* e->frequency * (INLINE_TIME_SCALE / CGRAPH_FREQ_BASE));
if (*time > MAX_TIME * INLINE_TIME_SCALE)
*time = MAX_TIME * INLINE_TIME_SCALE;
}
-/* Increase SIZE and TIME for size and time needed to handle all calls in NODE. */
+/* Estimate benefit devirtualizing indirect edge IE, provided KNOWN_VALS and
+ KNOWN_BINFOS. */
+
+static void
+estimate_edge_devirt_benefit (struct cgraph_edge *ie,
+ int *size, int *time, int prob,
+ VEC (tree, heap) *known_vals,
+ VEC (tree, heap) *known_binfos)
+{
+ tree target;
+ int time_diff, size_diff;
+
+ if (!known_vals && !known_binfos)
+ return;
+
+ target = ipa_get_indirect_edge_target (ie, known_vals, known_binfos);
+ if (!target)
+ return;
+
+ /* Account for difference in cost between indirect and direct calls. */
+ size_diff = ((eni_size_weights.indirect_call_cost - eni_size_weights.call_cost)
+ * INLINE_SIZE_SCALE);
+ *size -= size_diff;
+ time_diff = ((eni_time_weights.indirect_call_cost - eni_time_weights.call_cost)
+ * INLINE_TIME_SCALE * prob / REG_BR_PROB_BASE);
+ *time -= time_diff;
+
+ /* TODO: This code is trying to benefit indirect calls that will be inlined later.
+ The logic however do not belong into local size/time estimates and can not be
+ done here, or the accounting of changes will get wrong and we result with
+ negative function body sizes. We need to introduce infrastructure for independent
+ benefits to the inliner. */
+#if 0
+ struct cgraph_node *callee;
+ struct inline_summary *isummary;
+ int edge_size, edge_time, time_diff, size_diff;
+
+ callee = cgraph_get_node (target);
+ if (!callee || !callee->analyzed)
+ return;
+ isummary = inline_summary (callee);
+ if (!isummary->inlinable)
+ return;
+
+ estimate_edge_size_and_time (ie, &edge_size, &edge_time, prob);
+
+ /* Count benefit only from functions that definitely will be inlined
+ if additional context from NODE's caller were available.
+
+ We just account overall size change by inlining. TODO:
+ we really need to add sort of benefit metrics for these kind of
+ cases. */
+ if (edge_size - size_diff >= isummary->size * INLINE_SIZE_SCALE)
+ {
+ /* Subtract size and time that we added for edge IE. */
+ *size -= edge_size - size_diff;
+
+ /* Account inlined call. */
+ *size += isummary->size * INLINE_SIZE_SCALE;
+ }
+#endif
+}
+
+
+/* Increase SIZE and TIME for size and time needed to handle all calls in NODE.
+ POSSIBLE_TRUTHS, KNOWN_VALS and KNOWN_BINFOS describe context of the call
+ site. */
static void
-estimate_calls_size_and_time (struct cgraph_node *node, int *size, int *time)
+estimate_calls_size_and_time (struct cgraph_node *node, int *size, int *time,
+ clause_t possible_truths,
+ VEC (tree, heap) *known_vals,
+ VEC (tree, heap) *known_binfos)
{
struct cgraph_edge *e;
for (e = node->callees; e; e = e->next_callee)
- if (e->inline_failed)
- estimate_edge_size_and_time (e, size, time);
- else
- estimate_calls_size_and_time (e->callee, size, time);
- /* TODO: look for devirtualizing oppurtunities. */
+ {
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ if (!es->predicate || evaluate_predicate (es->predicate, possible_truths))
+ {
+ if (e->inline_failed)
+ {
+ /* Predicates of calls shall not use NOT_CHANGED codes,
+ sowe do not need to compute probabilities. */
+ estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE);
+ }
+ else
+ estimate_calls_size_and_time (e->callee, size, time,
+ possible_truths,
+ known_vals, known_binfos);
+ }
+ }
for (e = node->indirect_calls; e; e = e->next_callee)
- estimate_edge_size_and_time (e, size, time);
+ {
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ if (!es->predicate || evaluate_predicate (es->predicate, possible_truths))
+ {
+ estimate_edge_size_and_time (e, size, time, REG_BR_PROB_BASE);
+ estimate_edge_devirt_benefit (e, size, time, REG_BR_PROB_BASE,
+ known_vals, known_binfos);
+ }
+ }
}
-/* Estimate size and time needed to execute callee of EDGE assuming
- that parameters known to be constant at caller of EDGE are
- propagated. If INLINE_P is true, it is assumed that call will
- be inlined. */
+/* Estimate size and time needed to execute NODE assuming
+ POSSIBLE_TRUTHS clause, and KNOWN_VALS and KNOWN_BINFOS information
