escape.cpp revision 3226
/*
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* 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 "ci/bcEscapeAnalyzer.hpp"
#include "libadt/vectset.hpp"
#include "memory/allocation.hpp"
#include "opto/c2compiler.hpp"
#include "opto/callnode.hpp"
#include "opto/cfgnode.hpp"
#include "opto/compile.hpp"
#include "opto/escape.hpp"
#include "opto/phaseX.hpp"
#include "opto/rootnode.hpp"
}
_edges->append_if_missing(v);
}
}
#ifndef PRODUCT
static const char *node_type_names[] = {
"UnknownType",
"JavaObject",
"LocalVar",
"Field"
};
static const char *esc_names[] = {
"UnknownEscape",
"NoEscape",
"ArgEscape",
"GlobalEscape"
};
static const char *edge_type_suffix[] = {
"?", // UnknownEdge
"P", // PointsToEdge
"D", // DeferredEdge
"F" // FieldEdge
};
if (print_state) {
}
for (uint i = 0; i < edge_count(); i++) {
}
else
}
#endif
_processed(C->comp_arena()),
pt_ptset(C->comp_arena()),
pt_visited(C->comp_arena()),
_collecting(true),
_progress(false),
_compile(C),
_node_map(C->comp_arena()) {
// Add ConP(#NULL) and ConN(#NULL) nodes.
if (UseCompressedOops) {
} else {
}
}
assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of PointsTo edge");
if (f->has_unknown_ptr()) {
return;
} else {
f->set_has_unknown_ptr();
}
}
}
assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
assert(f->node_type() == PointsToNode::LocalVar || f->node_type() == PointsToNode::Field, "invalid source of Deferred edge");
assert(t->node_type() == PointsToNode::LocalVar || t->node_type() == PointsToNode::Field, "invalid destination of Deferred edge");
// don't add a self-referential edge, this can occur during removal of
// deferred edges
}
// We are computing a raw address for a store captured by an Initialize
// compute an appropriate address type. AddP cases #3 and #5 (see below).
"offset must be a constant or it is initialization of array");
return offs;
}
}
// Don't add fields to NULL pointer.
if (is_null_ptr(from_i))
return;
assert(f->node_type() != PointsToNode::UnknownType && t->node_type() != PointsToNode::UnknownType, "node types must be set");
t->set_offset(offset);
}
// Don't change non-escaping state of NULL pointer.
if (is_null_ptr(ni))
return;
}
// inline set_escape_state(idx, es);
if (done)
}
// If we are still collecting or there were no non-escaping allocations
// we don't know the answer yet
if (_collecting)
return PointsToNode::UnknownEscape;
// if the node was created after the escape computation, return
// UnknownEscape
if (idx >= nodes_size())
return PointsToNode::UnknownEscape;
// if we have already computed a value, return it
return es;
// PointsTo() calls n->uncast() which can return a new ideal node.
return PointsToNode::UnknownEscape;
// compute max escape state of anything this node could point to
}
// cache the computed escape state
} // orig_es could be PointsToNode::UnknownEscape
return es;
}
pt_visited.Reset();
pt_worklist.clear();
#ifdef ASSERT
#endif
n = n->uncast();
// If we have a JavaObject, return just that object
return &pt_ptset;
}
#ifdef ASSERT
if (orig_n != n)
n->dump();
}
#endif
while(pt_worklist.length() > 0) {
continue;
// ensure that all inputs of a Phi have been processed
int edges_processed = 0;
} else {
assert(false,"neither PointsToEdge or DeferredEdge");
}
}
if (edges_processed == 0) {
// no deferred or pointsto edges found. Assume the value was set
// outside this method. Add the phantom object to the pointsto set.
}
}
return &pt_ptset;
}
void ConnectionGraph::remove_deferred(uint ni, GrowableArray<uint>* deferred_edges, VectorSet* visited) {
// This method is most expensive during ConnectionGraph construction.
// Reuse vectorSet and an additional growable array for deferred edges.
deferred_edges->clear();
// Mark current edges as visited and move deferred edges to separate array.
#ifdef ASSERT
#else
#endif
deferred_edges->append(t);
} else {
i++;
}
}
continue;
} else {
assert(false,"invalid connection graph");
}
}
}
if (ptn->edge_count() == 0) {
// No pointsto edges found after deferred edges are removed.
// For example, in the next case where call is replaced
// with uncommon trap and as result array's load references
// itself through deferred edges:
//
// A a = b[i];
// if (c!=null) a = c.foo();
// b[i] = a;
//
// Assume the value was set outside this method and
// add edge to phantom object.
}
}
// Add an edge to node given by "to_i" from any field of adr_i whose offset
// matches "offset" A deferred edge is added if to_i is a LocalVar, and
// a pointsto edge is added if it is a JavaObject
// No fields for NULL pointer.
if (is_null_ptr(adr_i)) {
return;
}
if (escaped) {
// Values in fields escaped during call.
}
if (escaped) {
}
if (deferred)
else
}
}
}
// Add a deferred edge from node given by "from_i" to any field of adr_i
// whose offset matches "offset".
