/* * Copyright (c) 1998, 2022, 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. *
*/
//============================================================================= //--------------------------is_cloop_ind_var----------------------------------- // Determine if a node is a counted loop induction variable. // NOTE: The method is declared in "node.hpp". bool Node::is_cloop_ind_var() const { return (is_Phi() &&
as_Phi()->region()->is_CountedLoop() &&
as_Phi()->region()->as_CountedLoop()->phi() == this);
}
//============================================================================= //------------------------------dump_spec-------------------------------------- // Dump special per-node info #ifndef PRODUCT void LoopNode::dump_spec(outputStream *st) const { if (is_inner_loop()) st->print( "inner " ); if (is_partial_peel_loop()) st->print( "partial_peel " ); if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
} #endif
//------------------------------get_early_ctrl--------------------------------- // Compute earliest legal control
Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
uint i;
Node *early; if (n->in(0) && !n->is_expensive()) {
early = n->in(0); if (!early->is_CFG()) // Might be a non-CFG multi-def
early = get_ctrl(early); // So treat input as a straight data input
i = 1;
} else {
early = get_ctrl(n->in(1));
i = 2;
}
uint e_d = dom_depth(early);
assert( early, "" ); for (; i < n->req(); i++) {
Node *cin = get_ctrl(n->in(i));
assert( cin, "" ); // Keep deepest dominator depth
uint c_d = dom_depth(cin); if (c_d > e_d) { // Deeper guy?
early = cin; // Keep deepest found so far
e_d = c_d;
} elseif (c_d == e_d && // Same depth?
early != cin) { // If not equal, must use slower algorithm // If same depth but not equal, one _must_ dominate the other // and we want the deeper (i.e., dominated) guy.
Node *n1 = early;
Node *n2 = cin; while (1) {
n1 = idom(n1); // Walk up until break cycle
n2 = idom(n2); if (n1 == cin || // Walked early up to cin
dom_depth(n2) < c_d) break; // early is deeper; keep him if (n2 == early || // Walked cin up to early
dom_depth(n1) < c_d) {
early = cin; // cin is deeper; keep him break;
}
}
e_d = dom_depth(early); // Reset depth register cache
}
}
if (n->is_expensive() && !_verify_only && !_verify_me) {
assert(n->in(0), "should have control input");
early = get_early_ctrl_for_expensive(n, early);
}
return early;
}
//------------------------------get_early_ctrl_for_expensive--------------------------------- // Move node up the dominator tree as high as legal while still beneficial
Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
assert(OptimizeExpensiveOps, "optimization off?");
Node* ctl = n->in(0);
assert(ctl->is_CFG(), "expensive input 0 must be cfg");
uint min_dom_depth = dom_depth(earliest); #ifdef ASSERT if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
} #endif if (dom_depth(ctl) < min_dom_depth) { return earliest;
}
while (1) {
Node *next = ctl; // Moving the node out of a loop on the projection of a If // confuses loop predication. So once we hit a Loop in a If branch // that doesn't branch to an UNC, we stop. The code that process // expensive nodes will notice the loop and skip over it to try to // move the node further up. if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { break;
}
next = idom(ctl->in(1)->in(0));
} elseif (ctl->is_Proj()) { // We only move it up along a projection if the projection is // the single control projection for its parent: same code path, // if it's a If with UNC or fallthrough of a call.
Node* parent_ctl = ctl->in(0); if (parent_ctl == NULL) { break;
} elseif (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
} elseif (parent_ctl->is_If()) { if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { break;
}
assert(idom(ctl) == parent_ctl, "strange");
next = idom(parent_ctl);
} elseif (ctl->is_CatchProj()) { if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { break;
}
assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
next = parent_ctl->in(0)->in(0)->in(0);
} else { // Check if parent control has a single projection (this // control is the only possible successor of the parent // control). If so, we can try to move the node above the // parent control. int nb_ctl_proj = 0; for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
Node *p = parent_ctl->fast_out(i); if (p->is_Proj() && p->is_CFG()) {
nb_ctl_proj++; if (nb_ctl_proj > 1) { break;
}
}
}
//------------------------------set_subtree_ctrl------------------------------- // set missing _ctrl entries on new nodes void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) { // Already set? Get out. if (_nodes[n->_idx]) return; // Recursively set _nodes array to indicate where the Node goes
uint i; for (i = 0; i < n->req(); ++i) {
Node *m = n->in(i); if (m && m != C->root()) {
set_subtree_ctrl(m, update_body);
}
}
// Create a skeleton strip mined outer loop: a Loop head before the // inner strip mined loop, a safepoint and an exit condition guarded // by an opaque node after the inner strip mined loop with a backedge // to the loop head. The inner strip mined loop is left as it is. Only // once loop optimizations are over, do we adjust the inner loop exit // condition to limit its number of iterations, set the outer loop // exit condition and add Phis to the outer loop head. Some loop // optimizations that operate on the inner strip mined loop need to be // aware of the outer strip mined loop: loop unswitching needs to // clone the outer loop as well as the inner, unrolling needs to only // clone the inner loop etc. No optimizations need to change the outer // strip mined loop as it is only a skeleton.
IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
IdealLoopTree* loop, float cl_prob, float le_fcnt,
Node*& entry_control, Node*& iffalse) {
Node* outer_test = _igvn.intcon(0);
set_ctrl(outer_test, C->root());
Node *orig = iffalse;
iffalse = iffalse->clone();
_igvn.register_new_node_with_optimizer(iffalse);
set_idom(iffalse, idom(orig), dom_depth(orig));
#ifndef PRODUCT // report that the loop predication has been actually performed // for this loop if (TraceLoopLimitCheck) {
tty->print_cr("Counted Loop Limit Check generated:");
debug_only( bol->dump(2); )
} #endif
}
Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) { // Counted loop head must be a good RegionNode with only 3 not NULL // control input edges: Self, Entry, LoopBack. if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { return NULL;
}
Node *init_control = x->in(LoopNode::EntryControl);
Node *back_control = x->in(LoopNode::LoopBackControl); if (init_control == NULL || back_control == NULL) { // Partially dead return NULL;
} // Must also check for TOP when looking for a dead loop if (init_control->is_top() || back_control->is_top()) { return NULL;
}
// Allow funny placement of Safepoint if (back_control->Opcode() == Op_SafePoint) {
back_control = back_control->in(TypeFunc::Control);
}
// Controlling test for loop
Node *iftrue = back_control;
uint iftrue_op = iftrue->Opcode(); if (iftrue_op != Op_IfTrue &&
iftrue_op != Op_IfFalse) { // I have a weird back-control. Probably the loop-exit test is in // the middle of the loop and I am looking at some trailing control-flow // merge point. To fix this I would have to partially peel the loop. return NULL; // Obscure back-control
}
// Get boolean guarding loop-back test
Node *iff = iftrue->in(0); if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) { return NULL;
} return iftrue;
}
// Find the trip-counter increment & limit. Limit must be loop invariant.
incr = cmp->in(1);
limit = cmp->in(2);
// --------- // need 'loop()' test to tell if limit is loop invariant // ---------
if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
Node* tmp = incr; // Then reverse order into the CmpI
incr = limit;
limit = tmp;
bt = BoolTest(bt).commute(); // And commute the exit test
} if (is_member(loop, get_ctrl(limit))) { // Limit must be loop-invariant return NULL;
} if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant return NULL;
} return cmp;
}
Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) { if (incr->is_Phi()) { if (incr->as_Phi()->region() != x || incr->req() != 3) { return NULL; // Not simple trip counter expression
}
phi_incr = incr;
incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant return NULL;
}
} return incr;
}
Node* PhaseIdealLoop::loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi) {
assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp."); // Get merge point
xphi = incr->in(1);
Node *stride = incr->in(2); if (!stride->is_Con()) { // Oops, swap these if (!xphi->is_Con()) { // Is the other guy a constant? return NULL; // Nope, unknown stride, bail out
}
Node *tmp = xphi; // 'incr' is commutative, so ok to swap
xphi = stride;
stride = tmp;
} return stride;
}
PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop) { if (!xphi->is_Phi()) { return NULL; // Too much math on the trip counter
} if (phi_incr != NULL && phi_incr != xphi) { return NULL;
}
PhiNode *phi = xphi->as_Phi();
// Phi must be of loop header; backedge must wrap to increment if (phi->region() != x) { return NULL;
} return phi;
}
staticbool condition_stride_ok(BoolTest::mask bt, jlong stride_con) { // If the condition is inverted and we will be rolling // through MININT to MAXINT, then bail out. if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! // Odd stride
(bt == BoolTest::ne && stride_con != 1 && stride_con != -1) || // Count down loop rolls through MAXINT
((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) || // Count up loop rolls through MININT
((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) { returnfalse; // Bail out
} returntrue;
}
Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
BasicType bt) {
Node* iv_as_long; if (bt == T_LONG) {
iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
register_new_node(iv_as_long, inner_head);
} else {
iv_as_long = inner_iv;
}
Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
register_new_node(iv_replacement, inner_head); for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
Node* u = iv_to_replace->last_out(i); #ifdef ASSERT if (!is_dominator(inner_head, ctrl_or_self(u))) {
assert(u->is_Phi(), "should be a Phi"); for (uint j = 1; j < u->req(); j++) { if (u->in(j) == iv_to_replace) {
assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
}
}
} #endif
_igvn.rehash_node_delayed(u); int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
i -= nb;
} return iv_replacement;
}
void PhaseIdealLoop::add_empty_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop, SafePointNode* sfpt) { if (!C->too_many_traps(reason)) {
Node *cont = _igvn.intcon(1);
Node* opq = new Opaque1Node(C, cont);
_igvn.register_new_node_with_optimizer(opq);
Node *bol = new Conv2BNode(opq);
_igvn.register_new_node_with_optimizer(bol);
set_subtree_ctrl(bol, false);
IfNode* iff = new IfNode(inner_head->in(LoopNode::EntryControl), bol, PROB_MAX, COUNT_UNKNOWN);
register_control(iff, loop, inner_head->in(LoopNode::EntryControl));
Node* iffalse = new IfFalseNode(iff);
register_control(iffalse, _ltree_root, iff);
Node* iftrue = new IfTrueNode(iff);
register_control(iftrue, loop, iff);
C->add_predicate_opaq(opq);
for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
set_subtree_ctrl(unc->in(i), false);
}
register_control(unc, _ltree_root, iffalse);
Node* ctrl = new ProjNode(unc, TypeFunc::Control);
register_control(ctrl, _ltree_root, unc);
Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
register_control(halt, _ltree_root, ctrl);
_igvn.add_input_to(C->root(), halt);
// Find a safepoint node that dominates the back edge. We need a // SafePointNode so we can use its jvm state to create empty // predicates. staticbool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
SafePointNode* safepoint = NULL; for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
Node* u = x->fast_out(i); if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
Node* m = u->in(LoopNode::LoopBackControl); if (u->adr_type() == TypePtr::BOTTOM) { if (m->is_MergeMem() && mem->is_MergeMem()) { if (m != mem DEBUG_ONLY(|| true)) { // MergeMemStream can modify m, for example to adjust the length to mem. // This is unfortunate, and probably unnecessary. But as it is, we need // to add m to the igvn worklist, else we may have a modified node that // is not on the igvn worklist.
phase->igvn()._worklist.push(m); for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) { if (!mms.is_empty()) { if (mms.memory() != mms.memory2()) { returnfalse;
} #ifdef ASSERT if (mms.alias_idx() != Compile::AliasIdxBot) {
mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
} #endif
}
}
}
} elseif (mem->is_MergeMem()) { if (m != mem->as_MergeMem()->base_memory()) { returnfalse;
}
} else { returnfalse;
}
} else { if (mem->is_MergeMem()) { if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) { returnfalse;
} #ifdef ASSERT
mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory()); #endif
} else { if (m != mem) { returnfalse;
}
}
}
}
} returntrue;
}
SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
IfNode* exit_test = back_control->in(0)->as_If();
SafePointNode* safepoint = NULL; if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
safepoint = exit_test->in(0)->as_SafePoint();
} else {
Node* c = back_control; while (c != x && c->Opcode() != Op_SafePoint) {
c = idom(c);
}
if (c->Opcode() == Op_SafePoint) {
safepoint = c->as_SafePoint();
}
if (safepoint == NULL) { return NULL;
}
Node* mem = safepoint->in(TypeFunc::Memory);
// We can only use that safepoint if there's no side effect between the backedge and the safepoint.
// mm is used for book keeping
MergeMemNode* mm = NULL; #ifdef ASSERT if (mem->is_MergeMem()) {
mm = mem->clone()->as_MergeMem();
_igvn._worklist.push(mm); for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) { if (mms.alias_idx() != Compile::AliasIdxBot && loop != get_loop(ctrl_or_self(mms.memory()))) {
mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
}
}
} #endif if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
safepoint = NULL;
} else {
assert(mm == NULL|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
} #ifdef ASSERT if (mm != NULL) {
_igvn.remove_dead_node(mm);
} #endif
} return safepoint;
}
// If the loop has the shape of a counted loop but with a long // induction variable, transform the loop in a loop nest: an inner // loop that iterates for at most max int iterations with an integer // induction variable and an outer loop that iterates over the full // range of long values from the initial loop in (at most) max int // steps. That is: // // x: for (long phi = init; phi < limit; phi += stride) { // // phi := Phi(L, init, incr) // // incr := AddL(phi, longcon(stride)) // long incr = phi + stride; // ... use phi and incr ... // } // // OR: // // x: for (long phi = init; (phi += stride) < limit; ) { // // phi := Phi(L, AddL(init, stride), incr) // // incr := AddL(phi, longcon(stride)) // long incr = phi + stride; // ... use phi and (phi + stride) ... // } // // ==transform=> // // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0 // assert(stride <= inner_iters_limit); // else abort transform // assert((extralong)limit + stride <= LONG_MAX); // else deopt // outer_head: for (long outer_phi = init;;) { // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi))) // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi)); // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max); // assert(inner_iters_actual == (int)inner_iters_actual); // int inner_phi, inner_incr; // x: for (inner_phi = 0;; inner_phi = inner_incr) { // // inner_phi := Phi(x, intcon(0), inner_incr) // // inner_incr := AddI(inner_phi, intcon(stride)) // inner_incr = inner_phi + stride; // if (inner_incr < inner_iters_actual) { // ... use phi=>(outer_phi+inner_phi) and incr=>(outer_phi+inner_incr) ... // continue; // } // else break; // } // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit // continue; // else break; // } // // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int // loops with long range checks transformed to int range checks in the inner loop. bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
Node* x = loop->_head; // Only for inner loops if (loop->_child != NULL || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) { returnfalse;
}
if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) { returnfalse;
}
BaseCountedLoopNode* head = x->as_BaseCountedLoop();
BasicType bt = x->as_BaseCountedLoop()->bt();
check_counted_loop_shape(loop, x, bt);
#ifndef PRODUCT if (bt == T_LONG) {
Atomic::inc(&_long_loop_candidates);
} #endif
jlong stride_con = head->stride_con();
assert(stride_con != 0, "missed some peephole opt"); // We can't iterate for more than max int at a time. if (stride_con != (jint)stride_con) {
assert(bt == T_LONG, "only for long loops"); returnfalse;
} // The number of iterations for the integer count loop: guarantee no // overflow: max_jint - stride_con max. -1 so there's no need for a // loop limit check if the exit test is <= or >=. int iters_limit = max_jint - ABS(stride_con) - 1; #ifdef ASSERT if (bt == T_LONG && StressLongCountedLoop > 0) {
iters_limit = iters_limit / StressLongCountedLoop;
} #endif // At least 2 iterations so counted loop construction doesn't fail if (iters_limit/ABS(stride_con) < 2) { returnfalse;
}
// data nodes on back branch not supported if (back_control->outcnt() > 1) { returnfalse;
}
Node* limit = head->limit(); // We'll need to use the loop limit before the inner loop is entered if (!is_dominator(get_ctrl(limit), x)) { returnfalse;
}
IfNode* exit_test = head->loopexit();
assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
if (bt == T_INT) { // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further. if (range_checks.size() == 0) { returnfalse;
}
}
// Take what we know about the number of iterations of the long counted loop into account when computing the limit of // the inner loop. const Node* init = head->init_trip(); const TypeInteger* lo = _igvn.type(init)->is_integer(bt); const TypeInteger* hi = _igvn.type(limit)->is_integer(bt); if (stride_con < 0) {
swap(lo, hi);
} if (hi->hi_as_long() <= lo->lo_as_long()) { // not a loop after all returnfalse;
}
julong orig_iters = hi->hi_as_long() - lo->lo_as_long();
iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
// We need a safepoint to insert empty predicates for the inner loop.
SafePointNode* safepoint; if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) { // Loop is strip mined: use the safepoint of the outer strip mined loop
OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
assert(outer_loop != NULL, "no outer loop");
safepoint = outer_loop->outer_safepoint();
outer_loop->transform_to_counted_loop(&_igvn, this);
exit_test = head->loopexit();
} else {
safepoint = find_safepoint(back_control, x, loop);
}
// Clone the control flow of the loop to build an outer loop
Node* outer_back_branch = back_control->clone();
Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
Node* inner_exit_branch = exit_branch->clone();
// add an iv phi to the outer loop and use it to compute the inner // loop iteration limit
Node* outer_phi = phi->clone();
outer_phi->set_req(0, outer_head);
register_new_node(outer_phi, outer_head);
Node* inner_iters_limit = _igvn.integercon(iters_limit, bt); // inner_iters_max may not fit in a signed integer (iterating from // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned // min.
Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, TypeInteger::make(0, iters_limit, Type::WidenMin, bt), _igvn);
Node* inner_iters_actual_int; if (bt == T_LONG) {
inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
_igvn.register_new_node_with_optimizer(inner_iters_actual_int);
} else {
inner_iters_actual_int = inner_iters_actual;
}
Node* int_zero = _igvn.intcon(0);
set_ctrl(int_zero, C->root()); if (stride_con < 0) {
inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
_igvn.register_new_node_with_optimizer(inner_iters_actual_int);
}
// Clone the iv data nodes as an integer iv
Node* int_stride = _igvn.intcon(checked_cast<int>(stride_con));
set_ctrl(int_stride, C->root());
Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
Node* inner_incr = new AddINode(inner_phi, int_stride);
Node* inner_cmp = NULL;
inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
inner_phi->set_req(LoopNode::EntryControl, int_zero);
inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
register_new_node(inner_phi, x);
register_new_node(inner_incr, x);
register_new_node(inner_cmp, x);
register_new_node(inner_bol, x);
_igvn.replace_input_of(exit_test, 1, inner_bol);
// Clone inner loop phis to outer loop for (uint i = 0; i < head->outcnt(); i++) {
Node* u = head->raw_out(i); if (u->is_Phi() && u != inner_phi && u != phi) {
assert(u->in(0) == head, "inconsistent");
Node* clone = u->clone();
clone->set_req(0, outer_head);
register_new_node(clone, outer_head);
_igvn.replace_input_of(u, LoopNode::EntryControl, clone);
}
}
// Replace inner loop long iv phi as inner loop int iv phi + outer // loop iv phi
Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
// Replace inner loop long iv incr with inner loop int incr + outer // loop iv phi
loop_nest_replace_iv(incr, inner_incr, outer_phi, head, bt);
if (bt == T_INT) {
outer_phi = new ConvI2LNode(outer_phi);
register_new_node(outer_phi, outer_head);
}
transform_long_range_checks(checked_cast<int>(stride_con), range_checks, outer_phi, inner_iters_actual_int,
inner_phi, iv_add, inner_head); // Peel one iteration of the loop and use the safepoint at the end // of the peeled iteration to insert empty predicates. If no well // positioned safepoint peel to guarantee a safepoint in the outer // loop. if (safepoint != NULL || !loop->_has_call) {
old_new.clear();
do_peeling(loop, old_new);
} else {
C->set_major_progress();
}
if (safepoint != NULL) {
SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jlong stride_con, int iters_limit, PhiNode* phi,
Node_List& range_checks) { const jlong min_iters = 2;
jlong reduced_iters_limit = iters_limit;
jlong original_iters_limit = iters_limit; for (uint i = 0; i < loop->_body.size(); i++) {
Node* c = loop->_body.at(i); if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
CallStaticJavaNode* call = c->as_IfProj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); if (call != NULL) {
Node* range = NULL;
Node* offset = NULL;
jlong scale = 0;
RangeCheckNode* rc = c->in(0)->as_RangeCheck(); if (loop->is_range_check_if(rc, this, T_LONG, phi, range, offset, scale) &&
loop->is_invariant(range) && loop->is_invariant(offset) &&
original_iters_limit / ABS(scale * stride_con) >= min_iters) {
reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
range_checks.push(c);
}
}
}
}
return checked_cast<int>(reduced_iters_limit);
}
// One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init, // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its // first value, followed by A+S (S is the stride), next A+2S, etc. The limit is exclusive, so that the final value B of // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
// If |S|>1 the formula for the last value B would require a floor operation, specifically B=floor((Z-sgn(S)-A)/S)*S+A, // which is B=Z-sgn(S)U for some U in [1,|S|]. So when S>0, i ranges as i:[A,Z) or i:[A,B=Z-U], or else (in reverse) // as i:(Z,A] or i:[B=Z+U,A]. It will become important to reason about this inclusive range [A,B] or [B,A].
// Within the loop there may be many range checks. Each such range check (R.C.) is of the form 0 <= i*K+L < R, where K // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant. // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
// When a long loop over a 64-bit variable i (outer_iv) is decomposed into a series of shorter sub-loops over a 32-bit // variable j (inner_iv), j ranges over a shorter interval j:[0,B_2] or [0,Z_2) (assuming S > 0), where the limit is // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
// For S<0, j ranges (in reverse!) through j:[-|B_2|,0] or (-|Z_2|,0]. For either sign of S, we can say i=j+C and j // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
// The disjoint union of all the C+[A_2,B_2] ranges from the sub-loops must be identical to the whole range [A,B]. // Assuming S>0, the first C must be A itself, and the next C value is the previous C+B_2, plus S. If |S|=1, the next // C value is also the previous C+Z_2. In each sub-loop, j counts from j=A_2=0 and i counts from C+0 and exits at // j=B_2 (i=C+B_2), just before it gets to i=C+Z_2. Both i and j count up (from C and 0) if S>0; otherwise they count // down (from C and 0 again).
// Returning to range checks, we see that each i*K+L <u R expands to (C+j)*K+L <u R, or j*K+Q <u R, where Q=(C*K+L). // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see // PhaseIdealLoop::add_constraint.)
// We must transform this comparison so that it gets the same answer, but by means of a 32-bit R.C. (using j not i) of // the form j*K+L_2 <u32 R_2. Note that L_2 and R_2 must be loop-invariant, but only with respect to the sub-loop. Thus, the // problem reduces to computing values for L_2 and R_2 (for each R.C. in the loop) in the loop header for the sub-loop. // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum // values for the 32-bit counter j within which the range checks can be eliminated.
// So, given j*K+Q <u R, we need to find some j*K+L_2 <u32 R_2, where L_2 and R_2 fit in 32 bits, and the 32-bit operations do // not overflow. We also need to cover the cases where i*K+L (= j*K+Q) overflows to a 64-bit negative, since that is // allowed as an input to the R.C., as long as the R.C. as a whole fails.
// If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
// For each R.C. j*K+Q <u32 R, the range of mathematical values of j*K+Q in the sub-loop is [Q_min, Q_max], where // Q_min=Q and Q_max=B_2*K+Q (if S>0 and K>0), Q_min=A_2*K+Q and Q_max=Q (if S<0 and K>0), // Q_min=B_2*K+Q and Q_max=Q if (S>0 and K<0), Q_min=Q and Q_max=A_2*K+Q (if S<0 and K<0)
// Note that the first R.C. value is always Q=(S*K>0 ? Q_min : Q_max). Also Q_{min,max} = Q + {min,max}(A_2*K,B_2*K). // If S*K>0 then, as the loop iterations progress, each R.C. value i*K+L = j*K+Q goes up from Q=Q_min towards Q_max. // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
// Case A: Some Negatives (but no overflow). // Number line: // |s64_min . . . 0 . . . s64_max| // | . Q_min..Q_max . 0 . . . . | s64 negative // | . . . . R=0 R< R< R< R< | (against R values) // | . . . Q_min..0..Q_max . . . | small mixed // | . . . . R R R< R< R< | (against R values) // // R values which are out of range (>Q_max+1) are reduced to max(0,Q_max+1). They are marked on the number line as R<. // // So, if Q_min <s64 0, then use this test: // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward) // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward) // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q: // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
// If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
// Case B: No Negatives. // Number line: // |s64_min . . . 0 . . . s64_max| // | . . . . 0 Q_min..Q_max . . | small positive // | . . . . R> R R R< R< | (against R values) // | . . . . 0 . Q_min..Q_max . | s64 positive // | . . . . R> R> R R R< | (against R values) // // R values which are out of range (<Q_min or >Q_max+1) are reduced as marked: R> up to Q_min, R< down to Q_max+1. // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value. // // So, if both Q_min, Q_max+1 >=s64 0, then use this test: // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0 // More generally: // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
// Case C: Overflow in the 64-bit domain // Number line: // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow // | . . . . R> R> R> R> R | (against R values) // // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered. // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong. // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.) // // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test: // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0 // More generally: // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K // // Dropping the bad bound means only Q_min is used to reduce the range of R: // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K // // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]: // clamp(X, L, H) := max(L, min(X, H)) // When degenerately L > H, it returns L not H. // // All of the formulas above can be merged into a single one: // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L) // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range // replacement R.C.: // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp // or equivalently: // j*K + L_2 <u32 R_2 // where // L_2 = Q_first - L_clamp // R_2 = R_clamp - L_clamp // // Note on why R is never negative: // // Various details of this transformation would break badly if R could be negative, so this transformation only // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable // R.C., but R values from a.length and inside checkIndex are good to go. // void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, ;Node* outer_phi,
Node* inner_iters_actual_int, Node* inner_phi,
Node* iv_add, LoopNode* inner_head) {
Node* long_zero = _igvn.longcon(0);
set_ctrl(long_zero, C->root());
Node* int_zero = _igvn.intcon(0);
set_ctrl(int_zero, this->C->root());
Node* long_one = _igvn.longcon(1);
set_ctrl(long_one, this->C->root());
Node* int_stride = _igvn.intcon(checked_cast<int>(stride_con));
set_ctrl(int_stride, this->C->root());
for (uint i = 0; i < range_checks.size(); i++) {
ProjNode* proj = range_checks.at(i)->as_Proj();
ProjNode* unc_proj = proj->other_if_proj();
RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
jlong scale = 0;
Node* offset = NULL;
Node* rc_bol = rc->in(1);
Node* rc_cmp = rc_bol->in(1); if (rc_cmp->Opcode() == Op_CmpU) { // could be shared and have already been taken care of continue;
} bool short_scale = false; bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
assert(ok, "inconsistent: was tested before");
Node* range = rc_cmp->in(2);
Node* c = rc->in(0);
Node* entry_control = inner_head->in(LoopNode::EntryControl);
Node* R = range;
Node* K = _igvn.longcon(scale);
set_ctrl(K, this->C->root());
Node* L = offset;
if (short_scale) { // This converts: // (int)i*K + L <u64 R // with K an int into: // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) // to protect against an overflow of (int)i*K // // Because if (int)i*K overflows, there are K,L where: // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value. // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then // i*(long)K + L <u64 R is true. // // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution. // // It's always true that: // (int)i*K <u64 (long)max_jint + 1 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L // As a consequence: // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) // is always false in case of overflow of i*K // // Note, there are also K,L where i*K overflows and // i*K + L <u64 R is true, but // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false // So this transformation could cause spurious deoptimizations and failed range check elimination // (but not incorrect execution) for unlikely corner cases with overflow. // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
Node* max_jint_plus_one_long = _igvn.longcon((jlong)max_jint + 1);
set_ctrl(max_jint_plus_one_long, C->root());
Node* max_range = new AddLNode(max_jint_plus_one_long, L);
register_new_node(max_range, entry_control);
R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
set_subtree_ctrl(R, true);
}
Node* C = outer_phi;
// Start with 64-bit values: // i*K + L <u64 R // (C+j)*K + L <u64 R // j*K + Q <u64 R where Q = Q_first = C*K+L
Node* Q_first = new MulLNode(C, K);
register_new_node(Q_first, entry_control);
Q_first = new AddLNode(Q_first, L);
register_new_node(Q_first, entry_control);
// Compute endpoints of the range of values j*K + Q. // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
Node* Q_min = Q_first;
// Compute the exact ending value B_2 (which is really A_2 if S < 0)
Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
register_new_node(B_2, entry_control);
B_2 = new SubINode(B_2, int_stride);
register_new_node(B_2, entry_control);
B_2 = new ConvI2LNode(B_2);
register_new_node(B_2, entry_control);
Node* Q_max = new MulLNode(B_2, K);
register_new_node(Q_max, entry_control);
Q_max = new AddLNode(Q_max, Q_first);
register_new_node(Q_max, entry_control);
if (scale * stride_con < 0) {
swap(Q_min, Q_max);
} // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
// L_clamp = Q_min < 0 ? 0 : Q_min
Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
register_new_node(Q_min_cmp, entry_control);
Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
register_new_node(Q_min_bool, entry_control);
Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
register_new_node(L_clamp, entry_control); // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
register_new_node(Q_max_plus_one, entry_control);
// H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & ;Q_min >= 0.)
Node* max_jlong_long = _igvn.longcon(max_jlong);
set_ctrl(max_jlong_long, this->C->root());
Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
register_new_node(Q_max_cmp, entry_control);
Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
register_new_node(Q_max_bool, entry_control);
Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
register_new_node(H_clamp, entry_control); // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
// R_2 = clamp(R, L_clamp, H_clamp) - L_clamp // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow) // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
Node* R_2 = clamp(R, L_clamp, H_clamp);
R_2 = new SubLNode(R_2, L_clamp);
register_new_node(R_2, entry_control);
R_2 = new ConvL2INode(R_2, TypeInt::POS);
register_new_node(R_2, entry_control);
// L_2 = Q_first - L_clamp // We are subtracting L_clamp from both sides of the <u32 comparison. // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp. // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
Node* L_2 = new SubLNode(Q_first, L_clamp);
register_new_node(L_2, entry_control);
L_2 = new ConvL2INode(L_2, TypeInt::INT);
register_new_node(L_2, entry_control);
// Transform the range check using the computed values L_2/R_2 // from: i*K + L <u64 R // to: j*K + L_2 <u32 R_2 // that is: // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
K = _igvn.intcon(checked_cast<int>(scale));
set_ctrl(K, this->C->root());
Node* scaled_iv = new MulINode(inner_phi, K);
register_new_node(scaled_iv, c);
Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
register_new_node(scaled_iv_plus_offset, c);
Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
register_new_node(new_rc_cmp, c);
// Safepoint on backedge not supported
assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
} #endif
#ifdef ASSERT // convert an int counted loop to a long counted to stress handling of // long counted loops bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
Unique_Node_List iv_nodes;
Node_List old_new;
iv_nodes.push(cmp); bool failed = false;
for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
Node* n = iv_nodes.at(i); switch(n->Opcode()) { case Op_Phi: {
Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
old_new.map(n->_idx, clone); break;
} case Op_CmpI: {
Node* clone = new CmpLNode(NULL, NULL);
old_new.map(n->_idx, clone); break;
} case Op_AddI: {
Node* clone = new AddLNode(NULL, NULL);
old_new.map(n->_idx, clone); break;
} case Op_CastII: {
failed = true; break;
} default:
DEBUG_ONLY(n->dump());
fatal("unexpected");
}
for (uint i = 1; i < n->req(); i++) {
Node* in = n->in(i); if (in == NULL) { continue;
} if (loop->is_member(get_loop(get_ctrl(in)))) {
iv_nodes.push(in);
}
}
}
if (failed) { for (uint i = 0; i < iv_nodes.size(); i++) {
Node* n = iv_nodes.at(i);
Node* clone = old_new[n->_idx]; if (clone != NULL) {
_igvn.remove_dead_node(clone);
}
} returnfalse;
}
for (uint i = 0; i < iv_nodes.size(); i++) {
Node* n = iv_nodes.at(i);
Node* clone = old_new[n->_idx]; for (uint i = 1; i < n->req(); i++) {
Node* in = n->in(i); if (in == NULL) { continue;
}
Node* in_clone = old_new[in->_idx]; if (in_clone == NULL) {
assert(_igvn.type(in)->isa_int(), "");
in_clone = new ConvI2LNode(in);
_igvn.register_new_node_with_optimizer(in_clone);
set_subtree_ctrl(in_clone, false);
} if (in_clone->in(0) == NULL) {
in_clone->set_req(0, C->top());
clone->set_req(i, in_clone);
in_clone->set_req(0, NULL);
} else {
clone->set_req(i, in_clone);
}
}
_igvn.register_new_node_with_optimizer(clone);
}
set_ctrl(old_new[phi->_idx], phi->in(0));
for (uint i = 0; i < iv_nodes.size(); i++) {
Node* n = iv_nodes.at(i);
Node* clone = old_new[n->_idx];
set_subtree_ctrl(clone, false);
Node* m = n->Opcode() == Op_CmpI ? clone : NULL; for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* u = n->fast_out(i); if (iv_nodes.member(u)) { continue;
} if (m == NULL) {
m = new ConvL2INode(clone);
_igvn.register_new_node_with_optimizer(m);
set_subtree_ctrl(m, false);
}
_igvn.rehash_node_delayed(u); int nb = u->replace_edge(n, m, &_igvn);
--i, imax -= nb;
}
} returntrue;
} #endif
const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt); if (trunc1 != NULL) { // When there is a truncation, we must be sure that after the truncation // the trip counter will end up higher than the limit, otherwise we are looking // at an endless loop. Can happen with range checks.
// Example: // int i = 0; // while (true) // sum + = array[i]; // i++; // i = i && 0x7fff; // } // // If the array is shorter than 0x8000 this exits through a AIOOB // - Counted loop transformation is ok // If the array is longer then this is an endless loop // - No transformation can be done.
const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt); if (limit_t->hi_as_long() > incr_t->hi_as_long()) { // if the limit can have a higher value than the increment (before the phi) returnfalse;
}
}
// If iv trunc type is smaller than int, check for possible wrap. if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
assert(trunc1 != NULL, "must have found some truncation");
// Get a better type for the phi (filtered thru if's) const TypeInteger* phi_ft = filtered_type(phi);
// Can iv take on a value that will wrap? // // Ensure iv's limit is not within "stride" of the wrap value. // // Example for "short" type // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) // If the stride is +10, then the last value of the induction // variable before the increment (phi_ft->_hi) must be // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to // ensure no truncation occurs after the increment.
if (stride_con > 0) { if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) { returnfalse; // truncation may occur
}
} elseif (stride_con < 0) { if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) { returnfalse; // truncation may occur
}
} // No possibility of wrap so truncation can be discarded // Promote iv type to Int
} else {
assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
}
if (!condition_stride_ok(bt, stride_con)) { returnfalse;
}
if (phi_incr != NULL && bt != BoolTest::ne) { // check if there is a possibility of IV overflowing after the first increment if (stride_con > 0) { if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) { returnfalse;
}
} else { if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) { returnfalse;
}
}
}
// =================================================== // Generate loop limit check to avoid integer overflow // in cases like next (cyclic loops): // // for (i=0; i <= max_jint; i++) {} // for (i=0; i < max_jint; i+=2) {} // // // Limit check predicate depends on the loop test: // // for(;i != limit; i++) --> limit <= (max_jint) // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) //
// Check if limit is excluded to do more precise int overflow check. bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
jlong stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
// If compare points directly to the phi we need to adjust // the compare so that it points to the incr. Limit have // to be adjusted to keep trip count the same and the // adjusted limit should be checked for int overflow.
Node* adjusted_limit = limit; if (phi_incr != NULL) {
stride_m += stride_con;
}
int sov = check_stride_overflow(stride_m, limit_t, iv_bt); // If sov==0, limit's type always satisfies the condition, for // example, when it is an array length. if (sov != 0) { if (sov < 0) { returnfalse; // Bailout: integer overflow is certain.
}
assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed"); // Generate loop's limit check. // Loop limit check predicate should be near the loop.
ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); if (!limit_check_proj) { // The limit check predicate is not generated if this method trapped here before. #ifdef ASSERT if (TraceLoopLimitCheck) {
tty->print("missing loop limit check:");
loop->dump_head();
x->dump(1);
} #endif returnfalse;
}
// Now we need to canonicalize loop condition. if (bt == BoolTest::ne) {
assert(stride_con == 1 || stride_con == -1, "simple increment only"); if (stride_con > 0 && init_t->hi_as_long() < limit_t->lo_as_long()) { // 'ne' can be replaced with 'lt' only when init < limit.
bt = BoolTest::lt;
} elseif (stride_con < 0 && init_t->lo_as_long() > limit_t->hi_as_long()) { // 'ne' can be replaced with 'gt' only when init > limit.
bt = BoolTest::gt;
} else {
ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); if (!limit_check_proj) { // The limit check predicate is not generated if this method trapped here before. #ifdef ASSERT if (TraceLoopLimitCheck) {
tty->print("missing loop limit check:");
loop->dump_head();
x->dump(1);
} #endif returnfalse;
}
IfNode* check_iff = limit_check_proj->in(0)->as_If();
if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
!is_dominator(get_ctrl(init_trip), check_iff->in(0))) { returnfalse;
}
Node* cmp_limit;
Node* bol;
if (stride_con > 0) {
cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
bol = new BoolNode(cmp_limit, BoolTest::lt);
} else {
cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
bol = new BoolNode(cmp_limit, BoolTest::gt);
}
if (stride_con > 0) { // 'ne' can be replaced with 'lt' only when init < limit.
bt = BoolTest::lt;
} elseif (stride_con < 0) { // 'ne' can be replaced with 'gt' only when init > limit.
bt = BoolTest::gt;
}
}
}
if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
Node* backedge_sfpt = x->in(LoopNode::LoopBackControl); if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
iv_bt == T_LONG) &&
sfpt == NULL) { // Leaving the safepoint on the backedge and creating a // CountedLoop will confuse optimizations. We can't move the // safepoint around because its jvm state wouldn't match a new // location. Give up on that loop. returnfalse;
} if (is_deleteable_safept(backedge_sfpt)) {
lazy_replace(backedge_sfpt, iftrue); if (loop->_safepts != NULL) {
loop->_safepts->yank(backedge_sfpt);
}
loop->_tail = iftrue;
}
}
if (phi_incr != NULL) { // If compare points directly to the phi we need to adjust // the compare so that it points to the incr. Limit have // to be adjusted to keep trip count the same and we // should avoid int overflow. // // i = init; do {} while(i++ < limit); // is converted to // i = init; do {} while(++i < limit+1); //
adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
}
if (incl_limit) { // The limit check guaranties that 'limit <= (max_jint - stride)' so // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. //
Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt)); if (bt == BoolTest::le)
bt = BoolTest::lt; elseif (bt == BoolTest::ge)
bt = BoolTest::gt; else
ShouldNotReachHere();
}
set_subtree_ctrl(adjusted_limit, false);
// Build a canonical trip test. // Clone code, as old values may be in use.
incr = incr->clone();
incr->set_req(1,phi);
incr->set_req(2,stride);
incr = _igvn.register_new_node_with_optimizer(incr);
set_early_ctrl(incr, false);
_igvn.rehash_node_delayed(phi);
phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
// If phi type is more restrictive than Int, raise to // Int to prevent (almost) infinite recursion in igvn // which can only handle integer types for constants or minint..maxint. if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
nphi = _igvn.register_new_node_with_optimizer(nphi);
set_ctrl(nphi, get_ctrl(phi));
_igvn.replace_node(phi, nphi);
phi = nphi->as_Phi();
}
cmp = cmp->clone();
cmp->set_req(1,incr);
cmp->set_req(2, adjusted_limit);
cmp = _igvn.register_new_node_with_optimizer(cmp);
set_ctrl(cmp, iff->in(0));
test = test->clone()->as_Bool();
(*(BoolTest*)&test->_test)._test = bt;
test->set_req(1,cmp);
_igvn.register_new_node_with_optimizer(test);
set_ctrl(test, iff->in(0));
// Replace the old IfNode with a new LoopEndNode
Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
IfNode *le = lex->as_If();
uint dd = dom_depth(iff);
set_idom(le, le->in(0), dd); // Update dominance for loop exit
set_loop(le, loop);
// Get the loop-exit control
Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
// Need to swap loop-exit and loop-back control? if (iftrue_op == Op_IfFalse) {
Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
// Now setup a new CountedLoopNode to replace the existing LoopNode
BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve // The following assert is approximately true, and defines the intention // of can_be_counted_loop. It fails, however, because phase->type // is not yet initialized for this loop and its parts. //assert(l->can_be_counted_loop(this), "sanity");
_igvn.register_new_node_with_optimizer(l);
set_loop(l, loop);
loop->_head = l; // Fix all data nodes placed at the old loop head. // Uses the lazy-update mechanism of 'get_ctrl'.
lazy_replace( x, l );
set_idom(l, entry_control, dom_depth(entry_control) + 1);
if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) { // Check for immediately preceding SafePoint and remove if (sfpt != NULL && (strip_mine_loop || is_deleteable_safept(sfpt))) { if (strip_mine_loop) {
Node* outer_le = outer_ilt->_tail->in(0);
Node* sfpt_clone = sfpt->clone();
sfpt_clone->set_req(0, iffalse);
outer_le->set_req(0, sfpt_clone);
Node* polladdr = sfpt_clone->in(TypeFunc::Parms); if (polladdr != nullptr && polladdr->is_Load()) { // Polling load should be pinned outside inner loop.
Node* new_polladdr = polladdr->clone();
new_polladdr->set_req(0, iffalse);
_igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
set_ctrl(new_polladdr, iffalse);
sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
} // When this code runs, loop bodies have not yet been populated. constbool body_populated = false;
register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
}
lazy_replace(sfpt, sfpt->in(TypeFunc::Control)); if (loop->_safepts != NULL) {
loop->_safepts->yank(sfpt);
}
}
}
// Capture bounds of the loop in the induction variable Phi before // subsequent transformation (iteration splitting) obscures the // bounds
l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
if (strip_mine_loop) {
l->mark_strip_mined();
l->verify_strip_mined(1);
outer_ilt->_head->as_Loop()->verify_strip_mined(1);
loop = outer_ilt;
}
if (ABS(cl->stride_con()) == 1 ||
cl->limit()->Opcode() == Op_LoopLimit) { // Old code has exact limit (it could be incorrect in case of int overflow). // Loop limit is exact with stride == 1. And loop may already have exact limit. return cl->limit();
}
Node *limit = NULL; #ifdef ASSERT
BoolTest::mask bt = cl->loopexit()->test_trip();
assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); #endif if (cl->has_exact_trip_count()) { // Simple case: loop has constant boundaries. // Use jlongs to avoid integer overflow. int stride_con = cl->stride_con();
jlong init_con = cl->init_trip()->get_int();
jlong limit_con = cl->limit()->get_int();
julong trip_cnt = cl->trip_count();
jlong final_con = init_con + trip_cnt*stride_con; int final_int = (int)final_con; // The final value should be in integer range since the loop // is counted and the limit was checked for overflow.
assert(final_con == (jlong)final_int, "final value should be integer");
limit = _igvn.intcon(final_int);
} else { // Create new LoopLimit node to get exact limit (final iv value).
limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
register_new_node(limit, cl->in(LoopNode::EntryControl));
}
assert(limit != NULL, "sanity"); return limit;
}
//------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node. // Attempt to convert into a counted-loop.
Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
phase->C->set_major_progress();
} return RegionNode::Ideal(phase, can_reshape);
}
#ifdef ASSERT void LoopNode::verify_strip_mined(int expect_skeleton) const { const OuterStripMinedLoopNode* outer = NULL; const CountedLoopNode* inner = NULL; if (is_strip_mined()) { if (!is_valid_counted_loop(T_INT)) { return; // Skip malformed counted loop
}
assert(is_CountedLoop(), "no Loop should be marked strip mined");
inner = as_CountedLoop();
outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
} elseif (is_OuterStripMinedLoop()) {
outer = this->as_OuterStripMinedLoop();
inner = outer->unique_ctrl_out()->as_CountedLoop();
assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
} if (inner != NULL || outer != NULL) {
assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
Node* outer_tail = outer->in(LoopNode::LoopBackControl);
Node* outer_le = outer_tail->in(0);
assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
Node* sfpt = outer_le->in(0);
assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
Node* inner_out = sfpt->in(0);
CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
assert(cle == inner->loopexit_or_null(), "mismatch"); bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0; if (has_skeleton) {
assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
assert(outer->outcnt() == 2, "only control nodes");
} else {
assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
uint phis = 0;
uint be_loads = 0;
Node* be = inner->in(LoopNode::LoopBackControl); for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
Node* u = inner->fast_out(i); if (u->is_Phi()) {
phis++; for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
Node* n = be->fast_out(j); if (n->is_Load()) {
assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge"); do {
n = n->raw_out(0);
} while (!n->is_Phi()); if (n == u) {
be_loads++; break;
}
}
}
}
}
assert(be_loads <= phis, "wrong number phis that depends on a pinned load"); for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
Node* u = outer->fast_out(i);
assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
}
uint stores = 0; for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
Node* u = inner_out->fast_out(i); if (u->is_Store()) {
stores++;
}
} // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is // not guaranteed to be optimized out.
assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
}
assert(sfpt->outcnt() == 1, "no data node");
assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
}
} #endif
//============================================================================= //------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node. // Attempt to convert into a counted-loop.
Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { return RegionNode::Ideal(phase, can_reshape);
}
//------------------------------dump_spec-------------------------------------- // Dump special per-node info #ifndef PRODUCT void CountedLoopNode::dump_spec(outputStream *st) const {
LoopNode::dump_spec(st); if (stride_is_con()) {
st->print("stride: %d ",stride_con());
} if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); if (is_main_loop()) st->print("main of N%d", _idx); if (is_post_loop()) st->print("post of N%d", _main_idx); if (is_reduction_loop()) st->print(" reduction"); if (is_strip_mined()) st->print(" strip mined");
} #endif
//============================================================================= //------------------------------Value----------------------------------------- const Type* LoopLimitNode::Value(PhaseGVN* phase) const { const Type* init_t = phase->type(in(Init)); const Type* limit_t = phase->type(in(Limit)); const Type* stride_t = phase->type(in(Stride)); // Either input is TOP ==> the result is TOP if (init_t == Type::TOP) return Type::TOP; if (limit_t == Type::TOP) return Type::TOP; if (stride_t == Type::TOP) return Type::TOP;
int stride_con = stride_t->is_int()->get_con(); if (stride_con == 1) return bottom_type(); // Identity
if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { // Use jlongs to avoid integer overflow.
jlong init_con = init_t->is_int()->get_con();
jlong limit_con = limit_t->is_int()->get_con(); int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
jlong final_con = init_con + stride_con*trip_count; int final_int = (int)final_con; // The final value should be in integer range since the loop // is counted and the limit was checked for overflow.
assert(final_con == (jlong)final_int, "final value should be integer"); return TypeInt::make(final_int);
}
return bottom_type(); // TypeInt::INT
}
//------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node.
Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { if (phase->type(in(Init)) == Type::TOP ||
phase->type(in(Limit)) == Type::TOP ||
phase->type(in(Stride)) == Type::TOP) return NULL; // Dead
int stride_con = phase->type(in(Stride))->is_int()->get_con(); if (stride_con == 1) return NULL; // Identity
if (in(Init)->is_Con() && in(Limit)->is_Con()) return NULL; // Value
// Delay following optimizations until all loop optimizations // done to keep Ideal graph simple. if (!can_reshape || !phase->C->post_loop_opts_phase()) { return NULL;
}
const TypeInt* init_t = phase->type(in(Init) )->is_int(); const TypeInt* limit_t = phase->type(in(Limit))->is_int(); int stride_p;
jlong lim, ini;
julong max; if (stride_con > 0) {
stride_p = stride_con;
lim = limit_t->_hi;
ini = init_t->_lo;
max = (julong)max_jint;
} else {
stride_p = -stride_con;
lim = init_t->_hi;
ini = limit_t->_lo;
max = (julong)min_jint;
}
julong range = lim - ini + stride_p; if (range <= max) { // Convert to integer expression if it is not overflow.
Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
Node *bias = phase->transform(new AddINode(range, stride_m));
Node *trip = phase->transform(new DivINode(0, bias, in(Stride)));
Node *span = phase->transform(new MulINode(trip, in(Stride))); returnnew AddINode(span, in(Init)); // exact limit
}
if (is_power_of_2(stride_p) || // divisor is 2^n
!Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? // Convert to long expression to avoid integer overflow // and let igvn optimizer convert this division. //
Node* init = phase->transform( new ConvI2LNode(in(Init)));
Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
Node* stride = phase->longcon(stride_con);
Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
Node *range = phase->transform(new SubLNode(limit, init));
Node *bias = phase->transform(new AddLNode(range, stride_m));
Node *span; if (stride_con > 0 && is_power_of_2(stride_p)) { // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) // and avoid generating rounding for division. Zero trip guard should // guarantee that init < limit but sometimes the guard is missing and // we can get situation when init > limit. Note, for the empty loop // optimization zero trip guard is generated explicitly which leaves // only RCE predicate where exact limit is used and the predicate // will simply fail forcing recompilation.
Node* neg_stride = phase->longcon(-stride_con);
span = phase->transform(new AndLNode(bias, neg_stride));
} else {
Node *trip = phase->transform(new DivLNode(0, bias, stride));
span = phase->transform(new MulLNode(trip, stride));
} // Convert back to int
Node *span_int = phase->transform(new ConvL2INode(span)); returnnew AddINode(span_int, in(Init)); // exact limit
}
return NULL; // No progress
}
//------------------------------Identity--------------------------------------- // If stride == 1 return limit node.
Node* LoopLimitNode::Identity(PhaseGVN* phase) { int stride_con = phase->type(in(Stride))->is_int()->get_con(); if (stride_con == 1 || stride_con == -1) return in(Limit); returnthis;
}
//============================================================================= //----------------------match_incr_with_optional_truncation-------------------- // Match increment with optional truncation: // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 // Return NULL for failure. Success returns the increment node.
Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInteger** trunc_type,
BasicType bt) { // Quick cutouts: if (expr == NULL || expr->req() != 3) return NULL;
void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
PhaseIdealLoop* iloop) {
Node* cle_out = inner_cle->proj_out(false);
Node* cle_tail = inner_cle->proj_out(true); if (cle_out->outcnt() > 1) { // Look for chains of stores that were sunk // out of the inner loop and are in the outer loop for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
Node* u = cle_out->fast_out(i); if (u->is_Store()) { int alias_idx = igvn->C->get_alias_index(u->adr_type());
Node* first = u; for (;;) {
Node* next = first->in(MemNode::Memory); if (!next->is_Store() || next->in(0) != cle_out) { break;
}
assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
first = next;
}
Node* last = u; for (;;) {
Node* next = NULL; for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
Node* uu = last->fast_out(j); if (uu->is_Store() && uu->in(0) == cle_out) {
assert(next == NULL, "only one in the outer loop");
next = uu;
assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
}
} if (next == NULL) { break;
}
last = next;
}
Node* phi = NULL; for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
Node* uu = inner_cl->fast_out(j); if (uu->is_Phi()) {
Node* be = uu->in(LoopNode::LoopBackControl); if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
} if (be == last || be == first->in(MemNode::Memory)) {
assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
assert(phi == NULL, "only one phi");
phi = uu;
}
}
} #ifdef ASSERT for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
Node* uu = inner_cl->fast_out(j); if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
assert(phi == uu, "what's that phi?");
} elseif (uu->adr_type() == TypePtr::BOTTOM) {
Node* n = uu->in(LoopNode::LoopBackControl);
uint limit = igvn->C->live_nodes();
uint i = 0; while (n != uu) {
i++;
assert(i < limit, "infinite loop"); if (n->is_Proj()) {
n = n->in(0);
} elseif (n->is_SafePoint() || n->is_MemBar()) {
n = n->in(TypeFunc::Memory);
} elseif (n->is_Phi()) {
n = n->in(1);
} elseif (n->is_MergeMem()) {
n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
} elseif (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
n = n->in(MemNode::Memory);
} else {
n->dump();
ShouldNotReachHere();
}
}
}
}
} #endif if (phi == NULL) { // If an entire chains was sunk, the // inner loop has no phi for that memory // slice, create one for the outer loop
phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
phi->set_req(LoopNode::LoopBackControl, last);
phi = register_new_node(phi, inner_cl, igvn, iloop);
igvn->replace_input_of(first, MemNode::Memory, phi);
} else { // Or fix the outer loop fix to include // that chain of stores.
Node* be = phi->in(LoopNode::LoopBackControl);
assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported"); if (be == first->in(MemNode::Memory)) { if (be == phi->in(LoopNode::LoopBackControl)) {
igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
} else {
igvn->replace_input_of(be, MemNode::Memory, last);
}
} else { #ifdef ASSERT if (be == phi->in(LoopNode::LoopBackControl)) {
assert(phi->in(LoopNode::LoopBackControl) == last, "");
} else {
assert(be->in(MemNode::Memory) == last, "");
} #endif
}
}
}
}
}
}
void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) { // Look for the outer & inner strip mined loop, reduce number of // iterations of the inner loop, set exit condition of outer loop, // construct required phi nodes for outer loop.
CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
assert(inner_cl->is_strip_mined(), "inner loop should be strip mined"); if (LoopStripMiningIter == 0) {
remove_outer_loop_and_safepoint(igvn); return;
} if (LoopStripMiningIter == 1) {
transform_to_counted_loop(igvn, NULL); return;
}
Node* inner_iv_phi = inner_cl->phi(); if (inner_iv_phi == NULL) {
IfNode* outer_le = outer_loop_end();
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
igvn->replace_node(outer_le, iff);
inner_cl->clear_strip_mined(); return;
}
CountedLoopEndNode* inner_cle = inner_cl->loopexit();
int stride = inner_cl->stride_con();
jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride); int scaled_iters = (int)scaled_iters_long; int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride); const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
assert(iter_estimate > 0, "broken"); if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) { // Remove outer loop and safepoint (too few iterations)
remove_outer_loop_and_safepoint(igvn); return;
} if (iter_estimate <= scaled_iters_long) { // We would only go through one iteration of // the outer loop: drop the outer loop but // keep the safepoint so we don't run for // too long without a safepoint
IfNode* outer_le = outer_loop_end();
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
igvn->replace_node(outer_le, iff);
inner_cl->clear_strip_mined(); return;
}
Node* cle_tail = inner_cle->proj_out(true);
ResourceMark rm;
Node_List old_new; if (cle_tail->outcnt() > 1) { // Look for nodes on backedge of inner loop and clone them
Unique_Node_List backedge_nodes; for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
Node* u = cle_tail->fast_out(i); if (u != inner_cl) {
assert(!u->is_CFG(), "control flow on the backedge?");
backedge_nodes.push(u);
}
}
uint last = igvn->C->unique(); for (uint next = 0; next < backedge_nodes.size(); next++) {
Node* n = backedge_nodes.at(next);
old_new.map(n->_idx, n->clone()); for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* u = n->fast_out(i);
assert(!u->is_CFG(), "broken"); if (u->_idx >= last) { continue;
} if (!u->is_Phi()) {
backedge_nodes.push(u);
} else {
assert(u->in(0) == inner_cl, "strange phi on the backedge");
}
}
} // Put the clones on the outer loop backedge
Node* le_tail = outer_loop_tail(); for (uint next = 0; next < backedge_nodes.size(); next++) {
Node *n = old_new[backedge_nodes.at(next)->_idx]; for (uint i = 1; i < n->req(); i++) { if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
n->set_req(i, old_new[n->in(i)->_idx]);
}
} if (n->in(0) != NULL && n->in(0) == cle_tail) {
n->set_req(0, le_tail);
}
igvn->register_new_node_with_optimizer(n);
}
}
Node* iv_phi = NULL; // Make a clone of each phi in the inner loop // for the outer loop for (uint i = 0; i < inner_cl->outcnt(); i++) {
Node* u = inner_cl->raw_out(i); if (u->is_Phi()) {
assert(u->in(0) == inner_cl, "inconsistent");
Node* phi = u->clone();
phi->set_req(0, this);
Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx]; if (be != NULL) {
phi->set_req(LoopNode::LoopBackControl, be);
}
phi = igvn->transform(phi);
igvn->replace_input_of(u, LoopNode::EntryControl, phi); if (u == inner_iv_phi) {
iv_phi = phi;
}
}
}
if (iv_phi != NULL) { // Now adjust the inner loop's exit condition
Node* limit = inner_cl->limit(); // If limit < init for stride > 0 (or limit > init for stride < 0), // the loop body is run only once. Given limit - init (init - limit resp.) // would be negative, the unsigned comparison below would cause // the loop body to be run for LoopStripMiningIter.
Node* max = NULL; if (stride > 0) {
max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
} else {
max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
} // sub is positive and can be larger than the max signed int // value. Use an unsigned min.
Node* const_iters = igvn->intcon(scaled_iters);
Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn); // min is the number of iterations for the next inner loop execution: // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
Node* new_limit = NULL; if (stride > 0) {
new_limit = igvn->transform(new AddINode(min, iv_phi));
} else {
new_limit = igvn->transform(new SubINode(iv_phi, min));
}
Node* inner_cmp = inner_cle->cmp_node();
Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
Node* outer_bol = inner_bol; // cmp node for inner loop may be shared
inner_cmp = inner_cmp->clone();
inner_cmp->set_req(2, new_limit);
inner_bol = inner_bol->clone();
inner_bol->set_req(1, igvn->transform(inner_cmp));
igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol)); // Set the outer loop's exit condition too
igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
} else {
assert(false, "should be able to adjust outer loop");
IfNode* outer_le = outer_loop_end();
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
igvn->replace_node(outer_le, iff);
inner_cl->clear_strip_mined();
}
}
// make counted loop exit test always fail
ConINode* zero = igvn->intcon(0); if (iloop != NULL) {
iloop->set_ctrl(zero, igvn->C->root());
}
igvn->replace_input_of(cle, 1, zero); // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop); if (iloop == NULL) {
igvn->replace_node(outer_le, new_end);
} else {
iloop->lazy_replace(outer_le, new_end);
} // the backedge of the inner loop must be rewired to the new loop end
Node* backedge = cle->proj_out(true);
igvn->replace_input_of(backedge, 0, new_end); if (iloop != NULL) {
iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
} // make the outer loop go away
igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
inner_cl->clear_strip_mined(); if (iloop != NULL) {
Unique_Node_List wq;
wq.push(safepoint);
const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { if (!in(0)) return Type::TOP; if (phase->type(in(0)) == Type::TOP) return Type::TOP;
// Until expansion, the loop end condition is not set so this should not constant fold. if (is_expanded(phase)) { return IfNode::Value(phase);
}
return TypeTuple::IFBOTH;
}
bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const { // The outer strip mined loop head only has Phi uses after expansion if (phase->is_IterGVN()) {
Node* backedge = proj_out_or_null(true); if (backedge != NULL) {
Node* head = backedge->unique_ctrl_out_or_null(); if (head != NULL && head->is_OuterStripMinedLoop()) { if (head->find_out_with(Op_Phi) != NULL) { returntrue;
}
}
}
} returnfalse;
}
Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { if (remove_dead_region(phase, can_reshape)) returnthis;
return NULL;
}
//------------------------------filtered_type-------------------------------- // Return a type based on condition control flow // A successful return will be a type that is restricted due // to a series of dominating if-tests, such as: // if (i < 10) { // if (i > 0) { // here: "i" type is [1..10) // } // } // or a control flow merge // if (i < 10) { // do { // phi( , ) -- at top of loop type is [min_int..10) // i = ? // } while ( i < 10) // const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
assert(n && n->bottom_type()->is_int(), "must be int"); const TypeInt* filtered_t = NULL; if (!n->is_Phi()) {
assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
filtered_t = filtered_type_from_dominators(n, n_ctrl);
} else {
Node* phi = n->as_Phi();
Node* region = phi->in(0);
assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); if (region && region != C->top()) { for (uint i = 1; i < phi->req(); i++) {
Node* val = phi->in(i);
Node* use_c = region->in(i); const TypeInt* val_t = filtered_type_from_dominators(val, use_c); if (val_t != NULL) { if (filtered_t == NULL) {
filtered_t = val_t;
} else {
filtered_t = filtered_t->meet(val_t)->is_int();
}
}
}
}
} const TypeInt* n_t = _igvn.type(n)->is_int(); if (filtered_t != NULL) {
n_t = n_t->join(filtered_t)->is_int();
} return n_t;
}
//------------------------------filtered_type_from_dominators-------------------------------- // Return a possibly more restrictive type for val based on condition control flow of dominators const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { if (val->is_Con()) { return val->bottom_type()->is_int();
}
uint if_limit = 10; // Max number of dominating if's visited const TypeInt* rtn_t = NULL;
if (use_ctrl && use_ctrl != C->top()) {
Node* val_ctrl = get_ctrl(val);
uint val_dom_depth = dom_depth(val_ctrl);
Node* pred = use_ctrl;
uint if_cnt = 0; while (if_cnt < if_limit) { if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
if_cnt++; const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); if (if_t != NULL) { if (rtn_t == NULL) {
rtn_t = if_t;
} else {
rtn_t = rtn_t->join(if_t)->is_int();
}
}
}
pred = idom(pred); if (pred == NULL || pred == C->top()) { break;
} // Stop if going beyond definition block of val if (dom_depth(pred) < val_dom_depth) { break;
}
}
} return rtn_t;
}
//------------------------------dump_spec-------------------------------------- // Dump special per-node info #ifndef PRODUCT void CountedLoopEndNode::dump_spec(outputStream *st) const { if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
BoolTest bt( test_trip()); // Added this for g++.
//============================================================================= //------------------------------is_member-------------------------------------- // Is 'l' a member of 'this'? bool IdealLoopTree::is_member(const IdealLoopTree *l) const { while( l->_nest > _nest ) l = l->_parent; return l == this;
}
//------------------------------split_fall_in---------------------------------- // Split out multiple fall-in edges from the loop header. Move them to a // private RegionNode before the loop. This becomes the loop landing pad. void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
PhaseIterGVN &igvn = phase->_igvn;
uint i;
// Make a new RegionNode to be the landing pad.
Node *landing_pad = new RegionNode( fall_in_cnt+1 );
phase->set_loop(landing_pad,_parent); // Gather all the fall-in control paths into the landing pad
uint icnt = fall_in_cnt;
uint oreq = _head->req(); for( i = oreq-1; i>0; i-- ) if( !phase->is_member( this, _head->in(i) ) )
landing_pad->set_req(icnt--,_head->in(i));
// Peel off PhiNode edges as well for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *oj = _head->fast_out(j); if( oj->is_Phi() ) {
PhiNode* old_phi = oj->as_Phi();
assert( old_phi->region() == _head, "" );
igvn.hash_delete(old_phi); // Yank from hash before hacking edges
Node *p = PhiNode::make_blank(landing_pad, old_phi);
uint icnt = fall_in_cnt; for( i = oreq-1; i>0; i-- ) { if( !phase->is_member( this, _head->in(i) ) ) {
p->init_req(icnt--, old_phi->in(i)); // Go ahead and clean out old edges from old phi
old_phi->del_req(i);
}
} // Search for CSE's here, because ZKM.jar does a lot of // loop hackery and we need to be a little incremental // with the CSE to avoid O(N^2) node blow-up.
Node *p2 = igvn.hash_find_insert(p); // Look for a CSE if( p2 ) { // Found CSE
p->destruct(&igvn); // Recover useless new node
p = p2; // Use old node
} else {
igvn.register_new_node_with_optimizer(p, old_phi);
} // Make old Phi refer to new Phi.
old_phi->add_req(p); // Check for the special case of making the old phi useless and // disappear it. In JavaGrande I have a case where this useless // Phi is the loop limit and prevents recognizing a CountedLoop // which in turn prevents removing an empty loop.
Node *id_old_phi = old_phi->Identity(&igvn); if( id_old_phi != old_phi ) { // Found a simple identity? // Note that I cannot call 'replace_node' here, because // that will yank the edge from old_phi to the Region and // I'm mid-iteration over the Region's uses. for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
Node* use = old_phi->last_out(i);
igvn.rehash_node_delayed(use);
uint uses_found = 0; for (uint j = 0; j < use->len(); j++) { if (use->in(j) == old_phi) { if (j < use->req()) use->set_req (j, id_old_phi); else use->set_prec(j, id_old_phi);
uses_found++;
}
}
i -= uses_found; // we deleted 1 or more copies of this edge
}
}
igvn._worklist.push(old_phi);
}
} // Finally clean out the fall-in edges from the RegionNode for( i = oreq-1; i>0; i-- ) { if( !phase->is_member( this, _head->in(i) ) ) {
_head->del_req(i);
}
}
igvn.rehash_node_delayed(_head); // Transform landing pad
igvn.register_new_node_with_optimizer(landing_pad, _head); // Insert landing pad into the header
_head->add_req(landing_pad);
}
//------------------------------split_outer_loop------------------------------- // Split out the outermost loop from this shared header. void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
PhaseIterGVN &igvn = phase->_igvn;
// Find index of outermost loop; it should also be my tail.
uint outer_idx = 1; while( _head->in(outer_idx) != _tail ) outer_idx++;
// Make a LoopNode for the outermost loop.
Node *ctl = _head->in(LoopNode::EntryControl);
Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
outer = igvn.register_new_node_with_optimizer(outer, _head);
phase->set_created_loop_node();
// Outermost loop falls into '_head' loop
_head->set_req(LoopNode::EntryControl, outer);
_head->del_req(outer_idx); // Split all the Phis up between '_head' loop and 'outer' loop. for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *out = _head->fast_out(j); if( out->is_Phi() ) {
PhiNode *old_phi = out->as_Phi();
assert( old_phi->region() == _head, "" );
Node *phi = PhiNode::make_blank(outer, old_phi);
phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
phi = igvn.register_new_node_with_optimizer(phi, old_phi); // Make old Phi point to new Phi on the fall-in path
igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
old_phi->del_req(outer_idx);
}
}
// Use the new loop head instead of the old shared one
_head = outer;
phase->set_loop(_head, this);
}
//------------------------------estimate_path_freq----------------------------- staticfloat estimate_path_freq( Node *n ) { // Try to extract some path frequency info
IfNode *iff; for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
uint nop = n->Opcode(); if( nop == Op_SafePoint ) { // Skip any safepoint
n = n->in(0); continue;
} if( nop == Op_CatchProj ) { // Get count from a prior call // Assume call does not always throw exceptions: means the call-site // count is also the frequency of the fall-through path.
assert( n->is_CatchProj(), "" ); if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) return 0.0f; // Assume call exception path is rare
Node *call = n->in(0)->in(0)->in(0);
assert( call->is_Call(), "expect a call here" ); const JVMState *jvms = ((CallNode*)call)->jvms();
ciMethodData* methodData = jvms->method()->method_data(); if (!methodData->is_mature()) return 0.0f; // No call-site data
ciProfileData* data = methodData->bci_to_data(jvms->bci()); if ((data == NULL) || !data->is_CounterData()) { // no call profile available, try call's control input
n = n->in(0); continue;
} return data->as_CounterData()->count()/FreqCountInvocations;
} // See if there's a gating IF test
Node *n_c = n->in(0); if( !n_c->is_If() ) break; // No estimate available
iff = n_c->as_If(); if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? // Compute how much count comes on this path return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; // Have no count info. Skip dull uncommon-trap like branches. if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
(nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) break; // Skip through never-taken branch; look for a real loop exit.
n = iff->in(0);
} return 0.0f; // No estimate available
}
//------------------------------merge_many_backedges--------------------------- // Merge all the backedges from the shared header into a private Region. // Feed that region as the one backedge to this loop. void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
uint i;
// Scan for the top 2 hottest backedges float hotcnt = 0.0f; float warmcnt = 0.0f;
uint hot_idx = 0; // Loop starts at 2 because slot 1 is the fall-in path for( i = 2; i < _head->req(); i++ ) { float cnt = estimate_path_freq(_head->in(i)); if( cnt > hotcnt ) { // Grab hottest path
warmcnt = hotcnt;
hotcnt = cnt;
hot_idx = i;
} elseif( cnt > warmcnt ) { // And 2nd hottest path
warmcnt = cnt;
}
}
// See if the hottest backedge is worthy of being an inner loop // by being much hotter than the next hottest backedge. if( hotcnt <= 0.0001 ||
hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
// Peel out the backedges into a private merge point; peel // them all except optionally hot_idx.
PhaseIterGVN &igvn = phase->_igvn;
Node *hot_tail = NULL; // Make a Region for the merge point
Node *r = new RegionNode(1); for( i = 2; i < _head->req(); i++ ) { if( i != hot_idx )
r->add_req( _head->in(i) ); else hot_tail = _head->in(i);
}
igvn.register_new_node_with_optimizer(r, _head); // Plug region into end of loop _head, followed by hot_tail while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
igvn.replace_input_of(_head, 2, r); if( hot_idx ) _head->add_req(hot_tail);
// Split all the Phis up between '_head' loop and the Region 'r' for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
Node *out = _head->fast_out(j); if( out->is_Phi() ) {
PhiNode* n = out->as_Phi();
igvn.hash_delete(n); // Delete from hash before hacking edges
Node *hot_phi = NULL;
Node *phi = new PhiNode(r, n->type(), n->adr_type()); // Check all inputs for the ones to peel out
uint j = 1; for( uint i = 2; i < n->req(); i++ ) { if( i != hot_idx )
phi->set_req( j++, n->in(i) ); else hot_phi = n->in(i);
} // Register the phi but do not transform until whole place transforms
igvn.register_new_node_with_optimizer(phi, n); // Add the merge phi to the old Phi while( n->req() > 3 ) n->del_req( n->req()-1 );
igvn.replace_input_of(n, 2, phi); if( hot_idx ) n->add_req(hot_phi);
}
}
// Insert a new IdealLoopTree inserted below me. Turn it into a clone // of self loop tree. Turn self into a loop headed by _head and with // tail being the new merge point.
IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
phase->set_loop(_tail,ilt); // Adjust tail
_tail = r; // Self's tail is new merge point
phase->set_loop(r,this);
ilt->_child = _child; // New guy has my children
_child = ilt; // Self has new guy as only child
ilt->_parent = this; // new guy has self for parent
ilt->_nest = _nest; // Same nesting depth (for now)
// Starting with 'ilt', look for child loop trees using the same shared // header. Flatten these out; they will no longer be loops in the end.
IdealLoopTree **pilt = &_child; while( ilt ) { if( ilt->_head == _head ) {
uint i; for( i = 2; i < _head->req(); i++ ) if( _head->in(i) == ilt->_tail ) break; // Still a loop if( i == _head->req() ) { // No longer a loop // Flatten ilt. Hang ilt's "_next" list from the end of // ilt's '_child' list. Move the ilt's _child up to replace ilt.