+ about NODE's arguments. */
static void
-estimate_callee_size_and_time (struct cgraph_edge *edge, bool inline_p,
- int *ret_size, int *ret_time)
+estimate_node_size_and_time (struct cgraph_node *node,
+ clause_t possible_truths,
+ VEC (tree, heap) *known_vals,
+ VEC (tree, heap) *known_binfos,
+ int *ret_size, int *ret_time,
+ VEC (inline_param_summary_t, heap)
+ *inline_param_summary)
{
- struct inline_summary *info = inline_summary (edge->callee);
- clause_t clause = evaulate_conditions_for_edge (edge, inline_p);
+ struct inline_summary *info = inline_summary (node);
size_time_entry *e;
int size = 0, time = 0;
int i;
&& (dump_flags & TDF_DETAILS))
{
bool found = false;
- fprintf (dump_file, " Estimating callee body: %s/%i\n"
+ fprintf (dump_file, " Estimating body: %s/%i\n"
" Known to be false: ",
- cgraph_node_name (edge->callee),
- edge->callee->uid);
+ cgraph_node_name (node),
+ node->uid);
for (i = predicate_not_inlined_condition;
i < (predicate_first_dynamic_condition
+ (int)VEC_length (condition, info->conds)); i++)
- if (!(clause & (1 << i)))
+ if (!(possible_truths & (1 << i)))
{
if (found)
fprintf (dump_file, ", ");
}
for (i = 0; VEC_iterate (size_time_entry, info->entry, i, e); i++)
- if (evaulate_predicate (&e->predicate, clause))
- time += e->time, size += e->size;
+ if (evaluate_predicate (&e->predicate, possible_truths))
+ {
+ size += e->size;
+ if (!inline_param_summary)
+ time += e->time;
+ else
+ {
+ int prob = predicate_probability (info->conds,
+ &e->predicate,
+ possible_truths,
+ inline_param_summary);
+ time += e->time * prob / REG_BR_PROB_BASE;
+ }
+
+ }
if (time > MAX_TIME * INLINE_TIME_SCALE)
time = MAX_TIME * INLINE_TIME_SCALE;
- estimate_calls_size_and_time (edge->callee, &size, &time);
+ estimate_calls_size_and_time (node, &size, &time, possible_truths,
+ known_vals, known_binfos);
time = (time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
size = (size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
}
-/* Translate all conditions from callee representation into caller representaiton and
- symbolically evaulate predicate P into new predicate. */
+/* Estimate size and time needed to execute callee of EDGE assuming that
+ parameters known to be constant at caller of EDGE are propagated.
+ KNOWN_VALS and KNOWN_BINFOS are vectors of assumed known constant values
+ and types for parameters. */
+
+void
+estimate_ipcp_clone_size_and_time (struct cgraph_node *node,
+ VEC (tree, heap) *known_vals,
+ VEC (tree, heap) *known_binfos,
+ int *ret_size, int *ret_time)
+{
+ clause_t clause;
+
+ clause = evaluate_conditions_for_known_args (node, false, known_vals);
+ estimate_node_size_and_time (node, clause, known_vals, known_binfos,
+ ret_size, ret_time,
+ NULL);
+}
+
+
+/* Translate all conditions from callee representation into caller
+ representation and symbolically evaluate predicate P into new predicate.
+
+ INFO is inline_summary of function we are adding predicate into,
+ CALLEE_INFO is summary of function predicate P is from. OPERAND_MAP is
+ array giving callee formal IDs the caller formal IDs. POSSSIBLE_TRUTHS is
+ clausule of all callee conditions that may be true in caller context.
+ TOPLEV_PREDICATE is predicate under which callee is executed. */
static struct predicate
-remap_predicate (struct inline_summary *info, struct inline_summary *callee_info,
+remap_predicate (struct inline_summary *info,
+ struct inline_summary *callee_info,
struct predicate *p,
VEC (int, heap) *operand_map,
- clause_t possible_truths)
+ clause_t possible_truths,
+ struct predicate *toplev_predicate)
{
int i;
struct predicate out = true_predicate ();
/* True predicate is easy. */
- if (p->clause[0] == 0)
- return *p;
+ if (true_predicate_p (p))
+ return *toplev_predicate;
for (i = 0; p->clause[i]; i++)
{
clause_t clause = p->clause[i];
int cond;
struct predicate clause_predicate = false_predicate ();
+ gcc_assert (i < MAX_CLAUSES);
+
for (cond = 0; cond < NUM_CONDITIONS; cond ++)
/* Do we have condition we can't disprove? */
if (clause & possible_truths & (1 << cond))
/* See if we can remap condition operand to caller's operand.
