From: "Martin Liška" <mliska@suse.cz>
To: gcc-patches@gcc.gnu.org
Cc: Jan Hubicka <hubicka@ucw.cz>
Subject: Re: [PATCH] Improve profile handling in switch lowering.
Date: Thu, 24 Mar 2022 11:22:13 +0100 [thread overview]
Message-ID: <b8fbd09b-9e2a-7988-98f2-bcb295f464e5@suse.cz> (raw)
In-Reply-To: <8e839057-51b4-a036-c9f2-8049e41d5814@suse.cz>
PING^2
On 3/4/22 14:47, Martin Liška wrote:
> PING^1
>
> On 1/26/22 12:11, Martin Liška wrote:
>> Hello.
>>
>> Right now, switch lowering does not update basic_block::count values
>> so that they are uninitiliazed. Moreover, I've updated probability scaling
>> when a more complex expansion happens. There are still some situations where
>> the profile is a bit imprecise, but the patch improves rapidly the current situation.
>>
>> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
>>
>> Ready to be installed?
>> Thanks,
>> Martin
>>
>> PR tree-optimization/101301
>> PR tree-optimization/103680
>>
>> gcc/ChangeLog:
>>
>> * tree-switch-conversion.cc (bit_test_cluster::emit):
>> Handle correctly remaining probability.
>> (switch_decision_tree::try_switch_expansion): Fix BB's count
>> where a cluster expansion happens.
>> (switch_decision_tree::emit_cmp_and_jump_insns): Fill up also
>> BB count.
>> (switch_decision_tree::do_jump_if_equal): Likewise.
>> (switch_decision_tree::emit_case_nodes): Handle special case
>> for BT expansion which can also fallback to a default BB.
>> * tree-switch-conversion.h (cluster::cluster): Add
>> m_default_prob probability.
>> ---
>> gcc/tree-switch-conversion.cc | 51 ++++++++++++++++++++++++-----------
>> gcc/tree-switch-conversion.h | 8 +++++-
>> 2 files changed, 43 insertions(+), 16 deletions(-)
>>
>> diff --git a/gcc/tree-switch-conversion.cc b/gcc/tree-switch-conversion.cc
>> index 670397c87e4..d6679e8dee3 100644
>> --- a/gcc/tree-switch-conversion.cc
>> +++ b/gcc/tree-switch-conversion.cc
>> @@ -1538,10 +1538,12 @@ bit_test_cluster::emit (tree index_expr, tree index_type,
>> test[k].target_bb = n->m_case_bb;
>> test[k].label = n->m_case_label_expr;
>> test[k].bits = 0;
>> + test[k].prob = profile_probability::never ();
>> count++;
>> }
>>
>> test[k].bits += n->get_range (n->get_low (), n->get_high ());
>> + test[k].prob += n->m_prob;
>>
>> lo = tree_to_uhwi (int_const_binop (MINUS_EXPR, n->get_low (), minval));
>> if (n->get_high () == NULL_TREE)
>> @@ -1629,6 +1631,11 @@ bit_test_cluster::emit (tree index_expr, tree index_type,
>> /*simple=*/true, NULL_TREE,
>> /*before=*/true, GSI_SAME_STMT);
>>
>> + profile_probability subtree_prob = m_subtree_prob;
>> + profile_probability default_prob = m_default_prob;
>> + if (!default_prob.initialized_p ())
>> + default_prob = m_subtree_prob.invert ();
>> +
>> if (m_handles_entire_switch && entry_test_needed)
>> {
>> tree range = int_const_binop (MINUS_EXPR, maxval, minval);
>> @@ -1639,9 +1646,10 @@ bit_test_cluster::emit (tree index_expr, tree index_type,
>> /*simple=*/true, NULL_TREE,
>> /*before=*/true, GSI_SAME_STMT);
>> tmp = fold_build2 (GT_EXPR, boolean_type_node, idx, range);
>> + default_prob = default_prob.apply_scale (1, 2);
>> basic_block new_bb
>> = hoist_edge_and_branch_if_true (&gsi, tmp, default_bb,
>> - profile_probability::unlikely ());
>> + default_prob);
>> gsi = gsi_last_bb (new_bb);
