stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8223363: Bad node estimate assertion failure

Browse Source 
8223502: Node estimate for loop unswitching is not correct: assert(delta <= 2 * required) failed: Bad node estimate
8224648: assert(!exceeding_node_budget()) failed: Too many NODES required! failure with ctw

Tighten the node estimates. New est_loop_clone_sz() implementation that will compute a "fan-out" complexity estimate as part of the size estimate (to better estimate complex loop body size after cloning). New est_loop_unroll_sz() function, used to estimate the size of a loop body att full/maximal unrolling. Correction to node budget final tests and asserts.

Reviewed-by: neliasso, kvn
This commit is contained in:
Patric Hedlin 2019-05-28 14:56:58 +02:00
parent d36c7bad82
commit d222b01dee
6 changed files with 405 additions and 134 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

155
src/hotspot/share/opto/loopTransform.cpp
View File

@ -286,6 +286,9 @@ Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
Node* n2 = n1->in(3 - inv1_idx); Node* n2 = n1->in(3 - inv1_idx);
int inv2_idx = is_invariant_addition(n2, phase); int inv2_idx = is_invariant_addition(n2, phase);
if (!inv2_idx) return NULL; if (!inv2_idx) return NULL;
if (!phase->may_require_nodes(10, 10)) return NULL;
Node* x = n2->in(3 - inv2_idx); Node* x = n2->in(3 - inv2_idx);
Node* inv2 = n2->in(inv2_idx); Node* inv2 = n2->in(inv2_idx);
@ -337,61 +340,72 @@ void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
Node* nn = reassociate_add_sub(n, phase); Node* nn = reassociate_add_sub(n, phase);
if (nn == NULL) break; if (nn == NULL) break;
n = nn; // again n = nn; // again
}; }
} }
} }
//------------------------------policy_peeling--------------------------------- //------------------------------policy_peeling---------------------------------
// Return TRUE or FALSE if the loop should be peeled or not. Peel if we can // Return TRUE if the loop should be peeled, otherwise return FALSE. Peeling
// make some loop-invariant test (usually a null-check) happen before the loop. // is applicable if we can make a loop-invariant test (usually a null-check)
bool IdealLoopTree::policy_peeling(PhaseIdealLoop *phase) const { // execute before we enter the loop. When TRUE, the estimated node budget is
IdealLoopTree *loop = (IdealLoopTree*)this; // also requested.
bool IdealLoopTree::policy_peeling(PhaseIdealLoop *phase) {
uint estimate = estimate_peeling(phase);
return estimate == 0 ? false : phase->may_require_nodes(estimate);
}
// Perform actual policy and size estimate for the loop peeling transform, and
// return the estimated loop size if peeling is applicable, otherwise return
// zero. No node budget is allocated.
uint IdealLoopTree::estimate_peeling(PhaseIdealLoop *phase) {
// If nodes are depleted, some transform has miscalculated its needs. // If nodes are depleted, some transform has miscalculated its needs.
assert(!phase->exceeding_node_budget(), "sanity"); assert(!phase->exceeding_node_budget(), "sanity");
uint body_size = loop->_body.size(); // Peeling does loop cloning which can result in O(N^2) node construction.
// Peeling does loop cloning which can result in O(N^2) node construction if (_body.size() > 255) {
if (body_size > 255) { return 0; // Suppress too large body size.
return false; // Prevent overflow for large body size
} }
uint estimate = body_size * body_size; // Optimistic estimate that approximates loop body complexity via data and
// control flow fan-out (instead of using the more pessimistic: BodySize^2).
uint estimate = est_loop_clone_sz(2);
if (phase->exceeding_node_budget(estimate)) { if (phase->exceeding_node_budget(estimate)) {
return false; // Too large to safely clone return 0; // Too large to safely clone.
} }
// check for vectorized loops, any peeling done was already applied // Check for vectorized loops, any peeling done was already applied.
if (_head->is_CountedLoop()) { if (_head->is_CountedLoop()) {
CountedLoopNode* cl = _head->as_CountedLoop(); CountedLoopNode* cl = _head->as_CountedLoop();
if (cl->is_unroll_only() || cl->trip_count() == 1) { if (cl->is_unroll_only() || cl->trip_count() == 1) {
return false; return 0;
} }
} }
Node* test = loop->tail(); Node* test = tail();
while (test != _head) { // Scan till run off top of loop while (test != _head) { // Scan till run off top of loop
if (test->is_If()) { // Test? if (test->is_If()) { // Test?
Node *ctrl = phase->get_ctrl(test->in(1)); Node *ctrl = phase->get_ctrl(test->in(1));
if (ctrl->is_top()) { if (ctrl->is_top()) {
return false; // Found dead test on live IF? No peeling! return 0; // Found dead test on live IF? No peeling!
} }
// Standard IF only has one input value to check for loop invariance // Standard IF only has one input value to check for loop invariance.
assert(test->Opcode() == Op_If || assert(test->Opcode() == Op_If ||
test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_CountedLoopEnd ||
test->Opcode() == Op_RangeCheck, test->Opcode() == Op_RangeCheck,
"Check this code when new subtype is added"); "Check this code when new subtype is added");
// Condition is not a member of this loop? // Condition is not a member of this loop?
if (!is_member(phase->get_loop(ctrl)) && is_loop_exit(test)) { if (!is_member(phase->get_loop(ctrl)) && is_loop_exit(test)) {
// Found reason to peel! return estimate; // Found reason to peel!
return phase->may_require_nodes(estimate);
} }
} }
// Walk up dominators to loop _head looking for test which is // Walk up dominators to loop _head looking for test which is executed on
// executed on every path thru loop. // every path through the loop.
test = phase->idom(test); test = phase->idom(test);
} }
return false; return 0;
} }
//------------------------------peeled_dom_test_elim--------------------------- //------------------------------peeled_dom_test_elim---------------------------
@ -638,8 +652,8 @@ void PhaseIdealLoop::do_peeling(IdealLoopTree *loop, Node_List &old_new) {