Otherwise give up. */
if (!operand_map
+ || (int)VEC_length (int, operand_map) <= c->operand_num
|| VEC_index (int, operand_map, c->operand_num) == -1)
cond_predicate = true_predicate ();
else
cond_predicate.clause[0] = 1 << cond;
cond_predicate.clause[1] = 0;
}
- clause_predicate = or_predicates (&clause_predicate, &cond_predicate);
+ clause_predicate = or_predicates (info->conds, &clause_predicate,
+ &cond_predicate);
}
- out = and_predicates (&out, &clause_predicate);
+ out = and_predicates (info->conds, &out, &clause_predicate);
+ }
+ return and_predicates (info->conds, &out, toplev_predicate);
+}
+
+
+/* Update summary information of inline clones after inlining.
+ Compute peak stack usage. */
+
+static void
+inline_update_callee_summaries (struct cgraph_node *node,
+ int depth)
+{
+ struct cgraph_edge *e;
+ struct inline_summary *callee_info = inline_summary (node);
+ struct inline_summary *caller_info = inline_summary (node->callers->caller);
+ HOST_WIDE_INT peak;
+
+ callee_info->stack_frame_offset
+ = caller_info->stack_frame_offset
+ + caller_info->estimated_self_stack_size;
+ peak = callee_info->stack_frame_offset
+ + callee_info->estimated_self_stack_size;
+ if (inline_summary (node->global.inlined_to)->estimated_stack_size
+ < peak)
+ inline_summary (node->global.inlined_to)->estimated_stack_size = peak;
+ cgraph_propagate_frequency (node);
+ for (e = node->callees; e; e = e->next_callee)
+ {
+ if (!e->inline_failed)
+ inline_update_callee_summaries (e->callee, depth);
+ inline_edge_summary (e)->loop_depth += depth;
+ }
+ for (e = node->indirect_calls; e; e = e->next_callee)
+ inline_edge_summary (e)->loop_depth += depth;
+}
+
+/* Update change_prob of EDGE after INLINED_EDGE has been inlined.
+ When functoin A is inlined in B and A calls C with parameter that
+ changes with probability PROB1 and C is known to be passthroug
+ of argument if B that change with probability PROB2, the probability
+ of change is now PROB1*PROB2. */
+
+static void
+remap_edge_change_prob (struct cgraph_edge *inlined_edge,
+ struct cgraph_edge *edge)
+{
+ if (ipa_node_params_vector)
+ {
+ int i;
+ struct ipa_edge_args *args = IPA_EDGE_REF (edge);
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+ struct inline_edge_summary *inlined_es
+ = inline_edge_summary (inlined_edge);
+
+ for (i = 0; i < ipa_get_cs_argument_count (args); i++)
+ {
+ struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
+ if (jfunc->type == IPA_JF_PASS_THROUGH
+ && (jfunc->value.pass_through.formal_id
+ < (int) VEC_length (inline_param_summary_t,
+ inlined_es->param)))
+ {
+ int prob1 = VEC_index (inline_param_summary_t,
+ es->param, i)->change_prob;
+ int prob2 = VEC_index
+ (inline_param_summary_t,
+ inlined_es->param,
+ jfunc->value.pass_through.formal_id)->change_prob;
+ int prob = ((prob1 * prob2 + REG_BR_PROB_BASE / 2)
+ / REG_BR_PROB_BASE);
+
+ if (prob1 && prob2 && !prob)
+ prob = 1;
+
+ VEC_index (inline_param_summary_t,
+ es->param, i)->change_prob = prob;
+ }
+ }
+ }
+}
+
+/* Update edge summaries of NODE after INLINED_EDGE has been inlined.
+
+ Remap predicates of callees of NODE. Rest of arguments match
+ remap_predicate.
+
+ Also update change probabilities. */
+
+static void
+remap_edge_summaries (struct cgraph_edge *inlined_edge,
+ struct cgraph_node *node,
+ struct inline_summary *info,
+ struct inline_summary *callee_info,
+ VEC (int, heap) *operand_map,
+ clause_t possible_truths,
+ struct predicate *toplev_predicate)
+{
+ struct cgraph_edge *e;
+ for (e = node->callees; e; e = e->next_callee)
+ {
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ struct predicate p;
+
+ if (e->inline_failed)
+ {
+ remap_edge_change_prob (inlined_edge, e);
+
+ if (es->predicate)
+ {
+ p = remap_predicate (info, callee_info,
+ es->predicate, operand_map, possible_truths,
+ toplev_predicate);
+ edge_set_predicate (e, &p);
+ /* TODO: We should remove the edge for code that will be
+ optimized out, but we need to keep verifiers and tree-inline
+ happy. Make it cold for now. */
+ if (false_predicate_p (&p))
+ {
+ e->count = 0;
+ e->frequency = 0;
+ }
+ }
+ else
+ edge_set_predicate (e, toplev_predicate);
+ }
+ else
+ remap_edge_summaries (inlined_edge, e->callee, info, callee_info,
+ operand_map, possible_truths, toplev_predicate);
+ }
+ for (e = node->indirect_calls; e; e = e->next_callee)
+ {
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ struct predicate p;
+
+ remap_edge_change_prob (inlined_edge, e);
+ if (es->predicate)
+ {
+ p = remap_predicate (info, callee_info,
+ es->predicate, operand_map, possible_truths,
+ toplev_predicate);
+ edge_set_predicate (e, &p);
+ /* TODO: We should remove the edge for code that will be optimized
+ out, but we need to keep verifiers and tree-inline happy.
