8071302: assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo], def)) failed: after block local
Summary: Add merge nodes to node to block mapping
Reviewed-by: kvn, vlivanov
/*
* Copyright (c) 2000, 2013, 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.
*
*/
#include "precompiled.hpp"
#include "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/connode.hpp"
#include "opto/loopnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/runtime.hpp"
#include "opto/subnode.hpp"
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
extern int explicit_null_checks_elided;
//=============================================================================
//------------------------------Value------------------------------------------
// Return a tuple for whichever arm of the IF is reachable
const Type *IfNode::Value( PhaseTransform *phase ) const {
if( !in(0) ) return Type::TOP;
if( phase->type(in(0)) == Type::TOP )
return Type::TOP;
const Type *t = phase->type(in(1));
if( t == Type::TOP ) // data is undefined
return TypeTuple::IFNEITHER; // unreachable altogether
if( t == TypeInt::ZERO ) // zero, or false
return TypeTuple::IFFALSE; // only false branch is reachable
if( t == TypeInt::ONE ) // 1, or true
return TypeTuple::IFTRUE; // only true branch is reachable
assert( t == TypeInt::BOOL, "expected boolean type" );
return TypeTuple::IFBOTH; // No progress
}
const RegMask &IfNode::out_RegMask() const {
return RegMask::Empty;
}
//------------------------------split_if---------------------------------------
// Look for places where we merge constants, then test on the merged value.
// If the IF test will be constant folded on the path with the constant, we
// win by splitting the IF to before the merge point.
static Node* split_if(IfNode *iff, PhaseIterGVN *igvn) {
// I could be a lot more general here, but I'm trying to squeeze this
// in before the Christmas '98 break so I'm gonna be kinda restrictive
// on the patterns I accept. CNC
// Look for a compare of a constant and a merged value
Node *i1 = iff->in(1);
if( !i1->is_Bool() ) return NULL;
BoolNode *b = i1->as_Bool();
Node *cmp = b->in(1);
if( !cmp->is_Cmp() ) return NULL;
i1 = cmp->in(1);
if( i1 == NULL || !i1->is_Phi() ) return NULL;
PhiNode *phi = i1->as_Phi();
if( phi->is_copy() ) return NULL;
Node *con2 = cmp->in(2);
if( !con2->is_Con() ) return NULL;
// See that the merge point contains some constants
Node *con1=NULL;
uint i4;
for( i4 = 1; i4 < phi->req(); i4++ ) {
con1 = phi->in(i4);
if( !con1 ) return NULL; // Do not optimize partially collapsed merges
if( con1->is_Con() ) break; // Found a constant
// Also allow null-vs-not-null checks
const TypePtr *tp = igvn->type(con1)->isa_ptr();
if( tp && tp->_ptr == TypePtr::NotNull )
break;
}
if( i4 >= phi->req() ) return NULL; // Found no constants
igvn->C->set_has_split_ifs(true); // Has chance for split-if
// Make sure that the compare can be constant folded away
Node *cmp2 = cmp->clone();
cmp2->set_req(1,con1);
cmp2->set_req(2,con2);
const Type *t = cmp2->Value(igvn);
// This compare is dead, so whack it!
igvn->remove_dead_node(cmp2);
if( !t->singleton() ) return NULL;
// No intervening control, like a simple Call
Node *r = iff->in(0);
if( !r->is_Region() ) return NULL;
if( phi->region() != r ) return NULL;
// No other users of the cmp/bool
if (b->outcnt() != 1 || cmp->outcnt() != 1) {
//tty->print_cr("many users of cmp/bool");
return NULL;
}
// Make sure we can determine where all the uses of merged values go
for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
Node* u = r->fast_out(j);
if( u == r ) continue;
if( u == iff ) continue;
if( u->outcnt() == 0 ) continue; // use is dead & ignorable
if( !u->is_Phi() ) {
/*
if( u->is_Start() ) {
tty->print_cr("Region has inlined start use");
} else {
tty->print_cr("Region has odd use");
u->dump(2);
}*/
return NULL;
}
if( u != phi ) {
// CNC - do not allow any other merged value
//tty->print_cr("Merging another value");
//u->dump(2);
return NULL;
}
// Make sure we can account for all Phi uses
for (DUIterator_Fast kmax, k = u->fast_outs(kmax); k < kmax; k++) {
Node* v = u->fast_out(k); // User of the phi
// CNC - Allow only really simple patterns.
// In particular I disallow AddP of the Phi, a fairly common pattern
if( v == cmp ) continue; // The compare is OK
if( (v->is_ConstraintCast()) &&
v->in(0)->in(0) == iff )
continue; // CastPP/II of the IfNode is OK
// Disabled following code because I cannot tell if exactly one
// path dominates without a real dominator check. CNC 9/9/1999
//uint vop = v->Opcode();
//if( vop == Op_Phi ) { // Phi from another merge point might be OK
// Node *r = v->in(0); // Get controlling point
// if( !r ) return NULL; // Degraded to a copy
// // Find exactly one path in (either True or False doms, but not IFF)
// int cnt = 0;
// for( uint i = 1; i < r->req(); i++ )
// if( r->in(i) && r->in(i)->in(0) == iff )
// cnt++;
// if( cnt == 1 ) continue; // Exactly one of True or False guards Phi
//}
if( !v->is_Call() ) {
/*
if( v->Opcode() == Op_AddP ) {
tty->print_cr("Phi has AddP use");
} else if( v->Opcode() == Op_CastPP ) {
tty->print_cr("Phi has CastPP use");
} else if( v->Opcode() == Op_CastII ) {
tty->print_cr("Phi has CastII use");
} else {
tty->print_cr("Phi has use I cant be bothered with");
}
*/
}
return NULL;
/* CNC - Cut out all the fancy acceptance tests
// Can we clone this use when doing the transformation?
