ipa-cp: Multiple sweeps over the call-graph in the decision stage

Currently, IPA-CP makes only one sweep in the decision stage over the
call-graph, meaning that some clonin , even if relatively cheap, may
not be performed because the pass runs out of the overall growth
budget before it gets to evaluating it.  By making more (three by
default, but configurable with a parameter) sweeps over the call-graph
with progressivelly stricter cost limits, the more benefitial
candidates will have a better chance to be cloned before others.

gcc/ChangeLog:

2025-07-08  Martin Jambor  <mjambor@suse.cz>

	* params.opt (param_ipa_cp_sweeps): New.
	* doc/invoke.texi (ipa-cp-sweeps): New.
	* ipa-cp.cc (max_number_sweeps): New.
	(get_max_overall_size): New parameter cur_sweep, use it and the total
	number of sweeps from the NODE to calculate the result too.
	(ipcp_propagate_stage): Get the maximum number of sweeps specified in
	the corresponding parameter of any possibly affected node.
	(good_cloning_opportunity_p): Add parameter cur_sweep, adjust the
	threshold according to it.
	(decide_about_value): New parameter cur_sweep, pass it to
	get_max_overall_size and to good_cloning_opportunity_p.
	(decide_whether_version_node): New parameter cur_sweep, pass it to
	decide_about_value and get_max_overall_size.  Make sure the node is
	not dead.
	(ipcp_decision_stage): Make multiple sweeps over the call-graph.
This commit is contained in:
Martin Jambor
2026-01-07 11:53:15 +01:00
committed by Martin Jambor
parent f265e8e467
commit cfc746c211
3 changed files with 77 additions and 35 deletions

View File

@@ -17522,6 +17522,10 @@ are evaluated for cloning.
Percentage penalty functions containing a single call to another
function will receive when they are evaluated for cloning.
@item ipa-cp-sweeps
The number of times the interprocedural constant propagation will traverse
all functions to make cloning decisions.
@item ipa-max-agg-items
IPA-CP is also capable to propagate a number of scalar values passed
in an aggregate. @option{ipa-max-agg-items} controls the maximum