// No fields for NULL pointer.
if (is_null_ptr(adr_i)) {
return;
}
if (adr_i == _phantom_object) {
// Add only one edge for unknown object.
return;
}
if (!is_alloc) {
// Assume the field was set outside this method if it is not Allocation
}
}
}
// Some fields references (AddP) may still be missing
// until Connection Graph construction is complete.
// For example, loads from RAW pointers with offset 0
// which don't have AddP.
// A reference to phantom_object will be added if
// a field reference is still missing after completing
// Connection Graph (see remove_deferred()).
}
// Helper functions
//
// AddP cases for Base and Address inputs:
// case #1. Direct object's field reference:
// Allocate
// |
// Proj #5 ( oop result )
// |
// CheckCastPP (cast to instance type)
// | |
// AddP ( base == address )
//
// case #2. Indirect object's field reference:
// Phi
// |
// CastPP (cast to instance type)
// | |
// AddP ( base == address )
//
// case #3. Raw object's field reference for Initialize node:
// Allocate
// |
// Proj #5 ( oop result )
// top |
// \ |
// AddP ( base == top )
//
// case #4. Array's element reference:
// {CheckCastPP | CastPP}
// | | |
// | AddP ( array's element offset )
// | |
// AddP ( array's offset )
//
// case #5. Raw object's field reference for arraycopy stub call:
// The inline_native_clone() case when the arraycopy stub is called
// after the allocation before Initialize and CheckCastPP nodes.
// Allocate
// |
// Proj #5 ( oop result )
// | |
// AddP ( base == address )
//
// case #6. Constant Pool, ThreadLocal, CastX2P or
// Raw object's field reference:
// {ConP, ThreadLocal, CastX2P, raw Load}
// top |
// \ |
// AddP ( base == top )
//
// case #7. Klass's field reference.
// LoadKlass
// | |
// AddP ( base == address )
//
// case #8. narrow Klass's field reference.
// LoadNKlass
// |
// DecodeN
// | |
// AddP ( base == address )
//
// Case #6 (unsafe access) may have several chained AddP nodes.
}
}
return base;
}
//
// Find array's offset to push it on worklist first and
// as result process an array's element offset first (pushed second)
// to avoid CastPP for the array's offset.
// Otherwise the inserted CastPP (LocalVar) will point to what
// the AddP (Field) points to. Which would be wrong since
// the algorithm expects the CastPP has the same point as
// as AddP's base CheckCastPP (LocalVar).
//
// ArrayAllocation
// |
// CheckCastPP
// |
// memProj (from ArrayAllocation CheckCastPP)
// | ||
// | || Int (element index)
// | || | ConI (log(element size))
// | || | /
// | || LShift
// | || /
// | AddP (array's element offset)
// | |
// | | ConI (array's offset: #12(32-bits) or #24(64-bits))
// | / /
// AddP (array's offset)
// |
//
return addp2;
}
return NULL;
}
//
// Adjust the type and inputs of an AddP which computes the
// address of a field of an instance
//
if (t == NULL) {
// We are computing a raw address for a store captured by an Initialize
// compute an appropriate address type (cases #3 and #5).
assert(addp->in(AddPNode::Address)->is_Proj(), "base of raw address must be result projection from allocation");
}
"old type must be non-instance or match new type");
// The type 't' could be subclass of 'base_t'.
// As result t->offset() could be large then base_t's size and it will
// cause the failure in add_offset() with narrow oops since TypeOopPtr()
// constructor verifies correctness of the offset.
//
// It could happened on subclass's branch (from the type profiling
// inlining) which was not eliminated during parsing since the exactness
// of the allocation type was not propagated to the subclass type check.
//
// Or the type 't' could be not related to 'base_t' at all.
// It could happened when CHA type is different from MDO type on a dead path
// (for example, from instanceof check) which is not collapsed during parsing.
//
// Do nothing for such AddP node and don't process its users since
// this code branch will go away.
//
if (!t->is_known_instance() &&
return false; // bail out
}
// Do NOT remove the next line: ensure a new alias index is allocated
// for the instance type. Note: C++ will not remove it since the call
// has side effect.
// record the allocation in the node map
// Set addp's Base and Address to 'base'.
// Skip AddP cases #3 and #5.
} else {
} else {
// AddP case #4 (adr is array's element offset AddP node)
#ifdef ASSERT
#endif
}
}
}
// Put on IGVN worklist since at least addp's type was changed above.
return true;
}
//
// Create a new version of orig_phi if necessary. Returns either the newly
// created phi or an existing phi. Sets create_new to indicate whether a new
// phi was created. Cache the last newly created phi in the node map.
//
PhiNode *ConnectionGraph::create_split_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn, bool &new_created) {
new_created = false;
// nothing to do if orig_phi is bottom memory or matches alias_idx
if (phi_alias_idx == alias_idx) {
return orig_phi;
}
// Have we recently created a Phi for this alias index?
return result;
}
// Previous check may fail when the same wide memory Phi was split into Phis
// for different memory slices. Search all Phis for this region.
}
}
}
if (C->do_escape_analysis() == true && !C->failing()) {
// Retry compilation without escape analysis.
// If this is the first failure, the sentinel string will "stick"
// to the Compile object, and the C2Compiler will see it and retry.