// If all uses are from Phis at this merge or constants, then YES.
if( !v->in(0) && v != cmp ) {
tty->print_cr("Phi has free-floating use");
v->dump(2);
return NULL;
}
for( uint l = 1; l < v->req(); l++ ) {
if( (!v->in(l)->is_Phi() || v->in(l)->in(0) != r) &&
!v->in(l)->is_Con() ) {
tty->print_cr("Phi has use");
v->dump(2);
return NULL;
} // End of if Phi-use input is neither Phi nor Constant
} // End of for all inputs to Phi-use
*/
} // End of for all uses of Phi
} // End of for all uses of Region
// Only do this if the IF node is in a sane state
if (iff->outcnt() != 2)
return NULL;
// Got a hit! Do the Mondo Hack!
//
//ABC a1c def ghi B 1 e h A C a c d f g i
// R - Phi - Phi - Phi Rc - Phi - Phi - Phi Rx - Phi - Phi - Phi
// cmp - 2 cmp - 2 cmp - 2
// bool bool_c bool_x
// if if_c if_x
// T F T F T F
// ..s.. ..t .. ..s.. ..t.. ..s.. ..t..
//
// Split the paths coming into the merge point into 2 separate groups of
// merges. On the left will be all the paths feeding constants into the
// Cmp's Phi. On the right will be the remaining paths. The Cmp's Phi
// will fold up into a constant; this will let the Cmp fold up as well as
// all the control flow. Below the original IF we have 2 control
// dependent regions, 's' and 't'. Now we will merge the two paths
// just prior to 's' and 't' from the two IFs. At least 1 path (and quite
// likely 2 or more) will promptly constant fold away.
PhaseGVN *phase = igvn;
// Make a region merging constants and a region merging the rest
uint req_c = 0;
Node* predicate_proj = NULL;
for (uint ii = 1; ii < r->req(); ii++) {
if (phi->in(ii) == con1) {
req_c++;
}
Node* proj = PhaseIdealLoop::find_predicate(r->in(ii));
if (proj != NULL) {
assert(predicate_proj == NULL, "only one predicate entry expected");
predicate_proj = proj;
}
}
Node* predicate_c = NULL;
Node* predicate_x = NULL;
bool counted_loop = r->is_CountedLoop();
Node *region_c = new RegionNode(req_c + 1);
Node *phi_c = con1;
uint len = r->req();
Node *region_x = new RegionNode(len - req_c);
Node *phi_x = PhiNode::make_blank(region_x, phi);
for (uint i = 1, i_c = 1, i_x = 1; i < len; i++) {
if (phi->in(i) == con1) {
region_c->init_req( i_c++, r ->in(i) );
if (r->in(i) == predicate_proj)
predicate_c = predicate_proj;
} else {
region_x->init_req( i_x, r ->in(i) );
phi_x ->init_req( i_x++, phi->in(i) );
if (r->in(i) == predicate_proj)
predicate_x = predicate_proj;
}
}
if (predicate_c != NULL && (req_c > 1)) {
assert(predicate_x == NULL, "only one predicate entry expected");
predicate_c = NULL; // Do not clone predicate below merge point
}
if (predicate_x != NULL && ((len - req_c) > 2)) {
assert(predicate_c == NULL, "only one predicate entry expected");
predicate_x = NULL; // Do not clone predicate below merge point
}
// Register the new RegionNodes but do not transform them. Cannot
// transform until the entire Region/Phi conglomerate has been hacked
// as a single huge transform.
igvn->register_new_node_with_optimizer( region_c );
igvn->register_new_node_with_optimizer( region_x );
// Prevent the untimely death of phi_x. Currently he has no uses. He is
// about to get one. If this only use goes away, then phi_x will look dead.
// However, he will be picking up some more uses down below.
Node *hook = new Node(4);
hook->init_req(0, phi_x);
hook->init_req(1, phi_c);
phi_x = phase->transform( phi_x );
// Make the compare
Node *cmp_c = phase->makecon(t);
Node *cmp_x = cmp->clone();
cmp_x->set_req(1,phi_x);
cmp_x->set_req(2,con2);
cmp_x = phase->transform(cmp_x);
// Make the bool
Node *b_c = phase->transform(new BoolNode(cmp_c,b->_test._test));
Node *b_x = phase->transform(new BoolNode(cmp_x,b->_test._test));
// Make the IfNode
IfNode *iff_c = new IfNode(region_c,b_c,iff->_prob,iff->_fcnt);
igvn->set_type_bottom(iff_c);
igvn->_worklist.push(iff_c);
hook->init_req(2, iff_c);
IfNode *iff_x = new IfNode(region_x,b_x,iff->_prob, iff->_fcnt);
igvn->set_type_bottom(iff_x);
igvn->_worklist.push(iff_x);
hook->init_req(3, iff_x);
// Make the true/false arms
Node *iff_c_t = phase->transform(new IfTrueNode (iff_c));
Node *iff_c_f = phase->transform(new IfFalseNode(iff_c));
if (predicate_c != NULL) {
assert(predicate_x == NULL, "only one predicate entry expected");
// Clone loop predicates to each path
iff_c_t = igvn->clone_loop_predicates(predicate_c, iff_c_t, !counted_loop);
iff_c_f = igvn->clone_loop_predicates(predicate_c, iff_c_f, !counted_loop);
}
Node *iff_x_t = phase->transform(new IfTrueNode (iff_x));
Node *iff_x_f = phase->transform(new IfFalseNode(iff_x));
if (predicate_x != NULL) {
assert(predicate_c == NULL, "only one predicate entry expected");
// Clone loop predicates to each path
iff_x_t = igvn->clone_loop_predicates(predicate_x, iff_x_t, !counted_loop);
iff_x_f = igvn->clone_loop_predicates(predicate_x, iff_x_f, !counted_loop);
}
// Merge the TRUE paths
Node *region_s = new RegionNode(3);
igvn->_worklist.push(region_s);
region_s->init_req(1, iff_c_t);
region_s->init_req(2, iff_x_t);
igvn->register_new_node_with_optimizer( region_s );
// Merge the FALSE paths
Node *region_f = new RegionNode(3);
igvn->_worklist.push(region_f);
region_f->init_req(1, iff_c_f);
region_f->init_req(2, iff_x_f);
igvn->register_new_node_with_optimizer( region_f );
igvn->hash_delete(cmp);// Remove soon-to-be-dead node from hash table.