View File

@@ -151,6 +151,10 @@ object_allocator<ipcp_agg_lattice> ipcp_agg_lattice_pool
static long overall_size, orig_overall_size;
/* The maximum number of IPA-CP decision sweeps that any node requested in its
param. */
static int max_number_sweeps;
/* Node name to unique clone suffix number map. */
static hash_map<const char *, unsigned> *clone_num_suffixes;
@@ -3376,12 +3380,14 @@ incorporate_penalties (cgraph_node *node, ipa_node_params *info,
/* Return true if cloning NODE is a good idea, given the estimated TIME_BENEFIT
and SIZE_COST and with the sum of frequencies of incoming edges to the
potential new clone in FREQUENCIES. */
potential new clone in FREQUENCIES. CUR_SWEEP is the number of the current
sweep of IPA-CP over the call-graph in the decision stage. */
static bool
good_cloning_opportunity_p (struct cgraph_node *node, sreal time_benefit,
sreal freq_sum, profile_count count_sum,
int size_cost, bool called_without_ipa_profile)
int size_cost, bool called_without_ipa_profile,
int cur_sweep)
{
gcc_assert (count_sum.ipa () == count_sum);
if (count_sum.quality () == AFDO)
@@ -3396,7 +3402,9 @@ good_cloning_opportunity_p (struct cgraph_node *node, sreal time_benefit,
gcc_assert (size_cost > 0);
ipa_node_params *info = ipa_node_params_sum->get (node);
int num_sweeps = opt_for_fn (node->decl, param_ipa_cp_sweeps);
int eval_threshold = opt_for_fn (node->decl, param_ipa_cp_eval_threshold);
eval_threshold = (eval_threshold * num_sweeps) / cur_sweep;
/* If we know the execution IPA execution counts, we can estimate overall
speedup of the program. */
if (count_sum.nonzero_p ())
@@ -3551,20 +3559,25 @@ perform_estimation_of_a_value (cgraph_node *node,
val->local_size_cost = size;
}
/* Get the overall limit oof growth based on parameters extracted from growth.
it does not really make sense to mix functions with different overall growth
limits but it is possible and if it happens, we do not want to select one
limit at random. */
/* Get the overall limit of growth based on parameters extracted from growth,
and CUR_SWEEP, which is the number of the current sweep of IPA-CP over the
call-graph in the decision stage. It does not really make sense to mix
functions with different overall growth limits or even number of sweeps but
it is possible and if it happens, we do not want to select one limit at
random, so get the limits from NODE. */
static long
get_max_overall_size (cgraph_node *node)
get_max_overall_size (cgraph_node *node, int cur_sweep)
{
long max_new_size = orig_overall_size;
long large_unit = opt_for_fn (node->decl, param_ipa_cp_large_unit_insns);
if (max_new_size < large_unit)
max_new_size = large_unit;
int num_sweeps = opt_for_fn (node->decl, param_ipa_cp_sweeps);
gcc_assert (cur_sweep <= num_sweeps);
int unit_growth = opt_for_fn (node->decl, param_ipa_cp_unit_growth);
max_new_size += max_new_size * unit_growth / 100 + 1;
max_new_size += ((max_new_size * unit_growth * cur_sweep)
/ num_sweeps) / 100 + 1;
return max_new_size;
}
@@ -4022,6 +4035,10 @@ ipcp_propagate_stage (class ipa_topo_info *topo)
unsigned nlattices = ipa_get_param_count (info);
info->lattices.safe_grow_cleared (nlattices, true);
initialize_node_lattices (node);
int num_sweeps = opt_for_fn (node->decl, param_ipa_cp_sweeps);
if (max_number_sweeps < num_sweeps)
max_number_sweeps = num_sweeps;
}
ipa_size_summary *s = ipa_size_summaries->get (node);
if (node->definition && !node->alias && s != NULL)
@@ -5783,13 +5800,14 @@ ipcp_val_agg_replacement_ok_p (vec<ipa_argagg_value, va_gc> *,
parameter itself, otherwise it is stored at the given OFFSET of the
parameter. AVALS describes the other already known values. SELF_GEN_CLONES
is a vector which contains clones created for self-recursive calls with an
arithmetic pass-through jump function. */
arithmetic pass-through jump function. CUR_SWEEP is the number of the
current sweep of the call-graph during the decision stage. */
template <typename valtype>
static bool
decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
ipcp_value<valtype> *val, ipa_auto_call_arg_values *avals,
vec<cgraph_node *> *self_gen_clones)
vec<cgraph_node *> *self_gen_clones, int cur_sweep)
{
int caller_count;
sreal freq_sum;
@@ -5802,7 +5820,8 @@ decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
perhaps_add_new_callers (node, val);
return false;
}
else if (val->local_size_cost + overall_size > get_max_overall_size (node))
else if (val->local_size_cost + overall_size
> get_max_overall_size (node, cur_sweep))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, " Ignoring candidate value because "
