6837146: Should perform unswitch before maximally unroll in loop transformation
Summary: Move loop unswitch before maximally unroll
Reviewed-by: never
/*
* Copyright 2000-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
* CA 95054 USA or visit www.sun.com if you need additional information or
* have any questions.
*
*/
#include "incls/_precompiled.incl"
#include "incls/_loopTransform.cpp.incl"
//------------------------------is_loop_exit-----------------------------------
// Given an IfNode, return the loop-exiting projection or NULL if both
// arms remain in the loop.
Node *IdealLoopTree::is_loop_exit(Node *iff) const {
if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
PhaseIdealLoop *phase = _phase;
// Test is an IfNode, has 2 projections. If BOTH are in the loop
// we need loop unswitching instead of peeling.
if( !is_member(phase->get_loop( iff->raw_out(0) )) )
return iff->raw_out(0);
if( !is_member(phase->get_loop( iff->raw_out(1) )) )
return iff->raw_out(1);
return NULL;
}
//=============================================================================
//------------------------------record_for_igvn----------------------------
// Put loop body on igvn work list
void IdealLoopTree::record_for_igvn() {
for( uint i = 0; i < _body.size(); i++ ) {
Node *n = _body.at(i);
_phase->_igvn._worklist.push(n);
}
}
//------------------------------compute_profile_trip_cnt----------------------------
// Compute loop trip count from profile data as
// (backedge_count + loop_exit_count) / loop_exit_count
void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
if (!_head->is_CountedLoop()) {
return;
}
CountedLoopNode* head = _head->as_CountedLoop();
if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
return; // Already computed
}
float trip_cnt = (float)max_jint; // default is big
Node* back = head->in(LoopNode::LoopBackControl);
while (back != head) {
if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
back->in(0) &&
back->in(0)->is_If() &&
back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
break;
}
back = phase->idom(back);
}
if (back != head) {
assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
back->in(0), "if-projection exists");
IfNode* back_if = back->in(0)->as_If();
float loop_back_cnt = back_if->_fcnt * back_if->_prob;
// Now compute a loop exit count
float loop_exit_cnt = 0.0f;
for( uint i = 0; i < _body.size(); i++ ) {
Node *n = _body[i];
if( n->is_If() ) {
IfNode *iff = n->as_If();
if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
Node *exit = is_loop_exit(iff);
if( exit ) {
float exit_prob = iff->_prob;
if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
if (exit_prob > PROB_MIN) {
float exit_cnt = iff->_fcnt * exit_prob;
loop_exit_cnt += exit_cnt;
}
}
}
}
}
if (loop_exit_cnt > 0.0f) {
trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
} else {
// No exit count so use
trip_cnt = loop_back_cnt;
}
}
#ifndef PRODUCT
if (TraceProfileTripCount) {
tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
}
#endif
head->set_profile_trip_cnt(trip_cnt);
}
//---------------------is_invariant_addition-----------------------------
// Return nonzero index of invariant operand for an Add or Sub
// of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
int op = n->Opcode();
if (op == Op_AddI || op == Op_SubI) {
bool in1_invar = this->is_invariant(n->in(1));
bool in2_invar = this->is_invariant(n->in(2));
if (in1_invar && !in2_invar) return 1;
if (!in1_invar && in2_invar) return 2;
}
return 0;
}
//---------------------reassociate_add_sub-----------------------------
// Reassociate invariant add and subtract expressions:
//
// inv1 + (x + inv2) => ( inv1 + inv2) + x
// (x + inv2) + inv1 => ( inv1 + inv2) + x
// inv1 + (x - inv2) => ( inv1 - inv2) + x
// inv1 - (inv2 - x) => ( inv1 - inv2) + x
// (x + inv2) - inv1 => (-inv1 + inv2) + x
// (x - inv2) + inv1 => ( inv1 - inv2) + x
// (x - inv2) - inv1 => (-inv1 - inv2) + x
// inv1 + (inv2 - x) => ( inv1 + inv2) - x
// inv1 - (x - inv2) => ( inv1 + inv2) - x
// (inv2 - x) + inv1 => ( inv1 + inv2) - x
// (inv2 - x) - inv1 => (-inv1 + inv2) - x
// inv1 - (x + inv2) => ( inv1 - inv2) - x
//
Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
if (is_invariant(n1)) return NULL;
int inv1_idx = is_invariant_addition(n1, phase);
if (!inv1_idx) return NULL;
// Don't mess with add of constant (igvn moves them to expression tree root.)
if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
Node* inv1 = n1->in(inv1_idx);
Node* n2 = n1->in(3 - inv1_idx);
int inv2_idx = is_invariant_addition(n2, phase);
if (!inv2_idx) return NULL;
Node* x = n2->in(3 - inv2_idx);
Node* inv2 = n2->in(inv2_idx);
bool neg_x = n2->is_Sub() && inv2_idx == 1;
bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
if (n1->is_Sub() && inv1_idx == 1) {
neg_x = !neg_x;
neg_inv2 = !neg_inv2;
}
Node* inv1_c = phase->get_ctrl(inv1);
Node* inv2_c = phase->get_ctrl(inv2);
Node* n_inv1;
if (neg_inv1) {
Node *zero = phase->_igvn.intcon(0);
phase->set_ctrl(zero, phase->C->root());
n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
phase->register_new_node(n_inv1, inv1_c);
} else {
n_inv1 = inv1;
}
Node* inv;
if (neg_inv2) {
inv = new (phase->C, 3) SubINode(n_inv1, inv2);
} else {
inv = new (phase->C, 3) AddINode(n_inv1, inv2);
}
phase->register_new_node(inv, phase->get_early_ctrl(inv));
Node* addx;
if (neg_x) {
addx = new (phase->C, 3) SubINode(inv, x);
} else {
addx = new (phase->C, 3) AddINode(x, inv);
}
phase->register_new_node(addx, phase->get_ctrl(x));
phase->_igvn.hash_delete(n1);
phase->_igvn.subsume_node(n1, addx);
return addx;
}
//---------------------reassociate_invariants-----------------------------
// Reassociate invariant expressions:
void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
for (int i = _body.size() - 1; i >= 0; i--) {
Node *n = _body.at(i);
for (int j = 0; j < 5; j++) {
Node* nn = reassociate_add_sub(n, phase);
if (nn == NULL) break;
n = nn; // again
};
}
}
//------------------------------policy_peeling---------------------------------
// Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
// make some loop-invariant test (usually a null-check) happen before the loop.
bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
Node *test = ((IdealLoopTree*)this)->tail();
int body_size = ((IdealLoopTree*)this)->_body.size();
int uniq = phase->C->unique();
// Peeling does loop cloning which can result in O(N^2) node construction
if( body_size > 255 /* Prevent overflow for large body_size */
|| (body_size * body_size + uniq > MaxNodeLimit) ) {
return false; // too large to safely clone
}
while( test != _head ) { // Scan till run off top of loop
if( test->is_If() ) { // Test?
Node *ctrl = phase->get_ctrl(test->in(1));
if (ctrl->is_top())
return false; // Found dead test on live IF? No peeling!
