Hi,
c-ray has regressed somewhat. This is largely caused by bad luck - better
early optimizations results in different inling order which makes big_speedup
hint to barely not match.
This patch decreases big_speedup threshold and slightly improves the estimation
of speedups by considering loop structure when predicting loop invariant motion.
Bootstrapped/regtested x86_64-linux, will commit it in two parts so performance
consequenes are better tracked.
Honza
PR ipa/79224
* params.def (inline-min-speedup) Change from 10 to 8.
* ipa-inline-analysis.c (get_minimal_bb): New function.
(record_modified): Use it.
(remap_edge_change_prob): Handle also ancestor functions.
Index: params.def
===================================================================
--- params.def (revision 245307)
+++ params.def (working copy)
@@ -52,7 +52,7 @@ DEFPARAM (PARAM_PREDICTABLE_BRANCH_OUTCO
DEFPARAM (PARAM_INLINE_MIN_SPEEDUP,
"inline-min-speedup",
"The minimal estimated speedup allowing inliner to ignore
inline-insns-single and inline-insns-auto.",
- 10, 0, 0)
+ 8, 0, 0)
/* The single function inlining limit. This is the maximum size
of a function counted in internal gcc instructions (not in
Index: ipa-inline-analysis.c
===================================================================
--- ipa-inline-analysis.c (revision 245307)
+++ ipa-inline-analysis.c (working copy)
@@ -2151,6 +2151,23 @@ struct record_modified_bb_info
gimple *stmt;
};
+/* Value is initialized in INIT_BB and used in USE_BB. We want to copute
+ probability how often it changes between USE_BB.
+ INIT_BB->frequency/USE_BB->frequency is an estimate, but if INIT_BB
+ is in different loop nest, we can do better.
+ This is all just estimate. In theory we look for minimal cut separating
+ INIT_BB and USE_BB, but we only want to anticipate loop invariant motion
+ anyway. */
+
+static basic_block
+get_minimal_bb (basic_block init_bb, basic_block use_bb)
+{
+ struct loop *l = find_common_loop (init_bb->loop_father,
use_bb->loop_father);
+ if (l && l->header->frequency < init_bb->frequency)
+ return l->header;
+ return init_bb;
+}
+
/* Callback of walk_aliased_vdefs. Records basic blocks where the value may be
set except for info->stmt. */
@@ -2164,7 +2181,9 @@ record_modified (ao_ref *ao ATTRIBUTE_UN
bitmap_set_bit (info->bb_set,
SSA_NAME_IS_DEFAULT_DEF (vdef)
? ENTRY_BLOCK_PTR_FOR_FN (cfun)->index
- : gimple_bb (SSA_NAME_DEF_STMT (vdef))->index);
+ : get_minimal_bb
+ (gimple_bb (SSA_NAME_DEF_STMT (vdef)),
+ gimple_bb (info->stmt))->index);
return false;
}
@@ -2206,7 +2225,9 @@ param_change_prob (gimple *stmt, int i)
if (SSA_NAME_IS_DEFAULT_DEF (base))
init_freq = ENTRY_BLOCK_PTR_FOR_FN (cfun)->frequency;
else
- init_freq = gimple_bb (SSA_NAME_DEF_STMT (base))->frequency;
+ init_freq = get_minimal_bb
+ (gimple_bb (SSA_NAME_DEF_STMT (base)),
+ gimple_bb (stmt))->frequency;
if (!init_freq)
init_freq = 1;
@@ -3521,18 +3542,22 @@ remap_edge_change_prob (struct cgraph_ed
{
struct ipa_jump_func *jfunc = ipa_get_ith_jump_func (args, i);
if (jfunc->type == IPA_JF_PASS_THROUGH
- && (ipa_get_jf_pass_through_formal_id (jfunc)
- < (int) inlined_es->param.length ()))
+ || jfunc->type == IPA_JF_ANCESTOR)
{
- int jf_formal_id = ipa_get_jf_pass_through_formal_id (jfunc);
- int prob1 = es->param[i].change_prob;
- int prob2 = inlined_es->param[jf_formal_id].change_prob;
- int prob = combine_probabilities (prob1, prob2);
+ int id = jfunc->type == IPA_JF_PASS_THROUGH
+ ? ipa_get_jf_pass_through_formal_id (jfunc)
+ : ipa_get_jf_ancestor_formal_id (jfunc);
+ if (id < (int) inlined_es->param.length ())
+ {
+ int prob1 = es->param[i].change_prob;
+ int prob2 = inlined_es->param[id].change_prob;
+ int prob = combine_probabilities (prob1, prob2);
- if (prob1 && prob2 && !prob)
- prob = 1;
+ if (prob1 && prob2 && !prob)
+ prob = 1;
- es->param[i].change_prob = prob;
+ es->param[i].change_prob = prob;
+ }
}
}
}