} }
} }
// Step 4: Correct dom-depth info. Set to loop-head depth. // Step 4: Correct dom-depth info. Set to loop-head depth.
int dd = dom_depth(head); int dd = dom_depth(head);
set_idom(head, head->in(1), dd); set_idom(head, head->in(1), dd);
for (uint j3 = 0; j3 < loop->_body.size(); j3++) { for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
@ -657,11 +671,30 @@ void PhaseIdealLoop::do_peeling(IdealLoopTree *loop, Node_List &old_new) {
loop->record_for_igvn(); loop->record_for_igvn();
} }
#define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop // The Estimated Loop Unroll Size: UnrollFactor * (106% * BodySize + BC) + CC,
// where BC and CC are (totally) ad-hoc/magic "body" and "clone" constants,
// respectively, used to ensure that node usage estimates made are on the safe
// side, for the most part. This is a simplified version of the loop clone
// size calculation in est_loop_clone_sz(), defined for unroll factors larger
// than one (>1), performing an overflow check and returning 'UINT_MAX' in
// case of an overflow.
static uint est_loop_unroll_sz(uint factor, uint size) {
precond(0 < factor);
uint const bc = 5;
uint const cc = 7;
uint const sz = size + (size + 15) / 16;
uint estimate = factor * (sz + bc) + cc;
return (estimate - cc) / factor == sz + bc ? estimate : UINT_MAX;
}
#define EMPTY_LOOP_SIZE 7 // Number of nodes in an empty loop.
//------------------------------policy_maximally_unroll------------------------ //------------------------------policy_maximally_unroll------------------------
// Calculate exact loop trip count and return true if loop can be maximally // Calculate the exact loop trip-count and return TRUE if loop can be fully,
// unrolled. // i.e. maximally, unrolled, otherwise return FALSE. When TRUE, the estimated
// node budget is also requested.
bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop *phase) const { bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop *phase) const {
CountedLoopNode *cl = _head->as_CountedLoop(); CountedLoopNode *cl = _head->as_CountedLoop();
assert(cl->is_normal_loop(), ""); assert(cl->is_normal_loop(), "");
@ -693,7 +726,7 @@ bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop *phase) const {
// Take into account that after unroll conjoined heads and tails will fold, // Take into account that after unroll conjoined heads and tails will fold,
// otherwise policy_unroll() may allow more unrolling than max unrolling. // otherwise policy_unroll() may allow more unrolling than max unrolling.
uint new_body_size = est_loop_clone_sz(trip_count, body_size - EMPTY_LOOP_SIZE); uint new_body_size = est_loop_unroll_sz(trip_count, body_size - EMPTY_LOOP_SIZE);
if (new_body_size == UINT_MAX) { // Check for bad estimate (overflow). if (new_body_size == UINT_MAX) { // Check for bad estimate (overflow).
return false; return false;
@ -742,8 +775,9 @@ bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop *phase) const {
//------------------------------policy_unroll---------------------------------- //------------------------------policy_unroll----------------------------------
// Return TRUE or FALSE if the loop should be unrolled or not. Unroll if the // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll if
// loop is a CountedLoop and the body is small enough. // the loop is a counted loop and the loop body is small enough. When TRUE,
// the estimated node budget is also requested.
bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) { bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
CountedLoopNode *cl = _head->as_CountedLoop(); CountedLoopNode *cl = _head->as_CountedLoop();
@ -887,7 +921,7 @@ bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
LoopMaxUnroll = slp_max_unroll_factor; LoopMaxUnroll = slp_max_unroll_factor;
} }
uint estimate = est_loop_clone_sz(2, body_size); uint estimate = est_loop_clone_sz(2);
if (cl->has_passed_slp()) { if (cl->has_passed_slp()) {
if (slp_max_unroll_factor >= future_unroll_cnt) { if (slp_max_unroll_factor >= future_unroll_cnt) {
@ -958,8 +992,10 @@ bool IdealLoopTree::policy_align(PhaseIdealLoop *phase) const {
} }
//------------------------------policy_range_check----------------------------- //------------------------------policy_range_check-----------------------------
// Return TRUE or FALSE if the loop should be range-check-eliminated. // Return TRUE or FALSE if the loop should be range-check-eliminated or not.
// Actually we do iteration-splitting, a more powerful form of RCE. // When TRUE, the estimated node budget is also requested.
//
// We will actually perform iteration-splitting, a more powerful form of RCE.
bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const { bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const {
if (!RangeCheckElimination) return false; if (!RangeCheckElimination) return false;
@ -967,9 +1003,9 @@ bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const {
assert(!phase->exceeding_node_budget(), "sanity"); assert(!phase->exceeding_node_budget(), "sanity");
CountedLoopNode *cl = _head->as_CountedLoop(); CountedLoopNode *cl = _head->as_CountedLoop();
// If we unrolled with no intention of doing RCE and we later // If we unrolled with no intention of doing RCE and we later changed our
// changed our minds, we got no pre-loop. Either we need to // minds, we got no pre-loop. Either we need to make a new pre-loop, or we
// make a new pre-loop, or we gotta disallow RCE. // have to disallow RCE.
if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
Node *trip_counter = cl->phi(); Node *trip_counter = cl->phi();
@ -1016,13 +1052,13 @@ bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const {
if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
continue; continue;
} }
// Found a test like 'trip+off vs limit'. Test is an IfNode, has two // Found a test like 'trip+off vs limit'. Test is an IfNode, has two (2)
