summaryrefslogtreecommitdiff
path: root/0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch
diff options
context:
space:
mode:
Diffstat (limited to '0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch')
-rw-r--r--0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch353
1 files changed, 353 insertions, 0 deletions
diff --git a/0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch b/0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch
new file mode 100644
index 0000000..bbc98c6
--- /dev/null
+++ b/0026-AutoFDO-Enable-discriminator-and-MCF-algorithm-on-Au.patch
@@ -0,0 +1,353 @@
+From eb58d920a95696d8d5a7db9a6d640d4494fb023f Mon Sep 17 00:00:00 2001
+From: liyancheng <412998149@qq.com>
+Date: Tue, 25 Jan 2022 16:57:28 +0800
+Subject: [PATCH 26/28] [AutoFDO] Enable discriminator and MCF algorithm on
+ AutoFDO
+
+1. Support discriminator for distinguishes among several
+ basic blocks that share a common locus, allowing for
+ more accurate autofdo.
+
+2. Using option -fprofile-correction for calling MCF algorithm
+ to smooth non conservative BB counts.
+---
+ gcc/auto-profile.c | 172 ++++++++++++++++++++++++++++++++++++++++++++-
+ gcc/cfghooks.c | 7 ++
+ gcc/ipa-cp.c | 21 ++++++
+ gcc/opts.c | 5 +-
+ gcc/tree-inline.c | 14 ++++
+ 5 files changed, 215 insertions(+), 4 deletions(-)
+
+diff --git a/gcc/auto-profile.c b/gcc/auto-profile.c
+index aced8fca5..e6164b91b 100644
+--- a/gcc/auto-profile.c
++++ b/gcc/auto-profile.c
+@@ -678,6 +678,17 @@ string_table::get_index (const char *name) const
+ if (name == NULL)
+ return -1;
+ string_index_map::const_iterator iter = map_.find (name);
++ /* Function name may be duplicate. Try to distinguish by the
++ #file_name#function_name defined by the autofdo tool chain. */
++ if (iter == map_.end ())
++ {
++ char* file_name = get_original_name (lbasename (dump_base_name));
++ char* file_func_name
++ = concat ("#", file_name, "#", name, NULL);
++ iter = map_.find (file_func_name);
++ free (file_name);
++ free (file_func_name);
++ }
+ if (iter == map_.end ())
+ return -1;
+
+@@ -866,7 +877,7 @@ function_instance::read_function_instance (function_instance_stack *stack,
+
+ for (unsigned i = 0; i < num_pos_counts; i++)
+ {
+- unsigned offset = gcov_read_unsigned () & 0xffff0000;
++ unsigned offset = gcov_read_unsigned ();
+ unsigned num_targets = gcov_read_unsigned ();
+ gcov_type count = gcov_read_counter ();
+ s->pos_counts[offset].count = count;
+@@ -945,6 +956,10 @@ autofdo_source_profile::get_count_info (gimple *stmt, count_info *info) const
+ function_instance *s = get_function_instance_by_inline_stack (stack);
+ if (s == NULL)
+ return false;
++ if (s->get_count_info (stack[0].second + stmt->bb->discriminator, info))
++ {
++ return true;
++ }
+ return s->get_count_info (stack[0].second, info);
+ }
+
+@@ -1583,6 +1598,68 @@ afdo_propagate (bb_set *annotated_bb)
+ }
+ }
+
++/* Process the following scene when the branch probability
++ inversion when do function afdo_propagate (). E.g.
++ BB_NUM (sample count)
++ BB1 (1000)
++ / \
++ BB2 (10) BB3 (0)
++ \ /
++ BB4
++ In afdo_propagate(), count of BB3 is calculated by
++ COUNT (BB3) = 990 (990 = COUNT (BB1) - COUNT (BB2) = 1000 - 10)
++
++ In fact, BB3 may be colder than BB2 by sample count.
++
++ This function allocate source BB count to each succ BB by sample
++ rate, E.g.