}
return NULL;
}
new_created = true;
return result;
}
//
// Return a new version of Memory Phi "orig_phi" with the inputs having the
// specified alias index.
//
PhiNode *ConnectionGraph::split_memory_phi(PhiNode *orig_phi, int alias_idx, GrowableArray<PhiNode *> &orig_phi_worklist, PhaseGVN *igvn) {
bool new_phi_created;
if (!new_phi_created) {
return result;
}
bool finished = false;
while(!finished) {
PhiNode *newphi = create_split_phi(mem->as_Phi(), alias_idx, orig_phi_worklist, igvn, new_phi_created);
if (new_phi_created) {
// found an phi for which we created a new split, push current one on worklist and begin
// processing new one
idx = 1;
continue;
} else {
}
}
if (C->failing()) {
return NULL;
}
}
#ifdef ASSERT
// verify that the new Phi has an input for each input of the original
#endif
// Check if all new phi's inputs have specified alias index.
// Otherwise use old phi.
}
// we have finished processing a Phi, see if there are any more to do
if (!finished) {
}
}
return result;
}
//
// The next methods are derived from methods in MemNode.
//
// TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
// means an array I have not precisely typed yet. Do not do any
// alias stuff with it any time soon.
// Update input if it is progress over what we have now
}
return mem;
}
//
// Move memory users to their memory slices.
//
// Move users first
if (use->is_MergeMem()) {
continue; // Nothing to do
}
// Replace previous general reference to mem node.
--imax;
--i;
// Don't move related membars.
continue;
}
alias_idx == general_idx) {
continue; // Nothing to do
}
// Move to general memory slice.
--i;
#ifdef ASSERT
// Don't move related cardmark.
continue;
}
// Memory nodes should have new memory input.
"Following memory nodes should have new memory input or be on the same memory slice");
// Phi nodes should be split and moved already.
} else {
assert(false, "should not be here");
#endif
}
}
}
//
// Search memory chain of "mem" to find a MemNode whose address
// is the specified alias index.
//
Node* ConnectionGraph::find_inst_mem(Node *orig_mem, int alias_idx, GrowableArray<PhiNode *> &orig_phis, PhaseGVN *phase) {
return orig_mem;
break; // hit one of our sentinels
break; // Dead
break; // Found
break; // Do not skip store to general memory slice.
}
}
if (!is_instance)
continue; // don't search further for non-instance types
// skip over a call which does not affect this memory slice
break; // hit one of our sentinels
}
} else if (proj_in->is_Initialize()) {
// Stop if this is the initialization for the object instance which
// which contains this memory slice, otherwise skip over it.
}
}
} else if (result->is_MergeMem()) {
// Didn't find instance memory, search through general slice recursively.
if (C->failing()) {
return NULL;
}
}
} else {
break;
}
} else if (result->is_ClearArray()) {
// Can not bypass initialization of the instance
// we are looking for.
break;
}
// Otherwise skip it (the call updated 'result' value).
break;
}
}
}
if (!is_instance) {
// Push all non-instance Phis on the orig_phis worklist to update inputs
// during Phase 4 if needed.
} else if (C->get_alias_index(t) != alias_idx) {
// Create a new Phi with the specified alias index type.
}
}
// the result is either MemNode, PhiNode, InitializeNode.
return result;
}
//
// Convert the types of unescaped object to instance types where possible,
// propagate the new type information through the graph, and update memory
// edges and MergeMem inputs to reflect the new type.
//
// We start with allocations (and calls which may be allocations) on alloc_worklist.
// The processing is done in 4 phases:
//
// Phase 1: Process possible allocations from alloc_worklist. Create instance
// types for the CheckCastPP for allocations where possible.
// Propagate the the new types through users as follows:
// casts and Phi: push users on alloc_worklist
// AddP: cast Base and Address inputs to the instance type
// push any AddP users on alloc_worklist and push any memnode
// users onto memnode_worklist.
// Phase 2: Process MemNode's from memnode_worklist. compute new address type and
// search the Memory chain for a store with the appropriate type
// address type. If a Phi is found, create a new version with
// the appropriate memory slices from each of the Phi inputs.
// For stores, process the users as follows:
// MemNode: push on memnode_worklist
// MergeMem: push on mergemem_worklist
// Phase 3: Process MergeMem nodes from mergemem_worklist. Walk each memory slice
// moving the first node encountered of each instance type to the
// the input corresponding to its alias index.
// appropriate memory slice.
// Phase 4: Update the inputs of non-instance memory Phis and the Memory input of memnodes.
//
// In the following example, the CheckCastPP nodes are the cast of allocation
// results and the allocation of node 29 is unescaped and eligible to be an
// instance type.