// Standard IF only has one input value to check for loop invariance
assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
// Condition is not a member of this loop?
if( !is_member(phase->get_loop(ctrl)) &&
is_loop_exit(test) )
return true; // Found reason to peel!
}
// Walk up dominators to loop _head looking for test which is
// executed on every path thru loop.
test = phase->idom(test);
}
return false;
}
//------------------------------peeled_dom_test_elim---------------------------
// If we got the effect of peeling, either by actually peeling or by making
// a pre-loop which must execute at least once, we can remove all
// loop-invariant dominated tests in the main body.
void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
bool progress = true;
while( progress ) {
progress = false; // Reset for next iteration
Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
Node *test = prev->in(0);
while( test != loop->_head ) { // Scan till run off top of loop
int p_op = prev->Opcode();
if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
test->is_If() && // Test?
!test->in(1)->is_Con() && // And not already obvious?
// Condition is not a member of this loop?
!loop->is_member(get_loop(get_ctrl(test->in(1))))){
// Walk loop body looking for instances of this test
for( uint i = 0; i < loop->_body.size(); i++ ) {
Node *n = loop->_body.at(i);
if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
// IfNode was dominated by version in peeled loop body
progress = true;
dominated_by( old_new[prev->_idx], n );
}
}
}
prev = test;
test = idom(test);
} // End of scan tests in loop
} // End of while( progress )
}
//------------------------------do_peeling-------------------------------------
// Peel the first iteration of the given loop.
// Step 1: Clone the loop body. The clone becomes the peeled iteration.
// The pre-loop illegally has 2 control users (old & new loops).
// Step 2: Make the old-loop fall-in edges point to the peeled iteration.
// Do this by making the old-loop fall-in edges act as if they came
// around the loopback from the prior iteration (follow the old-loop
// backedges) and then map to the new peeled iteration. This leaves
// the pre-loop with only 1 user (the new peeled iteration), but the
// peeled-loop backedge has 2 users.
// Step 3: Cut the backedge on the clone (so its not a loop) and remove the
// extra backedge user.
void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
C->set_major_progress();
// Peeling a 'main' loop in a pre/main/post situation obfuscates the
// 'pre' loop from the main and the 'pre' can no longer have it's
// iterations adjusted. Therefore, we need to declare this loop as
// no longer a 'main' loop; it will need new pre and post loops before
// we can do further RCE.
Node *h = loop->_head;
if( h->is_CountedLoop() ) {
CountedLoopNode *cl = h->as_CountedLoop();
assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
cl->set_trip_count(cl->trip_count() - 1);
if( cl->is_main_loop() ) {
cl->set_normal_loop();
#ifndef PRODUCT
if( PrintOpto && VerifyLoopOptimizations ) {
tty->print("Peeling a 'main' loop; resetting to 'normal' ");
loop->dump_head();
}
#endif
}
}
// Step 1: Clone the loop body. The clone becomes the peeled iteration.
// The pre-loop illegally has 2 control users (old & new loops).
clone_loop( loop, old_new, dom_depth(loop->_head) );
// Step 2: Make the old-loop fall-in edges point to the peeled iteration.
// Do this by making the old-loop fall-in edges act as if they came
// around the loopback from the prior iteration (follow the old-loop
// backedges) and then map to the new peeled iteration. This leaves
// the pre-loop with only 1 user (the new peeled iteration), but the
// peeled-loop backedge has 2 users.
for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) {
Node* old = loop->_head->fast_out(j);
if( old->in(0) == loop->_head && old->req() == 3 &&
(old->is_Loop() || old->is_Phi()) ) {
Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
if( !new_exit_value ) // Backedge value is ALSO loop invariant?
// Then loop body backedge value remains the same.
new_exit_value = old->in(LoopNode::LoopBackControl);
_igvn.hash_delete(old);
old->set_req(LoopNode::EntryControl, new_exit_value);
}
}
// Step 3: Cut the backedge on the clone (so its not a loop) and remove the
// extra backedge user.
Node *nnn = old_new[loop->_head->_idx];
_igvn.hash_delete(nnn);
nnn->set_req(LoopNode::LoopBackControl, C->top());
for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) {
Node* use = nnn->fast_out(j2);
if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) {
_igvn.hash_delete(use);
use->set_req(LoopNode::LoopBackControl, C->top());
}
}
// Step 4: Correct dom-depth info. Set to loop-head depth.
int dd = dom_depth(loop->_head);
set_idom(loop->_head, loop->_head->in(1), dd);
for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
Node *old = loop->_body.at(j3);
Node *nnn = old_new[old->_idx];
if (!has_ctrl(nnn))
set_idom(nnn, idom(nnn), dd-1);
// While we're at it, remove any SafePoints from the peeled code
if( old->Opcode() == Op_SafePoint ) {
Node *nnn = old_new[old->_idx];
lazy_replace(nnn,nnn->in(TypeFunc::Control));
}
}
// Now force out all loop-invariant dominating tests. The optimizer
// finds some, but we _know_ they are all useless.
peeled_dom_test_elim(loop,old_new);
loop->record_for_igvn();
}
//------------------------------policy_maximally_unroll------------------------
// Return exact loop trip count, or 0 if not maximally unrolling
bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
CountedLoopNode *cl = _head->as_CountedLoop();
assert( cl->is_normal_loop(), "" );
Node *init_n = cl->init_trip();
Node *limit_n = cl->limit();
// Non-constant bounds
if( init_n == NULL || !init_n->is_Con() ||
limit_n == NULL || !limit_n->is_Con() ||
// protect against stride not being a constant
!cl->stride_is_con() ) {
return false;
}
int init = init_n->get_int();
int limit = limit_n->get_int();
int span = limit - init;
int stride = cl->stride_con();
if (init >= limit || stride > span) {
// return a false (no maximally unroll) and the regular unroll/peel
// route will make a small mess which CCP will fold away.
return false;
}
uint trip_count = span/stride; // trip_count can be greater than 2 Gig.
assert( (int)trip_count*stride == span, "must divide evenly" );
// Real policy: if we maximally unroll, does it get too big?
// Allow the unrolled mess to get larger than standard loop
// size. After all, it will no longer be a loop.
uint body_size = _body.size();
uint unroll_limit = (uint)LoopUnrollLimit * 4;
assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
cl->set_trip_count(trip_count);
if( trip_count <= unroll_limit && body_size <= unroll_limit ) {
uint new_body_size = body_size * trip_count;
if (new_body_size <= unroll_limit &&
body_size == new_body_size / trip_count &&
// Unrolling can result in a large amount of node construction
new_body_size < MaxNodeLimit - phase->C->unique()) {
return true; // maximally unroll
}
}
return false; // Do not maximally unroll
}
//------------------------------policy_unroll----------------------------------
// Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
// the loop is a CountedLoop and the body is small enough.
bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
CountedLoopNode *cl = _head->as_CountedLoop();
assert( cl->is_normal_loop() || cl->is_main_loop(), "" );
// protect against stride not being a constant
if( !cl->stride_is_con() ) return false;
// protect against over-unrolling
if( cl->trip_count() <= 1 ) return false;
int future_unroll_ct = cl->unrolled_count() * 2;
// Don't unroll if the next round of unrolling would push us
// over the expected trip count of the loop. One is subtracted
// from the expected trip count because the pre-loop normally
// executes 1 iteration.
if (UnrollLimitForProfileCheck > 0 &&
cl->profile_trip_cnt() != COUNT_UNKNOWN &&
future_unroll_ct > UnrollLimitForProfileCheck &&
(float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
return false;
}
// When unroll count is greater than LoopUnrollMin, don't unroll if:
// the residual iterations are more than 10% of the trip count
// and rounds of "unroll,optimize" are not making significant progress
// Progress defined as current size less than 20% larger than previous size.
if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
future_unroll_ct > LoopUnrollMin &&
(future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
return false;
}
Node *init_n = cl->init_trip();
Node *limit_n = cl->limit();
// Non-constant bounds.