// (2) projections. If BOTH are in the loop we need loop unswitching // projections. If BOTH are in the loop we need loop unswitching instead
// instead of iteration splitting. // of iteration splitting.
if (is_loop_exit(iff)) { if (is_loop_exit(iff)) {
// Found valid reason to split iterations (if there is room). // Found valid reason to split iterations (if there is room).
// NOTE: Usually a gross overestimate. // NOTE: Usually a gross overestimate.
return phase->may_require_nodes(est_loop_clone_sz(2, _body.size())); return phase->may_require_nodes(est_loop_clone_sz(2));
} }
} // End of is IF } // End of is IF
} }
@ -1521,9 +1557,6 @@ void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old
// only process vectorized main loops // only process vectorized main loops
if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return; if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
if (!may_require_nodes(est_loop_clone_sz(2, loop->_body.size()))) {
return;
}
int slp_max_unroll_factor = cl->slp_max_unroll(); int slp_max_unroll_factor = cl->slp_max_unroll();
int cur_unroll = cl->unrolled_count(); int cur_unroll = cl->unrolled_count();
@ -1535,6 +1568,10 @@ void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old
// we only ever process this one time // we only ever process this one time
if (cl->has_atomic_post_loop()) return; if (cl->has_atomic_post_loop()) return;
if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
return;
}
#ifndef PRODUCT #ifndef PRODUCT
if (TraceLoopOpts) { if (TraceLoopOpts) {
tty->print("PostVector "); tty->print("PostVector ");
@ -3178,9 +3215,6 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
AutoNodeBudget node_budget(phase); AutoNodeBudget node_budget(phase);
bool should_peel = policy_peeling(phase);
bool should_unswitch = policy_unswitching(phase);
// Non-counted loops may be peeled; exactly 1 iteration is peeled. // Non-counted loops may be peeled; exactly 1 iteration is peeled.
// This removes loop-invariant tests (usually null checks). // This removes loop-invariant tests (usually null checks).
if (!_head->is_CountedLoop()) { // Non-counted loop if (!_head->is_CountedLoop()) { // Non-counted loop
@ -3188,10 +3222,10 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
// Partial peel succeeded so terminate this round of loop opts // Partial peel succeeded so terminate this round of loop opts
return false; return false;
} }
if (should_peel) { // Should we peel? if (policy_peeling(phase)) { // Should we peel?
if (PrintOpto) { tty->print_cr("should_peel"); } if (PrintOpto) { tty->print_cr("should_peel"); }
phase->do_peeling(this,old_new); phase->do_peeling(this, old_new);
} else if (should_unswitch) { } else if (policy_unswitching(phase)) {
phase->do_unswitching(this, old_new); phase->do_unswitching(this, old_new);
} }
return true; return true;
@ -3209,12 +3243,11 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
// Before attempting fancy unrolling, RCE or alignment, see if we want // Before attempting fancy unrolling, RCE or alignment, see if we want
// to completely unroll this loop or do loop unswitching. // to completely unroll this loop or do loop unswitching.
if (cl->is_normal_loop()) { if (cl->is_normal_loop()) {
if (should_unswitch) { if (policy_unswitching(phase)) {
phase->do_unswitching(this, old_new); phase->do_unswitching(this, old_new);
return true; return true;
} }
bool should_maximally_unroll = policy_maximally_unroll(phase); if (policy_maximally_unroll(phase)) {
if (should_maximally_unroll) {
// Here we did some unrolling and peeling. Eventually we will // Here we did some unrolling and peeling. Eventually we will
// completely unroll this loop and it will no longer be a loop. // completely unroll this loop and it will no longer be a loop.
phase->do_maximally_unroll(this, old_new); phase->do_maximally_unroll(this, old_new);
@ -3222,6 +3255,9 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
} }
} }
uint est_peeling = estimate_peeling(phase);
bool should_peel = 0 < est_peeling;
// Counted loops may be peeled, may need some iterations run up // Counted loops may be peeled, may need some iterations run up
// front for RCE, and may want to align loop refs to a cache // front for RCE, and may want to align loop refs to a cache
// line. Thus we clone a full loop up front whose trip count is // line. Thus we clone a full loop up front whose trip count is
@ -3252,14 +3288,15 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
// peeling. // peeling.
if (should_rce || should_align || should_unroll) { if (should_rce || should_align || should_unroll) {
if (cl->is_normal_loop()) { // Convert to 'pre/main/post' loops if (cl->is_normal_loop()) { // Convert to 'pre/main/post' loops
if (!phase->may_require_nodes(est_loop_clone_sz(3, _body.size()))) { uint estimate = est_loop_clone_sz(3);
if (!phase->may_require_nodes(estimate)) {
return false; return false;
} }
phase->insert_pre_post_loops(this,old_new, !may_rce_align); phase->insert_pre_post_loops(this, old_new, !may_rce_align);
} }
// Adjust the pre- and main-loop limits to let the pre and post loops run // Adjust the pre- and main-loop limits to let the pre and post loops run
// with full checks, but the main-loop with no checks. Remove said // with full checks, but the main-loop with no checks. Remove said checks
// checks from the main body. // from the main body.
if (should_rce) { if (should_rce) {
if (phase->do_range_check(this, old_new) != 0) { if (phase->do_range_check(this, old_new) != 0) {
cl->mark_has_range_checks(); cl->mark_has_range_checks();
@ -3293,7 +3330,9 @@ bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_n
} }
} else { // Else we have an unchanged counted loop } else { // Else we have an unchanged counted loop
if (should_peel) { // Might want to peel but do nothing else if (should_peel) { // Might want to peel but do nothing else
phase->do_peeling(this,old_new); if (phase->may_require_nodes(est_peeling)) {
phase->do_peeling(this, old_new);
}
} }
} }
return true; return true;