IdealLoopTree **cp = &ilt->_child; while( *cp ) cp = &(*cp)->_next; // Find end of child list
*cp = ilt->_next; // Hang next list at end of child list
*pilt = ilt->_child; // Move child up to replace ilt
ilt->_head = NULL; // Flag as a loop UNIONED into parent
ilt = ilt->_child; // Repeat using new ilt continue; // do not advance over ilt->_child
}
assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
phase->set_loop(_head,ilt);
}
pilt = &ilt->_child; // Advance to next
ilt = *pilt;
}
if( _child ) fix_parent( _child, this );
}
//------------------------------beautify_loops--------------------------------- // Split shared headers and insert loop landing pads. // Insert a LoopNode to replace the RegionNode. // Return TRUE if loop tree is structurally changed. bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { bool result = false; // Cache parts in locals for easy
PhaseIterGVN &igvn = phase->_igvn;
igvn.hash_delete(_head); // Yank from hash before hacking edges
// Check for multiple fall-in paths. Peel off a landing pad if need be. int fall_in_cnt = 0; for( uint i = 1; i < _head->req(); i++ ) if( !phase->is_member( this, _head->in(i) ) )
fall_in_cnt++;
assert( fall_in_cnt, "at least 1 fall-in path" ); if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
split_fall_in( phase, fall_in_cnt );
// Swap inputs to the _head and all Phis to move the fall-in edge to // the left.
fall_in_cnt = 1; while( phase->is_member( this, _head->in(fall_in_cnt) ) )
fall_in_cnt++; if( fall_in_cnt > 1 ) { // Since I am just swapping inputs I do not need to update def-use info
Node *tmp = _head->in(1);
igvn.rehash_node_delayed(_head);
_head->set_req( 1, _head->in(fall_in_cnt) );
_head->set_req( fall_in_cnt, tmp ); // Swap also all Phis for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
Node* phi = _head->fast_out(i); if( phi->is_Phi() ) {
igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
tmp = phi->in(1);
phi->set_req( 1, phi->in(fall_in_cnt) );
phi->set_req( fall_in_cnt, tmp );
}
}
}
assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
// If I am a shared header (multiple backedges), peel off the many // backedges into a private merge point and use the merge point as // the one true backedge. if (_head->req() > 3) { // Merge the many backedges into a single backedge but leave // the hottest backedge as separate edge for the following peel. if (!_irreducible) {
merge_many_backedges( phase );
}
// When recursively beautify my children, split_fall_in can change // loop tree structure when I am an irreducible loop. Then the head // of my children has a req() not bigger than 3. Here we need to set // result to true to catch that case in order to tell the caller to // rebuild loop tree. See issue JDK-8244407 for details.
result = true;
}
// If I have one hot backedge, peel off myself loop. // I better be the outermost loop. if (_head->req() > 3 && !_irreducible) {
split_outer_loop( phase );
result = true;
} elseif (!_head->is_Loop() && !_irreducible) { // Make a new LoopNode to replace the old loop head
Node *l = new LoopNode( _head->in(1), _head->in(2) );
l = igvn.register_new_node_with_optimizer(l, _head);
phase->set_created_loop_node(); // Go ahead and replace _head
phase->_igvn.replace_node( _head, l );
_head = l;
phase->set_loop(_head, this);
}
// Now recursively beautify nested loops if( _child ) result |= _child->beautify_loops( phase ); if( _next ) result |= _next ->beautify_loops( phase ); return result;
}
//------------------------------allpaths_check_safepts---------------------------- // Allpaths backwards scan from loop tail, terminating each path at first safepoint // encountered. Helper for check_safepts. void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
assert(stack.size() == 0, "empty stack");
stack.push(_tail);
visited.clear();
visited.set(_tail->_idx); while (stack.size() > 0) {
Node* n = stack.pop(); if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { // Terminate this path
} elseif (n->Opcode() == Op_SafePoint) { if (_phase->get_loop(n) != this) { if (_required_safept == NULL) _required_safept = new Node_List();
_required_safept->push(n); // save the one closest to the tail
} // Terminate this path
} else {
uint start = n->is_Region() ? 1 : 0;
uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; for (uint i = start; i < end; i++) {
Node* in = n->in(i);
assert(in->is_CFG(), "must be"); if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
stack.push(in);
}
}
}
}
}
//------------------------------check_safepts---------------------------- // Given dominators, try to find loops with calls that must always be // executed (call dominates loop tail). These loops do not need non-call // safepoints (ncsfpt). // // A complication is that a safepoint in a inner loop may be needed // by an outer loop. In the following, the inner loop sees it has a // call (block 3) on every path from the head (block 2) to the // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) // in block 2, _but_ this leaves the outer loop without a safepoint. // // entry 0 // | // v // outer 1,2 +->1 // | | // | v // | 2<---+ ncsfpt in 2 // |_/|\ | // | v | // inner 2,3 / 3 | call in 3 // / | | // v +--+ // exit 4 // // // This method creates a list (_required_safept) of ncsfpt nodes that must // be protected is created for each loop. When a ncsfpt maybe deleted, it // is first looked for in the lists for the outer loops of the current loop. // // The insights into the problem: // A) counted loops are okay // B) innermost loops are okay (only an inner loop can delete // a ncsfpt needed by an outer loop) // C) a loop is immune from an inner loop deleting a safepoint // if the loop has a call on the idom-path // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the // idom-path that is not in a nested loop // E) otherwise, an ncsfpt on the idom-path that is nested in an inner // loop needs to be prevented from deletion by an inner loop // // There are two analyses: // 1) The first, and cheaper one, scans the loop body from // tail to head following the idom (immediate dominator) // chain, looking for the cases (C,D,E) above. // Since inner loops are scanned before outer loops, there is summary // information about inner loops. Inner loops can be skipped over // when the tail of an inner loop is encountered. // // 2) The second, invoked if the first fails to find a call or ncsfpt on // the idom path (which is rare), scans all predecessor control paths // from the tail to the head, terminating a path when a call or sfpt // is encountered, to find the ncsfpt's that are closest to the tail. // void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { // Bottom up traversal
IdealLoopTree* ch = _child; if (_child) _child->check_safepts(visited, stack); if (_next) _next ->check_safepts(visited, stack);
if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { bool has_call = false; // call on dom-path bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth // Scan the dom-path nodes from tail to head for (Node* n = tail(); n != _head; n = _phase->idom(n)) { if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
has_call = true;
_has_sfpt = 1; // Then no need for a safept! break;
} elseif (n->Opcode() == Op_SafePoint) { if (_phase->get_loop(n) == this) {
has_local_ncsfpt = true; break;
} if (nonlocal_ncsfpt == NULL) {
nonlocal_ncsfpt = n; // save the one closest to the tail
}
} else {
IdealLoopTree* nlpt = _phase->get_loop(n); if (this != nlpt) { // If at an inner loop tail, see if the inner loop has already // recorded seeing a call on the dom-path (and stop.) If not, // jump to the head of the inner loop.
assert(is_member(nlpt), "nested loop");
Node* tail = nlpt->_tail; if (tail->in(0)->is_If()) tail = tail->in(0); if (n == tail) { // If inner loop has call on dom-path, so does outer loop if (nlpt->_has_sfpt) {
has_call = true;
_has_sfpt = 1; break;
} // Skip to head of inner loop
assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
n = nlpt->_head;
}
}
}
} // Record safept's that this loop needs preserved when an // inner loop attempts to delete it's safepoints. if (_child != NULL && !has_call && !has_local_ncsfpt) { if (nonlocal_ncsfpt != NULL) { if (_required_safept == NULL) _required_safept = new Node_List();
_required_safept->push(nonlocal_ncsfpt);
} else { // Failed to find a suitable safept on the dom-path. Now use // an all paths walk from tail to head, looking for safepoints to preserve.
allpaths_check_safepts(visited, stack);
}
}
}
}
//---------------------------is_deleteable_safept---------------------------- // Is safept not required by an outer loop? bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
assert(sfpt->Opcode() == Op_SafePoint, "");
IdealLoopTree* lp = get_loop(sfpt)->_parent; while (lp != NULL) {
Node_List* sfpts = lp->_required_safept; if (sfpts != NULL) { for (uint i = 0; i < sfpts->size(); i++) { if (sfpt == sfpts->at(i)) returnfalse;
}
}
lp = lp->_parent;
} returntrue;
}
// Visit all children, looking for Phis for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
Node *out = cl->out(i); // Look for other phis (secondary IVs). Skip dead ones if (!out->is_Phi() || out == phi || !has_node(out)) { continue;
}
PhiNode* phi2 = out->as_Phi();
Node* incr2 = phi2->in(LoopNode::LoopBackControl); // Look for induction variables of the form: X += constant if (phi2->region() != loop->_head ||
incr2->req() != 3 ||
incr2->in(1)->uncast() != phi2 ||
incr2 == incr ||
incr2->Opcode() != Op_AddI ||
!incr2->in(2)->is_Con()) { continue;
}
if (incr2->in(1)->is_ConstraintCast() &&
!(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) { // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck continue;
} // Check for parallel induction variable (parallel to trip counter) // via an affine function. In particular, count-down loops with // count-up array indices are common. We only RCE references off // the trip-counter, so we need to convert all these to trip-counter // expressions.
Node* init2 = phi2->in(LoopNode::EntryControl); int stride_con2 = incr2->in(2)->get_int();
// The ratio of the two strides cannot be represented as an int // if stride_con2 is min_int and stride_con is -1. if (stride_con2 == min_jint && stride_con == -1) { continue;
}
// The general case here gets a little tricky. We want to find the // GCD of all possible parallel IV's and make a new IV using this // GCD for the loop. Then all possible IVs are simple multiples of // the GCD. In practice, this will cover very few extra loops. // Instead we require 'stride_con2' to be a multiple of 'stride_con', // where +/-1 is the common case, but other integer multiples are // also easy to handle. int ratio_con = stride_con2/stride_con;
if ((ratio_con * stride_con) == stride_con2) { // Check for exact #ifndef PRODUCT if (TraceLoopOpts) {
tty->print("Parallel IV: %d ", phi2->_idx);
loop->dump_head();
} #endif // Convert to using the trip counter. The parallel induction // variable differs from the trip counter by a loop-invariant // amount, the difference between their respective initial values. // It is scaled by the 'ratio_con'.
Node* ratio = _igvn.intcon(ratio_con);
set_ctrl(ratio, C->root());
Node* ratio_init = new MulINode(init, ratio);
_igvn.register_new_node_with_optimizer(ratio_init, init);
set_early_ctrl(ratio_init, false);
Node* diff = new SubINode(init2, ratio_init);
_igvn.register_new_node_with_optimizer(diff, init2);
set_early_ctrl(diff, false);
Node* ratio_idx = new MulINode(phi, ratio);
_igvn.register_new_node_with_optimizer(ratio_idx, phi);
set_ctrl(ratio_idx, cl);
Node* add = new AddINode(ratio_idx, diff);
_igvn.register_new_node_with_optimizer(add);
set_ctrl(add, cl);
_igvn.replace_node( phi2, add ); // Sometimes an induction variable is unused if (add->outcnt() == 0) {
_igvn.remove_dead_node(add);
}
--i; // deleted this phi; rescan starting with next position continue;
}
}
}
void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
Node* keep = NULL; if (keep_one) { // Look for a safepoint on the idom-path. for (Node* i = tail(); i != _head; i = phase->idom(i)) { if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
keep = i; break; // Found one
}
}
}
// Don't remove any safepoints if it is requested to keep a single safepoint and // no safepoint was found on idom-path. It is not safe to remove any safepoint // in this case since there's no safepoint dominating all paths in the loop body. bool prune = !keep_one || keep != NULL;
// Delete other safepoints in this loop.
Node_List* sfpts = _safepts; if (prune && sfpts != NULL) {
assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); for (uint i = 0; i < sfpts->size(); i++) {
Node* n = sfpts->at(i);
assert(phase->get_loop(n) == this, ""); if (n != keep && phase->is_deleteable_safept(n)) {
phase->lazy_replace(n, n->in(TypeFunc::Control));
}
}
}
}
//------------------------------counted_loop----------------------------------- // Convert to counted loops where possible void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
// For grins, set the inner-loop flag here if (!_child) { if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
}
// Look for induction variables
phase->replace_parallel_iv(this);
} elseif (_head->is_LongCountedLoop() ||
phase->is_counted_loop(_head, loop, T_LONG)) {
remove_safepoints(phase, true);
} else {
assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail"); if (_parent != NULL && !_irreducible) { // Not a counted loop. Keep one safepoint. bool keep_one_sfpt = true;
remove_safepoints(phase, keep_one_sfpt);
}
}
// Recursively
assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); if (loop->_child && loop->_child != this) loop->_child->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 {
// The Estimated Loop (full-) Unroll Size: // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm, // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that // node usage estimates made are on the safe side, for the most part. This is // a "light" version of the loop clone size calculation (above), based on the // assumption that most of the loop-construct overhead will be unraveled when // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1), // including an overflow check and returning UINT_MAX in case of an overflow.
uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
precond(factor > 0);
// Take into account that after unroll conjoined heads and tails will fold.
uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
uint const cc = 7;
uint const sz = b0 + (b0 + 15) / 16;
uint estimate = factor * sz + cc;
// Estimate the growth effect (in nodes) of merging control and data flow when // cloning a loop body, based on the amount of control and data flow reaching // outside of the (current) loop body.
uint IdealLoopTree::est_loop_flow_merge_sz() const {
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 == NULL) continue; if (out->is_CFG()) { if (!is_member(_phase->get_loop(out))) {
ctrl_edge_out_cnt++;
}
} elseif (_phase->has_ctrl(out)) {
Node* ctrl = _phase->get_ctrl(out);
assert(ctrl != NULL, "must be");
assert(ctrl->is_CFG(), "must be"); if (!is_member(_phase->get_loop(ctrl))) {
data_edge_out_cnt++;
}
}
}
} // Use data and control count (x2.0) in estimate iff both are > 0. This is // a rather pessimistic estimate for the most part, in particular for some // complex loops, but still not enough to capture all loops. if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) { return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
} return 0;
}
if (cl->is_pre_loop ()) tty->print(" pre" ); if (cl->is_main_loop()) tty->print(" main"); if (cl->is_post_loop()) tty->print(" post"); if (cl->is_reduction_loop()) tty->print(" reduction"); if (cl->is_vectorized_loop()) tty->print(" vector"); if (cl->range_checks_present()) tty->print(" rc "); if (cl->is_multiversioned()) tty->print(" multi ");
} if (_has_call) tty->print(" has_call"); if (_has_sfpt) tty->print(" has_sfpt"); if (_rce_candidate) tty->print(" rce"); if (_safepts != NULL && _safepts->size() > 0) {
tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
} if (_required_safept != NULL && _required_safept->size() > 0) {
tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
} if (Verbose) {
tty->print(" body={"); _body.dump_simple(); tty->print(" }");
} if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
tty->print(" strip_mined");
}
tty->cr();
}
//------------------------------dump------------------------------------------- // Dump loops by loop tree void IdealLoopTree::dump() const {
dump_head(); if (_child) _child->dump(); if (_next) _next ->dump();
}
#endif
staticvoid log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { if (loop == root) { if (loop->_child != NULL) {
log->begin_head("loop_tree");
log->end_head();
log_loop_tree_helper(root, loop->_child, log);
log->tail("loop_tree");
assert(loop->_next == NULL, "what?");
}
} elseif (loop != NULL) {
Node* head = loop->_head;
log->begin_head("loop");
log->print(" idx='%d' ", head->_idx); if (loop->_irreducible) log->print("irreducible='1' "); if (head->is_Loop()) { if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
} elseif (head->is_CountedLoop()) {
CountedLoopNode* cl = head->as_CountedLoop(); if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
}
log->end_head();
log_loop_tree_helper(root, loop->_child, log);
log->tail("loop");
log_loop_tree_helper(root, loop->_next, log);
}
}
//---------------------collect_potentially_useful_predicates----------------------- // Helper function to collect potentially useful predicates to prevent them from // being eliminated by PhaseIdealLoop::eliminate_useless_predicates void PhaseIdealLoop::collect_potentially_useful_predicates(IdealLoopTree* loop, Unique_Node_List &useful_predicates) { if (loop->_child) { // child
collect_potentially_useful_predicates(loop->_child, useful_predicates);
}
// self (only loops that we can apply loop predication may use their predicates) if (loop->_head->is_Loop() &&
!loop->_irreducible &&
!loop->tail()->is_top()) {
LoopNode* lpn = loop->_head->as_Loop();
Node* entry = lpn->in(LoopNode::EntryControl);
Node* predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); if (predicate != NULL) { // right pattern that can be used by loop predication
assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
entry = skip_loop_predicates(entry);
} if (UseProfiledLoopPredicate) {
predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); if (predicate != NULL) { // right pattern that can be used by loop predication
useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
get_skeleton_predicates(entry, useful_predicates, true);
entry = skip_loop_predicates(entry);
}
}
if (UseLoopPredicate) {
predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); if (predicate != NULL) { // right pattern that can be used by loop predication
useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
get_skeleton_predicates(entry, useful_predicates, true);
}
}
}
if (loop->_next) { // sibling
collect_potentially_useful_predicates(loop->_next, useful_predicates);
}
}
//------------------------eliminate_useless_predicates----------------------------- // Eliminate all inserted predicates if they could not be used by loop predication. // Note: it will also eliminates loop limits check predicate since it also uses // Opaque1 node (see Parse::add_predicate()). void PhaseIdealLoop::eliminate_useless_predicates() { if (C->predicate_count() == 0 && C->skeleton_predicate_count() == 0) { return; // no predicate left
}
Unique_Node_List useful_predicates; // to store useful predicates if (C->has_loops()) {
collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
}
for (int i = C->predicate_count(); i > 0; i--) {
Node* n = C->predicate_opaque1_node(i - 1);
assert(n->Opcode() == Op_Opaque1, "must be"); if (!useful_predicates.member(n)) { // not in the useful list
_igvn.replace_node(n, n->in(1));
}
}
for (int i = C->skeleton_predicate_count(); i > 0; i--) {
Node* n = C->skeleton_predicate_opaque4_node(i - 1);
assert(n->Opcode() == Op_Opaque4, "must be"); if (!useful_predicates.member(n)) { // not in the useful list
_igvn.replace_node(n, n->in(2));
}
}
}
//------------------------process_expensive_nodes----------------------------- // Expensive nodes have their control input set to prevent the GVN // from commoning them and as a result forcing the resulting node to // be in a more frequent path. Use CFG information here, to change the // control inputs so that some expensive nodes can be commoned while // not executed more frequently. bool PhaseIdealLoop::process_expensive_nodes() {
assert(OptimizeExpensiveOps, "optimization off?");
// Sort nodes to bring similar nodes together
C->sort_expensive_nodes();
bool progress = false;
for (int i = 0; i < C->expensive_count(); ) {
Node* n = C->expensive_node(i); int start = i; // Find nodes similar to n
i++; for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); int end = i; // And compare them two by two for (int j = start; j < end; j++) {
Node* n1 = C->expensive_node(j); if (is_node_unreachable(n1)) { continue;
} for (int k = j+1; k < end; k++) {
Node* n2 = C->expensive_node(k); if (is_node_unreachable(n2)) { continue;
}
assert(n1 != n2, "should be pair of nodes");
Node* c1 = n1->in(0);
Node* c2 = n2->in(0);
Node* parent_c1 = c1;
Node* parent_c2 = c2;
// The call to get_early_ctrl_for_expensive() moves the // expensive nodes up but stops at loops that are in a if // branch. See whether we can exit the loop and move above the // If. if (c1->is_Loop()) {
parent_c1 = c1->in(1);
} if (c2->is_Loop()) {
parent_c2 = c2->in(1);
}
if (parent_c1 == parent_c2) {
_igvn._worklist.push(n1);
_igvn._worklist.push(n2); continue;
}
// Look for identical expensive node up the dominator chain. if (is_dominator(c1, c2)) {
c2 = c1;
} elseif (is_dominator(c2, c1)) {
c1 = c2;
} elseif (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { // Both branches have the same expensive node so move it up // before the if.
c1 = c2 = idom(parent_c1->in(0));
} // Do the actual moves if (n1->in(0) != c1) {
_igvn.replace_input_of(n1, 0, c1);
progress = true;
} if (n2->in(0) != c2) {
_igvn.replace_input_of(n2, 0, c2);
progress = true;
}
}
}
}
return progress;
}
#ifdef ASSERT // Goes over all children of the root of the loop tree. Check if any of them have a path // down to Root, that does not go via a NeverBranch exit. bool PhaseIdealLoop::only_has_infinite_loops() {
ResourceMark rm;
Unique_Node_List worklist; // start traversal at all loop heads of first-level loops for (IdealLoopTree* l = _ltree_root->_child; l != NULL; l = l->_next) {
Node* head = l->_head;
assert(head->is_Region(), "");
worklist.push(head);
} // BFS traversal down the CFG, except through NeverBranch exits for (uint i = 0; i < worklist.size(); ++i) {
Node* n = worklist.at(i);
assert(n->is_CFG(), "only traverse CFG"); if (n->is_Root()) { // Found root -> there was an exit! returnfalse;
} elseif (n->Opcode() == Op_NeverBranch) { // Only follow the loop-internal projection, not the NeverBranch exit
ProjNode* proj = n->as_Multi()->proj_out_or_null(0);
assert(proj != nullptr, "must find loop-internal projection of NeverBranch");
worklist.push(proj);
} else { // Traverse all CFG outputs for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* use = n->fast_out(i); if (use->is_CFG()) {
worklist.push(use);
}
}
}
} // No exit found for any loop -> all are infinite returntrue;
} #endif
//============================================================================= //----------------------------build_and_optimize------------------------------- // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. void PhaseIdealLoop::build_and_optimize() {
assert(!C->post_loop_opts_phase(), "no loop opts allowed");
int old_progress = C->major_progress();
uint orig_worklist_size = _igvn._worklist.size();
// Reset major-progress flag for the driver's heuristics
C->clear_major_progress();
#ifndef PRODUCT // Capture for later assert
uint unique = C->unique();
_loop_invokes++;
_loop_work += unique; #endif
// True if the method has at least 1 irreducible loop
_has_irreducible_loops = false;
_created_loop_node = false;
VectorSet visited; // Pre-grow the mapping from Nodes to IdealLoopTrees.