// Protect against over-unrolling when init or/and limit are not constant
// (so that trip_count's init value is maxint) but iv range is known.
if( init_n == NULL || !init_n->is_Con() ||
limit_n == NULL || !limit_n->is_Con() ) {
Node* phi = cl->phi();
if( phi != NULL ) {
assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
int next_stride = cl->stride_con() * 2; // stride after this unroll
if( next_stride > 0 ) {
if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow
iv_type->_lo + next_stride > iv_type->_hi ) {
return false; // over-unrolling
}
} else if( next_stride < 0 ) {
if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow
iv_type->_hi + next_stride < iv_type->_lo ) {
return false; // over-unrolling
}
}
}
}
// Adjust body_size to determine if we unroll or not
uint body_size = _body.size();
// Key test to unroll CaffeineMark's Logic test
int xors_in_loop = 0;
// Also count ModL, DivL and MulL which expand mightly
for( uint k = 0; k < _body.size(); k++ ) {
switch( _body.at(k)->Opcode() ) {
case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
case Op_ModL: body_size += 30; break;
case Op_DivL: body_size += 30; break;
case Op_MulL: body_size += 10; break;
}
}
// Check for being too big
if( body_size > (uint)LoopUnrollLimit ) {
if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
// Normal case: loop too big
return false;
}
// Check for stride being a small enough constant
if( abs(cl->stride_con()) > (1<<3) ) return false;
// Unroll once! (Each trip will soon do double iterations)
return true;
}
//------------------------------policy_align-----------------------------------
// Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
// expression that does the alignment. Note that only one array base can be
// aligned in a loop (unless the VM guarantees mutual alignment). Note that
// if we vectorize short memory ops into longer memory ops, we may want to
// increase alignment.
bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
return false;
}
//------------------------------policy_range_check-----------------------------
// Return TRUE or FALSE if the loop should be range-check-eliminated.
// Actually we do iteration-splitting, a more powerful form of RCE.
bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
if( !RangeCheckElimination ) return false;
CountedLoopNode *cl = _head->as_CountedLoop();
// If we unrolled with no intention of doing RCE and we later
// changed our minds, we got no pre-loop. Either we need to
// make a new pre-loop, or we gotta disallow RCE.
if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
Node *trip_counter = cl->phi();
// Check loop body for tests of trip-counter plus loop-invariant vs
// loop-invariant.
for( uint i = 0; i < _body.size(); i++ ) {
Node *iff = _body[i];
if( iff->Opcode() == Op_If ) { // Test?
// Comparing trip+off vs limit
Node *bol = iff->in(1);
if( bol->req() != 2 ) continue; // dead constant test
Node *cmp = bol->in(1);
Node *rc_exp = cmp->in(1);
Node *limit = cmp->in(2);
Node *limit_c = phase->get_ctrl(limit);
if( limit_c == phase->C->top() )
return false; // Found dead test on live IF? No RCE!
if( is_member(phase->get_loop(limit_c) ) ) {
// Compare might have operands swapped; commute them
rc_exp = cmp->in(2);
limit = cmp->in(1);
limit_c = phase->get_ctrl(limit);
if( is_member(phase->get_loop(limit_c) ) )
continue; // Both inputs are loop varying; cannot RCE
}
if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
continue;
}
// Yeah! Found a test like 'trip+off vs limit'
// Test is an IfNode, has 2 projections. If BOTH are in the loop
// we need loop unswitching instead of iteration splitting.
if( is_loop_exit(iff) )
return true; // Found reason to split iterations
} // End of is IF
}
return false;
}
//------------------------------policy_peel_only-------------------------------
// Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
// for unrolling loops with NO array accesses.
bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
for( uint i = 0; i < _body.size(); i++ )
if( _body[i]->is_Mem() )
return false;
// No memory accesses at all!
return true;
}
//------------------------------clone_up_backedge_goo--------------------------
// If Node n lives in the back_ctrl block and cannot float, we clone a private
// version of n in preheader_ctrl block and return that, otherwise return n.
Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
if( get_ctrl(n) != back_ctrl ) return n;
Node *x = NULL; // If required, a clone of 'n'
// Check for 'n' being pinned in the backedge.
if( n->in(0) && n->in(0) == back_ctrl ) {
x = n->clone(); // Clone a copy of 'n' to preheader
x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
}
// Recursive fixup any other input edges into x.
// If there are no changes we can just return 'n', otherwise
// we need to clone a private copy and change it.
for( uint i = 1; i < n->req(); i++ ) {
Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
if( g != n->in(i) ) {
if( !x )
x = n->clone();
x->set_req(i, g);
}
}
if( x ) { // x can legally float to pre-header location
register_new_node( x, preheader_ctrl );
return x;
} else { // raise n to cover LCA of uses
set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
}
return n;
}
//------------------------------insert_pre_post_loops--------------------------
// Insert pre and post loops. If peel_only is set, the pre-loop can not have
// more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
// alignment. Useful to unroll loops that do no array accesses.
void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
C->set_major_progress();
// Find common pieces of the loop being guarded with pre & post loops
CountedLoopNode *main_head = loop->_head->as_CountedLoop();
assert( main_head->is_normal_loop(), "" );
CountedLoopEndNode *main_end = main_head->loopexit();
assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
uint dd_main_head = dom_depth(main_head);
uint max = main_head->outcnt();
Node *pre_header= main_head->in(LoopNode::EntryControl);
Node *init = main_head->init_trip();
Node *incr = main_end ->incr();
Node *limit = main_end ->limit();
Node *stride = main_end ->stride();
Node *cmp = main_end ->cmp_node();
BoolTest::mask b_test = main_end->test_trip();
// Need only 1 user of 'bol' because I will be hacking the loop bounds.
Node *bol = main_end->in(CountedLoopEndNode::TestValue);
if( bol->outcnt() != 1 ) {
bol = bol->clone();
register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
_igvn.hash_delete(main_end);
main_end->set_req(CountedLoopEndNode::TestValue, bol);
}
// Need only 1 user of 'cmp' because I will be hacking the loop bounds.
if( cmp->outcnt() != 1 ) {
cmp = cmp->clone();
register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
_igvn.hash_delete(bol);
bol->set_req(1, cmp);
}
//------------------------------
// Step A: Create Post-Loop.