+ Make it cold for now. */
+ if (false_predicate_p (&p))
+ {
+ e->count = 0;
+ e->frequency = 0;
+ }
+ }
+ else
+ edge_set_predicate (e, toplev_predicate);
}
- return out;
}
size_time_entry *e;
VEC (int, heap) *operand_map = NULL;
int i;
+ struct predicate toplev_predicate;
+ struct predicate true_p = true_predicate ();
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+
+ if (es->predicate)
+ toplev_predicate = *es->predicate;
+ else
+ toplev_predicate = true_predicate ();
- if (ipa_node_params_vector && callee_info->conds
- /* FIXME: it seems that we forget to get argument count in some cases,
- probaby for previously indirect edges or so.
- Removing the test leads to ICE on tramp3d. */
- && ipa_get_cs_argument_count (IPA_EDGE_REF (edge)))
+ if (ipa_node_params_vector && callee_info->conds)
{
struct ipa_edge_args *args = IPA_EDGE_REF (edge);
int count = ipa_get_cs_argument_count (args);
int i;
- clause = evaulate_conditions_for_edge (edge, true);
- VEC_safe_grow_cleared (int, heap, operand_map, count);
+ evaluate_properties_for_edge (edge, true, &clause, NULL, NULL);
+ if (count)
+ VEC_safe_grow_cleared (int, heap, operand_map, count);
for (i = 0; i < count; i++)
{
struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
&& jfunc->value.pass_through.operation == NOP_EXPR)
map = jfunc->value.pass_through.formal_id;
VEC_replace (int, operand_map, i, map);
+ gcc_assert (map < ipa_get_param_count (IPA_NODE_REF (to)));
}
}
for (i = 0; VEC_iterate (size_time_entry, callee_info->entry, i, e); i++)
{
struct predicate p = remap_predicate (info, callee_info,
- &e->predicate, operand_map, clause);
- gcov_type add_time = ((gcov_type)e->time * edge->frequency
- + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
- if (add_time > MAX_TIME)
- add_time = MAX_TIME;
- account_size_time (info, e->size, add_time, &p);
+ &e->predicate, operand_map, clause,
+ &toplev_predicate);
+ if (!false_predicate_p (&p))
+ {
+ gcov_type add_time = ((gcov_type)e->time * edge->frequency
+ + CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
+ int prob = predicate_probability (callee_info->conds,
+ &e->predicate,
+ clause, es->param);
+ add_time = add_time * prob / REG_BR_PROB_BASE;
+ if (add_time > MAX_TIME * INLINE_TIME_SCALE)
+ add_time = MAX_TIME * INLINE_TIME_SCALE;
+ if (prob != REG_BR_PROB_BASE
+ && dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "\t\tScaling time by probability:%f\n",
+ (double)prob / REG_BR_PROB_BASE);
+ }
+ account_size_time (info, e->size, add_time, &p);
+ }
}
+ remap_edge_summaries (edge, edge->callee, info, callee_info, operand_map,
+ clause, &toplev_predicate);
info->size = 0;
info->time = 0;
for (i = 0; VEC_iterate (size_time_entry, info->entry, i, e); i++)
info->size += e->size, info->time += e->time;
- estimate_calls_size_and_time (to, &info->size, &info->time);
+ estimate_calls_size_and_time (to, &info->size, &info->time,
+ ~(clause_t)(1 << predicate_false_condition),
+ NULL, NULL);
+
+ inline_update_callee_summaries (edge->callee,
+ inline_edge_summary (edge)->loop_depth);
+
+ /* We do not maintain predicates of inlined edges, free it. */
+ edge_set_predicate (edge, &true_p);
+ /* Similarly remove param summaries. */
+ VEC_free (inline_param_summary_t, heap, es->param);
+
info->time = (info->time + INLINE_TIME_SCALE / 2) / INLINE_TIME_SCALE;
info->size = (info->size + INLINE_SIZE_SCALE / 2) / INLINE_SIZE_SCALE;
}
int time;
int size;
gcov_type ret;
+ struct cgraph_node *callee;
+ clause_t clause;
+ VEC (tree, heap) *known_vals;
+ VEC (tree, heap) *known_binfos;
+ struct inline_edge_summary *es = inline_edge_summary (edge);
- gcc_checking_assert (edge->inline_failed);
- estimate_callee_size_and_time (edge, true, &size, &time);
+ callee = cgraph_function_or_thunk_node (edge->callee, NULL);
- ret = (((gcov_type)time - edge->call_stmt_time) * edge->frequency
+ gcc_checking_assert (edge->inline_failed);
+ evaluate_properties_for_edge (edge, true,
+ &clause, &known_vals, &known_binfos);
+ estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
+ &size, &time, es->param);
+ VEC_free (tree, heap, known_vals);
+ VEC_free (tree, heap, known_binfos);
+
+ ret = (((gcov_type)time
+ - es->call_stmt_time) * edge->frequency
+ CGRAPH_FREQ_BASE / 2) / CGRAPH_FREQ_BASE;
- if (ret > MAX_TIME)
- ret = MAX_TIME;
/* When caching, update the cache entry. */
if (edge_growth_cache)
VEC_index (edge_growth_cache_entry, edge_growth_cache, edge->uid)->time