cmp->set_req(1,NULL); // Whack the inputs to cmp because it will be dead
cmp->set_req(2,NULL);
// Check for all uses of the Phi and give them a new home.
// The 'cmp' got cloned, but CastPP/IIs need to be moved.
Node *phi_s = NULL; // do not construct unless needed
Node *phi_f = NULL; // do not construct unless needed
for (DUIterator_Last i2min, i2 = phi->last_outs(i2min); i2 >= i2min; --i2) {
Node* v = phi->last_out(i2);// User of the phi
igvn->rehash_node_delayed(v); // Have to fixup other Phi users
uint vop = v->Opcode();
Node *proj = NULL;
if( vop == Op_Phi ) { // Remote merge point
Node *r = v->in(0);
for (uint i3 = 1; i3 < r->req(); i3++)
if (r->in(i3) && r->in(i3)->in(0) == iff) {
proj = r->in(i3);
break;
}
} else if( v->is_ConstraintCast() ) {
proj = v->in(0); // Controlling projection
} else {
assert( 0, "do not know how to handle this guy" );
}
Node *proj_path_data, *proj_path_ctrl;
if( proj->Opcode() == Op_IfTrue ) {
if( phi_s == NULL ) {
// Only construct phi_s if needed, otherwise provides
// interfering use.
phi_s = PhiNode::make_blank(region_s,phi);
phi_s->init_req( 1, phi_c );
phi_s->init_req( 2, phi_x );
hook->add_req(phi_s);
phi_s = phase->transform(phi_s);
}
proj_path_data = phi_s;
proj_path_ctrl = region_s;
} else {
if( phi_f == NULL ) {
// Only construct phi_f if needed, otherwise provides
// interfering use.
phi_f = PhiNode::make_blank(region_f,phi);
phi_f->init_req( 1, phi_c );
phi_f->init_req( 2, phi_x );
hook->add_req(phi_f);
phi_f = phase->transform(phi_f);
}
proj_path_data = phi_f;
proj_path_ctrl = region_f;
}
// Fixup 'v' for for the split
if( vop == Op_Phi ) { // Remote merge point
uint i;
for( i = 1; i < v->req(); i++ )
if( v->in(i) == phi )
break;
v->set_req(i, proj_path_data );
} else if( v->is_ConstraintCast() ) {
v->set_req(0, proj_path_ctrl );
v->set_req(1, proj_path_data );
} else
ShouldNotReachHere();
}
// Now replace the original iff's True/False with region_s/region_t.
// This makes the original iff go dead.
for (DUIterator_Last i3min, i3 = iff->last_outs(i3min); i3 >= i3min; --i3) {
Node* p = iff->last_out(i3);
assert( p->Opcode() == Op_IfTrue || p->Opcode() == Op_IfFalse, "" );
Node *u = (p->Opcode() == Op_IfTrue) ? region_s : region_f;
// Replace p with u
igvn->add_users_to_worklist(p);
for (DUIterator_Last lmin, l = p->last_outs(lmin); l >= lmin;) {
Node* x = p->last_out(l);
igvn->hash_delete(x);
uint uses_found = 0;
for( uint j = 0; j < x->req(); j++ ) {
if( x->in(j) == p ) {
x->set_req(j, u);
uses_found++;
}
}
l -= uses_found; // we deleted 1 or more copies of this edge
}
igvn->remove_dead_node(p);
}
// Force the original merge dead
igvn->hash_delete(r);
// First, remove region's dead users.
for (DUIterator_Last lmin, l = r->last_outs(lmin); l >= lmin;) {
Node* u = r->last_out(l);
if( u == r ) {
r->set_req(0, NULL);
} else {
assert(u->outcnt() == 0, "only dead users");
igvn->remove_dead_node(u);
}
l -= 1;
}
igvn->remove_dead_node(r);
// Now remove the bogus extra edges used to keep things alive
igvn->remove_dead_node( hook );
// Must return either the original node (now dead) or a new node
// (Do not return a top here, since that would break the uniqueness of top.)
return new ConINode(TypeInt::ZERO);
}
//------------------------------is_range_check---------------------------------
// Return 0 if not a range check. Return 1 if a range check and set index and
// offset. Return 2 if we had to negate the test. Index is NULL if the check
// is versus a constant.
int IfNode::is_range_check(Node* &range, Node* &index, jint &offset) {
Node* b = in(1);
if (b == NULL || !b->is_Bool()) return 0;
BoolNode* bn = b->as_Bool();
Node* cmp = bn->in(1);
if (cmp == NULL) return 0;
if (cmp->Opcode() != Op_CmpU) return 0;
Node* l = cmp->in(1);
Node* r = cmp->in(2);
int flip_test = 1;
if (bn->_test._test == BoolTest::le) {
l = cmp->in(2);
r = cmp->in(1);
flip_test = 2;
} else if (bn->_test._test != BoolTest::lt) {
return 0;
}
if (l->is_top()) return 0; // Top input means dead test
if (r->Opcode() != Op_LoadRange) return 0;
// We have recognized one of these forms:
// Flip 1: If (Bool[<] CmpU(l, LoadRange)) ...