6
src/hotspot/share/opto/loopUnswitch.cpp
View File

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 2006, 2012, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2006, 2019, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* *
* This code is free software; you can redistribute it and/or modify it * This code is free software; you can redistribute it and/or modify it
@ -79,7 +79,7 @@ bool IdealLoopTree::policy_unswitching( PhaseIdealLoop *phase ) const {
} }
// Too speculative if running low on nodes. // Too speculative if running low on nodes.
return phase->may_require_nodes(est_loop_clone_sz(3, _body.size())); return phase->may_require_nodes(est_loop_clone_sz(2));
} }
//------------------------------find_unswitching_candidate----------------------------- //------------------------------find_unswitching_candidate-----------------------------
@ -116,7 +116,7 @@ IfNode* PhaseIdealLoop::find_unswitching_candidate(const IdealLoopTree *loop) co
// Clone loop with an invariant test (that does not exit) and // Clone loop with an invariant test (that does not exit) and
// insert a clone of the test that selects which version to // insert a clone of the test that selects which version to
// execute. // execute.
void PhaseIdealLoop::do_unswitching (IdealLoopTree *loop, Node_List &old_new) { void PhaseIdealLoop::do_unswitching(IdealLoopTree *loop, Node_List &old_new) {
// Find first invariant test that doesn't exit the loop // Find first invariant test that doesn't exit the loop
LoopNode *head = loop->_head->as_Loop(); LoopNode *head = loop->_head->as_Loop();