//
// We start with:
//
// 7 Parm #memory
// 10 ConI "12"
// 19 CheckCastPP "Foo"
// 20 AddP _ 19 19 10 Foo+12 alias_index=4
// 29 CheckCastPP "Foo"
// 30 AddP _ 29 29 10 Foo+12 alias_index=4
//
// 40 StoreP 25 7 20 ... alias_index=4
// 50 StoreP 35 40 30 ... alias_index=4
// 60 StoreP 45 50 20 ... alias_index=4
// 70 LoadP _ 60 30 ... alias_index=4
// 80 Phi 75 50 60 Memory alias_index=4
// 90 LoadP _ 80 30 ... alias_index=4
// 100 LoadP _ 80 20 ... alias_index=4
//
//
// Phase 1 creates an instance type for node 29 assigning it an instance id of 24
// and creating a new alias index for node 30. This gives:
//
// 7 Parm #memory
// 10 ConI "12"
// 19 CheckCastPP "Foo"
// 20 AddP _ 19 19 10 Foo+12 alias_index=4
// 29 CheckCastPP "Foo" iid=24
// 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
//
// 40 StoreP 25 7 20 ... alias_index=4
// 50 StoreP 35 40 30 ... alias_index=6
// 60 StoreP 45 50 20 ... alias_index=4
// 70 LoadP _ 60 30 ... alias_index=6
// 80 Phi 75 50 60 Memory alias_index=4
// 90 LoadP _ 80 30 ... alias_index=6
// 100 LoadP _ 80 20 ... alias_index=4
//
// In phase 2, new memory inputs are computed for the loads and stores,
// And a new version of the phi is created. In phase 4, the inputs to
// node 80 are updated and then the memory nodes are updated with the
// values computed in phase 2. This results in:
//
// 7 Parm #memory
// 10 ConI "12"
// 19 CheckCastPP "Foo"
// 20 AddP _ 19 19 10 Foo+12 alias_index=4
// 29 CheckCastPP "Foo" iid=24
// 30 AddP _ 29 29 10 Foo+12 alias_index=6 iid=24
//
// 40 StoreP 25 7 20 ... alias_index=4
// 50 StoreP 35 7 30 ... alias_index=6
// 60 StoreP 45 40 20 ... alias_index=4
// 70 LoadP _ 50 30 ... alias_index=6
// 80 Phi 75 40 60 Memory alias_index=4
// 120 Phi 75 50 50 Memory alias_index=6
// 90 LoadP _ 120 30 ... alias_index=6
// 100 LoadP _ 80 20 ... alias_index=4
//
// Phase 1: Process possible allocations from alloc_worklist.
// Create instance types for the CheckCastPP for allocations where possible.
//
// (Note: don't forget to change the order of the second AddP node on
// the alloc_worklist if the order of the worklist processing is changed,
// see the comment in find_second_addp().)
//
while (alloc_worklist.length() != 0) {
if (n->is_Call()) {
// copy escape information to call node
// We have an allocation or call which returns a Java object,
// see if it is unescaped.
continue;
// Find CheckCastPP for the allocate or for the return value of a call
n = alloc->result_cast();
if (n == NULL) { // No uses except Initialize node
if (alloc->is_Allocate()) {
// Set the scalar_replaceable flag for allocation
// so it could be eliminated if it has no uses.
}
continue;
}
if (!n->is_CheckCastPP()) { // not unique CheckCastPP.
continue;
}
// The inline code for Object.clone() casts the allocation result to
// java.lang.Object and then to the actual type of the allocated
// object. Detect this case and use the second cast.
// Also detect j.l.reflect.Array.newInstance(jobject, jint) case when
// the allocation result is cast to java.lang.Object and then
// to the actual Array type.
&& (alloc->is_AllocateArray() ||
if (use->is_CheckCastPP()) {
break;
}
}
n = cast2;
} else {
// Non-scalar replaceable if the allocation type is unknown statically
// (reflection allocation), the object can't be restored during
// deoptimization without precise type.
continue;
}
}
if (alloc->is_Allocate()) {
// Set the scalar_replaceable flag for allocation
// so it could be eliminated.
}
// in order for an object to be scalar-replaceable, it must be:
// - a direct allocation (not a call returning an object)
// - non-escaping
// - eligible to be a unique type
// - not determined to be ineligible by escape analysis
if (t == NULL)
continue; // not a TypeOopPtr
igvn->hash_delete(n);
n->raise_bottom_type(tinst);
igvn->hash_insert(n);
// First, put on the worklist all Field edges from Connection Graph
// which is more accurate then putting immediate users from Ideal Graph.
"only AddP nodes are Field edges in CG");
}
}
}
// An allocation may have an Initialize which has raw stores. Scan
// the users of the raw allocation result and push AddP users
// on alloc_worklist.
}
}
}
}
} else if (n->is_AddP()) {
if (elem == _phantom_object) {
assert(false, "escaped allocation");
continue; // Assume the value was set outside this method.
}
} else if (n->is_Phi() ||
n->is_CheckCastPP() ||
n->is_EncodeP() ||
n->is_DecodeN() ||
continue; // already processed
}
if (elem == _phantom_object) {
assert(false, "escaped allocation");
continue; // Assume the value was set outside this method.