_nodes.map(C->unique(), NULL);
memset(_nodes.adr(), 0, wordSize * C->unique());
// Pre-build the top-level outermost loop tree entry
_ltree_root = new IdealLoopTree( this, C->root(), C->root() ); // Do not need a safepoint at the top level
_ltree_root->_has_sfpt = 1;
// Build a loop tree on the fly. Build a mapping from CFG nodes to // IdealLoopTree entries. Data nodes are NOT walked.
build_loop_tree(); // Check for bailout, and return if (C->failing()) { return;
}
// Verify that the has_loops() flag set at parse time is consistent // with the just built loop tree. With infinite loops, it could be // that one pass of loop opts only finds infinite loops, clears the // has_loops() flag but adds NeverBranch nodes so the next loop opts // verification pass finds a non empty loop tree. When the back edge // is an exception edge, parsing doesn't set has_loops().
assert(_ltree_root->_child == NULL || C->has_loops() || only_has_infinite_loops() || C->has_exception_backedge(), "parsing found no loops but there are some"); // No loops after all if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
// There should always be an outer loop containing the Root and Return nodes. // If not, we have a degenerate empty program. Bail out in this case. if (!has_node(C->root())) { if (!_verify_only) {
C->clear_major_progress();
C->record_method_not_compilable("empty program detected during loop optimization");
} return;
}
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); // Nothing to do, so get out bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
!_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode); bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode); if (stop_early && !do_expensive_nodes) { return;
}
// Set loop nesting depth
_ltree_root->set_nest( 0 );
// Split shared headers and insert loop landing pads. // Do not bother doing this on the Root loop of course. if( !_verify_me && !_verify_only && _ltree_root->_child ) {
C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); if( _ltree_root->_child->beautify_loops( this ) ) { // Re-build loop tree!
_ltree_root->_child = NULL;
_nodes.clear();
reallocate_preorders();
build_loop_tree(); // Check for bailout, and return if (C->failing()) { return;
} // Reset loop nesting depth
_ltree_root->set_nest( 0 );
// Build Dominators for elision of NULL checks & loop finding. // Since nodes do not have a slot for immediate dominator, make // a persistent side array for that info indexed on node->_idx.
_idom_size = C->unique();
_idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
_dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
_dom_stk = NULL; // Allocated on demand in recompute_dom_depth
memset( _dom_depth, 0, _idom_size * sizeof(uint) );
Dominators();
if (!_verify_only) { // As a side effect, Dominators removed any unreachable CFG paths // into RegionNodes. It doesn't do this test against Root, so // we do it here. for( uint i = 1; i < C->root()->req(); i++ ) { if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root?
_igvn.delete_input_of(C->root(), i);
i--; // Rerun same iteration on compressed edges
}
}
// Given dominators, try to find inner loops with calls that must // always be executed (call dominates loop tail). These loops do // not need a separate safepoint.
Node_List cisstack;
_ltree_root->check_safepts(visited, cisstack);
}
// Walk the DATA nodes and place into loops. Find earliest control // node. For CFG nodes, the _nodes array starts out and remains // holding the associated IdealLoopTree pointer. For DATA nodes, the // _nodes array holds the earliest legal controlling CFG node.
// Allocate stack with enough space to avoid frequent realloc int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
Node_Stack nstack(stack_size);
visited.clear();
Node_List worklist; // Don't need C->root() on worklist since // it will be processed among C->top() inputs
worklist.push(C->top());
visited.set(C->top()->_idx); // Set C->top() as visited now
build_loop_early( visited, worklist, nstack );
// Given early legal placement, try finding counted loops. This placement // is good enough to discover most loop invariants. if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
_ltree_root->counted_loop( this );
}
// Find latest loop placement. Find ideal loop placement.
visited.clear();
init_dom_lca_tags(); // Need C->root() on worklist when processing outs
worklist.push(C->root());
NOT_PRODUCT( C->verify_graph_edges(); )
worklist.push(C->top());
build_loop_late( visited, worklist, nstack );
// clear out the dead code after build_loop_late while (_deadlist.size()) {
_igvn.remove_globally_dead_node(_deadlist.pop());
}
if (stop_early) {
assert(do_expensive_nodes, "why are we here?"); if (process_expensive_nodes()) { // If we made some progress when processing expensive nodes then // the IGVN may modify the graph in a way that will allow us to // make some more progress: we need to try processing expensive // nodes again.
C->set_major_progress();
} return;
}
// Some parser-inserted loop predicates could never be used by loop // predication or they were moved away from loop during some optimizations. // For example, peeling. Eliminate them before next loop optimizations.
eliminate_useless_predicates();
#ifndef PRODUCT
C->verify_graph_edges(); if (_verify_me) { // Nested verify pass? // Check to see if the verify _mode is broken
assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?"); return;
} if (VerifyLoopOptimizations) verify(); if (TraceLoopOpts && C->has_loops()) {
_ltree_root->dump();
} #endif
if (skip_loop_opts) {
C->restore_major_progress(old_progress); return;
}
if (do_max_unroll) { for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current(); if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
lpt->compute_trip_count(this); if (!lpt->do_one_iteration_loop(this) &&
!lpt->do_remove_empty_loop(this)) {
AutoNodeBudget node_budget(this); if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
lpt->policy_maximally_unroll(this)) {
memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
do_maximally_unroll(lpt, worklist);
}
}
}
}
if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) { return;
}
if (ReassociateInvariants && !C->major_progress()) { // Reassociate invariants and prep for split_thru_phi for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current(); if (!lpt->is_loop()) { continue;
}
Node* head = lpt->_head; if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
// check for vectorized loops, any reassociation of invariants was already done if (head->is_CountedLoop()) { if (head->as_CountedLoop()->is_unroll_only()) { continue;
} else {
AutoNodeBudget node_budget(this);
lpt->reassociate_invariants(this);
}
} // Because RCE opportunities can be masked by split_thru_phi, // look for RCE candidates and inhibit split_thru_phi // on just their loop-phi's for this pass of loop opts if (SplitIfBlocks && do_split_ifs &&
head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
(lpt->policy_range_check(this, true, T_LONG) ||
(head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
lpt->_rce_candidate = 1; // = true
}
}
}
// Check for aggressive application of split-if and other transforms // that require basic-block info (like cloning through Phi's) if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
visited.clear();
split_if_with_blocks( visited, nstack);
NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
}
if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
C->set_major_progress();
}
// Perform loop predication before iteration splitting if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
_ltree_root->_child->loop_predication(this);
}
if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { if (do_intrinsify_fill()) {
C->set_major_progress();
}
}
// Perform iteration-splitting on inner loops. Split iterations to avoid // range checks or one-shot null checks.
// If split-if's didn't hack the graph too bad (no CFG changes) // then do loop opts. if (C->has_loops() && !C->major_progress()) {
memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
_ltree_root->_child->iteration_split( this, worklist ); // No verify after peeling! GCM has hoisted code out of the loop. // After peeling, the hoisted code could sink inside the peeled area. // The peeling code does not try to recompute the best location for // all the code before the peeled area, so the verify pass will always // complain about it.
}
// Check for bailout, and return if (C->failing()) { return;
}
// Do verify graph edges in any case
NOT_PRODUCT( C->verify_graph_edges(); );
if (!do_split_ifs) { // We saw major progress in Split-If to get here. We forced a // pass with unrolling and not split-if, however more split-if's // might make progress. If the unrolling didn't make progress // then the major-progress flag got cleared and we won't try // another round of Split-If. In particular the ever-common // instance-of/check-cast pattern requires at least 2 rounds of // Split-If to clear out.
C->set_major_progress();
}
// Repeat loop optimizations if new loops were seen if (created_loop_node()) {
C->set_major_progress();
}
// Keep loop predicates and perform optimizations with them // until no more loop optimizations could be done. // After that switch predicates off and do more loop optimizations. if (!C->major_progress() && (C->predicate_count() > 0)) {
C->cleanup_loop_predicates(_igvn); if (TraceLoopOpts) {
tty->print_cr("PredicatesOff");
}
C->set_major_progress();
}
// Convert scalar to superword operations at the end of all loop opts. if (UseSuperWord && C->has_loops() && !C->major_progress()) { // SuperWord transform
SuperWord sw(this); for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
IdealLoopTree* lpt = iter.current(); if (lpt->is_counted()) {
CountedLoopNode *cl = lpt->_head->as_CountedLoop();
if (cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
assert(PostLoopMultiversioning, "multiversioning must be enabled"); // Check that the rce'd post loop is encountered first, multiversion after all // major main loop optimization are concluded if (!C->major_progress()) {
IdealLoopTree *lpt_next = lpt->_next; if (lpt_next && lpt_next->is_counted()) {
CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
has_range_checks(lpt_next); if (cl->is_post_loop() && cl->range_checks_present()) { if (!cl->is_multiversioned()) { if (multi_version_post_loops(lpt, lpt_next) == false) { // Cause the rce loop to be optimized away if we fail
cl->mark_is_multiversioned();
cl->set_slp_max_unroll(0);
poison_rce_post_loop(lpt);
}
}
}
}
sw.transform_loop(lpt, true);
}
} elseif (cl->is_main_loop()) { if (!sw.transform_loop(lpt, true)) { // Instigate more unrolling for optimization when vectorization fails. if (cl->has_passed_slp()) {
C->set_major_progress();
cl->set_notpassed_slp();
cl->mark_do_unroll_only();
}
}
}
}
}
}
// disable assert until issue with split_flow_path is resolved (6742111) // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), // "shouldn't introduce irreducible loops");
}
#ifndef PRODUCT //------------------------------print_statistics------------------------------- int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique volatileint PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen volatileint PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest volatileint PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop void PhaseIdealLoop::print_statistics() {
tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d, long loops=%d/%d/%d", _loop_invokes, _loop_work, _long_loop_counted_loops, _long_loop_nests, _long_loop_candidates);
}
//------------------------------verify----------------------------------------- // Build a verify-only PhaseIdealLoop, and see that it agrees with me. staticint fail; // debug only, so its multi-thread dont care void PhaseIdealLoop::verify() const { int old_progress = C->major_progress();
ResourceMark rm;
PhaseIdealLoop loop_verify(_igvn, this);
VectorSet visited;
fail = 0;
verify_compare(C->root(), &loop_verify, visited);
assert(fail == 0, "verify loops failed"); // Verify loop structure is the same
_ltree_root->verify_tree(loop_verify._ltree_root, NULL); // Reset major-progress. It was cleared by creating a verify version of // PhaseIdealLoop.
C->restore_major_progress(old_progress);
}
//------------------------------verify_compare--------------------------------- // Make sure me and the given PhaseIdealLoop agree on key data structures void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { if( !n ) return; if( visited.test_set( n->_idx ) ) return; if( !_nodes[n->_idx] ) { // Unreachable
assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); return;
}
uint i; for( i = 0; i < n->req(); i++ )
verify_compare( n->in(i), loop_verify, visited );
// Check the '_nodes' block/loop structure
i = n->_idx; if( has_ctrl(n) ) { // We have control; verify has loop or ctrl if( _nodes[i] != loop_verify->_nodes[i] &&
get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
tty->print("Mismatched control setting for: ");
n->dump(); if( fail++ > 10 ) return;
Node *c = get_ctrl_no_update(n);
tty->print("We have it as: "); if( c->in(0) ) c->dump(); else tty->print_cr("N%d",c->_idx);
tty->print("Verify thinks: "); if( loop_verify->has_ctrl(n) )
loop_verify->get_ctrl_no_update(n)->dump(); else
loop_verify->get_loop_idx(n)->dump();
tty->cr();
}
} else { // We have a loop
IdealLoopTree *us = get_loop_idx(n); if( loop_verify->has_ctrl(n) ) {
tty->print("Mismatched loop setting for: ");
n->dump(); if( fail++ > 10 ) return;
tty->print("We have it as: ");
us->dump();
tty->print("Verify thinks: ");
loop_verify->get_ctrl_no_update(n)->dump();
tty->cr();
} elseif (!C->major_progress()) { // Loop selection can be messed up if we did a major progress // operation, like split-if. Do not verify in that case.
IdealLoopTree *them = loop_verify->get_loop_idx(n); if( us->_head != them->_head || us->_tail != them->_tail ) {
tty->print("Unequals loops for: ");
n->dump(); if( fail++ > 10 ) return;
tty->print("We have it as: ");
us->dump();
tty->print("Verify thinks: ");
them->dump();
tty->cr();
}
}
}
// Check for immediate dominators being equal if( i >= _idom_size ) { if( !n->is_CFG() ) return;
tty->print("CFG Node with no idom: ");
n->dump(); return;
} if( !n->is_CFG() ) return; if( n == C->root() ) return; // No IDOM here
assert(n->_idx == i, "sanity");
Node *id = idom_no_update(n); if( id != loop_verify->idom_no_update(n) ) {
tty->print("Unequals idoms for: ");
n->dump(); if( fail++ > 10 ) return;
tty->print("We have it as: ");
id->dump();
tty->print("Verify thinks: ");
loop_verify->idom_no_update(n)->dump();
tty->cr();
}
}
//------------------------------verify_tree------------------------------------ // Verify that tree structures match. Because the CFG can change, siblings // within the loop tree can be reordered. We attempt to deal with that by // reordering the verify's loop tree if possible. void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
assert( _parent == parent, "Badly formed loop tree" );
// Siblings not in same order? Attempt to re-order. if( _head != loop->_head ) { // Find _next pointer to update
IdealLoopTree **pp = &loop->_parent->_child; while( *pp != loop )
pp = &((*pp)->_next); // Find proper sibling to be next
IdealLoopTree **nn = &loop->_next; while( (*nn) && (*nn)->_head != _head )
nn = &((*nn)->_next);
// Check for no match. if( !(*nn) ) { // Annoyingly, irreducible loops can pick different headers // after a major_progress operation, so the rest of the loop // tree cannot be matched. if (_irreducible && Compile::current()->major_progress()) return;
assert( 0, "failed to match loop tree" );
}
// Move (*nn) to (*pp)
IdealLoopTree *hit = *nn;
*nn = hit->_next;
hit->_next = loop;
*pp = loop;
loop = hit; // Now try again to verify
}
assert( _head == loop->_head , "mismatched loop head" );
Node *tail = _tail; // Inline a non-updating version of while( !tail->in(0) ) // the 'tail()' call.
tail = tail->in(1);
assert( tail == loop->_tail, "mismatched loop tail" );
if (_child != NULL) _child->verify_tree(loop->_child, this); if (_next != NULL) _next ->verify_tree(loop->_next, parent); // Innermost loops need to verify loop bodies, // but only if no 'major_progress' int fail = 0; if (!Compile::current()->major_progress() && _child == NULL) { for( uint i = 0; i < _body.size(); i++ ) {
Node *n = _body.at(i); if (n->outcnt() == 0) continue; // Ignore dead
uint j; for( j = 0; j < loop->_body.size(); j++ ) if( loop->_body.at(j) == n ) break; if( j == loop->_body.size() ) { // Not found in loop body // Last ditch effort to avoid assertion: Its possible that we // have some users (so outcnt not zero) but are still dead. // Try to find from root. if (Compile::current()->root()->find(n->_idx)) {
fail++;
tty->print("We have that verify does not: ");
n->dump();
}
}
} for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
Node *n = loop->_body.at(i2); if (n->outcnt() == 0) continue; // Ignore dead
uint j; for( j = 0; j < _body.size(); j++ ) if( _body.at(j) == n ) break; if( j == _body.size() ) { // Not found in loop body // Last ditch effort to avoid assertion: Its possible that we // have some users (so outcnt not zero) but are still dead. // Try to find from root. if (Compile::current()->root()->find(n->_idx)) {
fail++;
tty->print("Verify has that we do not: ");
n->dump();
}
}
}
assert( !fail, "loop body mismatch" );
}
}
//------------------------------recompute_dom_depth--------------------------------------- // The dominator tree is constructed with only parent pointers. // This recomputes the depth in the tree by first tagging all // nodes as "no depth yet" marker. The next pass then runs up // the dom tree from each node marked "no depth yet", and computes // the depth on the way back down. void PhaseIdealLoop::recompute_dom_depth() {
uint no_depth_marker = C->unique();
uint i; // Initialize depth to "no depth yet" and realize all lazy updates for (i = 0; i < _idom_size; i++) { // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitialized). if (_dom_depth[i] > 0 && _idom[i] != NULL) {
_dom_depth[i] = no_depth_marker;
// heal _idom if it has a fwd mapping in _nodes if (_idom[i]->in(0) == NULL) {
idom(i);
}
}
} if (_dom_stk == NULL) {
uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. if (init_size < 10) init_size = 10;
_dom_stk = new GrowableArray<uint>(init_size);
} // Compute new depth for each node. for (i = 0; i < _idom_size; i++) {
uint j = i; // Run up the dom tree to find a node with a depth while (_dom_depth[j] == no_depth_marker) {
_dom_stk->push(j);
j = _idom[j]->_idx;
} // Compute the depth on the way back down this tree branch
uint dd = _dom_depth[j] + 1; while (_dom_stk->length() > 0) {
uint j = _dom_stk->pop();
_dom_depth[j] = dd;
dd++;
}
}
}
//------------------------------sort------------------------------------------- // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the // loop tree, not the root.
IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { if( !innermost ) return loop; // New innermost loop
int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
assert( loop_preorder, "not yet post-walked loop" );
IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
IdealLoopTree *l = *pp; // Do I go before or after 'l'?
// Insert at start of list while( l ) { // Insertion sort based on pre-order if( l == loop ) return innermost; // Already on list! int l_preorder = get_preorder(l->_head); // Cache pre-order number
assert( l_preorder, "not yet post-walked l" ); // Check header pre-order number to figure proper nesting if( loop_preorder > l_preorder ) break; // End of insertion // If headers tie (e.g., shared headers) check tail pre-order numbers. // Since I split shared headers, you'd think this could not happen. // BUT: I must first do the preorder numbering before I can discover I // have shared headers, so the split headers all get the same preorder // number as the RegionNode they split from. if( loop_preorder == l_preorder &&
get_preorder(loop->_tail) < get_preorder(l->_tail) ) break; // Also check for shared headers (same pre#)
pp = &l->_parent; // Chain up list
l = *pp;
} // Link into list // Point predecessor to me
*pp = loop; // Point me to successor
IdealLoopTree *p = loop->_parent;
loop->_parent = l; // Point me to successor if( p ) sort( p, innermost ); // Insert my parents into list as well return innermost;
}
//------------------------------build_loop_tree-------------------------------- // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit // bits. The _nodes[] array is mapped by Node index and holds a NULL for // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the // tightest enclosing IdealLoopTree for post-walked. // // During my forward walk I do a short 1-layer lookahead to see if I can find // a loop backedge with that doesn't have any work on the backedge. This // helps me construct nested loops with shared headers better. // // Once I've done the forward recursion, I do the post-work. For each child // I check to see if there is a backedge. Backedges define a loop! I // insert an IdealLoopTree at the target of the backedge. // // During the post-work I also check to see if I have several children // belonging to different loops. If so, then this Node is a decision point // where control flow can choose to change loop nests. It is at this // decision point where I can figure out how loops are nested. At this // time I can properly order the different loop nests from my children. // Note that there may not be any backedges at the decision point! // // Since the decision point can be far removed from the backedges, I can't // order my loops at the time I discover them. Thus at the decision point // I need to inspect loop header pre-order numbers to properly nest my // loops. This means I need to sort my childrens' loops by pre-order. // The sort is of size number-of-control-children, which generally limits // it to size 2 (i.e., I just choose between my 2 target loops). void PhaseIdealLoop::build_loop_tree() { // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
Node *n = C->root();
bltstack.push(n); int pre_order = 1; int stack_size;
while ( ( stack_size = bltstack.length() ) != 0 ) {
n = bltstack.top(); // Leave node on stack if ( !is_visited(n) ) { // ---- Pre-pass Work ---- // Pre-walked but not post-walked nodes need a pre_order number.
set_preorder_visited( n, pre_order ); // set as visited
// ---- Scan over children ---- // Scan first over control projections that lead to loop headers. // This helps us find inner-to-outer loops with shared headers better.
// Scan children's children for loop headers. for ( int i = n->outcnt() - 1; i >= 0; --i ) {
Node* m = n->raw_out(i); // Child if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children // Scan over children's children to find loop for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
Node* l = m->fast_out(j); if( is_visited(l) && // Been visited?
!is_postvisited(l) && // But not post-visited
get_preorder(l) < pre_order ) { // And smaller pre-order // Found! Scan the DFS down this path before doing other paths
bltstack.push(m); break;
}
}
}
}
pre_order++;
} elseif ( !is_postvisited(n) ) { // Note: build_loop_tree_impl() adds out edges on rare occasions, // such as com.sun.rsasign.am::a. // For non-recursive version, first, process current children. // On next iteration, check if additional children were added. for ( int k = n->outcnt() - 1; k >= 0; --k ) {
Node* u = n->raw_out(k); if ( u->is_CFG() && !is_visited(u) ) {
bltstack.push(u);
}
} if ( bltstack.length() == stack_size ) { // There were no additional children, post visit node now
(void)bltstack.pop(); // Remove node from stack
pre_order = build_loop_tree_impl( n, pre_order ); // Check for bailout if (C->failing()) { return;
} // Check to grow _preorders[] array for the case when // build_loop_tree_impl() adds new nodes.
check_grow_preorders();
}
} else {
(void)bltstack.pop(); // Remove post-visited node from stack
}
}
}
//------------------------------build_loop_tree_impl--------------------------- int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { // ---- Post-pass Work ---- // Pre-walked but not post-walked nodes need a pre_order number.
// Tightest enclosing loop for this Node
IdealLoopTree *innermost = NULL;
// For all children, see if any edge is a backedge. If so, make a loop // for it. Then find the tightest enclosing loop for the self Node. for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* m = n->fast_out(i); // Child if( n == m ) continue; // Ignore control self-cycles if( !m->is_CFG() ) continue;// Ignore non-CFG edges
IdealLoopTree *l; // Child's loop if( !is_postvisited(m) ) { // Child visited but not post-visited? // Found a backedge
assert( get_preorder(m) < pre_order, "should be backedge" ); // Check for the RootNode, which is already a LoopNode and is allowed // to have multiple "backedges". if( m == C->root()) { // Found the root?
l = _ltree_root; // Root is the outermost LoopNode
} else { // Else found a nested loop // Insert a LoopNode to mark this loop.
l = new IdealLoopTree(this, m, n);
} // End of Else found a nested loop if( !has_loop(m) ) // If 'm' does not already have a loop set
set_loop(m, l); // Set loop header to loop now
} else { // Else not a nested loop if( !_nodes[m->_idx] ) continue; // Dead code has no loop
l = get_loop(m); // Get previously determined loop // If successor is header of a loop (nest), move up-loop till it // is a member of some outer enclosing loop. Since there are no // shared headers (I've split them already) I only need to go up // at most 1 level. while( l && l->_head == m ) // Successor heads loop?
l = l->_parent; // Move up 1 for me // If this loop is not properly parented, then this loop // has no exit path out, i.e. its an infinite loop. if( !l ) { // Make loop "reachable" from root so the CFG is reachable. Basically // insert a bogus loop exit that is never taken. 'm', the loop head, // points to 'n', one (of possibly many) fall-in paths. There may be // many backedges as well.
// Here I set the loop to be the root loop. I could have, after // inserting a bogus loop exit, restarted the recursion and found my // new loop exit. This would make the infinite loop a first-class // loop and it would then get properly optimized. What's the use of // optimizing an infinite loop?
l = _ltree_root; // Oops, found infinite loop
if (!_verify_only) { // Insert the NeverBranch between 'm' and it's control user.
NeverBranchNode *iff = new NeverBranchNode( m );
_igvn.register_new_node_with_optimizer(iff);
set_loop(iff, l);
Node *if_t = new CProjNode( iff, 0 );
_igvn.register_new_node_with_optimizer(if_t);
set_loop(if_t, l);
Node* cfg = NULL; // Find the One True Control User of m for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
Node* x = m->fast_out(j); if (x->is_CFG() && x != m && x != iff)
{ cfg = x; break; }
}
assert(cfg != NULL, "must find the control user of m");
uint k = 0; // Probably cfg->in(0) while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
_igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
// Now create the never-taken loop exit
Node *if_f = new CProjNode( iff, 1 );
_igvn.register_new_node_with_optimizer(if_f);
set_loop(if_f, l); // Find frame ptr for Halt. Relies on the optimizer // V-N'ing. Easier and quicker than searching through // the program structure.
Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
_igvn.register_new_node_with_optimizer(frame); // Halt & Catch Fire
Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
_igvn.register_new_node_with_optimizer(halt);
set_loop(halt, l);
_igvn.add_input_to(C->root(), halt);
}
set_loop(C->root(), _ltree_root);
}
} // Weeny check for irreducible. This child was already visited (this // IS the post-work phase). Is this child's loop header post-visited // as well? If so, then I found another entry into the loop. if (!_verify_only) { while( is_postvisited(l->_head) ) { // found irreducible
l->_irreducible = 1; // = true
l = l->_parent;
_has_irreducible_loops = true; // Check for bad CFG here to prevent crash, and bailout of compile if (l == NULL) {
C->record_method_not_compilable("unhandled CFG detected during loop optimization"); return pre_order;
}
}
C->set_has_irreducible_loop(_has_irreducible_loops);
}
// This Node might be a decision point for loops. It is only if // it's children belong to several different loops. The sort call // does a trivial amount of work if there is only 1 child or all // children belong to the same loop. If however, the children // belong to different loops, the sort call will properly set the // _parent pointers to show how the loops nest. // // In any case, it returns the tightest enclosing loop.
innermost = sort( l, innermost );
}
// Def-use info will have some dead stuff; dead stuff will have no // loop decided on.
// Am I a loop header? If so fix up my parent's child and next ptrs. if( innermost && innermost->_head == n ) {
assert( get_loop(n) == innermost, "" );
IdealLoopTree *p = innermost->_parent;
IdealLoopTree *l = innermost; while( p && l->_head == n ) {
l->_next = p->_child; // Put self on parents 'next child'
p->_child = l; // Make self as first child of parent
l = p; // Now walk up the parent chain
p = l->_parent;
}
} else { // Note that it is possible for a LoopNode to reach here, if the // backedge has been made unreachable (hence the LoopNode no longer // denotes a Loop, and will eventually be removed).
// Record tightest enclosing loop for self. Mark as post-visited.
set_loop(n, innermost); // Also record has_call flag early on if( innermost ) { if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { // Do not count uncommon calls if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
Node *iff = n->in(0)->in(0); // No any calls for vectorized loops. if( UseSuperWord || !iff->is_If() ||
(n->in(0)->Opcode() == Op_IfFalse &&
(1.0 - iff->as_If()->_prob) >= 0.01) ||
(iff->as_If()->_prob >= 0.01) )
innermost->_has_call = 1;
}
} elseif( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { // Disable loop optimizations if the loop has a scalar replaceable // allocation. This disabling may cause a potential performance lost // if the allocation is not eliminated for some reason.
innermost->_allow_optimizations = false;
innermost->_has_call = 1; // = true
} elseif (n->Opcode() == Op_SafePoint) { // Record all safepoints in this loop. if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
innermost->_safepts->push(n);
}
}
}
// Flag as post-visited now
set_postvisited(n); return pre_order;
}
//------------------------------build_loop_early------------------------------- // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. // First pass computes the earliest controlling node possible. This is the // controlling input with the deepest dominating depth. void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { while (worklist.size() != 0) { // Use local variables nstack_top_n & nstack_top_i to cache values // on nstack's top.
Node *nstack_top_n = worklist.pop();
uint nstack_top_i = 0; //while_nstack_nonempty: while (true) { // Get parent node and next input's index from stack's top.
Node *n = nstack_top_n;
uint i = nstack_top_i;
uint cnt = n->req(); // Count of inputs if (i == 0) { // Pre-process the node. if( has_node(n) && // Have either loop or control already?
!has_ctrl(n) ) { // Have loop picked out already? // During "merge_many_backedges" we fold up several nested loops // into a single loop. This makes the members of the original // loop bodies pointing to dead loops; they need to move up // to the new UNION'd larger loop. I set the _head field of these // dead loops to NULL and the _parent field points to the owning // loop. Shades of UNION-FIND algorithm.
IdealLoopTree *ilt; while( !(ilt = get_loop(n))->_head ) { // Normally I would use a set_loop here. But in this one special // case, it is legal (and expected) to change what loop a Node // belongs to.
_nodes.map(n->_idx, (Node*)(ilt->_parent) );
} // Remove safepoints ONLY if I've already seen I don't need one. // (the old code here would yank a 2nd safepoint after seeing a // first one, even though the 1st did not dominate in the loop body // and thus could be avoided indefinitely) if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
is_deleteable_safept(n)) {
Node *in = n->in(TypeFunc::Control);
lazy_replace(n,in); // Pull safepoint now if (ilt->_safepts != NULL) {
ilt->_safepts->yank(n);
} // Carry on with the recursion "as if" we are walking // only the control input if( !visited.test_set( in->_idx ) ) {
worklist.push(in); // Visit this guy later, using worklist
} // Get next node from nstack: // - skip n's inputs processing by setting i > cnt; // - we also will not call set_early_ctrl(n) since // has_node(n) == true (see the condition above).
i = cnt + 1;
}
}
} // if (i == 0)
// Visit all inputs bool done = true; // Assume all n's inputs will be processed while (i < cnt) {
Node *in = n->in(i);
++i; if (in == NULL) continue; if (in->pinned() && !in->is_CFG())
set_ctrl(in, in->in(0)); int is_visited = visited.test_set( in->_idx ); if (!has_node(in)) { // No controlling input yet?
assert( !in->is_CFG(), "CFG Node with no controlling input?" );
assert( !is_visited, "visit only once" );
nstack.push(n, i); // Save parent node and next input's index.
nstack_top_n = in; // Process current input now.
nstack_top_i = 0;
done = false; // Not all n's inputs processed. break; // continue while_nstack_nonempty;
} elseif (!is_visited) { // This guy has a location picked out for him, but has not yet // been visited. Happens to all CFG nodes, for instance. // Visit him using the worklist instead of recursion, to break // cycles. Since he has a location already we do not need to // find his location before proceeding with the current Node.
worklist.push(in); // Visit this guy later, using worklist
}
} if (done) { // All of n's inputs have been processed, complete post-processing.
// Compute earliest point this Node can go. // CFG, Phi, pinned nodes already know their controlling input. if (!has_node(n)) { // Record earliest legal location
set_early_ctrl(n, false);
} if (nstack.is_empty()) { // Finished all nodes on stack. // Process next node on the worklist. break;
} // Get saved parent node and next input's index.
nstack_top_n = nstack.node();
nstack_top_i = nstack.index();
nstack.pop();
}
} // while (true)
}
}
//------------------------------dom_lca_internal-------------------------------- // Pair-wise LCA
Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { if( !n1 ) return n2; // Handle NULL original LCA
assert( n1->is_CFG(), "" );
assert( n2->is_CFG(), "" ); // find LCA of all uses
uint d1 = dom_depth(n1);
uint d2 = dom_depth(n2); while (n1 != n2) { if (d1 > d2) {
n1 = idom(n1);
d1 = dom_depth(n1);
} elseif (d1 < d2) {
n2 = idom(n2);
d2 = dom_depth(n2);
} else { // Here d1 == d2. Due to edits of the dominator-tree, sections // of the tree might have the same depth. These sections have // to be searched more carefully.
// Scan up all the n1's with equal depth, looking for n2.
Node *t1 = idom(n1); while (dom_depth(t1) == d1) { if (t1 == n2) return n2;
t1 = idom(t1);
} // Scan up all the n2's with equal depth, looking for n1.
Node *t2 = idom(n2); while (dom_depth(t2) == d2) { if (t2 == n1) return n1;
t2 = idom(t2);
} // Move up to a new dominator-depth value as well as up the dom-tree.
n1 = t1;
n2 = t2;
d1 = dom_depth(n1);
d2 = dom_depth(n2);
}
} return n1;
}
//------------------------------compute_idom----------------------------------- // Locally compute IDOM using dom_lca call. Correct only if the incoming // IDOMs are correct.
Node *PhaseIdealLoop::compute_idom( Node *region ) const {
assert( region->is_Region(), "" );
Node *LCA = NULL; for( uint i = 1; i < region->req(); i++ ) { if( region->in(i) != C->top() )
LCA = dom_lca( LCA, region->in(i) );
} return LCA;
}
bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { bool had_error = false; #ifdef ASSERT if (early != C->root()) { // Make sure that there's a dominance path from LCA to early
Node* d = LCA; while (d != early) { if (d == C->root()) {
dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
had_error = true; break;
}
d = idom(d);
}
} #endif return had_error;
}
Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { // Compute LCA over list of uses bool had_error = false;
Node *LCA = NULL; for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
Node* c = n->fast_out(i); if (_nodes[c->_idx] == NULL) continue; // Skip the occasional dead node if( c->is_Phi() ) { // For Phis, we must land above on the path for( uint j=1; j<c->req(); j++ ) {// For all inputs if( c->in(j) == n ) { // Found matching input?
Node *use = c->in(0)->in(j); if (_verify_only && use->is_top()) continue;
LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
}
}
} else { // For CFG data-users, use is in the block just prior
Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
}
}
assert(!had_error, "bad dominance"); return LCA;
}
// Check the shape of the graph at the loop entry. In some cases, // the shape of the graph does not match the shape outlined below. // That is caused by the Opaque1 node "protecting" the shape of // the graph being removed by, for example, the IGVN performed // in PhaseIdealLoop::build_and_optimize(). // // After the Opaque1 node has been removed, optimizations (e.g., split-if, // loop unswitching, and IGVN, or a combination of them) can freely change // the graph's shape. As a result, the graph shape outlined below cannot // be guaranteed anymore.