Node* main_exit = main_end->proj_out(false);
assert( main_exit->Opcode() == Op_IfFalse, "" );
int dd_main_exit = dom_depth(main_exit);
// Step A1: Clone the loop body. The clone becomes the post-loop. The main
// loop pre-header illegally has 2 control users (old & new loops).
clone_loop( loop, old_new, dd_main_exit );
assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
post_head->set_post_loop(main_head);
// Reduce the post-loop trip count.
CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
post_end->_prob = PROB_FAIR;
// Build the main-loop normal exit.
IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
_igvn.register_new_node_with_optimizer( new_main_exit );
set_idom(new_main_exit, main_end, dd_main_exit );
set_loop(new_main_exit, loop->_parent);
// Step A2: Build a zero-trip guard for the post-loop. After leaving the
// main-loop, the post-loop may not execute at all. We 'opaque' the incr
// (the main-loop trip-counter exit value) because we will be changing
// the exit value (via unrolling) so we cannot constant-fold away the zero
// trip guard until all unrolling is done.
Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
register_new_node( zer_opaq, new_main_exit );
register_new_node( zer_cmp , new_main_exit );
register_new_node( zer_bol , new_main_exit );
// Build the IfNode
IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
_igvn.register_new_node_with_optimizer( zer_iff );
set_idom(zer_iff, new_main_exit, dd_main_exit);
set_loop(zer_iff, loop->_parent);
// Plug in the false-path, taken if we need to skip post-loop
_igvn.hash_delete( main_exit );
main_exit->set_req(0, zer_iff);
_igvn._worklist.push(main_exit);
set_idom(main_exit, zer_iff, dd_main_exit);
set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
// Make the true-path, must enter the post loop
Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
_igvn.register_new_node_with_optimizer( zer_taken );
set_idom(zer_taken, zer_iff, dd_main_exit);
set_loop(zer_taken, loop->_parent);
// Plug in the true path
_igvn.hash_delete( post_head );
post_head->set_req(LoopNode::EntryControl, zer_taken);
set_idom(post_head, zer_taken, dd_main_exit);
// Step A3: Make the fall-in values to the post-loop come from the
// fall-out values of the main-loop.
for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
Node* main_phi = main_head->fast_out(i);
if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
Node *post_phi = old_new[main_phi->_idx];
Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
post_head->init_control(),
main_phi->in(LoopNode::LoopBackControl));
_igvn.hash_delete(post_phi);
post_phi->set_req( LoopNode::EntryControl, fallmain );
}
}
// Update local caches for next stanza
main_exit = new_main_exit;
//------------------------------
// Step B: Create Pre-Loop.
// Step B1: Clone the loop body. The clone becomes the pre-loop. The main
// loop pre-header illegally has 2 control users (old & new loops).
clone_loop( loop, old_new, dd_main_head );
CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
pre_head->set_pre_loop(main_head);
Node *pre_incr = old_new[incr->_idx];
// Reduce the pre-loop trip count.
pre_end->_prob = PROB_FAIR;
// Find the pre-loop normal exit.
Node* pre_exit = pre_end->proj_out(false);
assert( pre_exit->Opcode() == Op_IfFalse, "" );
IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
_igvn.register_new_node_with_optimizer( new_pre_exit );
set_idom(new_pre_exit, pre_end, dd_main_head);
set_loop(new_pre_exit, loop->_parent);
// Step B2: Build a zero-trip guard for the main-loop. After leaving the
// pre-loop, the main-loop may not execute at all. Later in life this
// zero-trip guard will become the minimum-trip guard when we unroll
// the main-loop.
Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
register_new_node( min_opaq, new_pre_exit );
register_new_node( min_cmp , new_pre_exit );
register_new_node( min_bol , new_pre_exit );
// Build the IfNode (assume the main-loop is executed always).
IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
_igvn.register_new_node_with_optimizer( min_iff );
set_idom(min_iff, new_pre_exit, dd_main_head);
set_loop(min_iff, loop->_parent);
// Plug in the false-path, taken if we need to skip main-loop
_igvn.hash_delete( pre_exit );
pre_exit->set_req(0, min_iff);
set_idom(pre_exit, min_iff, dd_main_head);
set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
// Make the true-path, must enter the main loop
Node *min_taken = new (C, 1) IfTrueNode( min_iff );
_igvn.register_new_node_with_optimizer( min_taken );
set_idom(min_taken, min_iff, dd_main_head);
set_loop(min_taken, loop->_parent);
// Plug in the true path
_igvn.hash_delete( main_head );
main_head->set_req(LoopNode::EntryControl, min_taken);
set_idom(main_head, min_taken, dd_main_head);
// Step B3: Make the fall-in values to the main-loop come from the
// fall-out values of the pre-loop.
for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
Node* main_phi = main_head->fast_out(i2);
if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
Node *pre_phi = old_new[main_phi->_idx];
Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
main_head->init_control(),
pre_phi->in(LoopNode::LoopBackControl));
_igvn.hash_delete(main_phi);
main_phi->set_req( LoopNode::EntryControl, fallpre );
}
}
// Step B4: Shorten the pre-loop to run only 1 iteration (for now).
// RCE and alignment may change this later.
Node *cmp_end = pre_end->cmp_node();
assert( cmp_end->in(2) == limit, "" );
Node *pre_limit = new (C, 3) AddINode( init, stride );
// Save the original loop limit in this Opaque1 node for
// use by range check elimination.
Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
register_new_node( pre_limit, pre_head->in(0) );
register_new_node( pre_opaq , pre_head->in(0) );
// Since no other users of pre-loop compare, I can hack limit directly
assert( cmp_end->outcnt() == 1, "no other users" );
_igvn.hash_delete(cmp_end);
cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
// Special case for not-equal loop bounds:
// Change pre loop test, main loop test, and the
// main loop guard test to use lt or gt depending on stride
// direction:
// positive stride use <
// negative stride use >
if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
// Modify pre loop end condition
Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
register_new_node( new_bol0, pre_head->in(0) );
_igvn.hash_delete(pre_end);
pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
// Modify main loop guard condition
assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
register_new_node( new_bol1, new_pre_exit );
_igvn.hash_delete(min_iff);
min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
// Modify main loop end condition
BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
_igvn.hash_delete(main_end);
main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
}
// Flag main loop
main_head->set_main_loop();
if( peel_only ) main_head->set_main_no_pre_loop();
// It's difficult to be precise about the trip-counts
// for the pre/post loops. They are usually very short,
// so guess that 4 trips is a reasonable value.
post_head->set_profile_trip_cnt(4.0);
pre_head->set_profile_trip_cnt(4.0);
// Now force out all loop-invariant dominating tests. The optimizer
// finds some, but we _know_ they are all useless.