}
const TypeOopPtr *tn_t;
if (tn_type->isa_narrowoop()) {
} else {
}
if (tn_type->isa_narrowoop()) {
} else {
}
} else {
"unexpected type");
continue; // Skip dead path with different type
}
}
} else {
debug_only(n->dump();)
assert(false, "EA: unexpected node");
continue;
}
// push allocation's users on appropriate worklist
}
use->is_CheckCastPP() ||
use->is_EncodeP() ||
use->is_DecodeN() ||
#ifdef ASSERT
} else if (use->is_MergeMem()) {
assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
} else if (use->is_SafePoint()) {
// Look for MergeMem nodes for calls which reference unique allocation
// (through CheckCastPP nodes) even for debug info.
if (m->is_MergeMem()) {
}
} else {
n->dump();
assert(false, "EA: missing allocation reference path");
}
#endif
}
}
}
// New alias types were created in split_AddP().
// Phase 2: Process MemNode's from memnode_worklist. compute new address type and
// compute new values for Memory inputs (the Memory inputs are not
// actually updated until phase 4.)
if (memnode_worklist.length() == 0)
return; // nothing to do
while (memnode_worklist.length() != 0) {
continue;
if (n->is_Phi() || n->is_ClearArray()) {
// we don't need to do anything, but the users must be pushed
} else if (n->is_MemBar()) { // Initialize, MemBar nodes
// we don't need to do anything, but the users must be pushed
if (n == NULL)
continue;
} else {
continue;
return;
}
// We delay the memory edge update since we need old one in
// MergeMem code below when instances memory slices are separated.
}
if (n->is_Load()) {
continue; // don't push users
} else if (n->is_LoadStore()) {
// get the memory projection
n = use;
break;
}
}
}
}
// push user on appropriate worklist
continue;
#ifdef ASSERT
} else if (use->is_MergeMem()) {
assert(_mergemem_worklist.contains(use->as_MergeMem()), "EA: missing MergeMem node in the worklist");
} else {
if (!(op == Op_StoreCM ||
n->dump();
assert(false, "EA: missing memory path");
}
#endif
}
}
}
// Phase 3: Process MergeMem nodes from mergemem_worklist.
// Walk each memory slice moving the first node encountered of each
// instance type to the the input corresponding to its alias index.
// Note: we don't want to use MergeMemStream here because we only want to
// scan inputs which exist at the start, not ones we add during processing.
// Note 2: MergeMem may already contains instance memory slices added
// during find_inst_mem() call when memory nodes were processed above.
continue;
// First, update mergemem by moving memory nodes to corresponding slices
// if their type became more precise since this mergemem was created.
if (idx == i) {
} else {
}
}
}
}
// Find any instance of the current type if we haven't encountered
// already a memory slice of the instance along the memory chain.
if (nmm->is_empty_memory(m)) {
return;
}
}
}
}
}
// Find the rest of instances values
// Didn't find instance memory, search through general slice recursively.
return;
}
}
}
}
// Phase 4: Update the inputs of non-instance memory Phis and
// the Memory input of memnodes
// First update the inputs of any non-instance Phi's from
// which we split out an instance Phi. Note we don't have
// to recursively process Phi's encounted on the input memory
// chains as is done in split_memory_phi() since they will
// also be processed here.
return;
}
}
}
}
// Update the memory inputs of MemNodes with the value we computed
// in Phase 2 and move stores memory users to corresponding memory slices.
// Disable memory split verification code until the fix for 6984348.
// Currently it produces false negative results since it does not cover all cases.
#if 0 // ifdef ASSERT
#endif
for (uint i = 0; i < nodes_size(); i++) {
if (n->is_Mem()) {
#if 0 // ifdef ASSERT
}
#endif
if (!n->is_Load()) {
// Move memory users of a store first.
}
// Now update memory input
igvn->hash_delete(n);
igvn->hash_insert(n);
} else {
}
}
}
#if 0 // ifdef ASSERT
// Verify that memory was split correctly
while (old_mems.is_nonempty()) {
}
#endif
}
// are represented by ideal Macro nodes.
int cnt = C->macro_count();
for( int i=0; i < cnt; i++ ) {
Node *n = C->macro_node(i);
if ( n->is_Allocate() )
return true;
if( n->is_Lock() ) {
return true;
}
}
return false;
}
// Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction
// to create space for them in ConnectionGraph::_nodes[].
// Perform escape analysis
if (congraph->compute_escape()) {
// There are non escaping objects.
C->set_congraph(congraph);
}
// Cleanup.
}
bool ConnectionGraph::compute_escape() {
// 1. Populate Connection Graph (CG) with Ideal nodes.
// Initialize worklist
}
// Push all useful nodes onto CG list and set their type.
// Only allocations and java static calls results are checked
// for an escape status. See process_call_result() below.
if (n->is_Allocate() || n->is_CallStaticJava() &&
alloc_worklist.append(n);
} else if(n->is_AddP()) {
// Collect address nodes. Use them during stage 3 below
// to build initial connection graph field edges.
addp_worklist.append(n);
} else if (n->is_MergeMem()) {
// Collect all MergeMem nodes to add memory slices for
// scalar replaceable objects in split_unique_types().
} else if (OptimizePtrCompare && n->is_Cmp() &&
// Compare pointers nodes
} else if (n->is_MemBarStoreStore()) {
// Collect all MemBarStoreStore nodes so that depending on the
// escape status of the associated Allocate node some of them
// may be eliminated.