75
src/hotspot/share/opto/loopnode.cpp
View File

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* *
* This code is free software; you can redistribute it and/or modify it * This code is free software; you can redistribute it and/or modify it
@ -2439,12 +2439,63 @@ void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
if (loop->_next) loop->_next ->counted_loop(phase); if (loop->_next) loop->_next ->counted_loop(phase);
} }
// The Estimated Loop Clone Size:
// CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
// where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
// respectively, used to ensure that the node usage estimates made are on the
// safe side, for the most part. The FanOutTerm is an attempt to estimate the
// possible additional/excessive nodes generated due to data and control flow
// merging, for edges reaching outside the loop.
uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
precond(0 < factor && factor < 16);
uint const bc = 13;
uint const cc = 17;
uint const sz = _body.size() + (_body.size() + 7) / 8;
uint estimate = factor * (sz + bc) + cc;
assert((estimate - cc) / factor == sz + bc, "overflow");
uint ctrl_edge_out_cnt = 0;
uint data_edge_out_cnt = 0;
for (uint i = 0; i < _body.size(); i++) {
Node* node = _body.at(i);
uint outcnt = node->outcnt();
for (uint k = 0; k < outcnt; k++) {
Node* out = node->raw_out(k);
if (out->is_CFG()) {
if (!is_member(_phase->get_loop(out))) {
ctrl_edge_out_cnt++;
}
} else {
Node* ctrl = _phase->get_ctrl(out);
assert(ctrl->is_CFG(), "must be");
if (!is_member(_phase->get_loop(ctrl))) {
data_edge_out_cnt++;
}
}
}
}
// Add data (x1.5) and control (x1.0) count to estimate iff both are > 0.
if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
estimate += ctrl_edge_out_cnt + data_edge_out_cnt + data_edge_out_cnt / 2;
}
return estimate;
}
#ifndef PRODUCT #ifndef PRODUCT
//------------------------------dump_head-------------------------------------- //------------------------------dump_head--------------------------------------
// Dump 1 liner for loop header info // Dump 1 liner for loop header info
void IdealLoopTree::dump_head( ) const { void IdealLoopTree::dump_head() const {
for (uint i=0; i<_nest; i++) for (uint i = 0; i < _nest; i++) {
tty->print(" "); tty->print(" ");
}
tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
if (_irreducible) tty->print(" IRREDUCIBLE"); if (_irreducible) tty->print(" IRREDUCIBLE");
Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl); Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
@ -2513,7 +2564,7 @@ void IdealLoopTree::dump_head( ) const {
//------------------------------dump------------------------------------------- //------------------------------dump-------------------------------------------
// Dump loops by loop tree // Dump loops by loop tree
void IdealLoopTree::dump( ) const { void IdealLoopTree::dump() const {
dump_head(); dump_head();
if (_child) _child->dump(); if (_child) _child->dump();
if (_next) _next ->dump(); if (_next) _next ->dump();
@ -2908,8 +2959,8 @@ void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
return; return;
} }
if(VerifyLoopOptimizations) verify(); if (VerifyLoopOptimizations) verify();
if(TraceLoopOpts && C->has_loops()) { if (TraceLoopOpts && C->has_loops()) {
_ltree_root->dump(); _ltree_root->dump();
} }
#endif #endif
@ -2938,7 +2989,6 @@ void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
} }
if (ReassociateInvariants) { if (ReassociateInvariants) {
AutoNodeBudget node_budget(this, AutoNodeBudget::NO_BUDGET_CHECK);
// Reassociate invariants and prep for split_thru_phi // Reassociate invariants and prep for split_thru_phi
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current(); IdealLoopTree* lpt = iter.current();
@ -2946,14 +2996,17 @@ void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
if (!is_counted || !lpt->is_innermost()) continue; if (!is_counted || !lpt->is_innermost()) continue;
// check for vectorized loops, any reassociation of invariants was already done // check for vectorized loops, any reassociation of invariants was already done
if (is_counted && lpt->_head->as_CountedLoop()->is_unroll_only()) continue; if (is_counted && lpt->_head->as_CountedLoop()->is_unroll_only()) {
continue;
lpt->reassociate_invariants(this); } else {
AutoNodeBudget node_budget(this);
lpt->reassociate_invariants(this);
}
// Because RCE opportunities can be masked by split_thru_phi, // Because RCE opportunities can be masked by split_thru_phi,
// look for RCE candidates and inhibit split_thru_phi // look for RCE candidates and inhibit split_thru_phi
// on just their loop-phi's for this pass of loop opts // on just their loop-phi's for this pass of loop opts
if (SplitIfBlocks && do_split_ifs) { if (SplitIfBlocks && do_split_ifs) {
AutoNodeBudget node_budget(this, AutoNodeBudget::NO_BUDGET_CHECK);
if (lpt->policy_range_check(this)) { if (lpt->policy_range_check(this)) {
lpt->_rce_candidate = 1; // = true lpt->_rce_candidate = 1; // = true
} }