= ret + (ret >= 0);
- ret_size = size - edge->call_stmt_size;
- gcc_checking_assert (edge->call_stmt_size);
+ ret_size = size - es->call_stmt_size;
+ gcc_checking_assert (es->call_stmt_size);
VEC_index (edge_growth_cache_entry, edge_growth_cache, edge->uid)->size
= ret_size + (ret_size >= 0);
}
do_estimate_edge_growth (struct cgraph_edge *edge)
{
int size;
+ struct cgraph_node *callee;
+ clause_t clause;
+ VEC (tree, heap) *known_vals;
+ VEC (tree, heap) *known_binfos;
/* When we do caching, use do_estimate_edge_time to populate the entry. */
return size - (size > 0);
}
+ callee = cgraph_function_or_thunk_node (edge->callee, NULL);
+
/* Early inliner runs without caching, go ahead and do the dirty work. */
gcc_checking_assert (edge->inline_failed);
- estimate_callee_size_and_time (edge, true, &size, NULL);
- gcc_checking_assert (edge->call_stmt_size);
- return size - edge->call_stmt_size;
+ evaluate_properties_for_edge (edge, true,
+ &clause, &known_vals, &known_binfos);
+ estimate_node_size_and_time (callee, clause, known_vals, known_binfos,
+ &size, NULL, NULL);
+ VEC_free (tree, heap, known_vals);
+ VEC_free (tree, heap, known_binfos);
+ gcc_checking_assert (inline_edge_summary (edge)->call_stmt_size);
+ return size - inline_edge_summary (edge)->call_stmt_size;
}
estimate_time_after_inlining (struct cgraph_node *node,
struct cgraph_edge *edge)
{
- gcov_type time = inline_summary (node)->time + estimate_edge_time (edge);
- if (time < 0)
- time = 0;
- if (time > MAX_TIME)
- time = MAX_TIME;
- return time;
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+ if (!es->predicate || !false_predicate_p (es->predicate))
+ {
+ gcov_type time = inline_summary (node)->time + estimate_edge_time (edge);
+ if (time < 0)
+ time = 0;
+ if (time > MAX_TIME)
+ time = MAX_TIME;
+ return time;
+ }
+ return inline_summary (node)->time;
}
estimate_size_after_inlining (struct cgraph_node *node,
struct cgraph_edge *edge)
{
- int size = inline_summary (node)->size + estimate_edge_growth (edge);
- gcc_assert (size >= 0);
- return size;
+ struct inline_edge_summary *es = inline_edge_summary (edge);
+ if (!es->predicate || !false_predicate_p (es->predicate))
+ {
+ int size = inline_summary (node)->size + estimate_edge_growth (edge);
+ gcc_assert (size >= 0);
+ return size;
+ }
+ return inline_summary (node)->size;
}
-/* Estimate the growth caused by inlining NODE into all callees. */
+struct growth_data
+{
+ bool self_recursive;
+ int growth;
+};
-int
-do_estimate_growth (struct cgraph_node *node)
+
+/* Worker for do_estimate_growth. Collect growth for all callers. */
+
+static bool
+do_estimate_growth_1 (struct cgraph_node *node, void *data)
{
- int growth = 0;
struct cgraph_edge *e;
- bool self_recursive = false;
- struct inline_summary *info = inline_summary (node);
+ struct growth_data *d = (struct growth_data *) data;
for (e = node->callers; e; e = e->next_caller)
{
if (e->caller == node
|| (e->caller->global.inlined_to
&& e->caller->global.inlined_to == node))
- self_recursive = true;
- growth += estimate_edge_growth (e);
+ d->self_recursive = true;
+ d->growth += estimate_edge_growth (e);
}
-
+ return false;
+}
+
+
+/* Estimate the growth caused by inlining NODE into all callees. */
+
+int
+do_estimate_growth (struct cgraph_node *node)
+{
+ struct growth_data d = {0, false};
+ struct inline_summary *info = inline_summary (node);
+
+ cgraph_for_node_and_aliases (node, do_estimate_growth_1, &d, true);
/* For self recursive functions the growth estimation really should be
infinity. We don't want to return very large values because the growth
plays various roles in badness computation fractions. Be sure to not
return zero or negative growths. */
- if (self_recursive)
- growth = growth < info->size ? info->size : growth;
+ if (d.self_recursive)
+ d.growth = d.growth < info->size ? info->size : d.growth;
else
{
- if (cgraph_will_be_removed_from_program_if_no_direct_calls (node)
- && !DECL_EXTERNAL (node->decl))
- growth -= info->size;
- /* COMDAT functions are very often not shared across multiple units since they
- come from various template instantiations. Take this into account. */
+ if (!DECL_EXTERNAL (node->decl)
+ && cgraph_will_be_removed_from_program_if_no_direct_calls (node))
+ d.growth -= info->size;
+ /* COMDAT functions are very often not shared across multiple units
+ since they come from various template instantiations.