// Flip 2: If (Bool[<=] CmpU(LoadRange, l)) ...
// Make sure it's a real range check by requiring an uncommon trap
// along the OOB path. Otherwise, it's possible that the user wrote
// something which optimized to look like a range check but behaves
// in some other way.
Node* iftrap = proj_out(flip_test == 2 ? true : false);
bool found_trap = false;
if (iftrap != NULL) {
Node* u = iftrap->unique_ctrl_out();
if (u != NULL) {
// It could be a merge point (Region) for uncommon trap.
if (u->is_Region()) {
Node* c = u->unique_ctrl_out();
if (c != NULL) {
iftrap = u;
u = c;
}
}
if (u->in(0) == iftrap && u->is_CallStaticJava()) {
int req = u->as_CallStaticJava()->uncommon_trap_request();
if (Deoptimization::trap_request_reason(req) ==
Deoptimization::Reason_range_check) {
found_trap = true;
}
}
}
}
if (!found_trap) return 0; // sorry, no cigar
// Look for index+offset form
Node* ind = l;
jint off = 0;
if (l->is_top()) {
return 0;
} else if (l->Opcode() == Op_AddI) {
if ((off = l->in(1)->find_int_con(0)) != 0) {
ind = l->in(2);
} else if ((off = l->in(2)->find_int_con(0)) != 0) {
ind = l->in(1);
}
} else if ((off = l->find_int_con(-1)) >= 0) {
// constant offset with no variable index
ind = NULL;
} else {
// variable index with no constant offset (or dead negative index)
off = 0;
}
// Return all the values:
index = ind;
offset = off;
range = r;
return flip_test;
}
//------------------------------adjust_check-----------------------------------
// Adjust (widen) a prior range check
static void adjust_check(Node* proj, Node* range, Node* index,
int flip, jint off_lo, PhaseIterGVN* igvn) {
PhaseGVN *gvn = igvn;
// Break apart the old check
Node *iff = proj->in(0);
Node *bol = iff->in(1);
if( bol->is_top() ) return; // In case a partially dead range check appears
// bail (or bomb[ASSERT/DEBUG]) if NOT projection-->IfNode-->BoolNode
DEBUG_ONLY( if( !bol->is_Bool() ) { proj->dump(3); fatal("Expect projection-->IfNode-->BoolNode"); } )
if( !bol->is_Bool() ) return;
Node *cmp = bol->in(1);
// Compute a new check
Node *new_add = gvn->intcon(off_lo);
if( index ) {
new_add = off_lo ? gvn->transform(new AddINode( index, new_add )) : index;
}
Node *new_cmp = (flip == 1)
? new CmpUNode( new_add, range )
: new CmpUNode( range, new_add );
new_cmp = gvn->transform(new_cmp);
// See if no need to adjust the existing check
if( new_cmp == cmp ) return;
// Else, adjust existing check
Node *new_bol = gvn->transform( new BoolNode( new_cmp, bol->as_Bool()->_test._test ) );
igvn->rehash_node_delayed( iff );
iff->set_req_X( 1, new_bol, igvn );
}
//------------------------------up_one_dom-------------------------------------
// Walk up the dominator tree one step. Return NULL at root or true
// complex merges. Skips through small diamonds.
Node* IfNode::up_one_dom(Node *curr, bool linear_only) {
Node *dom = curr->in(0);
if( !dom ) // Found a Region degraded to a copy?
return curr->nonnull_req(); // Skip thru it
if( curr != dom ) // Normal walk up one step?
return dom;
// Use linear_only if we are still parsing, since we cannot
// trust the regions to be fully filled in.
if (linear_only)
return NULL;
if( dom->is_Root() )
return NULL;
// Else hit a Region. Check for a loop header
if( dom->is_Loop() )
return dom->in(1); // Skip up thru loops
// Check for small diamonds
Node *din1, *din2, *din3, *din4;
if( dom->req() == 3 && // 2-path merge point
(din1 = dom ->in(1)) && // Left path exists
(din2 = dom ->in(2)) && // Right path exists
(din3 = din1->in(0)) && // Left path up one
(din4 = din2->in(0)) ) { // Right path up one
if( din3->is_Call() && // Handle a slow-path call on either arm
(din3 = din3->in(0)) )
din3 = din3->in(0);
if( din4->is_Call() && // Handle a slow-path call on either arm
(din4 = din4->in(0)) )
din4 = din4->in(0);
if( din3 == din4 && din3->is_If() )
return din3; // Skip around diamonds
}
// Give up the search at true merges
return NULL; // Dead loop? Or hit root?