}
worklist_init.push(m);
}
}
if (alloc_worklist.length() == 0) {
_collecting = false;
return false; // Nothing to do.
}
// 2. First pass to create simple CG edges (doesn't require to walk CG).
}
// 3. Pass to create initial fields edges (JavaObject -F-> AddP)
// to reduce number of iterations during stage 4 below.
}
}
// 4. Build Connection Graph which need
// to walk the connection graph.
_progress = false;
if (n != NULL) { // Call, AddP, LoadP, StoreP
}
}
// After IGVN user nodes may have smaller _idx than
// their inputs so they will be processed first in
// previous loop. Because of that not all Graph
// edges will be created. Walk over interesting
// nodes again until no new edges are created.
//
// Normally only 1-3 passes needed to build
// Connection Graph depending on graph complexity.
// Observed 8 passes in jvm2008 compiler.compiler.
// Set limit to 20 to catch situation when something
// did go wrong and recompile the method without EA.
// Also limit build time to 30 sec (60 in debug VM).
#define CG_BUILD_ITER_LIMIT 20
#ifdef ASSERT
#define CG_BUILD_TIME_LIMIT 60.0
#else
#define CG_BUILD_TIME_LIMIT 30.0
#endif
int iterations = 0;
while(_progress &&
(iterations++ < CG_BUILD_ITER_LIMIT) &&
_progress = false;
}
}
if ((iterations >= CG_BUILD_ITER_LIMIT) ||
assert(false, err_msg("infinite EA connection graph build (%f sec, %d iterations) with %d nodes and worklist size %d",
// Possible infinite build_connection_graph loop,
// bailout (no changes to ideal graph were made).
_collecting = false;
return false;
}
// 5. Propagate escaped states.
// mark all nodes reachable from GlobalEscape nodes
// mark all nodes reachable from ArgEscape nodes
bool has_non_escaping_obj = propagate_escape_state(&cg_worklist, &worklist, PointsToNode::ArgEscape);
// 6. Find fields initializing values for not escaped allocations
has_non_escaping_obj = true;
if (n->is_Allocate()) {
// The object allocated by this Allocate node will never be
// seen by an other thread. Mark it so that when it is
// expanded no MemBarStoreStore is added.
}
// Same as above. Mark this Allocate node so that when it is
// expanded no MemBarStoreStore is added.
}
}
// Skip the rest of code if all objects escaped.
if (!has_non_escaping_obj) {
cg_length = 0;
addp_length = 0;
}
if (ptn->edge_count() == 0) {
// No values were found. Assume the value was set
// outside this method - add edge to phantom object.
}
}
}
// 7. Remove deferred edges from the graph.
}
}
// 8. Adjust escape state of nonescaping objects.
}
// push all NoEscape nodes on the worklist
!is_null_ptr(nk))
}
// Propagate scalar_replaceable value.
if (n->is_Allocate() && scalar_replaceable) {
// Push scalar replaceable allocations on alloc_worklist
// for processing in split_unique_types(). Note,
// following code may change scalar_replaceable value.
alloc_worklist.append(n);
}
if (is_null_ptr(npi))
continue;
if (!scalar_replaceable) {
np->set_scalar_replaceable(false);
}
np->set_scalar_replaceable(false);
}
}
}
_collecting = false;
assert(C->unique() == nodes_size(), "there should be no new ideal nodes during ConnectionGraph build");
if (UseCompressedOops) {
}
if (EliminateLocks && has_non_escaping_obj) {
// Mark locks before changing ideal graph.
int cnt = C->macro_count();
for( int i=0; i < cnt; i++ ) {
Node *n = C->macro_node(i);
if (n->is_AbstractLock()) { // Lock and Unlock nodes
if (!alock->is_non_esc_obj()) {
// The lock could be marked eliminated by lock coarsening
// code during first IGVN before EA. Replace coarsened flag
alock->set_non_esc_obj();
}
}
}
}
}
if (OptimizePtrCompare && has_non_escaping_obj) {
// Add ConI(#CC_GT) and ConI(#CC_EQ).
// Optimize objects compare.
while (ptr_cmp_worklist.length() != 0) {
#ifndef PRODUCT
if (PrintOptimizePtrCompare) {
tty->print_cr("++++ Replaced: %d %s(%d,%d) --> %s", n->_idx, (n->Opcode() == Op_CmpP ? "CmpP" : "CmpN"), n->in(1)->_idx, n->in(2)->_idx, (res == _pcmp_eq ? "EQ" : "NotEQ"));
if (Verbose) {
n->dump(1);
}
}
#endif
}
}
// cleanup
}
// For MemBarStoreStore nodes added in library_call.cpp, check
// escape status of associated AllocateNode and optimize out
// MemBarStoreStore node if the allocated object never escapes.
while (storestore_worklist.length() != 0) {
}
}
#ifndef PRODUCT
if (PrintEscapeAnalysis) {
dump(); // Dump ConnectionGraph
}
#endif
bool has_scalar_replaceable_candidates = false;
if (ptn->scalar_replaceable()) {
break;
}
}
// Now use the escape information to create unique types for
// scalar replaceable objects.
if (C->failing()) return false;
#ifdef ASSERT
C->method()->print_short_name();
if(!EliminateAllocations) {
} else if(!has_scalar_replaceable_candidates) {
} else if(C->AliasLevel() < 3) {
}
#endif
}
return has_non_escaping_obj;
}
// Find fields initializing values for allocations.