>> }
>>
>> @@ -1662,14 +1670,12 @@ bit_test_cluster::emit (tree index_expr, tree index_type,
>> else
>> csui = tmp;
>>
>> - profile_probability prob = profile_probability::always ();
>> -
>> /* for each unique set of cases:
>> if (const & csui) goto target */
>> for (k = 0; k < count; k++)
>> {
>> - prob = profile_probability::always ().apply_scale (test[k].bits,
>> - bt_range);
>> + profile_probability prob = test[k].prob / (subtree_prob + default_prob);
>> + subtree_prob -= test[k].prob;
>> bt_range -= test[k].bits;
>> tmp = wide_int_to_tree (word_type_node, test[k].mask);
>> tmp = fold_build2 (BIT_AND_EXPR, word_type_node, csui, tmp);
>> @@ -1908,9 +1914,13 @@ switch_decision_tree::try_switch_expansion (vec<cluster *> &clusters)
>> /* Emit cluster-specific switch handling. */
>> for (unsigned i = 0; i < clusters.length (); i++)
>> if (clusters[i]->get_type () != SIMPLE_CASE)
>> - clusters[i]->emit (index_expr, index_type,
>> - gimple_switch_default_label (m_switch),
>> - m_default_bb, gimple_location (m_switch));
>> + {
>> + edge e = single_pred_edge (clusters[i]->m_case_bb);
>> + e->dest->count = e->src->count.apply_probability (e->probability);
>> + clusters[i]->emit (index_expr, index_type,
>> + gimple_switch_default_label (m_switch),
>> + m_default_bb, gimple_location (m_switch));
>> + }
>> }
>>
>> fix_phi_operands_for_edges ();
>> @@ -2162,6 +2172,7 @@ switch_decision_tree::emit_cmp_and_jump_insns (basic_block bb, tree op0,
>> edge false_edge = split_block (bb, cond);
>> false_edge->flags = EDGE_FALSE_VALUE;
>> false_edge->probability = prob.invert ();
>> + false_edge->dest->count = bb->count.apply_probability (prob.invert ());
>>
>> edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE);
>> true_edge->probability = prob;
>> @@ -2192,6 +2203,7 @@ switch_decision_tree::do_jump_if_equal (basic_block bb, tree op0, tree op1,
>> edge false_edge = split_block (bb, cond);
>> false_edge->flags = EDGE_FALSE_VALUE;
>> false_edge->probability = prob.invert ();
>> + false_edge->dest->count = bb->count.apply_probability (prob.invert ());
>>
>> edge true_edge = make_edge (bb, label_bb, EDGE_TRUE_VALUE);
>> true_edge->probability = prob;
>> @@ -2227,7 +2239,7 @@ switch_decision_tree::emit_case_nodes (basic_block bb, tree index,
>> node->m_c->m_case_bb, p, loc);
>> /* Since this case is taken at this point, reduce its weight from
>> subtree_weight. */
>> - node->m_c->m_subtree_prob -= p;
>> + node->m_c->m_subtree_prob -= node->m_c->m_prob;
>>
>> if (node->m_left != NULL && node->m_right != NULL)
>> {
>> @@ -2246,6 +2258,7 @@ switch_decision_tree::emit_case_nodes (basic_block bb, tree index,
>> / (node->m_c->m_subtree_prob + default_prob));
>> bb = do_jump_if_equal (bb, index, node->m_right->m_c->get_low (),
>> node->m_right->m_c->m_case_bb, p, loc);
>> + node->m_c->m_subtree_prob -= node->m_right->m_c->m_prob;
>>
>> p = (node->m_left->m_c->m_prob
>> / (node->m_c->m_subtree_prob + default_prob));
>> @@ -2262,6 +2275,7 @@ switch_decision_tree::emit_case_nodes (basic_block bb, tree index,
>> p = ((node->m_right->m_c->m_subtree_prob
>> + default_prob.apply_scale (1, 2))
>> / (node->m_c->m_subtree_prob + default_prob));
>> + test_bb->count = bb->count.apply_probability (p);
>> bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (),