@@ -5850,10 +5869,10 @@ decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
if (!good_cloning_opportunity_p (node, val->local_time_benefit,
freq_sum, count_sum,
val->local_size_cost,
called_without_ipa_profile)
called_without_ipa_profile, cur_sweep)
&& !good_cloning_opportunity_p (node, val->prop_time_benefit,
freq_sum, count_sum, val->prop_size_cost,
called_without_ipa_profile))
called_without_ipa_profile, cur_sweep))
return false;
if (dump_file)
@@ -5914,16 +5933,18 @@ ipa_range_contains_p (const vrange &r, tree val)
return r.contains_p (val);
}
/* Decide whether and what specialized clones of NODE should be created. */
/* Decide whether and what specialized clones of NODE should be created.
CUR_SWEEP is the number of the current sweep of the call-graph during the
decision stage. */
static bool
decide_whether_version_node (struct cgraph_node *node)
decide_whether_version_node (struct cgraph_node *node, int cur_sweep)
{
ipa_node_params *info = ipa_node_params_sum->get (node);
int i, count = ipa_get_param_count (info);
bool ret = false;
if (count == 0)
if (info->node_dead || count == 0)
return false;
if (dump_file && (dump_flags & TDF_DETAILS))
@@ -5974,7 +5995,7 @@ decide_whether_version_node (struct cgraph_node *node)
continue;
}
ret |= decide_about_value (node, i, -1, val, &avals,
&self_gen_clones);
&self_gen_clones, cur_sweep);
}
}
@@ -5990,7 +6011,7 @@ decide_whether_version_node (struct cgraph_node *node)
|| !aglat->is_single_const ()))
for (val = aglat->values; val; val = val->next)
ret |= decide_about_value (node, i, aglat->offset, val, &avals,
&self_gen_clones);
&self_gen_clones, cur_sweep);
}
if (!ctxlat->bottom
@@ -5999,7 +6020,7 @@ decide_whether_version_node (struct cgraph_node *node)
ipcp_value<ipa_polymorphic_call_context> *val;
for (val = ctxlat->values; val; val = val->next)
ret |= decide_about_value (node, i, -1, val, &avals,
&self_gen_clones);
&self_gen_clones, cur_sweep);
}
}
@@ -6049,9 +6070,10 @@ decide_whether_version_node (struct cgraph_node *node)
}
else if (good_cloning_opportunity_p (node, time, stats.freq_sum,
stats.count_sum, size,
stats.called_without_ipa_profile))
stats.called_without_ipa_profile,
cur_sweep))
{
if (size + overall_size <= get_max_overall_size (node))
if (size + overall_size <= get_max_overall_size (node, cur_sweep))
{
if (!dbg_cnt (ipa_cp_values))
return ret;
@@ -6277,26 +6299,38 @@ ipcp_decision_stage (class ipa_topo_info *topo)
int i;
if (dump_file)
fprintf (dump_file, "\nIPA decision stage:\n\n");
fprintf (dump_file, "\nIPA decision stage (%i sweeps):\n",
max_number_sweeps);
for (i = topo->nnodes - 1; i >= 0; i--)
for (int cur_sweep = 1; cur_sweep <= max_number_sweeps; cur_sweep++)
{
struct cgraph_node *node = topo->order[i];
bool change = false, iterate = true;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "\nIPA decision sweep number %i (out of %i):\n",
cur_sweep, max_number_sweeps);
while (iterate)
for (i = topo->nnodes - 1; i >= 0; i--)
{
struct cgraph_node *v;
iterate = false;
for (v = node; v; v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
if (v->has_gimple_body_p ()
&& ipcp_versionable_function_p (v))
iterate |= decide_whether_version_node (v);
struct cgraph_node *node = topo->order[i];
bool change = false, iterate = true;
change |= iterate;
while (iterate)
{
struct cgraph_node *v;
iterate = false;
for (v = node;
v;
v = ((struct ipa_dfs_info *) v->aux)->next_cycle)
if (v->has_gimple_body_p ()
&& ipcp_versionable_function_p (v)
&& (cur_sweep
<= opt_for_fn (node->decl, param_ipa_cp_sweeps)))
iterate |= decide_whether_version_node (v, cur_sweep);
change |= iterate;
}
if (change)
identify_dead_nodes (node);
}
if (change)
identify_dead_nodes (node);
}
/* Currently, the primary use of callback edges is constant propagation.

View File

@@ -273,6 +273,10 @@ Percentage penalty the recursive functions will receive when they are evaluated
Common Joined UInteger Var(param_ipa_cp_single_call_penalty) Init(15) IntegerRange(0, 100) Param Optimization
Percentage penalty functions containing a single call to another function will receive when they are evaluated for cloning.
-param=ipa-cp-sweeps=
Common Joined UInteger Var(param_ipa_cp_sweeps) Init(3) IntegerRange(1, 100) Param Optimization
The number of times the interprocedural constant propagation will traverse all functions to make cloning decisions.
-param=ipa-cp-unit-growth=
Common Joined UInteger Var(param_ipa_cp_unit_growth) Init(10) Param Optimization
How much can given compilation unit grow because of the interprocedural constant propagation (in percent).