+ Take this into account. */
else if (DECL_COMDAT (node->decl)
&& cgraph_can_remove_if_no_direct_calls_p (node))
- growth -= (info->size
- * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY)) + 50) / 100;
+ d.growth -= (info->size
+ * (100 - PARAM_VALUE (PARAM_COMDAT_SHARING_PROBABILITY))
+ + 50) / 100;
}
if (node_growth_cache)
{
if ((int)VEC_length (int, node_growth_cache) <= node->uid)
VEC_safe_grow_cleared (int, heap, node_growth_cache, cgraph_max_uid);
- VEC_replace (int, node_growth_cache, node->uid, growth + (growth >= 0));
+ VEC_replace (int, node_growth_cache, node->uid,
+ d.growth + (d.growth >= 0));
}
- return growth;
+ return d.growth;
}
if (dump_file)
fprintf (dump_file, "\nAnalyzing function: %s/%u\n",
cgraph_node_name (node), node->uid);
- /* FIXME: We should remove the optimize check after we ensure we never run
- IPA passes when not optimizing. */
- if (flag_indirect_inlining && optimize)
+ if (optimize && !node->thunk.thunk_p)
inline_indirect_intraprocedural_analysis (node);
compute_inline_parameters (node, false);
function_insertion_hook_holder =
cgraph_add_function_insertion_hook (&add_new_function, NULL);
- if (flag_indirect_inlining)
- ipa_register_cgraph_hooks ();
+ ipa_register_cgraph_hooks ();
+ inline_free_summary ();
- for (node = cgraph_nodes; node; node = node->next)
- if (node->analyzed)
+ FOR_EACH_DEFINED_FUNCTION (node)
+ if (!node->alias)
inline_analyze_function (node);
}
+/* Read predicate from IB. */
+
+static struct predicate
+read_predicate (struct lto_input_block *ib)
+{
+ struct predicate out;
+ clause_t clause;
+ int k = 0;
+
+ do
+ {
+ gcc_assert (k <= MAX_CLAUSES);
+ clause = out.clause[k++] = streamer_read_uhwi (ib);
+ }
+ while (clause);
+
+ /* Zero-initialize the remaining clauses in OUT. */
+ while (k <= MAX_CLAUSES)
+ out.clause[k++] = 0;
+
+ return out;
+}
+
+
+/* Write inline summary for edge E to OB. */
+
+static void
+read_inline_edge_summary (struct lto_input_block *ib, struct cgraph_edge *e)
+{
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ struct predicate p;
+ int length, i;
+
+ es->call_stmt_size = streamer_read_uhwi (ib);
+ es->call_stmt_time = streamer_read_uhwi (ib);
+ es->loop_depth = streamer_read_uhwi (ib);
+ p = read_predicate (ib);
+ edge_set_predicate (e, &p);
+ length = streamer_read_uhwi (ib);
+ if (length)
+ {
+ VEC_safe_grow_cleared (inline_param_summary_t, heap, es->param, length);
+ for (i = 0; i < length; i++)
+ VEC_index (inline_param_summary_t, es->param, i)->change_prob
+ = streamer_read_uhwi (ib);
+ }
+}
+
+
/* Stream in inline summaries from the section. */
static void
{
const struct lto_function_header *header =
(const struct lto_function_header *) data;
- const int32_t cfg_offset = sizeof (struct lto_function_header);
- const int32_t main_offset = cfg_offset + header->cfg_size;
- const int32_t string_offset = main_offset + header->main_size;
+ const int cfg_offset = sizeof (struct lto_function_header);
+ const int main_offset = cfg_offset + header->cfg_size;
+ const int string_offset = main_offset + header->main_size;
struct data_in *data_in;
struct lto_input_block ib;
unsigned int i, count2, j;
data_in =
lto_data_in_create (file_data, (const char *) data + string_offset,
header->string_size, NULL);
- f_count = lto_input_uleb128 (&ib);
+ f_count = streamer_read_uhwi (&ib);
for (i = 0; i < f_count; i++)
{
unsigned int index;
struct inline_summary *info;
lto_cgraph_encoder_t encoder;
struct bitpack_d bp;
+ struct cgraph_edge *e;
- index = lto_input_uleb128 (&ib);
+ index = streamer_read_uhwi (&ib);
encoder = file_data->cgraph_node_encoder;
node = lto_cgraph_encoder_deref (encoder, index);