peeled_dom_test_elim(loop,old_new);
}
//------------------------------is_invariant-----------------------------
// Return true if n is invariant
bool IdealLoopTree::is_invariant(Node* n) const {
Node *n_c = _phase->get_ctrl(n);
if (n_c->is_top()) return false;
return !is_member(_phase->get_loop(n_c));
}
//------------------------------do_unroll--------------------------------------
// Unroll the loop body one step - make each trip do 2 iterations.
void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
assert( LoopUnrollLimit, "" );
#ifndef PRODUCT
if( PrintOpto && VerifyLoopOptimizations ) {
tty->print("Unrolling ");
loop->dump_head();
}
#endif
CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
CountedLoopEndNode *loop_end = loop_head->loopexit();
assert( loop_end, "" );
// Remember loop node count before unrolling to detect
// if rounds of unroll,optimize are making progress
loop_head->set_node_count_before_unroll(loop->_body.size());
Node *ctrl = loop_head->in(LoopNode::EntryControl);
Node *limit = loop_head->limit();
Node *init = loop_head->init_trip();
Node *strid = loop_head->stride();
Node *opaq = NULL;
if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
assert( loop_head->is_main_loop(), "" );
assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
Node *iff = ctrl->in(0);
assert( iff->Opcode() == Op_If, "" );
Node *bol = iff->in(1);
assert( bol->Opcode() == Op_Bool, "" );
Node *cmp = bol->in(1);
assert( cmp->Opcode() == Op_CmpI, "" );
opaq = cmp->in(2);
// Occasionally it's possible for a pre-loop Opaque1 node to be
// optimized away and then another round of loop opts attempted.
// We can not optimize this particular loop in that case.
if( opaq->Opcode() != Op_Opaque1 )
return; // Cannot find pre-loop! Bail out!
}
C->set_major_progress();
// Adjust max trip count. The trip count is intentionally rounded
// down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
// the main, unrolled, part of the loop will never execute as it is protected
// by the min-trip test. See bug 4834191 for a case where we over-unrolled
// and later determined that part of the unrolled loop was dead.
loop_head->set_trip_count(loop_head->trip_count() / 2);
// Double the count of original iterations in the unrolled loop body.
loop_head->double_unrolled_count();
// -----------
// Step 2: Cut back the trip counter for an unroll amount of 2.
// Loop will normally trip (limit - init)/stride_con. Since it's a
// CountedLoop this is exact (stride divides limit-init exactly).
// We are going to double the loop body, so we want to knock off any
// odd iteration: (trip_cnt & ~1). Then back compute a new limit.
Node *span = new (C, 3) SubINode( limit, init );
register_new_node( span, ctrl );
Node *trip = new (C, 3) DivINode( 0, span, strid );
register_new_node( trip, ctrl );
Node *mtwo = _igvn.intcon(-2);
set_ctrl(mtwo, C->root());
Node *rond = new (C, 3) AndINode( trip, mtwo );
register_new_node( rond, ctrl );
Node *spn2 = new (C, 3) MulINode( rond, strid );
register_new_node( spn2, ctrl );
Node *lim2 = new (C, 3) AddINode( spn2, init );
register_new_node( lim2, ctrl );
// Hammer in the new limit
Node *ctrl2 = loop_end->in(0);
Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
register_new_node( cmp2, ctrl2 );
Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
register_new_node( bol2, ctrl2 );
_igvn.hash_delete(loop_end);
loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
// Step 3: Find the min-trip test guaranteed before a 'main' loop.
// Make it a 1-trip test (means at least 2 trips).
if( adjust_min_trip ) {
// Guard test uses an 'opaque' node which is not shared. Hence I
// can edit it's inputs directly. Hammer in the new limit for the
// minimum-trip guard.
assert( opaq->outcnt() == 1, "" );
_igvn.hash_delete(opaq);
opaq->set_req(1, lim2);
}
// ---------
// Step 4: Clone the loop body. Move it inside the loop. This loop body
// represents the odd iterations; since the loop trips an even number of
// times its backedge is never taken. Kill the backedge.
uint dd = dom_depth(loop_head);
clone_loop( loop, old_new, dd );
// Make backedges of the clone equal to backedges of the original.
// Make the fall-in from the original come from the fall-out of the clone.
for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
Node* phi = loop_head->fast_out(j);
if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
Node *newphi = old_new[phi->_idx];
_igvn.hash_delete( phi );
_igvn.hash_delete( newphi );
phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
phi ->set_req(LoopNode::LoopBackControl, C->top());
}
}
Node *clone_head = old_new[loop_head->_idx];
_igvn.hash_delete( clone_head );
loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
loop_head ->set_req(LoopNode::LoopBackControl, C->top());
loop->_head = clone_head; // New loop header
set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
// Kill the clone's backedge
Node *newcle = old_new[loop_end->_idx];
_igvn.hash_delete( newcle );
Node *one = _igvn.intcon(1);
set_ctrl(one, C->root());
newcle->set_req(1, one);
// Force clone into same loop body
uint max = loop->_body.size();
for( uint k = 0; k < max; k++ ) {
Node *old = loop->_body.at(k);
Node *nnn = old_new[old->_idx];
loop->_body.push(nnn);
if (!has_ctrl(old))
set_loop(nnn, loop);
}
loop->record_for_igvn();
}
//------------------------------do_maximally_unroll----------------------------
void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
CountedLoopNode *cl = loop->_head->as_CountedLoop();
assert( cl->trip_count() > 0, "");
// If loop is tripping an odd number of times, peel odd iteration
if( (cl->trip_count() & 1) == 1 ) {
do_peeling( loop, old_new );
}
// Now its tripping an even number of times remaining. Double loop body.
// Do not adjust pre-guards; they are not needed and do not exist.
if( cl->trip_count() > 0 ) {
do_unroll( loop, old_new, false );
}
}
//------------------------------dominates_backedge---------------------------------
// Returns true if ctrl is executed on every complete iteration
bool IdealLoopTree::dominates_backedge(Node* ctrl) {
assert(ctrl->is_CFG(), "must be control");
Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
}
//------------------------------add_constraint---------------------------------
// Constrain the main loop iterations so the condition:
// scale_con * I + offset < limit
// always holds true. That is, either increase the number of iterations in
// the pre-loop or the post-loop until the condition holds true in the main
// loop. Stride, scale, offset and limit are all loop invariant. Further,
// stride and scale are constants (offset and limit often are).
void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
// Compute "I :: (limit-offset)/scale_con"
Node *con = new (C, 3) SubINode( limit, offset );
register_new_node( con, pre_ctrl );
Node *scale = _igvn.intcon(scale_con);
set_ctrl(scale, C->root());
Node *X = new (C, 3) DivINode( 0, con, scale );
register_new_node( X, pre_ctrl );
// For positive stride, the pre-loop limit always uses a MAX function
// and the main loop a MIN function. For negative stride these are
// reversed.
// Also for positive stride*scale the affine function is increasing, so the
// pre-loop must check for underflow and the post-loop for overflow.