++ BB2_COUNT = BB1_COUNT * (BB2_COUNT / (BB2_COUNT + BB3_COUNT)) */
++
++static void
++afdo_preprocess_bb_count ()
++{
++ basic_block bb;
++ FOR_ALL_BB_FN (bb, cfun)
++ {
++ if (bb->count.ipa_p () && EDGE_COUNT (bb->succs) > 1
++ && bb->count > profile_count::zero ().afdo ())
++ {
++ basic_block bb1 = EDGE_SUCC (bb, 0)->dest;
++ basic_block bb2 = EDGE_SUCC (bb, 1)->dest;
++ if (single_succ_p (bb1) && single_succ_p (bb2)
++ && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
++ {
++ gcov_type max_count = 0;
++ gcov_type total_count = 0;
++ edge e;
++ edge_iterator ei;
++ FOR_EACH_EDGE (e, ei, bb->succs)
++ {
++ if (!e->dest->count.ipa_p ())
++ {
++ continue;
++ }
++ max_count = MAX(max_count, e->dest->count.to_gcov_type ());
++ total_count += e->dest->count.to_gcov_type ();
++ }
++ /* Only bb_count > max_count * 2, branch probability will
++ inversion. */
++ if (max_count > 0
++ && bb->count.to_gcov_type () > max_count * 2)
++ {
++ FOR_EACH_EDGE (e, ei, bb->succs)
++ {
++ gcov_type target_count = bb->count.to_gcov_type ()
++ * e->dest->count.to_gcov_type () / total_count;
++ e->dest->count
++ = profile_count::from_gcov_type (target_count).afdo ();
++ }
++ }
++ }
++ }
++ }
++}
++
+ /* Propagate counts on control flow graph and calculate branch
+ probabilities. */
+
+@@ -1608,6 +1685,7 @@ afdo_calculate_branch_prob (bb_set *annotated_bb)
+ }
+
+ afdo_find_equiv_class (annotated_bb);
++ afdo_preprocess_bb_count ();
+ afdo_propagate (annotated_bb);
+
+ FOR_EACH_BB_FN (bb, cfun)
+@@ -1711,6 +1789,82 @@ afdo_vpt_for_early_inline (stmt_set *promoted_stmts)
+ return false;
+ }
+
++/* Preparation before executing MCF algorithm. */
++
++static void
++afdo_init_mcf ()
++{
++ basic_block bb;
++ edge e;
++ edge_iterator ei;
++
++ if (dump_file)
++ {
++ fprintf (dump_file, "\n init calling mcf_smooth_cfg (). \n");
++ }
++
++ /* Step1: when use mcf, BB id must be continous,
++ so we need compact_blocks (). */
++ compact_blocks ();
++
++ /* Step2: allocate memory for MCF input data. */
++ bb_gcov_counts.safe_grow_cleared (cfun->cfg->x_last_basic_block);
++ edge_gcov_counts = new hash_map<edge, gcov_type>;
++
++ /* Step3: init MCF input data from cfg. */
++ FOR_ALL_BB_FN (bb, cfun)
++ {
++ /* Init BB count for MCF. */
++ bb_gcov_count (bb) = bb->count.to_gcov_type ();
++
++ gcov_type total_count = 0;
++ FOR_EACH_EDGE (e, ei, bb->succs)
++ {
++ total_count += e->dest->count.to_gcov_type ();
++ }
++
++ /* If there is no sample in each successor blocks, source
++ BB samples are allocated to each edge by branch static prob. */
++
++ FOR_EACH_EDGE (e, ei, bb->succs)
++ {
++ if (total_count == 0)
++ {
++ edge_gcov_count (e) = e->src->count.to_gcov_type ()
++ * e->probability.to_reg_br_prob_base () / REG_BR_PROB_BASE;
++ }
++ else
++ {
++ edge_gcov_count (e) = e->src->count.to_gcov_type ()
++ * e->dest->count.to_gcov_type () / total_count;
++ }
++ }
++ }
++}
++
++/* Free the resources used by MCF and reset BB count from MCF result,
++ branch probability has been updated in mcf_smooth_cfg (). */
++
++static void
++afdo_process_after_mcf ()
++{
++ basic_block bb;
++ /* Reset BB count from MCF result. */
++ FOR_EACH_BB_FN (bb, cfun)
++ {
++ if (bb_gcov_count (bb))
++ {
++ bb->count
++ = profile_count::from_gcov_type (bb_gcov_count (bb)).afdo ();
++ }
++ }
++
++ /* Clean up MCF resource. */
++ bb_gcov_counts.release ();
++ delete edge_gcov_counts;
++ edge_gcov_counts = NULL;
++}
++
+ /* Annotate auto profile to the control flow graph. Do not annotate value
+ profile for stmts in PROMOTED_STMTS. */
+
+@@ -1762,8 +1916,20 @@ afdo_annotate_cfg (const stmt_set &promoted_stmts)
+ afdo_source_profile->mark_annotated (cfun->function_end_locus);
+ if (max_count > profile_count::zero ())
+ {
+- /* Calculate, propagate count and probability information on CFG. */
+- afdo_calculate_branch_prob (&annotated_bb);
++ /* 1 means -fprofile-correction is enabled manually, and MCF
++ algorithm will be used to calculate count and probability.