Node* CountedLoopNode::is_canonical_loop_entry() { if (!is_main_loop() && !is_post_loop()) { return NULL;
}
Node* ctrl = skip_predicates();
assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); return LCA;
}
// if this is a load, check for anti-dependent stores // We use a conservative algorithm to identify potential interfering // instructions and for rescheduling the load. The users of the memory // input of this load are examined. Any use which is not a load and is // dominated by early is considered a potentially interfering store. // This can produce false positives.
Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) { int load_alias_idx = C->get_alias_index(n->adr_type()); if (C->alias_type(load_alias_idx)->is_rewritable()) {
Unique_Node_List worklist;
Node* mem = n->in(MemNode::Memory); for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
Node* s = mem->fast_out(i);
worklist.push(s);
} for (uint i = 0; i < worklist.size() && LCA != early; i++) {
Node* s = worklist.at(i); if (s->is_Load() || s->Opcode() == Op_SafePoint ||
(s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
s->is_Phi()) { continue;
} elseif (s->is_MergeMem()) { for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
Node* s1 = s->fast_out(i);
worklist.push(s1);
}
} else {
Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
assert(sctrl != NULL || !s->is_reachable_from_root(), "must have control"); if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { const TypePtr* adr_type = s->adr_type(); if (s->is_ArrayCopy()) { // Copy to known instance needs destination type to test for aliasing const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type; if (dest_type != TypeOopPtr::BOTTOM) {
adr_type = dest_type;
}
} if (C->can_alias(adr_type, load_alias_idx)) {
LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
} elseif (s->is_CFG() && s->is_Multi()) { // Look for the memory use of s (that is the use of its memory projection) for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
Node* s1 = s->fast_out(i);
assert(s1->is_Proj(), "projection expected"); if (_igvn.type(s1) == Type::MEMORY) { for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
Node* s2 = s1->fast_out(j);
worklist.push(s2);
}
}
}
}
}
}
} // For Phis only consider Region's inputs that were reached by following the memory edges if (LCA != early) { for (uint i = 0; i < worklist.size(); i++) {
Node* s = worklist.at(i); if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
Node* r = s->in(0); for (uint j = 1; j < s->req(); j++) {
Node* in = s->in(j);
Node* r_in = r->in(j); // We can't reach any node from a Phi because we don't enqueue Phi's uses above if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
}
}
}
}
}
} return LCA;
}
// true if CFG node d dominates CFG node n bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { if (d == n) returntrue;
assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
uint dd = dom_depth(d); while (dom_depth(n) >= dd) { if (n == d) returntrue;
n = idom(n);
} returnfalse;
}
//------------------------------dom_lca_for_get_late_ctrl_internal------------- // Pair-wise LCA with tags. // Tag each index with the node 'tag' currently being processed // before advancing up the dominator chain using idom(). // Later calls that find a match to 'tag' know that this path has already // been considered in the current LCA (which is input 'n1' by convention). // Since get_late_ctrl() is only called once for each node, the tag array // does not need to be cleared between calls to get_late_ctrl(). // Algorithm trades a larger constant factor for better asymptotic behavior //
Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
uint d1 = dom_depth(n1);
uint d2 = dom_depth(n2);
jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
do { if (d1 > d2) { // current lca is deeper than n2
_dom_lca_tags.at_put_grow(n1->_idx, tag);
n1 = idom(n1);
d1 = dom_depth(n1);
} elseif (d1 < d2) { // n2 is deeper than current lca
jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0); if (memo == tag) { return n1; // Return the current LCA
}
_dom_lca_tags.at_put_grow(n2->_idx, tag);
n2 = idom(n2);
d2 = dom_depth(n2);
} else { // Here d1 == d2. Due to edits of the dominator-tree, sections // of the tree might have the same depth. These sections have // to be searched more carefully.
// Scan up all the n1's with equal depth, looking for n2.
_dom_lca_tags.at_put_grow(n1->_idx, tag);
Node *t1 = idom(n1); while (dom_depth(t1) == d1) { if (t1 == n2) return n2;
_dom_lca_tags.at_put_grow(t1->_idx, tag);
t1 = idom(t1);
} // Scan up all the n2's with equal depth, looking for n1.
_dom_lca_tags.at_put_grow(n2->_idx, tag);
Node *t2 = idom(n2); while (dom_depth(t2) == d2) { if (t2 == n1) return n1;
_dom_lca_tags.at_put_grow(t2->_idx, tag);
t2 = idom(t2);
} // Move up to a new dominator-depth value as well as up the dom-tree.
n1 = t1;
n2 = t2;
d1 = dom_depth(n1);
d2 = dom_depth(n2);
}
} while (n1 != n2); return n1;
}
//------------------------------init_dom_lca_tags------------------------------ // Tag could be a node's integer index, 32bits instead of 64bits in some cases // Intended use does not involve any growth for the array, so it could // be of fixed size. void PhaseIdealLoop::init_dom_lca_tags() {
uint limit = C->unique() + 1;
_dom_lca_tags.at_grow(limit, 0);
_dom_lca_tags_round = 0; #ifdef ASSERT for (uint i = 0; i < limit; ++i) {
assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
} #endif// ASSERT
}
//------------------------------build_loop_late-------------------------------- // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. // Second pass finds latest legal placement, and ideal loop placement. void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { while (worklist.size() != 0) {
Node *n = worklist.pop(); // Only visit once if (visited.test_set(n->_idx)) continue;
uint cnt = n->outcnt();
uint i = 0; while (true) {
assert( _nodes[n->_idx], "no dead nodes" ); // Visit all children if (i < cnt) {
Node* use = n->raw_out(i);
++i; // Check for dead uses. Aggressively prune such junk. It might be // dead in the global sense, but still have local uses so I cannot // easily call 'remove_dead_node'. if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? // Due to cycles, we might not hit the same fixed point in the verify // pass as we do in the regular pass. Instead, visit such phis as // simple uses of the loop head. if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { if( !visited.test(use->_idx) )
worklist.push(use);
} elseif( !visited.test_set(use->_idx) ) {
nstack.push(n, i); // Save parent and next use's index.
n = use; // Process all children of current use.
cnt = use->outcnt();
i = 0;
}
} else { // Do not visit around the backedge of loops via data edges. // push dead code onto a worklist
_deadlist.push(use);
}
} else { // All of n's children have been processed, complete post-processing.
build_loop_late_post(n); if (nstack.is_empty()) { // Finished all nodes on stack. // Process next node on the worklist. break;
} // Get saved parent node and next use's index. Visit the rest of uses.
n = nstack.node();
cnt = n->outcnt();
i = nstack.index();
nstack.pop();
}
}
}
}
// Verify that no data node is scheduled in the outer loop of a strip // mined loop. void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { #ifdef ASSERT if (get_loop(least)->_nest == 0) { return;
}
IdealLoopTree* loop = get_loop(least);
Node* head = loop->_head; if (head->is_OuterStripMinedLoop() && // Verification can't be applied to fully built strip mined loops
head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
Node* sfpt = head->as_Loop()->outer_safepoint();
ResourceMark rm;
Unique_Node_List wq;
wq.push(sfpt); for (uint i = 0; i < wq.size(); i++) {
Node *m = wq.at(i); for (uint i = 1; i < m->req(); i++) {
Node* nn = m->in(i); if (nn == n) { return;
} if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
wq.push(nn);
}
}
}
ShouldNotReachHere();
} #endif
}
//------------------------------build_loop_late_post--------------------------- // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. // Second pass finds latest legal placement, and ideal loop placement. void PhaseIdealLoop::build_loop_late_post(Node *n) {
build_loop_late_post_work(n, true);
}
// We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads // _must_ be pinned (they have to observe their control edge of course). // Unlike Stores (which modify an unallocable resource, the memory // state), Mods/Loads can float around. So free them up. switch( n->Opcode() ) { case Op_DivI: case Op_DivF: case Op_DivD: case Op_ModI: case Op_ModF: case Op_ModD: case Op_LoadB: // Same with Loads; they can sink case Op_LoadUB: // during loop optimizations. case Op_LoadUS: case Op_LoadD: case Op_LoadF: case Op_LoadI: case Op_LoadKlass: case Op_LoadNKlass: case Op_LoadL: case Op_LoadS: case Op_LoadP: case Op_LoadN: case Op_LoadRange: case Op_LoadD_unaligned: case Op_LoadL_unaligned: case Op_StrComp: // Does a bunch of load-like effects case Op_StrEquals: case Op_StrIndexOf: case Op_StrIndexOfChar: case Op_AryEq: case Op_CountPositives:
pinned = false;
} if (n->is_CMove() || n->is_ConstraintCast()) {
pinned = false;
} if( pinned ) {
IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); if( !chosen_loop->_child ) // Inner loop?
chosen_loop->_body.push(n); // Collect inner loops return;
}
} else { // No slot zero if( n->is_CFG() ) { // CFG with no slot 0 is dead
_nodes.map(n->_idx,0); // No block setting, it's globally dead return;
}
assert(!n->is_CFG() || n->outcnt() == 0, "");
}
// Do I have a "safe range" I can select over?
Node *early = get_ctrl(n);// Early location already computed
// Compute latest point this Node can go
Node *LCA = get_late_ctrl( n, early ); // LCA is NULL due to uses being dead if( LCA == NULL ) { #ifdef ASSERT for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
} #endif
_nodes.map(n->_idx, 0); // This node is useless
_deadlist.push(n); return;
}
assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
Node *legal = LCA; // Walk 'legal' up the IDOM chain
Node *least = legal; // Best legal position so far while( early != legal ) { // While not at earliest legal #ifdef ASSERT if (legal->is_Start() && !early->is_Root()) { // Bad graph. Print idom path and fail.
dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
assert(false, "Bad graph detected in build_loop_late");
} #endif // Find least loop nesting depth
legal = idom(legal); // Bump up the IDOM tree // Check for lower nesting depth if( get_loop(legal)->_nest < get_loop(least)->_nest )
least = legal;
}
assert(early == legal || legal != C->root(), "bad dominance of inputs");
if (least != early) { // Move the node above predicates as far up as possible so a // following pass of loop predication doesn't hoist a predicate // that depends on it above that node.
Node* new_ctrl = least; for (;;) { if (!new_ctrl->is_Proj()) { break;
}
CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); if (call == NULL) { break;
} int req = call->uncommon_trap_request();
Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); if (trap_reason != Deoptimization::Reason_loop_limit_check &&
trap_reason != Deoptimization::Reason_predicate &&
trap_reason != Deoptimization::Reason_profile_predicate) { break;
}
Node* c = new_ctrl->in(0)->in(0); if (is_dominator(c, early) && c != early) { break;
}
new_ctrl = c;
}
least = new_ctrl;
} // Try not to place code on a loop entry projection // which can inhibit range check elimination. if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
Node* ctrl_out = least->unique_ctrl_out_or_null(); if (ctrl_out != NULL && ctrl_out->is_Loop() &&
least == ctrl_out->in(LoopNode::EntryControl) &&
(ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
Node* least_dom = idom(least); if (get_loop(least_dom)->is_member(get_loop(least))) {
least = least_dom;
}
}
} // Don't extend live ranges of raw oops if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
!n->bottom_type()->isa_rawptr()) {
least = early;
}
#ifdef ASSERT // If verifying, verify that 'verify_me' has a legal location // and choose it as our location. if( _verify_me ) {
Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
Node *legal = LCA; while( early != legal ) { // While not at earliest legal if( legal == v_ctrl ) break; // Check for prior good location
legal = idom(legal) ;// Bump up the IDOM tree
} // Check for prior good location if( legal == v_ctrl ) least = legal; // Keep prior if found
} #endif
// Assign discovered "here or above" point
least = find_non_split_ctrl(least);
verify_strip_mined_scheduling(n, least);
set_ctrl(n, least);
#ifdef ASSERT void PhaseIdealLoop::dump_bad_graph(constchar* msg, Node* n, Node* early, Node* LCA) {
tty->print_cr("%s", msg);
tty->print("n: "); n->dump();
tty->print("early(n): "); early->dump(); if (n->in(0) != NULL && !n->in(0)->is_top() &&
n->in(0) != early && !n->in(0)->is_Root()) {
tty->print("n->in(0): "); n->in(0)->dump();
} for (uint i = 1; i < n->req(); i++) {
Node* in1 = n->in(i); if (in1 != NULL && in1 != n && !in1->is_top()) {
tty->print("n->in(%d): ", i); in1->dump();
Node* in1_early = get_ctrl(in1);
tty->print("early(n->in(%d)): ", i); in1_early->dump(); if (in1->in(0) != NULL && !in1->in(0)->is_top() &&
in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
} for (uint j = 1; j < in1->req(); j++) {
Node* in2 = in1->in(j); if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
Node* in2_early = get_ctrl(in2);
tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); if (in2->in(0) != NULL && !in2->in(0)->is_top() &&
in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
}
}
}
}
}
tty->cr();
tty->print("LCA(n): "); LCA->dump(); for (uint i = 0; i < n->outcnt(); i++) {
Node* u1 = n->raw_out(i); if (u1 == n) continue;
tty->print("n->out(%d): ", i); u1->dump(); if (u1->is_CFG()) { for (uint j = 0; j < u1->outcnt(); j++) {
Node* u2 = u1->raw_out(j); if (u2 != u1 && u2 != n && u2->is_CFG()) {
tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
}
}
} else {
Node* u1_later = get_ctrl(u1);
tty->print("later(n->out(%d)): ", i); u1_later->dump(); if (u1->in(0) != NULL && !u1->in(0)->is_top() &&
u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
} for (uint j = 0; j < u1->outcnt(); j++) {
Node* u2 = u1->raw_out(j); if (u2 == n || u2 == u1) continue;
tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); if (!u2->is_CFG()) {
Node* u2_later = get_ctrl(u2);
tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); if (u2->in(0) != NULL && !u2->in(0)->is_top() &&
u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
}
}
}
}
}
dump_idoms(early, LCA);
tty->cr();
}
// Class to compute the real LCA given an early node and a wrong LCA in a bad graph. class RealLCA { const PhaseIdealLoop* _phase;
Node* _early;
Node* _wrong_lca;
uint _early_index; int _wrong_lca_index;
// Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which // is different: Return the previously visited node which must be the real LCA. // The node lists also contain _early and _wrong_lca, respectively.
Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) { int early_index = early_with_idoms.size() - 1; int wrong_lca_index = wrong_lca_with_idoms.size() - 1; bool found_difference = false; do { if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) { // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
found_difference = true; break;
}
early_index--;
wrong_lca_index--;
} while (wrong_lca_index >= 0);
assert(early_index >= 0, "must always find an LCA - cannot be early");
_early_index = early_index;
_wrong_lca_index = wrong_lca_index;
Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA"); return real_lca;
}
void dump(Node* real_lca) {
tty->cr();
tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
_phase->dump_idom(_early, _early_index + 1);
// Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA. void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
// Now scan for CFG nodes in the same loop for (uint j = idx; j > 0; j--) {
Node* n = rpo_list[j-1]; if (!_nodes[n->_idx]) // Skip dead nodes continue;
if (get_loop(n) != loop) { // Wrong loop nest if (get_loop(n)->_head == n && // Found nested loop?
get_loop(n)->_parent == loop)
dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly continue;
}
// Dump controlling node
tty->sp(2 * loop->_nest);
tty->print("C"); if (n == C->root()) {
n->dump();
} else {
Node* cached_idom = idom_no_update(n);
Node* computed_idom = n->in(0); if (n->is_Region()) {
computed_idom = compute_idom(n); // computed_idom() will return n->in(0) when idom(n) is an IfNode (or // any MultiBranch ctrl node), so apply a similar transform to // the cached idom returned from idom_no_update.
cached_idom = find_non_split_ctrl(cached_idom);
}
tty->print(" ID:%d", computed_idom->_idx);
n->dump(); if (cached_idom != computed_idom) {
tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
computed_idom->_idx, cached_idom->_idx);
}
} // Dump nodes it controls for (uint k = 0; k < _nodes.Size(); k++) { // (k < C->unique() && get_ctrl(find(k)) == n) if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
Node* m = C->root()->find(k); if (m && m->outcnt() > 0) { if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p",
_nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
}
tty->sp(2 * loop->_nest + 1);
m->dump();
}
}
}
}
}
void PhaseIdealLoop::dump_idom(Node* n, const uint count) const { if (has_ctrl(n)) {
tty->print_cr("No idom for data nodes");
} else {
ResourceMark rm;
Unique_Node_List idoms;
get_idoms(n, count, idoms);
dump_idoms_in_reverse(n, idoms);
}
}
void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
Node* next = n; for (uint i = 0; !next->is_Start() && i < count; i++) {
next = idom(next);
assert(!idoms.member(next), "duplicated idom is not possible");
idoms.push(next);
}
}
void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
Node* next;
uint padding = 3;
uint node_index_padding_width = static_cast<int>(log10(C->unique())) + 1; for (int i = idom_list.size() - 1; i >= 0; i--) { if (i == 9 || i == 99) {
padding++;
}
next = idom_list[i];
tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
}
tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
} #endif// NOT PRODUCT
// Collect a R-P-O for the whole CFG. // Result list is in post-order (scan backwards for RPO) void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
stk.push(start, 0);
visited.set(start->_idx);
while (stk.is_nonempty()) {
Node* m = stk.node();
uint idx = stk.index(); if (idx < m->outcnt()) {
stk.set_index(idx + 1);
Node* n = m->raw_out(idx); if (n->is_CFG() && !visited.test_set(n->_idx)) {
stk.push(n, 0);
}
} else {
rpo_list.push(m);
stk.pop();
}
}
}
// Advance to next loop tree using a preorder, left-to-right traversal. void LoopTreeIterator::next() {
assert(!done(), "must not be done."); if (_curnt->_child != NULL) {
_curnt = _curnt->_child;
} elseif (_curnt->_next != NULL) {
_curnt = _curnt->_next;
} else { while (_curnt != _root && _curnt->_next == NULL) {
_curnt = _curnt->_parent;
} if (_curnt == _root) {
_curnt = NULL;
assert(done(), "must be done.");
} else {
assert(_curnt->_next != NULL, "must be more to do");
_curnt = _curnt->_next;
}
}
}
Messung V0.5 in Prozent
¤ Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.0.131Bemerkung:
(vorverarbeitet am 2026-05-02)
¤
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