// Negative stride*scale reverses this; pre-loop checks for overflow and
// post-loop for underflow.
if( stride_con*scale_con > 0 ) {
// Compute I < (limit-offset)/scale_con
// Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
*main_limit = (stride_con > 0)
? (Node*)(new (C, 3) MinINode( *main_limit, X ))
: (Node*)(new (C, 3) MaxINode( *main_limit, X ));
register_new_node( *main_limit, pre_ctrl );
} else {
// Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
// Add the negation of the main-loop constraint to the pre-loop.
// See footnote [++] below for a derivation of the limit expression.
Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
set_ctrl(incr, C->root());
Node *adj = new (C, 3) AddINode( X, incr );
register_new_node( adj, pre_ctrl );
*pre_limit = (scale_con > 0)
? (Node*)new (C, 3) MinINode( *pre_limit, adj )
: (Node*)new (C, 3) MaxINode( *pre_limit, adj );
register_new_node( *pre_limit, pre_ctrl );
// [++] Here's the algebra that justifies the pre-loop limit expression:
//
// NOT( scale_con * I + offset < limit )
// ==
// scale_con * I + offset >= limit
// ==
// SGN(scale_con) * I >= (limit-offset)/|scale_con|
// ==
// (limit-offset)/|scale_con| <= I * SGN(scale_con)
// ==
// (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
// ==
// ( if (scale_con > 0) /*common case*/
// (limit-offset)/scale_con - 1 < I
// else
// (limit-offset)/scale_con + 1 > I
// )
// ( if (scale_con > 0) /*common case*/
// (limit-offset)/scale_con + SGN(-scale_con) < I
// else
// (limit-offset)/scale_con + SGN(-scale_con) > I
}
}
//------------------------------is_scaled_iv---------------------------------
// Return true if exp is a constant times an induction var
bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
if (exp == iv) {
if (p_scale != NULL) {
*p_scale = 1;
}
return true;
}
int opc = exp->Opcode();
if (opc == Op_MulI) {
if (exp->in(1) == iv && exp->in(2)->is_Con()) {
if (p_scale != NULL) {
*p_scale = exp->in(2)->get_int();
}
return true;
}
if (exp->in(2) == iv && exp->in(1)->is_Con()) {
if (p_scale != NULL) {
*p_scale = exp->in(1)->get_int();
}
return true;
}
} else if (opc == Op_LShiftI) {
if (exp->in(1) == iv && exp->in(2)->is_Con()) {
if (p_scale != NULL) {
*p_scale = 1 << exp->in(2)->get_int();
}
return true;
}
}
return false;
}
//-----------------------------is_scaled_iv_plus_offset------------------------------
// Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
if (is_scaled_iv(exp, iv, p_scale)) {
if (p_offset != NULL) {
Node *zero = _igvn.intcon(0);
set_ctrl(zero, C->root());
*p_offset = zero;
}
return true;
}
int opc = exp->Opcode();
if (opc == Op_AddI) {
if (is_scaled_iv(exp->in(1), iv, p_scale)) {
if (p_offset != NULL) {
*p_offset = exp->in(2);
}
return true;
}
if (exp->in(2)->is_Con()) {
Node* offset2 = NULL;
if (depth < 2 &&
is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
p_offset != NULL ? &offset2 : NULL, depth+1)) {
if (p_offset != NULL) {
Node *ctrl_off2 = get_ctrl(offset2);
Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
register_new_node(offset, ctrl_off2);
*p_offset = offset;
}
return true;
}
}
} else if (opc == Op_SubI) {
if (is_scaled_iv(exp->in(1), iv, p_scale)) {
if (p_offset != NULL) {
Node *zero = _igvn.intcon(0);
set_ctrl(zero, C->root());
Node *ctrl_off = get_ctrl(exp->in(2));
Node* offset = new (C, 3) SubINode(zero, exp->in(2));
register_new_node(offset, ctrl_off);
*p_offset = offset;
}
return true;
}
if (is_scaled_iv(exp->in(2), iv, p_scale)) {
if (p_offset != NULL) {
*p_scale *= -1;
*p_offset = exp->in(1);
}
return true;
}
}
return false;
}
//------------------------------do_range_check---------------------------------
// Eliminate range-checks and other trip-counter vs loop-invariant tests.
void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
#ifndef PRODUCT
if( PrintOpto && VerifyLoopOptimizations ) {
tty->print("Range Check Elimination ");
loop->dump_head();
}
#endif
assert( RangeCheckElimination, "" );
CountedLoopNode *cl = loop->_head->as_CountedLoop();
assert( cl->is_main_loop(), "" );
// Find the trip counter; we are iteration splitting based on it
Node *trip_counter = cl->phi();
// Find the main loop limit; we will trim it's iterations
// to not ever trip end tests
Node *main_limit = cl->limit();
// Find the pre-loop limit; we will expand it's iterations to
// not ever trip low tests.
Node *ctrl = cl->in(LoopNode::EntryControl);
assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
Node *iffm = ctrl->in(0);
assert( iffm->Opcode() == Op_If, "" );
Node *p_f = iffm->in(0);
assert( p_f->Opcode() == Op_IfFalse, "" );
CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
assert( pre_end->loopnode()->is_pre_loop(), "" );
Node *pre_opaq1 = pre_end->limit();
// Occasionally it's possible for a pre-loop Opaque1 node to be
// optimized away and then another round of loop opts attempted.
// We can not optimize this particular loop in that case.
if( pre_opaq1->Opcode() != Op_Opaque1 )
return;
Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
Node *pre_limit = pre_opaq->in(1);
// Where do we put new limit calculations
Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
// Ensure the original loop limit is available from the
// pre-loop Opaque1 node.
Node *orig_limit = pre_opaq->original_loop_limit();
if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP )
return;
// Need to find the main-loop zero-trip guard
Node *bolzm = iffm->in(1);
assert( bolzm->Opcode() == Op_Bool, "" );
Node *cmpzm = bolzm->in(1);
assert( cmpzm->is_Cmp(), "" );
Node *opqzm = cmpzm->in(2);
if( opqzm->Opcode() != Op_Opaque1 )
return;
assert( opqzm->in(1) == main_limit, "do not understand situation" );
// Must know if its a count-up or count-down loop
// protect against stride not being a constant
if ( !cl->stride_is_con() ) {
return;
}
int stride_con = cl->stride_con();
Node *zero = _igvn.intcon(0);
Node *one = _igvn.intcon(1);
set_ctrl(zero, C->root());
set_ctrl(one, C->root());
// Range checks that do not dominate the loop backedge (ie.
// conditionally executed) can lengthen the pre loop limit beyond
// the original loop limit. To prevent this, the pre limit is
// (for stride > 0) MINed with the original loop limit (MAXed
// stride < 0) when some range_check (rc) is conditionally
// executed.
bool conditional_rc = false;
// Check loop body for tests of trip-counter plus loop-invariant vs
// loop-invariant.
for( uint i = 0; i < loop->_body.size(); i++ ) {
Node *iff = loop->_body[i];
if( iff->Opcode() == Op_If ) { // Test?
// Test is an IfNode, has 2 projections. If BOTH are in the loop
// we need loop unswitching instead of iteration splitting.