// Check if a oop field's initializing value is recorded and add
// a corresponding NULL field's value if it is not recorded.
// Connection Graph does not record a default initialization by NULL
// captured by Initialize node.
//
bool visited_bottom_offset = false;
if (!visited_bottom_offset) {
visited_bottom_offset = true;
// Check only oop fields.
if (!adr_type->isa_aryptr() ||
// OffsetBot is used to reference array's element,
// always add reference to NULL since we don't
// known which element is referenced.
}
}
// Check only oop fields.
if (adr_type->isa_instptr()) {
} else {
// Ignore non field load (for example, klass load)
}
} else if (adr_type->isa_aryptr()) {
} else {
// Ignore array length load
}
#ifdef ASSERT
} else {
// Raw pointers are used for initializing stores so skip it
// since it should be recorded already
#endif
}
if (basic_field_type == T_OBJECT ||
basic_field_type == T_NARROWOOP ||
basic_field_type == T_ARRAY) {
// Check for a store which follows allocation without branches.
// For example, a volatile field store is not collected
// by Initialize node. TODO: it would be nice to use idom() here.
//
// Search all references to the same field which use different
// AddP nodes, for example, in the next case:
//
// Point p[] = new Point[1];
// if ( x ) { p[0] = new Point(); p[0].x = x; }
// if ( p[0] != null ) { y = p[0].x; } // has CastPP
//
}
break;
}
}
}
}
}
}
}
// A field's initializing value was not recorded. Add NULL.
}
}
}
}
}
// Adjust escape state after Connection Graph is built.
// Search for objects which are not scalar replaceable
// and mark them to propagate the state to referenced objects.
//
// An object is not scalar replaceable if the field which may point
// to it has unknown offset (unknown element of an array of objects).
//
}
}
// Currently an object is not scalar replaceable if a LoadStore node
// access its field since the field value is unknown after it.
//
bool has_LoadStore = false;
if (use->is_LoadStore()) {
has_LoadStore = true;
break;
}
}
// An object is not scalar replaceable if the address points
// to unknown field (unknown element for arrays, offset is OffsetBot).
//
// Or the address may point to more then one object. This may produce
// the false positive result (set not scalar replaceable)
// since the flow-insensitive escape analysis can't separate
// the case when stores overwrite the field's value from the case
// when stores happened on different control branches.
//
// Note: it will disable scalar replacement in some cases:
//
// Point p[] = new Point[1];
// p[0] = new Point(); // Will be not scalar replaced
//
// but it will save us from incorrect optimizations in next cases:
//
// Point p[] = new Point[1];
// if ( x ) p[0] = new Point(); // Will be not scalar replaced
//
}
}
}
// Propagate escape states to referenced nodes.
bool has_java_obj = false;
// push all nodes with the same escape state on the worklist
}
// mark all reachable nodes
!is_null_ptr(pt)) {
has_java_obj = true;
// fields values are unknown if object escapes
}
}
if (is_null_ptr(npi))
continue;
}
}
}
// Has not escaping java objects
}
// Optimize objects compare.
// Clone returned Set since PointsTo() returns pointer
// to the same structure ConnectionGraph.pt_ptset.
// Check simple cases first.
// Comparing the same not escaping object.
return _pcmp_eq;
}
// Comparing not escaping allocation.
return _pcmp_neq; // This includes nullness check.
}
}
// Comparing not escaping allocation.
return _pcmp_neq; // This includes nullness check.
}
}
}
return NULL; // Sets are not disjoint
}
// Sets are disjoint.
if (set1_has_unknown_ptr && set2_has_null_ptr ||
// Check nullness of unknown object.
return NULL;
}
// Disjointness by itself is not sufficient since
// alias analysis is not complete for escaped objects.
// Disjoint sets are definitely unrelated only when
// at least one set has only not escaping objects.
if (!set1_has_unknown_ptr && !set1_has_null_ptr) {
bool has_only_non_escaping_alloc = true;
has_only_non_escaping_alloc = false;
break;
}
}
if (has_only_non_escaping_alloc) {
return _pcmp_neq;
}
}
if (!set2_has_unknown_ptr && !set2_has_null_ptr) {
bool has_only_non_escaping_alloc = true;
has_only_non_escaping_alloc = false;
break;
}
}
if (has_only_non_escaping_alloc) {
return _pcmp_neq;
}
}
return NULL;
}
bool is_arraycopy = false;
#ifdef ASSERT
case Op_Allocate:
case Op_AllocateArray:
case Op_Lock:
case Op_Unlock:
assert(false, "should be done already");
break;
#endif
case Op_CallLeafNoFP:
// fall through
case Op_CallLeaf:
{
// Stub calls, objects do not escape but they are not scale replaceable.
// Adjust escape state for outgoing arguments.
bool src_has_oops = false;
#ifdef ASSERT
if (!(is_arraycopy ||
) {
assert(false, "EA: unexpected CallLeaf");
}
#endif
//
// The inline_native_clone() case when the arraycopy stub is called
// after the allocation before Initialize and CheckCastPP nodes.