}
//------------------------------filtered_int_type--------------------------------
// Return a possibly more restrictive type for val based on condition control flow for an if
const TypeInt* IfNode::filtered_int_type(PhaseGVN* gvn, Node *val, Node* if_proj) {
assert(if_proj &&
(if_proj->Opcode() == Op_IfTrue || if_proj->Opcode() == Op_IfFalse), "expecting an if projection");
if (if_proj->in(0) && if_proj->in(0)->is_If()) {
IfNode* iff = if_proj->in(0)->as_If();
if (iff->in(1) && iff->in(1)->is_Bool()) {
BoolNode* bol = iff->in(1)->as_Bool();
if (bol->in(1) && bol->in(1)->is_Cmp()) {
const CmpNode* cmp = bol->in(1)->as_Cmp();
if (cmp->in(1) == val) {
const TypeInt* cmp2_t = gvn->type(cmp->in(2))->isa_int();
if (cmp2_t != NULL) {
jint lo = cmp2_t->_lo;
jint hi = cmp2_t->_hi;
BoolTest::mask msk = if_proj->Opcode() == Op_IfTrue ? bol->_test._test : bol->_test.negate();
switch (msk) {
case BoolTest::ne:
// Can't refine type
return NULL;
case BoolTest::eq:
return cmp2_t;
case BoolTest::lt:
lo = TypeInt::INT->_lo;
if (hi - 1 < hi) {
hi = hi - 1;
}
break;
case BoolTest::le:
lo = TypeInt::INT->_lo;
break;
case BoolTest::gt:
if (lo + 1 > lo) {
lo = lo + 1;
}
hi = TypeInt::INT->_hi;
break;
case BoolTest::ge:
// lo unchanged
hi = TypeInt::INT->_hi;
break;
}
const TypeInt* rtn_t = TypeInt::make(lo, hi, cmp2_t->_widen);
return rtn_t;
}
}
}
}
}
return NULL;
}
//------------------------------fold_compares----------------------------
// See if a pair of CmpIs can be converted into a CmpU. In some cases
// the direction of this if is determined by the preceding if so it
// can be eliminate entirely. Given an if testing (CmpI n c) check
// for an immediately control dependent if that is testing (CmpI n c2)
// and has one projection leading to this if and the other projection
// leading to a region that merges one of this ifs control
// projections.
//
// If
// / |
// / |
// / |
// If |
// /\ |
// / \ |
// / \ |
// / Region
//
Node* IfNode::fold_compares(PhaseGVN* phase) {
if (Opcode() != Op_If) return NULL;
Node* this_cmp = in(1)->in(1);
if (this_cmp != NULL && this_cmp->Opcode() == Op_CmpI &&
this_cmp->in(2)->is_Con() && this_cmp->in(2) != phase->C->top()) {
Node* ctrl = in(0);
BoolNode* this_bool = in(1)->as_Bool();
Node* n = this_cmp->in(1);
int hi = this_cmp->in(2)->get_int();
if (ctrl != NULL && ctrl->is_Proj() && ctrl->outcnt() == 1 &&
ctrl->in(0)->is_If() &&
ctrl->in(0)->outcnt() == 2 &&
ctrl->in(0)->in(1)->is_Bool() &&
ctrl->in(0)->in(1)->in(1)->Opcode() == Op_CmpI &&
ctrl->in(0)->in(1)->in(1)->in(2)->is_Con() &&
ctrl->in(0)->in(1)->in(1)->in(2) != phase->C->top() &&
ctrl->in(0)->in(1)->in(1)->in(1) == n) {
IfNode* dom_iff = ctrl->in(0)->as_If();
Node* otherproj = dom_iff->proj_out(!ctrl->as_Proj()->_con);
if (otherproj->outcnt() == 1 && otherproj->unique_out()->is_Region() &&
this_bool->_test._test != BoolTest::ne && this_bool->_test._test != BoolTest::eq) {
// Identify which proj goes to the region and which continues on
RegionNode* region = otherproj->unique_out()->as_Region();
Node* success = NULL;
Node* fail = NULL;
for (int i = 0; i < 2; i++) {
Node* proj = proj_out(i);
if (success == NULL && proj->outcnt() == 1 && proj->unique_out() == region) {
success = proj;
} else if (fail == NULL) {
fail = proj;
} else {
success = fail = NULL;
}
}
if (success != NULL && fail != NULL && !region->has_phi()) {
int lo = dom_iff->in(1)->in(1)->in(2)->get_int();
BoolNode* dom_bool = dom_iff->in(1)->as_Bool();
Node* dom_cmp = dom_bool->in(1);
const TypeInt* failtype = filtered_int_type(phase, n, ctrl);
if (failtype != NULL) {
const TypeInt* type2 = filtered_int_type(phase, n, fail);
if (type2 != NULL) {
failtype = failtype->join(type2)->is_int();
} else {
failtype = NULL;
}
}
if (failtype != NULL &&
dom_bool->_test._test != BoolTest::ne && dom_bool->_test._test != BoolTest::eq) {
int bound = failtype->_hi - failtype->_lo + 1;
if (failtype->_hi != max_jint && failtype->_lo != min_jint && bound > 1) {
// Merge the two compares into a single unsigned compare by building (CmpU (n - lo) hi)
BoolTest::mask cond = fail->as_Proj()->_con ? BoolTest::lt : BoolTest::ge;
Node* adjusted = phase->transform(new SubINode(n, phase->intcon(failtype->_lo)));
Node* newcmp = phase->transform(new CmpUNode(adjusted, phase->intcon(bound)));
Node* newbool = phase->transform(new BoolNode(newcmp, cond));
phase->is_IterGVN()->replace_input_of(dom_iff, 1, phase->intcon(ctrl->as_Proj()->_con));
phase->hash_delete(this);
set_req(1, newbool);
return this;
}
if (failtype->_lo > failtype->_hi) {
// previous if determines the result of this if so
// replace Bool with constant
phase->hash_delete(this);
set_req(1, phase->intcon(success->as_Proj()->_con));
return this;
}
}
}
}
}
}
return NULL;
}
//------------------------------remove_useless_bool----------------------------
// Check for people making a useless boolean: things like
// if( (x < y ? true : false) ) { ... }
// Replace with if( x < y ) { ... }
static Node *remove_useless_bool(IfNode *iff, PhaseGVN *phase) {
Node *i1 = iff->in(1);
if( !i1->is_Bool() ) return NULL;
BoolNode *bol = i1->as_Bool();
Node *cmp = bol->in(1);
if( cmp->Opcode() != Op_CmpI ) return NULL;
// Must be comparing against a bool
const Type *cmp2_t = phase->type( cmp->in(2) );
if( cmp2_t != TypeInt::ZERO &&
cmp2_t != TypeInt::ONE )
return NULL;
// Find a prior merge point merging the boolean
i1 = cmp->in(1);
if( !i1->is_Phi() ) return NULL;
PhiNode *phi = i1->as_Phi();
if( phase->type( phi ) != TypeInt::BOOL )
return NULL;
// Check for diamond pattern
int true_path = phi->is_diamond_phi();
if( true_path == 0 ) return NULL;
// Make sure that iff and the control of the phi are different. This
// should really only happen for dead control flow since it requires
// an illegal cycle.