info = inline_summary (node);
info->estimated_stack_size
- = info->estimated_self_stack_size = lto_input_uleb128 (&ib);
- info->size = info->self_size = lto_input_uleb128 (&ib);
- info->time = info->self_time = lto_input_uleb128 (&ib);
+ = info->estimated_self_stack_size = streamer_read_uhwi (&ib);
+ info->size = info->self_size = streamer_read_uhwi (&ib);
+ info->time = info->self_time = streamer_read_uhwi (&ib);
- bp = lto_input_bitpack (&ib);
+ bp = streamer_read_bitpack (&ib);
info->inlinable = bp_unpack_value (&bp, 1);
- info->versionable = bp_unpack_value (&bp, 1);
- count2 = lto_input_uleb128 (&ib);
+ count2 = streamer_read_uhwi (&ib);
gcc_assert (!info->conds);
for (j = 0; j < count2; j++)
{
struct condition c;
- c.operand_num = lto_input_uleb128 (&ib);
- c.code = (enum tree_code) lto_input_uleb128 (&ib);
- c.val = lto_input_tree (&ib, data_in);
+ c.operand_num = streamer_read_uhwi (&ib);
+ c.code = (enum tree_code) streamer_read_uhwi (&ib);
+ c.val = stream_read_tree (&ib, data_in);
VEC_safe_push (condition, gc, info->conds, &c);
}
- count2 = lto_input_uleb128 (&ib);
+ count2 = streamer_read_uhwi (&ib);
gcc_assert (!info->entry);
for (j = 0; j < count2; j++)
{
struct size_time_entry e;
- clause_t clause;
- int k = 0;
- e.size = lto_input_uleb128 (&ib);
- e.time = lto_input_uleb128 (&ib);
- do
- {
- clause = e.predicate.clause[k++] = lto_input_uleb128 (&ib);
- }
- while (clause);
+ e.size = streamer_read_uhwi (&ib);
+ e.time = streamer_read_uhwi (&ib);
+ e.predicate = read_predicate (&ib);
VEC_safe_push (size_time_entry, gc, info->entry, &e);
}
+ for (e = node->callees; e; e = e->next_callee)
+ read_inline_edge_summary (&ib, e);
+ for (e = node->indirect_calls; e; e = e->next_callee)
+ read_inline_edge_summary (&ib, e);
}
lto_free_section_data (file_data, LTO_section_inline_summary, NULL, data,
while ((file_data = file_data_vec[j++]))
{
size_t len;
- const char *data = lto_get_section_data (file_data, LTO_section_inline_summary, NULL, &len);
+ const char *data = lto_get_section_data (file_data,
+ LTO_section_inline_summary,
+ NULL, &len);
if (data)
inline_read_section (file_data, data, len);
else
- /* Fatal error here. We do not want to support compiling ltrans units with
- different version of compiler or different flags than the WPA unit, so
- this should never happen. */
+ /* Fatal error here. We do not want to support compiling ltrans units
+ with different version of compiler or different flags than the WPA
+ unit, so this should never happen. */
fatal_error ("ipa inline summary is missing in input file");
}
- if (flag_indirect_inlining)
+ if (optimize)
{
ipa_register_cgraph_hooks ();
if (!flag_ipa_cp)
}
+/* Write predicate P to OB. */
+
+static void
+write_predicate (struct output_block *ob, struct predicate *p)
+{
+ int j;
+ if (p)
+ for (j = 0; p->clause[j]; j++)
+ {
+ gcc_assert (j < MAX_CLAUSES);
+ streamer_write_uhwi (ob, p->clause[j]);
+ }
+ streamer_write_uhwi (ob, 0);
+}
+
+
+/* Write inline summary for edge E to OB. */
+
+static void
+write_inline_edge_summary (struct output_block *ob, struct cgraph_edge *e)
+{
+ struct inline_edge_summary *es = inline_edge_summary (e);
+ int i;
+
+ streamer_write_uhwi (ob, es->call_stmt_size);
+ streamer_write_uhwi (ob, es->call_stmt_time);
+ streamer_write_uhwi (ob, es->loop_depth);
+ write_predicate (ob, es->predicate);
+ streamer_write_uhwi (ob, VEC_length (inline_param_summary_t, es->param));
+ for (i = 0; i < (int)VEC_length (inline_param_summary_t, es->param); i++)
+ streamer_write_uhwi (ob, VEC_index (inline_param_summary_t,
+ es->param, i)->change_prob);
+}
+
+
/* Write inline summary for node in SET.