116
src/hotspot/share/opto/loopnode.hpp
View File

@ -589,17 +589,18 @@ public:
// Convert one iteration loop into normal code. // Convert one iteration loop into normal code.
bool do_one_iteration_loop( PhaseIdealLoop *phase ); bool do_one_iteration_loop( PhaseIdealLoop *phase );
// Return TRUE or FALSE if the loop should be peeled or not. Peel if we can // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
// make some loop-invariant test (usually a null-check) happen before the // move some loop-invariant test (usually a null-check) before the loop.
// loop. bool policy_peeling(PhaseIdealLoop *phase);
bool policy_peeling( PhaseIdealLoop *phase ) const;
uint estimate_peeling(PhaseIdealLoop *phase);
// Return TRUE or FALSE if the loop should be maximally unrolled. Stash any // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
// known trip count in the counted loop node. // known trip count in the counted loop node.
bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
// Return TRUE or FALSE if the loop should be unrolled or not. Unroll if // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
// the loop is a CountedLoop and the body is small enough. // if the loop is a counted loop and the loop body is small enough.
bool policy_unroll(PhaseIdealLoop *phase); bool policy_unroll(PhaseIdealLoop *phase);
// Loop analyses to map to a maximal superword unrolling for vectorization. // Loop analyses to map to a maximal superword unrolling for vectorization.
@ -620,6 +621,9 @@ public:
// Return TRUE if "iff" is a range check. // Return TRUE if "iff" is a range check.
bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
// Estimate the number of nodes required when cloning a loop (body).
uint est_loop_clone_sz(uint factor) const;
// Compute loop trip count if possible // Compute loop trip count if possible
void compute_trip_count(PhaseIdealLoop* phase); void compute_trip_count(PhaseIdealLoop* phase);
@ -1356,50 +1360,66 @@ private:
// < UINT_MAX Nodes currently requested (estimate). // < UINT_MAX Nodes currently requested (estimate).
uint _nodes_required; uint _nodes_required;
enum { REQUIRE_MIN = 70 };
uint nodes_required() const { return _nodes_required; }
// Given the _currently_ available number of nodes, check whether there is
// "room" for an additional request or not, considering the already required
// number of nodes. Return TRUE if the new request is exceeding the node
// budget limit, otherwise return FALSE. Note that this interpretation will
// act pessimistic on additional requests when new nodes have already been
// generated since the 'begin'. This behaviour fits with the intention that
// node estimates/requests should be made upfront.
bool exceeding_node_budget(uint required = 0) { bool exceeding_node_budget(uint required = 0) {
assert(C->live_nodes() < C->max_node_limit(), "sanity"); assert(C->live_nodes() < C->max_node_limit(), "sanity");
uint available = C->max_node_limit() - C->live_nodes(); uint available = C->max_node_limit() - C->live_nodes();
return available < required + _nodes_required; return available < required + _nodes_required;
} }
uint require_nodes(uint require) { uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
precond(require > 0); precond(require > 0);
_nodes_required += MAX2(100u, require); // Keep requests at minimum 100. _nodes_required += MAX2(require, minreq);
return _nodes_required; return _nodes_required;
} }
bool may_require_nodes(uint require) { bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
return !exceeding_node_budget(require) && require_nodes(require) > 0; return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
} }
void require_nodes_begin() { uint require_nodes_begin() {
assert(_nodes_required == UINT_MAX, "Bad state (begin)."); assert(_nodes_required == UINT_MAX, "Bad state (begin).");
_nodes_required = 0; _nodes_required = 0;
return C->live_nodes();
} }
// Final check that the requested nodes did not exceed the limit and that // When a node request is final, optionally check that the requested number
// the request was reasonably correct with respect to the number of new // of nodes was reasonably correct with respect to the number of new nodes
// nodes introduced by any transform since the last 'begin'. // introduced since the last 'begin'. Always check that we have not exceeded
void require_nodes_final_check(uint live_at_begin) { // the maximum node limit.
uint required = _nodes_required; void require_nodes_final(uint live_at_begin, bool check_estimate) {
require_nodes_final();
uint delta = C->live_nodes() - live_at_begin;
// Assert is disabled, see JDK-8223911 and related issues.
assert(true || delta <= 2 * required, "Bad node estimate (actual: %d, request: %d)",
delta, required);
}
void require_nodes_final() {
assert(_nodes_required < UINT_MAX, "Bad state (final)."); assert(_nodes_required < UINT_MAX, "Bad state (final).");
assert(!exceeding_node_budget(), "Too many NODES required!");
if (check_estimate) {
// Assert that the node budget request was not off by too much (x2).
// Should this be the case we _surely_ need to improve the estimates
// used in our budget calculations.
assert(C->live_nodes() - live_at_begin <= 2 * _nodes_required,
"Bad node estimate: actual = %d >> request = %d",
C->live_nodes() - live_at_begin, _nodes_required);
}
// Assert that we have stayed within the node budget limit.
assert(C->live_nodes() < C->max_node_limit(),
"Exceeding node budget limit: %d + %d > %d (request = %d)",
C->live_nodes() - live_at_begin, live_at_begin,
C->max_node_limit(), _nodes_required);
_nodes_required = UINT_MAX; _nodes_required = UINT_MAX;
} }
bool _created_loop_node; bool _created_loop_node;
public: public:
uint nodes_required() const { return _nodes_required; }
void set_created_loop_node() { _created_loop_node = true; } void set_created_loop_node() { _created_loop_node = true; }
bool created_loop_node() { return _created_loop_node; } bool created_loop_node() { return _created_loop_node; }
void register_new_node( Node *n, Node *blk ); void register_new_node( Node *n, Node *blk );
@ -1438,29 +1458,30 @@ public:
{ {
precond(_phase != NULL); precond(_phase != NULL);
_nodes_at_begin = _phase->C->live_nodes(); _nodes_at_begin = _phase->require_nodes_begin();
_phase->require_nodes_begin();
} }
~AutoNodeBudget() { ~AutoNodeBudget() {
if (_check_at_final) {
#ifndef PRODUCT #ifndef PRODUCT
if (TraceLoopOpts) { if (TraceLoopOpts) {
uint request = _phase->nodes_required(); uint request = _phase->nodes_required();
uint delta = _phase->C->live_nodes() - _nodes_at_begin;
if (request > 0) { if (request < delta) {
uint delta = _phase->C->live_nodes() - _nodes_at_begin; tty->print_cr("Exceeding node budget: %d < %d", request, delta);
} else {
if (request < delta) { uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
tty->print_cr("Exceeding node budget: %d < %d", request, delta); // Identify the worst estimates as "poor" ones.
if (request > REQUIRE_MIN && delta > 0) {
if ((delta > REQUIRE_MIN && request > 3 * delta) ||
(delta <= REQUIRE_MIN && request > 10 * delta)) {
tty->print_cr("Poor node estimate: %d >> %d", request, delta);
} }
} }
} }
#endif
_phase->require_nodes_final_check(_nodes_at_begin);
} else {
_phase->require_nodes_final();
} }
#endif // PRODUCT
_phase->require_nodes_final(_nodes_at_begin, _check_at_final);
} }
private: private:
@ -1469,17 +1490,6 @@ private:
uint _nodes_at_begin; uint _nodes_at_begin;
}; };
// The Estimated Loop Clone Size: CloneFactor * (BodySize + BC) + CC, where BC
// and CC are totally ad-hoc/magic "body" and "clone" constants, respectively,
// used to ensure that node usage estimates made are on the safe side, for the
// most part.
static inline uint est_loop_clone_sz(uint fact, uint size) {
uint const bc = 31;
uint const cc = 41;
uint estimate = fact * (size + bc) + cc;
return (estimate - cc) / fact == size + bc ? estimate : UINT_MAX;
}
// This kit may be used for making of a reserved copy of a loop before this loop // This kit may be used for making of a reserved copy of a loop before this loop
// goes under non-reversible changes. // goes under non-reversible changes.