if (phi->in(0)->in(1)->in(0) == iff) return NULL;
// phi->region->if_proj->ifnode->bool->cmp
BoolNode *bol2 = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
// Now get the 'sense' of the test correct so we can plug in
// either iff2->in(1) or its complement.
int flip = 0;
if( bol->_test._test == BoolTest::ne ) flip = 1-flip;
else if( bol->_test._test != BoolTest::eq ) return NULL;
if( cmp2_t == TypeInt::ZERO ) flip = 1-flip;
const Type *phi1_t = phase->type( phi->in(1) );
const Type *phi2_t = phase->type( phi->in(2) );
// Check for Phi(0,1) and flip
if( phi1_t == TypeInt::ZERO ) {
if( phi2_t != TypeInt::ONE ) return NULL;
flip = 1-flip;
} else {
// Check for Phi(1,0)
if( phi1_t != TypeInt::ONE ) return NULL;
if( phi2_t != TypeInt::ZERO ) return NULL;
}
if( true_path == 2 ) {
flip = 1-flip;
}
Node* new_bol = (flip ? phase->transform( bol2->negate(phase) ) : bol2);
assert(new_bol != iff->in(1), "must make progress");
iff->set_req(1, new_bol);
// Intervening diamond probably goes dead
phase->C->set_major_progress();
return iff;
}
static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff);
struct RangeCheck {
Node* ctl;
jint off;
};
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node. Strip out
// control copies
Node *IfNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (remove_dead_region(phase, can_reshape)) return this;
// No Def-Use info?
if (!can_reshape) return NULL;
PhaseIterGVN *igvn = phase->is_IterGVN();
// Don't bother trying to transform a dead if
if (in(0)->is_top()) return NULL;
// Don't bother trying to transform an if with a dead test
if (in(1)->is_top()) return NULL;
// Another variation of a dead test
if (in(1)->is_Con()) return NULL;
// Another variation of a dead if
if (outcnt() < 2) return NULL;
// Canonicalize the test.
Node* idt_if = idealize_test(phase, this);
if (idt_if != NULL) return idt_if;
// Try to split the IF
Node *s = split_if(this, igvn);
if (s != NULL) return s;
// Check for people making a useless boolean: things like
// if( (x < y ? true : false) ) { ... }
// Replace with if( x < y ) { ... }
Node *bol2 = remove_useless_bool(this, phase);
if( bol2 ) return bol2;
// Setup to scan up the CFG looking for a dominating test
Node *dom = in(0);
Node *prev_dom = this;
// Check for range-check vs other kinds of tests
Node *index1, *range1;
jint offset1;
int flip1 = is_range_check(range1, index1, offset1);
if( flip1 ) {
// Try to remove extra range checks. All 'up_one_dom' gives up at merges
// so all checks we inspect post-dominate the top-most check we find.
// If we are going to fail the current check and we reach the top check
// then we are guaranteed to fail, so just start interpreting there.
// We 'expand' the top 3 range checks to include all post-dominating
// checks.
// The top 3 range checks seen
const int NRC =3;
RangeCheck prev_checks[NRC];
int nb_checks = 0;
// Low and high offsets seen so far
jint off_lo = offset1;
jint off_hi = offset1;
bool found_immediate_dominator = false;
// Scan for the top checks and collect range of offsets
for (int dist = 0; dist < 999; dist++) { // Range-Check scan limit
if (dom->Opcode() == Op_If && // Not same opcode?
prev_dom->in(0) == dom) { // One path of test does dominate?
if (dom == this) return NULL; // dead loop
// See if this is a range check
Node *index2, *range2;
jint offset2;
int flip2 = dom->as_If()->is_range_check(range2, index2, offset2);
// See if this is a _matching_ range check, checking against
// the same array bounds.
if (flip2 == flip1 && range2 == range1 && index2 == index1 &&
dom->outcnt() == 2) {
if (nb_checks == 0 && dom->in(1) == in(1)) {
// Found an immediately dominating test at the same offset.