Node *exit = loop->is_loop_exit(iff);
if( !exit ) continue;
int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
// Get boolean condition to test
Node *i1 = iff->in(1);
if( !i1->is_Bool() ) continue;
BoolNode *bol = i1->as_Bool();
BoolTest b_test = bol->_test;
// Flip sense of test if exit condition is flipped
if( flip )
b_test = b_test.negate();
// Get compare
Node *cmp = bol->in(1);
// Look for trip_counter + offset vs limit
Node *rc_exp = cmp->in(1);
Node *limit = cmp->in(2);
jint scale_con= 1; // Assume trip counter not scaled
Node *limit_c = get_ctrl(limit);
if( loop->is_member(get_loop(limit_c) ) ) {
// Compare might have operands swapped; commute them
b_test = b_test.commute();
rc_exp = cmp->in(2);
limit = cmp->in(1);
limit_c = get_ctrl(limit);
if( loop->is_member(get_loop(limit_c) ) )
continue; // Both inputs are loop varying; cannot RCE
}
// Here we know 'limit' is loop invariant
// 'limit' maybe pinned below the zero trip test (probably from a
// previous round of rce), in which case, it can't be used in the
// zero trip test expression which must occur before the zero test's if.
if( limit_c == ctrl ) {
continue; // Don't rce this check but continue looking for other candidates.
}
// Check for scaled induction variable plus an offset
Node *offset = NULL;
if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
continue;
}
Node *offset_c = get_ctrl(offset);
if( loop->is_member( get_loop(offset_c) ) )
continue; // Offset is not really loop invariant
// Here we know 'offset' is loop invariant.
// As above for the 'limit', the 'offset' maybe pinned below the
// zero trip test.
if( offset_c == ctrl ) {
continue; // Don't rce this check but continue looking for other candidates.
}
// At this point we have the expression as:
// scale_con * trip_counter + offset :: limit
// where scale_con, offset and limit are loop invariant. Trip_counter
// monotonically increases by stride_con, a constant. Both (or either)
// stride_con and scale_con can be negative which will flip about the
// sense of the test.
// Adjust pre and main loop limits to guard the correct iteration set
if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
if( b_test._test == BoolTest::lt ) { // Range checks always use lt
// The overflow limit: scale*I+offset < limit
add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
// The underflow limit: 0 <= scale*I+offset.
// Some math yields: -scale*I-(offset+1) < 0
Node *plus_one = new (C, 3) AddINode( offset, one );
register_new_node( plus_one, pre_ctrl );
Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
register_new_node( neg_offset, pre_ctrl );
add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
if (!conditional_rc) {
conditional_rc = !loop->dominates_backedge(iff);
}
} else {
#ifndef PRODUCT
if( PrintOpto )
tty->print_cr("missed RCE opportunity");
#endif
continue; // In release mode, ignore it
}
} else { // Otherwise work on normal compares
switch( b_test._test ) {
case BoolTest::ge: // Convert X >= Y to -X <= -Y
scale_con = -scale_con;
offset = new (C, 3) SubINode( zero, offset );
register_new_node( offset, pre_ctrl );
limit = new (C, 3) SubINode( zero, limit );
register_new_node( limit, pre_ctrl );
// Fall into LE case
case BoolTest::le: // Convert X <= Y to X < Y+1
limit = new (C, 3) AddINode( limit, one );
register_new_node( limit, pre_ctrl );
// Fall into LT case
case BoolTest::lt:
add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
if (!conditional_rc) {
conditional_rc = !loop->dominates_backedge(iff);
}
break;
default:
#ifndef PRODUCT
if( PrintOpto )
tty->print_cr("missed RCE opportunity");
#endif
continue; // Unhandled case
}
}
// Kill the eliminated test
C->set_major_progress();
Node *kill_con = _igvn.intcon( 1-flip );
set_ctrl(kill_con, C->root());
_igvn.hash_delete(iff);
iff->set_req(1, kill_con);
_igvn._worklist.push(iff);
// Find surviving projection
assert(iff->is_If(), "");
ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
// Find loads off the surviving projection; remove their control edge
for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
Node* cd = dp->fast_out(i); // Control-dependent node
if( cd->is_Load() ) { // Loads can now float around in the loop
_igvn.hash_delete(cd);
// Allow the load to float around in the loop, or before it
// but NOT before the pre-loop.
cd->set_req(0, ctrl); // ctrl, not NULL
_igvn._worklist.push(cd);
--i;
--imax;
}
}
} // End of is IF
}
// Update loop limits
if (conditional_rc) {
pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
: (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
register_new_node(pre_limit, pre_ctrl);
}
_igvn.hash_delete(pre_opaq);
pre_opaq->set_req(1, pre_limit);
// Note:: we are making the main loop limit no longer precise;
// need to round up based on stride.
if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
// "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
// Hopefully, compiler will optimize for powers of 2.
Node *ctrl = get_ctrl(main_limit);
Node *stride = cl->stride();
Node *init = cl->init_trip();
Node *span = new (C, 3) SubINode(main_limit,init);
register_new_node(span,ctrl);
Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
Node *add = new (C, 3) AddINode(span,rndup);
register_new_node(add,ctrl);
Node *div = new (C, 3) DivINode(0,add,stride);
register_new_node(div,ctrl);
Node *mul = new (C, 3) MulINode(div,stride);
register_new_node(mul,ctrl);
Node *newlim = new (C, 3) AddINode(mul,init);
register_new_node(newlim,ctrl);
main_limit = newlim;
}
Node *main_cle = cl->loopexit();
Node *main_bol = main_cle->in(1);
// Hacking loop bounds; need private copies of exit test
if( main_bol->outcnt() > 1 ) {// BoolNode shared?
_igvn.hash_delete(main_cle);
main_bol = main_bol->clone();// Clone a private BoolNode
register_new_node( main_bol, main_cle->in(0) );
main_cle->set_req(1,main_bol);
}
Node *main_cmp = main_bol->in(1);
if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
_igvn.hash_delete(main_bol);
main_cmp = main_cmp->clone();// Clone a private CmpNode
register_new_node( main_cmp, main_cle->in(0) );
main_bol->set_req(1,main_cmp);
}
// Hack the now-private loop bounds
_igvn.hash_delete(main_cmp);
main_cmp->set_req(2, main_limit);
_igvn._worklist.push(main_cmp);
// The OpaqueNode is unshared by design
_igvn.hash_delete(opqzm);
assert( opqzm->outcnt() == 1, "cannot hack shared node" );
opqzm->set_req(1,main_limit);
_igvn._worklist.push(opqzm);
}
//------------------------------DCE_loop_body----------------------------------
// Remove simplistic dead code from loop body
void IdealLoopTree::DCE_loop_body() {
for( uint i = 0; i < _body.size(); i++ )
if( _body.at(i)->outcnt() == 0 )
_body.map( i--, _body.pop() );
}
//------------------------------adjust_loop_exit_prob--------------------------
// Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
// Replace with a 1-in-10 exit guess.