Jump functions are shared among ipa-cp and inliner, so when ipa-cp is
active, we don't need to write them twice. */
for (i = 0; i < lto_cgraph_encoder_size (encoder); i++)
if (lto_cgraph_encoder_deref (encoder, i)->analyzed)
count++;
- lto_output_uleb128_stream (ob->main_stream, count);
+ streamer_write_uhwi (ob, count);
for (i = 0; i < lto_cgraph_encoder_size (encoder); i++)
{
{
struct inline_summary *info = inline_summary (node);
struct bitpack_d bp;
+ struct cgraph_edge *edge;
int i;
size_time_entry *e;
struct condition *c;
-
-
- lto_output_uleb128_stream (ob->main_stream,
- lto_cgraph_encoder_encode (encoder, node));
- lto_output_sleb128_stream (ob->main_stream,
- info->estimated_self_stack_size);
- lto_output_sleb128_stream (ob->main_stream,
- info->self_size);
- lto_output_sleb128_stream (ob->main_stream,
- info->self_time);
+
+ streamer_write_uhwi (ob, lto_cgraph_encoder_encode (encoder, node));
+ streamer_write_hwi (ob, info->estimated_self_stack_size);
+ streamer_write_hwi (ob, info->self_size);
+ streamer_write_hwi (ob, info->self_time);
bp = bitpack_create (ob->main_stream);
bp_pack_value (&bp, info->inlinable, 1);
- bp_pack_value (&bp, info->versionable, 1);
- lto_output_bitpack (&bp);
- lto_output_uleb128_stream (ob->main_stream,
- VEC_length (condition, info->conds));
+ streamer_write_bitpack (&bp);
+ streamer_write_uhwi (ob, VEC_length (condition, info->conds));
for (i = 0; VEC_iterate (condition, info->conds, i, c); i++)
{
- lto_output_uleb128_stream (ob->main_stream,
- c->operand_num);
- lto_output_uleb128_stream (ob->main_stream,
- c->code);
- lto_output_tree (ob, c->val, true);
+ streamer_write_uhwi (ob, c->operand_num);
+ streamer_write_uhwi (ob, c->code);
+ stream_write_tree (ob, c->val, true);
}
- lto_output_uleb128_stream (ob->main_stream,
- VEC_length (size_time_entry, info->entry));
+ streamer_write_uhwi (ob, VEC_length (size_time_entry, info->entry));
for (i = 0;
VEC_iterate (size_time_entry, info->entry, i, e);
i++)
{
- int j;
- lto_output_uleb128_stream (ob->main_stream,
- e->size);
- lto_output_uleb128_stream (ob->main_stream,
- e->time);
- for (j = 0; e->predicate.clause[j]; j++)
- lto_output_uleb128_stream (ob->main_stream,
- e->predicate.clause[j]);
- lto_output_uleb128_stream (ob->main_stream, 0);
+ streamer_write_uhwi (ob, e->size);
+ streamer_write_uhwi (ob, e->time);
+ write_predicate (ob, &e->predicate);
}
+ for (edge = node->callees; edge; edge = edge->next_callee)
+ write_inline_edge_summary (ob, edge);
+ for (edge = node->indirect_calls; edge; edge = edge->next_callee)
+ write_inline_edge_summary (ob, edge);
}
}
- lto_output_1_stream (ob->main_stream, 0);
+ streamer_write_char_stream (ob->main_stream, 0);
produce_asm (ob, NULL);
destroy_output_block (ob);
- if (flag_indirect_inlining && !flag_ipa_cp)
+ if (optimize && !flag_ipa_cp)
ipa_prop_write_jump_functions (set);
}
void
inline_free_summary (void)
{
+ struct cgraph_node *node;
+ FOR_EACH_DEFINED_FUNCTION (node)
+ reset_inline_summary (node);
if (function_insertion_hook_holder)
cgraph_remove_function_insertion_hook (function_insertion_hook_holder);
function_insertion_hook_holder = NULL;
if (node_removal_hook_holder)
cgraph_remove_node_removal_hook (node_removal_hook_holder);
node_removal_hook_holder = NULL;
+ if (edge_removal_hook_holder)
+ cgraph_remove_edge_removal_hook (edge_removal_hook_holder);
+ edge_removal_hook_holder = NULL;
if (node_duplication_hook_holder)
cgraph_remove_node_duplication_hook (node_duplication_hook_holder);
node_duplication_hook_holder = NULL;
+ if (edge_duplication_hook_holder)
+ cgraph_remove_edge_duplication_hook (edge_duplication_hook_holder);
+ edge_duplication_hook_holder = NULL;
VEC_free (inline_summary_t, gc, inline_summary_vec);
inline_summary_vec = NULL;
+ VEC_free (inline_edge_summary_t, heap, inline_edge_summary_vec);
+ inline_edge_summary_vec = NULL;
+ if (edge_predicate_pool)
+ free_alloc_pool (edge_predicate_pool);
+ edge_predicate_pool = 0;
}
OSDN Git Service