++ Otherwise, use the default calculate algorithm. */
++ if (flag_profile_correction == 1)
++ {
++ afdo_init_mcf ();
++ mcf_smooth_cfg ();
++ afdo_process_after_mcf ();
++ }
++ else
++ {
++ /* Calculate, propagate count and probability information on CFG. */
++ afdo_calculate_branch_prob (&annotated_bb);
++ }
+ }
+ update_max_bb_count ();
+ profile_status_for_fn (cfun) = PROFILE_READ;
+diff --git a/gcc/cfghooks.c b/gcc/cfghooks.c
+index ea558b469..4ea490a8a 100644
+--- a/gcc/cfghooks.c
++++ b/gcc/cfghooks.c
+@@ -526,6 +526,9 @@ split_block_1 (basic_block bb, void *i)
+ return NULL;
+
+ new_bb->count = bb->count;
++ /* Copy discriminator from original bb for distinguishes among
++ several basic blocks that share a common locus, allowing for
++ more accurate autofdo. */
+ new_bb->discriminator = bb->discriminator;
+
+ if (dom_info_available_p (CDI_DOMINATORS))
+@@ -1091,6 +1094,10 @@ duplicate_block (basic_block bb, edge e, basic_block after, copy_bb_data *id)
+ move_block_after (new_bb, after);
+
+ new_bb->flags = (bb->flags & ~BB_DUPLICATED);
++ /* Copy discriminator from original bb for distinguishes among
++ several basic blocks that share a common locus, allowing for
++ more accurate autofdo. */
++ new_bb->discriminator = bb->discriminator;
+ FOR_EACH_EDGE (s, ei, bb->succs)
+ {
+ /* Since we are creating edges from a new block to successors
+diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
+index b1f0881bd..c208070c9 100644
+--- a/gcc/ipa-cp.c
++++ b/gcc/ipa-cp.c
+@@ -4365,6 +4365,27 @@ update_profiling_info (struct cgraph_node *orig_node,
+ orig_node_count.dump (dump_file);
+ fprintf (dump_file, "\n");
+ }
++
++ /* When autofdo uses PMU as the sampling unit, the count of
++ cgraph_node->count cannot be obtained directly and will
++ be zero. It using for apply_scale will cause the node
++ count incorrectly overestimated. So set orig_new_node_count
++ equal to orig_node_count, which is same as known error
++ handling. */
++ if (orig_node->count == profile_count::zero ().afdo ()
++ && new_node->count == profile_count::zero ().global0adjusted ())
++ {
++ orig_new_node_count = (orig_sum + new_sum).apply_scale (12, 10);
++
++ if (dump_file)
++ {
++ fprintf (dump_file, " node %s with zero count from afdo ",
++ new_node->dump_name ());
++ fprintf (dump_file, " proceeding by pretending it was ");
++ orig_new_node_count.dump (dump_file);
++ fprintf (dump_file, "\n");
++ }
++ }
+ }
+
+ remainder = orig_node_count.combine_with_ipa_count (orig_node_count.ipa ()
+diff --git a/gcc/opts.c b/gcc/opts.c
+index 642327296..7a39f618b 100644
+--- a/gcc/opts.c
++++ b/gcc/opts.c
+@@ -2606,7 +2606,10 @@ common_handle_option (struct gcc_options *opts,
+ /* FALLTHRU */
+ case OPT_fauto_profile:
+ enable_fdo_optimizations (opts, opts_set, value);
+- SET_OPTION_IF_UNSET (opts, opts_set, flag_profile_correction, value);
++ /* 2 is special and means flag_profile_correction trun on by
++ -fauto-profile. */
++ SET_OPTION_IF_UNSET (opts, opts_set, flag_profile_correction,
++ (value ? 2 : 0));
+ SET_OPTION_IF_UNSET (opts, opts_set,
+ param_early_inliner_max_iterations, 10);
+ break;
+diff --git a/gcc/tree-inline.c b/gcc/tree-inline.c
+index efde5d158..8405a959c 100644
+--- a/gcc/tree-inline.c
++++ b/gcc/tree-inline.c
+@@ -2015,6 +2015,10 @@ copy_bb (copy_body_data *id, basic_block bb,
+ basic_block_info automatically. */
+ copy_basic_block = create_basic_block (NULL, (basic_block) prev->aux);
+ copy_basic_block->count = bb->count.apply_scale (num, den);
++ /* Copy discriminator from original bb for distinguishes among
++ several basic blocks that share a common locus, allowing for
++ more accurate autofdo. */
++ copy_basic_block->discriminator = bb->discriminator;
+
+ copy_gsi = gsi_start_bb (copy_basic_block);
+
+@@ -3028,6 +3032,16 @@ copy_cfg_body (copy_body_data * id,
+ den += e->count ();
+ ENTRY_BLOCK_PTR_FOR_FN (cfun)->count = den;
+ }
++ /* When autofdo uses PMU as the sampling unit, the number of
++ ENTRY_BLOCK_PTR_FOR_FN cannot be obtained directly and will
++ be zero. It using for adjust_for_ipa_scaling will cause the
++ inlined BB count incorrectly overestimated. So set den equal
++ to num, which is the source inline BB count to avoid
++ overestimated. */
++ if (den == profile_count::zero ().afdo ())
++ {
++ den = num;
++ }
+
+ profile_count::adjust_for_ipa_scaling (&num, &den);
+
+--
+2.27.0.windows.1
+