19
src/hotspot/share/opto/loopopts.cpp
View File

@ -1,5 +1,5 @@
/* /*
* Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
* *
* This code is free software; you can redistribute it and/or modify it * This code is free software; you can redistribute it and/or modify it
@ -3029,16 +3029,16 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
assert(!loop->_head->is_CountedLoop(), "Non-counted loop only"); assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
if (!loop->_head->is_Loop()) { if (!loop->_head->is_Loop()) {
return false; } return false;
}
LoopNode *head = loop->_head->as_Loop(); LoopNode *head = loop->_head->as_Loop();
if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) { if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
return false; return false;
} }
// Check for complex exit control // Check for complex exit control
for(uint ii = 0; ii < loop->_body.size(); ii++ ) { for (uint ii = 0; ii < loop->_body.size(); ii++) {
Node *n = loop->_body.at(ii); Node *n = loop->_body.at(ii);
int opc = n->Opcode(); int opc = n->Opcode();
if (n->is_Call() || if (n->is_Call() ||
@ -3065,12 +3065,12 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
IfNode *peel_if_cmpu = NULL; IfNode *peel_if_cmpu = NULL;
Node *iff = loop->tail(); Node *iff = loop->tail();
while( iff != head ) { while (iff != head) {
if( iff->is_If() ) { if (iff->is_If()) {
Node *ctrl = get_ctrl(iff->in(1)); Node *ctrl = get_ctrl(iff->in(1));
if (ctrl->is_top()) return false; // Dead test on live IF. if (ctrl->is_top()) return false; // Dead test on live IF.
// If loop-varying exit-test, check for induction variable // If loop-varying exit-test, check for induction variable
if( loop->is_member(get_loop(ctrl)) && if (loop->is_member(get_loop(ctrl)) &&
loop->is_loop_exit(iff) && loop->is_loop_exit(iff) &&
is_possible_iv_test(iff)) { is_possible_iv_test(iff)) {
Node* cmp = iff->in(1)->in(1); Node* cmp = iff->in(1)->in(1);
@ -3084,6 +3084,7 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
} }
iff = idom(iff); iff = idom(iff);
} }
// Prefer signed compare over unsigned compare. // Prefer signed compare over unsigned compare.
IfNode* new_peel_if = NULL; IfNode* new_peel_if = NULL;
if (peel_if == NULL) { if (peel_if == NULL) {
@ -3131,7 +3132,7 @@ bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
Node_List worklist(area); Node_List worklist(area);
Node_List sink_list(area); Node_List sink_list(area);
if (!may_require_nodes(est_loop_clone_sz(2, loop->_body.size()))) { if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
return false; return false;
} }

168
test/hotspot/jtreg/compiler/loopopts/LoopUnswitchingBadNodeBudget.java Normal file
View File