void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
Node *test = tail();
while( test != _head ) {
uint top = test->Opcode();
if( top == Op_IfTrue || top == Op_IfFalse ) {
int test_con = ((ProjNode*)test)->_con;
assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
IfNode *iff = test->in(0)->as_If();
if( iff->outcnt() == 2 ) { // Ignore dead tests
Node *bol = iff->in(1);
if( bol && bol->req() > 1 && bol->in(1) &&
((bol->in(1)->Opcode() == Op_StorePConditional ) ||
(bol->in(1)->Opcode() == Op_StoreIConditional ) ||
(bol->in(1)->Opcode() == Op_StoreLConditional ) ||
(bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
(bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
(bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
(bol->in(1)->Opcode() == Op_CompareAndSwapN )))
return; // Allocation loops RARELY take backedge
// Find the OTHER exit path from the IF
Node* ex = iff->proj_out(1-test_con);
float p = iff->_prob;
if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
if( top == Op_IfTrue ) {
if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
iff->_prob = PROB_STATIC_FREQUENT;
}
} else {
if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
iff->_prob = PROB_STATIC_INFREQUENT;
}
}
}
}
}
test = phase->idom(test);
}
}
//------------------------------policy_do_remove_empty_loop--------------------
// Micro-benchmark spamming. Policy is to always remove empty loops.
// The 'DO' part is to replace the trip counter with the value it will
// have on the last iteration. This will break the loop.
bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
// Minimum size must be empty loop
if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false;
if( !_head->is_CountedLoop() ) return false; // Dead loop
CountedLoopNode *cl = _head->as_CountedLoop();
if( !cl->loopexit() ) return false; // Malformed loop
if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) )
return false; // Infinite loop
#ifndef PRODUCT
if( PrintOpto )
tty->print_cr("Removing empty loop");
#endif
#ifdef ASSERT
// Ensure only one phi which is the iv.
Node* iv = NULL;
for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
Node* n = cl->fast_out(i);
if (n->Opcode() == Op_Phi) {
assert(iv == NULL, "Too many phis" );
iv = n;
}
}
assert(iv == cl->phi(), "Wrong phi" );
#endif
// Replace the phi at loop head with the final value of the last
// iteration. Then the CountedLoopEnd will collapse (backedge never
// taken) and all loop-invariant uses of the exit values will be correct.
Node *phi = cl->phi();
Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
phase->register_new_node(final,cl->in(LoopNode::EntryControl));
phase->_igvn.hash_delete(phi);
phase->_igvn.subsume_node(phi,final);
phase->C->set_major_progress();
return true;
}
//=============================================================================
//------------------------------iteration_split_impl---------------------------
bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
// Check and remove empty loops (spam micro-benchmarks)
if( policy_do_remove_empty_loop(phase) )
return true; // Here we removed an empty loop
bool should_peel = policy_peeling(phase); // Should we peel?
bool should_unswitch = policy_unswitching(phase);
// Non-counted loops may be peeled; exactly 1 iteration is peeled.
// This removes loop-invariant tests (usually null checks).
if( !_head->is_CountedLoop() ) { // Non-counted loop
if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
// Partial peel succeeded so terminate this round of loop opts
return false;
}
if( should_peel ) { // Should we peel?
#ifndef PRODUCT
if (PrintOpto) tty->print_cr("should_peel");
#endif
phase->do_peeling(this,old_new);
} else if( should_unswitch ) {
phase->do_unswitching(this, old_new);
}
return true;
}
CountedLoopNode *cl = _head->as_CountedLoop();
if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
// Do nothing special to pre- and post- loops
if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
// Compute loop trip count from profile data
compute_profile_trip_cnt(phase);
// Before attempting fancy unrolling, RCE or alignment, see if we want
// to completely unroll this loop or do loop unswitching.
if( cl->is_normal_loop() ) {
if (should_unswitch) {
phase->do_unswitching(this, old_new);
return true;
}
bool should_maximally_unroll = policy_maximally_unroll(phase);
if( should_maximally_unroll ) {
// Here we did some unrolling and peeling. Eventually we will
// completely unroll this loop and it will no longer be a loop.
phase->do_maximally_unroll(this,old_new);
return true;
}
}
// Counted loops may be peeled, may need some iterations run up
// front for RCE, and may want to align loop refs to a cache
// line. Thus we clone a full loop up front whose trip count is
// at least 1 (if peeling), but may be several more.
// The main loop will start cache-line aligned with at least 1
// iteration of the unrolled body (zero-trip test required) and
// will have some range checks removed.
// A post-loop will finish any odd iterations (leftover after
// unrolling), plus any needed for RCE purposes.
bool should_unroll = policy_unroll(phase);
bool should_rce = policy_range_check(phase);
bool should_align = policy_align(phase);
// If not RCE'ing (iteration splitting) or Aligning, then we do not
// need a pre-loop. We may still need to peel an initial iteration but
// we will not be needing an unknown number of pre-iterations.
//
// Basically, if may_rce_align reports FALSE first time through,
// we will not be able to later do RCE or Aligning on this loop.
bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
// If we have any of these conditions (RCE, alignment, unrolling) met, then
// we switch to the pre-/main-/post-loop model. This model also covers
// peeling.
if( should_rce || should_align || should_unroll ) {
if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
phase->insert_pre_post_loops(this,old_new, !may_rce_align);
// Adjust the pre- and main-loop limits to let the pre and post loops run
// with full checks, but the main-loop with no checks. Remove said
// checks from the main body.
if( should_rce )
phase->do_range_check(this,old_new);
// Double loop body for unrolling. Adjust the minimum-trip test (will do
// twice as many iterations as before) and the main body limit (only do
// an even number of trips). If we are peeling, we might enable some RCE
// and we'd rather unroll the post-RCE'd loop SO... do not unroll if
// peeling.
if( should_unroll && !should_peel )
phase->do_unroll(this,old_new, true);
// Adjust the pre-loop limits to align the main body
// iterations.
if( should_align )
Unimplemented();
} else { // Else we have an unchanged counted loop
if( should_peel ) // Might want to peel but do nothing else
phase->do_peeling(this,old_new);
}
return true;
}
//=============================================================================
//------------------------------iteration_split--------------------------------
bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
// Recursively iteration split nested loops
if( _child && !_child->iteration_split( phase, old_new ))
return false;
// Clean out prior deadwood
DCE_loop_body();
// Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
// Replace with a 1-in-10 exit guess.
if( _parent /*not the root loop*/ &&
!_irreducible &&
// Also ignore the occasional dead backedge
!tail()->is_top() ) {
adjust_loop_exit_prob(phase);
}
// Gate unrolling, RCE and peeling efforts.
if( !_child && // If not an inner loop, do not split
!_irreducible &&
_allow_optimizations &&
!tail()->is_top() ) { // Also ignore the occasional dead backedge
if (!_has_call) {
if (!iteration_split_impl( phase, old_new )) {
return false;
}
} else if (policy_unswitching(phase)) {
phase->do_unswitching(this, old_new);
}
}
// Minor offset re-organization to remove loop-fallout uses of
// trip counter.
if( _head->is_CountedLoop() ) phase->reorg_offsets( this );
if( _next && !_next->iteration_split( phase, old_new ))
return false;
return true;
}