>> GT_EXPR, test_bb, p, loc);
>> default_prob = default_prob.apply_scale (1, 2);
>> @@ -2348,21 +2362,28 @@ switch_decision_tree::emit_case_nodes (basic_block bb, tree index,
>> is the one to branch to. */
>> if (node->has_child () || node->m_c->get_type () != SIMPLE_CASE)
>> {
>> + bool is_bt = node->m_c->get_type () == BIT_TEST;
>> + int parts = is_bt ? 3 : 2;
>> +
>> /* Branch to a label where we will handle it later. */
>> basic_block test_bb = split_edge (single_succ_edge (bb));
>> redirect_edge_succ (single_pred_edge (test_bb),
>> single_succ_edge (bb)->dest);
>>
>> -
>> - profile_probability right_prob = profile_probability::never ();
>> - if (node->m_right)
>> - right_prob = node->m_right->m_c->m_subtree_prob;
>> - p = ((right_prob + default_prob.apply_scale (1, 2))
>> + profile_probability right_prob = profile_probability::never ();
>> + if (node->m_right)
>> + right_prob = node->m_right->m_c->m_subtree_prob;
>> + p = ((right_prob + default_prob.apply_scale (1, parts))
>> / (node->m_c->m_subtree_prob + default_prob));
>> + test_bb->count = bb->count.apply_probability (p);
>>
>> bb = emit_cmp_and_jump_insns (bb, index, node->m_c->get_high (),
>> GT_EXPR, test_bb, p, loc);
>> - default_prob = default_prob.apply_scale (1, 2);
>> +
>> + default_prob = default_prob.apply_scale (1, parts);
>> + node->m_c->m_subtree_prob -= right_prob;
>> + if (is_bt)
>> + node->m_c->m_default_prob = default_prob;
>>
>> /* Value belongs to this node or to the left-hand subtree. */
>> p = node->m_c->m_prob / (node->m_c->m_subtree_prob + default_prob);
>> diff --git a/gcc/tree-switch-conversion.h b/gcc/tree-switch-conversion.h
>> index e969c051a05..5f3f99353ba 100644
>> --- a/gcc/tree-switch-conversion.h
>> +++ b/gcc/tree-switch-conversion.h
>> @@ -102,6 +102,10 @@ public:
>> /* Probability of reaching subtree rooted at this node. */
>> profile_probability m_subtree_prob;
>>
>> + /* Probability of default case when reaching the node.
>> + It is used by bit-test right now. */
>> + profile_probability m_default_prob;
>> +
>> protected:
>> /* Default constructor. */
>> cluster () {}
>> @@ -110,7 +114,8 @@ protected:
>> cluster::cluster (tree case_label_expr, basic_block case_bb,
>> profile_probability prob, profile_probability subtree_prob):
>> m_case_label_expr (case_label_expr), m_case_bb (case_bb), m_prob (prob),
>> - m_subtree_prob (subtree_prob)
>> + m_subtree_prob (subtree_prob),
>> + m_default_prob (profile_probability::uninitialized ())
>> {
>> }
>>
>> @@ -542,6 +547,7 @@ public:
>> basic_block target_bb;
>> tree label;
>> int bits;
>> + profile_probability prob;
>>
>> /* Comparison function for qsort to order bit tests by decreasing
>> probability of execution. */
>
next prev parent reply other threads:[~2022-03-24 10:22 UTC|newest]
Thread overview: 6+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-01-26 11:11 Martin Liška
2022-03-04 13:47 ` Martin Liška
2022-03-24 10:22 ` Martin Liška [this message]
2022-04-07 14:00 ` Martin Liška
2022-11-28 9:53 ` Martin Liška
2022-11-28 17:38 ` Jeff Law
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=b8fbd09b-9e2a-7988-98f2-bcb295f464e5@suse.cz \
--to=mliska@suse.cz \
--cc=gcc-patches@gcc.gnu.org \
--cc=hubicka@ucw.cz \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).