@ -0,0 +1,168 @@
/*
* Copyright (c) 2019, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*/
/*
* @test
* @bug 8223502
* @summary Node estimate for loop unswitching is not correct:
* assert(delta <= 2 * required) failed: Bad node estimate
*
* @requires !vm.graal.enabled
*
* @run main/othervm -XX:-TieredCompilation -XX:-BackgroundCompilation
* -XX:-UseOnStackReplacement -XX:CompileOnly=LoopUnswitchingBadNodeBudget::test
* -XX:CompileCommand=dontinline,LoopUnswitchingBadNodeBudget::helper
* -XX:+UnlockExperimentalVMOptions -XX:-UseSwitchProfiling LoopUnswitchingBadNodeBudget
*
*/
public class LoopUnswitchingBadNodeBudget {
public static void main(String[] args) {
for (int i = 0; i < 20_000; i++) {
for (int j = 0; j < 100; j++) {
test(j, true, 0, 0, 0);
test(j, false, 0, 0, 0);
}
}
}
private static int test(int j, boolean flag, int k, int l, int m) {
int res = 0;
for (int i = 0; i < 24; i++) {
if (flag) {
k = k / 2;
l = l * 2;
m = m + 2;
}
switch (j) {
case 0: break;
case 1: return helper(j, k, l, m);
case 2: return helper(j, k, l, m);
case 3: return helper(j, k, l, m);
case 4: return helper(j, k, l, m);
case 5: return helper(j, k, l, m);
case 6: return helper(j, k, l, m);
case 7: return helper(j, k, l, m);
case 8: return helper(j, k, l, m);
case 9: return helper(j, k, l, m);
case 10: return helper(j, k, l, m);
case 11: return helper(j, k, l, m);
case 12: return helper(j, k, l, m);
case 13: return helper(j, k, l, m);
case 14: return helper(j, k, l, m);
case 15: return helper(j, k, l, m);
case 16: return helper(j, k, l, m);
case 17: return helper(j, k, l, m);
case 18: return helper(j, k, l, m);
case 19: return helper(j, k, l, m);
case 20: return helper(j, k, l, m);
case 21: return helper(j, k, l, m);
case 22: return helper(j, k, l, m);
case 23: return helper(j, k, l, m);
case 24: return helper(j, k, l, m);
case 25: return helper(j, k, l, m);
case 26: return helper(j, k, l, m);
case 27: return helper(j, k, l, m);
case 28: return helper(j, k, l, m);
case 29: return helper(j, k, l, m);
case 30: return helper(j, k, l, m);
case 31: return helper(j, k, l, m);
case 32: return helper(j, k, l, m);
case 33: return helper(j, k, l, m);
case 34: return helper(j, k, l, m);
case 35: return helper(j, k, l, m);
case 36: return helper(j, k, l, m);
case 37: return helper(j, k, l, m);
case 38: return helper(j, k, l, m);
case 39: return helper(j, k, l, m);
case 40: return helper(j, k, l, m);
case 41: return helper(j, k, l, m);
case 42: return helper(j, k, l, m);
case 43: return helper(j, k, l, m);
case 44: return helper(j, k, l, m);
case 45: return helper(j, k, l, m);
case 46: return helper(j, k, l, m);
case 47: return helper(j, k, l, m);
case 48: return helper(j, k, l, m);
case 49: return helper(j, k, l, m);
case 50: return helper(j, k, l, m);
case 51: return helper(j, k, l, m);
case 52: return helper(j, k, l, m);
case 53: return helper(j, k, l, m);
case 54: return helper(j, k, l, m);
case 55: return helper(j, k, l, m);
case 56: return helper(j, k, l, m);
case 57: return helper(j, k, l, m);
case 58: return helper(j, k, l, m);
case 59: return helper(j, k, l, m);
case 60: return helper(j, k, l, m);
case 61: return helper(j, k, l, m);
case 62: return helper(j, k, l, m);
case 63: return helper(j, k, l, m);
case 64: return helper(j, k, l, m);
case 65: return helper(j, k, l, m);
case 66: return helper(j, k, l, m);
case 67: return helper(j, k, l, m);
case 68: return helper(j, k, l, m);
case 69: return helper(j, k, l, m);
case 70: return helper(j, k, l, m);
case 71: return helper(j, k, l, m);
case 72: return helper(j, k, l, m);
case 73: return helper(j, k, l, m);
case 74: return helper(j, k, l, m);
case 75: return helper(j, k, l, m);
case 76: return helper(j, k, l, m);
case 77: return helper(j, k, l, m);
case 78: return helper(j, k, l, m);
case 79: return helper(j, k, l, m);
case 80: return helper(j, k, l, m);
case 81: return helper(j, k, l, m);
case 82: return helper(j, k, l, m);
case 83: return helper(j, k, l, m);
case 84: return helper(j, k, l, m);
case 85: return helper(j, k, l, m);
case 86: return helper(j, k, l, m);
case 87: return helper(j, k, l, m);
case 88: return helper(j, k, l, m);
case 89: return helper(j, k, l, m);
case 90: return helper(j, k, l, m);
case 91: return helper(j, k, l, m);
case 92: return helper(j, k, l, m);
case 93: return helper(j, k, l, m);
case 94: return helper(j, k, l, m);
case 95: return helper(j, k, l, m);
case 96: return helper(j, k, l, m);
case 97: return helper(j, k, l, m);
case 98: return helper(j, k, l, m);
case 99: return helper(j, k, l, m);
}
res += helper(j, k, l, m);
}
return res;
}
private static int helper(int j, int k, int l, int m) {
return j + k;
}
}