// This kind of back-to-back test can be eliminated locally,
// and there is no need to search further for dominating tests.
assert(offset2 == offset1, "Same test but different offsets");
found_immediate_dominator = true;
break;
}
// Gather expanded bounds
off_lo = MIN2(off_lo,offset2);
off_hi = MAX2(off_hi,offset2);
// Record top NRC range checks
prev_checks[nb_checks%NRC].ctl = prev_dom;
prev_checks[nb_checks%NRC].off = offset2;
nb_checks++;
}
}
prev_dom = dom;
dom = up_one_dom(dom);
if (!dom) break;
}
if (!found_immediate_dominator) {
// Attempt to widen the dominating range check to cover some later
// ones. Since range checks "fail" by uncommon-trapping to the
// interpreter, widening a check can make us speculatively enter
// the interpreter. If we see range-check deopt's, do not widen!
if (!phase->C->allow_range_check_smearing()) return NULL;
// Didn't find prior covering check, so cannot remove anything.
if (nb_checks == 0) {
return NULL;
}
// Constant indices only need to check the upper bound.
// Non-constant indices must check both low and high.
int chk0 = (nb_checks - 1) % NRC;
if (index1) {
if (nb_checks == 1) {
return NULL;
} else {
// If the top range check's constant is the min or max of
// all constants we widen the next one to cover the whole
// range of constants.
RangeCheck rc0 = prev_checks[chk0];
int chk1 = (nb_checks - 2) % NRC;
RangeCheck rc1 = prev_checks[chk1];
if (rc0.off == off_lo) {
adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
prev_dom = rc1.ctl;
} else if (rc0.off == off_hi) {
adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
prev_dom = rc1.ctl;
} else {
// If the top test's constant is not the min or max of all
// constants, we need 3 range checks. We must leave the
// top test unchanged because widening it would allow the
// accesses it protects to successfully read/write out of
// bounds.
if (nb_checks == 2) {
return NULL;
}
int chk2 = (nb_checks - 3) % NRC;
RangeCheck rc2 = prev_checks[chk2];
// The top range check a+i covers interval: -a <= i < length-a
// The second range check b+i covers interval: -b <= i < length-b
if (rc1.off <= rc0.off) {
// if b <= a, we change the second range check to:
// -min_of_all_constants <= i < length-min_of_all_constants
// Together top and second range checks now cover:
// -min_of_all_constants <= i < length-a
// which is more restrictive than -b <= i < length-b:
// -b <= -min_of_all_constants <= i < length-a <= length-b
// The third check is then changed to:
// -max_of_all_constants <= i < length-max_of_all_constants
// so 2nd and 3rd checks restrict allowed values of i to:
// -min_of_all_constants <= i < length-max_of_all_constants
adjust_check(rc1.ctl, range1, index1, flip1, off_lo, igvn);
adjust_check(rc2.ctl, range1, index1, flip1, off_hi, igvn);
} else {
// if b > a, we change the second range check to:
// -max_of_all_constants <= i < length-max_of_all_constants
// Together top and second range checks now cover:
// -a <= i < length-max_of_all_constants
// which is more restrictive than -b <= i < length-b:
// -b < -a <= i < length-max_of_all_constants <= length-b
// The third check is then changed to:
// -max_of_all_constants <= i < length-max_of_all_constants
// so 2nd and 3rd checks restrict allowed values of i to:
// -min_of_all_constants <= i < length-max_of_all_constants
adjust_check(rc1.ctl, range1, index1, flip1, off_hi, igvn);
adjust_check(rc2.ctl, range1, index1, flip1, off_lo, igvn);
}
prev_dom = rc2.ctl;
}
}
} else {
RangeCheck rc0 = prev_checks[chk0];
// 'Widen' the offset of the 1st and only covering check
adjust_check(rc0.ctl, range1, index1, flip1, off_hi, igvn);
// Test is now covered by prior checks, dominate it out
prev_dom = rc0.ctl;
}
}
} else { // Scan for an equivalent test
Node *cmp;
int dist = 0; // Cutoff limit for search
int op = Opcode();
if( op == Op_If &&
(cmp=in(1)->in(1))->Opcode() == Op_CmpP ) {
if( cmp->in(2) != NULL && // make sure cmp is not already dead
cmp->in(2)->bottom_type() == TypePtr::NULL_PTR ) {
dist = 64; // Limit for null-pointer scans
} else {
dist = 4; // Do not bother for random pointer tests
}
} else {
dist = 4; // Limit for random junky scans
}
// Normal equivalent-test check.
if( !dom ) return NULL; // Dead loop?
Node* result = fold_compares(phase);
if (result != NULL) {
return result;
}
// Search up the dominator tree for an If with an identical test
while( dom->Opcode() != op || // Not same opcode?
dom->in(1) != in(1) || // Not same input 1?
(req() == 3 && dom->in(2) != in(2)) || // Not same input 2?
prev_dom->in(0) != dom ) { // One path of test does not dominate?
if( dist < 0 ) return NULL;
dist--;
prev_dom = dom;
dom = up_one_dom( dom );
if( !dom ) return NULL;
}
// Check that we did not follow a loop back to ourselves
if( this == dom )
return NULL;
if( dist > 2 ) // Add to count of NULL checks elided
explicit_null_checks_elided++;
} // End of Else scan for an equivalent test
// Hit! Remove this IF
#ifndef PRODUCT
if( TraceIterativeGVN ) {
tty->print(" Removing IfNode: "); this->dump();
}
if( VerifyOpto && !phase->allow_progress() ) {
// Found an equivalent dominating test,
// we can not guarantee reaching a fix-point for these during iterativeGVN
// since intervening nodes may not change.
return NULL;
}
#endif
// Replace dominated IfNode
dominated_by( prev_dom, igvn );
// Must return either the original node (now dead) or a new node
// (Do not return a top here, since that would break the uniqueness of top.)
return new ConINode(TypeInt::ZERO);
}
//------------------------------dominated_by-----------------------------------
void IfNode::dominated_by( Node *prev_dom, PhaseIterGVN *igvn ) {
igvn->hash_delete(this); // Remove self to prevent spurious V-N
Node *idom = in(0);
// Need opcode to decide which way 'this' test goes
int prev_op = prev_dom->Opcode();
Node *top = igvn->C->top(); // Shortcut to top
// Loop predicates may have depending checks which should not
// be skipped. For example, range check predicate has two checks
// for lower and upper bounds.
ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj();
if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate))
prev_dom = idom;
// Now walk the current IfNode's projections.
// Loop ends when 'this' has no more uses.
for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) {
Node *ifp = last_out(i); // Get IfTrue/IfFalse
igvn->add_users_to_worklist(ifp);
// Check which projection it is and set target.
// Data-target is either the dominating projection of the same type
// or TOP if the dominating projection is of opposite type.
// Data-target will be used as the new control edge for the non-CFG
// nodes like Casts and Loads.
Node *data_target = (ifp->Opcode() == prev_op) ? prev_dom : top;
// Control-target is just the If's immediate dominator or TOP.
Node *ctrl_target = (ifp->Opcode() == prev_op) ? idom : top;
// For each child of an IfTrue/IfFalse projection, reroute.
// Loop ends when projection has no more uses.
for (DUIterator_Last jmin, j = ifp->last_outs(jmin); j >= jmin; --j) {
Node* s = ifp->last_out(j); // Get child of IfTrue/IfFalse
if( !s->depends_only_on_test() ) {
// Find the control input matching this def-use edge.
// For Regions it may not be in slot 0.
uint l;
for( l = 0; s->in(l) != ifp; l++ ) { }
igvn->replace_input_of(s, l, ctrl_target);
} else { // Else, for control producers,
igvn->replace_input_of(s, 0, data_target); // Move child to data-target
}
} // End for each child of a projection
igvn->remove_dead_node(ifp);
} // End for each IfTrue/IfFalse child of If
// Kill the IfNode
igvn->remove_dead_node(this);
}
//------------------------------Identity---------------------------------------
// If the test is constant & we match, then we are the input Control
Node *IfProjNode::Identity(PhaseTransform *phase) {
// Can only optimize if cannot go the other way
const TypeTuple *t = phase->type(in(0))->is_tuple();
if (t == TypeTuple::IFNEITHER ||
// kill dead branch first otherwise the IfNode's control will
// have 2 control uses (the IfNode that doesn't go away because
// it still has uses and this branch of the
// If). Node::has_special_unique_user() will cause this node to
// be reprocessed once the dead branch is killed.
(always_taken(t) && in(0)->outcnt() == 1)) {
// IfNode control
return in(0)->in(0);
}
// no progress
return this;
}
//------------------------------dump_spec--------------------------------------
#ifndef PRODUCT
void IfNode::dump_spec(outputStream *st) const {
st->print("P=%f, C=%f",_prob,_fcnt);
}
#endif
//------------------------------idealize_test----------------------------------
// Try to canonicalize tests better. Peek at the Cmp/Bool/If sequence and
// come up with a canonical sequence. Bools getting 'eq', 'gt' and 'ge' forms
// converted to 'ne', 'le' and 'lt' forms. IfTrue/IfFalse get swapped as
// needed.
static IfNode* idealize_test(PhaseGVN* phase, IfNode* iff) {
assert(iff->in(0) != NULL, "If must be live");
if (iff->outcnt() != 2) return NULL; // Malformed projections.
Node* old_if_f = iff->proj_out(false);
Node* old_if_t = iff->proj_out(true);
// CountedLoopEnds want the back-control test to be TRUE, irregardless of
// whether they are testing a 'gt' or 'lt' condition. The 'gt' condition
// happens in count-down loops
if (iff->is_CountedLoopEnd()) return NULL;
if (!iff->in(1)->is_Bool()) return NULL; // Happens for partially optimized IF tests
BoolNode *b = iff->in(1)->as_Bool();
BoolTest bt = b->_test;
// Test already in good order?
if( bt.is_canonical() )
return NULL;
// Flip test to be canonical. Requires flipping the IfFalse/IfTrue and
// cloning the IfNode.
Node* new_b = phase->transform( new BoolNode(b->in(1), bt.negate()) );
if( !new_b->is_Bool() ) return NULL;
b = new_b->as_Bool();
PhaseIterGVN *igvn = phase->is_IterGVN();
assert( igvn, "Test is not canonical in parser?" );
// The IF node never really changes, but it needs to be cloned
iff = new IfNode( iff->in(0), b, 1.0-iff->_prob, iff->_fcnt);
Node *prior = igvn->hash_find_insert(iff);
if( prior ) {
igvn->remove_dead_node(iff);
iff = (IfNode*)prior;
} else {
// Cannot call transform on it just yet
igvn->set_type_bottom(iff);
}
igvn->_worklist.push(iff);
// Now handle projections. Cloning not required.
Node* new_if_f = (Node*)(new IfFalseNode( iff ));
Node* new_if_t = (Node*)(new IfTrueNode ( iff ));
igvn->register_new_node_with_optimizer(new_if_f);
igvn->register_new_node_with_optimizer(new_if_t);
// Flip test, so flip trailing control
igvn->replace_node(old_if_f, new_if_t);
igvn->replace_node(old_if_t, new_if_f);
// Progress
return iff;
}