// Or normal arraycopy for object arrays case.
//
// Set AddP's base (Allocate) as not scalar replaceable since
// pointer to the base (with offset) is passed as argument.
//
}
}
}
//
// src or dst could be j.l.Object when other is basic type array:
//
// arraycopy(char[],0,Object*,0,size);
// arraycopy(Object*,0,char[],0,size);
//
// Do nothing special in such cases.
//
src_has_oops && arg_has_oops) {
// Destination object's fields reference an unknown object.
}
}
// Conservatively all values in source object fields globally escape
// since we don't know if values in destination object fields
// escape (it could be traced but it is too expensive).
}
// Use OffsetTop to indicate fields global escape.
}
}
}
}
break;
}
case Op_CallStaticJava:
// For a static call, we know exactly what method is being called.
// Use bytecode estimator to record the call's escape affects
{
// fall-through if not a Java method or no analyzer information
if (call_analyzer != NULL) {
bool copy_dependencies = false;
bool global_escapes = false;
bool fields_escapes = false;
if (!call_analyzer->is_arg_stack(k)) {
// The argument global escapes, mark everything it could point to
global_escapes = true;
} else {
if (!call_analyzer->is_arg_local(k)) {
// The argument itself doesn't escape, but any fields might
fields_escapes = true;
}
copy_dependencies = true;
}
if (global_escapes) {
// The argument global escapes, mark everything it could point to
} else {
if (fields_escapes) {
// The argument itself doesn't escape, but any fields might.
// Use OffsetTop to indicate such case.
}
}
}
}
}
if (copy_dependencies)
break;
}
}
default:
// Fall-through here if not a Java method or no analyzer information
// or some other type of call, assume the worst case: all arguments
// globally escape.
{
// adjust escape state for outgoing arguments
}
}
}
}
}
}
case Op_Allocate:
{
} else {
}
break;
}
case Op_AllocateArray:
{
} else {
// Not scalar replaceable if the length is not constant or too big.
}
}
break;
}
case Op_CallStaticJava:
// For a static call, we know exactly what method is being called.
// Use bytecode estimator to record whether the call's return value escapes
{
bool done = true;
// Note: we use isa_ptr() instead of isa_oopptr() here because the
// _multianewarray functions return a TypeRawPtr.
break; // doesn't return a pointer type
}
// not a Java method, assume global escape
} else {
bool copy_dependencies = false;
if (call_analyzer->is_return_allocated()) {
// Returns a newly allocated unescaped object, simply
// update dependency information.
// Mark it as NoEscape so that objects referenced by
// it's fields will be marked as NoEscape at least.
// Fields values are unknown
copy_dependencies = true;
} else {
// determine whether any arguments are returned
bool ret_arg = false;
ret_arg = true;
done = false;
else
}
}
}
if (done) {
copy_dependencies = true;
// is_return_local() is true when only arguments are returned.
// Returns unknown object.
}
}
}
if (copy_dependencies)
}
if (done)
break;
}
default:
// Some other type of call, assume the worst case that the
// returned value, if any, globally escapes.
{
// Note: we use isa_ptr() instead of isa_oopptr() here because the
// _multianewarray functions return a TypeRawPtr.
}
}
}
}
}
// Populate Connection Graph with Ideal nodes and create simple
// connection graph edges (do not need to check the node_type of inputs
// or to call PointsTo() to walk the connection graph).
return; // No need to redefine node's state.
if (n->is_Call()) {
// Arguments to allocation and locking don't escape.
if (n->is_Allocate()) {
// Put Lock and Unlock nodes on IGVN worklist to process them during
// the first IGVN optimization when escape information is still available.
} else {
// Don't mark as processed since call's arguments have to be processed.
// Check if a call returns an object.
if (!n->is_CallStaticJava()) {
// Since the called mathod is statically unknown assume
// the worst case that the returned value globally escapes.
}
}
}
return;
}
// Using isa_ptr() instead of isa_oopptr() for LoadP and Phi because
// ThreadLocal has RawPrt type.
switch (n->Opcode()) {
case Op_AddP:
{
break;
}
case Op_CastX2P:
{ // "Unsafe" memory access.
break;
}
case Op_CastPP:
case Op_CheckCastPP:
case Op_EncodeP:
case Op_DecodeN:
{
break;
} else {
}
break;
}
case Op_ConP:
{
// assume all pointer constants globally escape except for null
else
break;
}
case Op_ConN:
{
// assume all narrow oop constants globally escape except for null
else
break;
}
case Op_CreateEx:
{
// assume that all exception objects globally escape
break;
}
case Op_LoadKlass:
case Op_LoadNKlass:
{
break;
}
case Op_LoadP:
case Op_LoadN:
{
return;
}
break;
}
case Op_Parm:
{
return;
return;
// We have to assume all input parameters globally escape
// (Note: passing 'false' since _processed is already set).
break;
}
case Op_PartialSubtypeCheck:
{ // Produces Null or notNull and is used in CmpP.
break;
}
case Op_Phi:
{
// nothing to do if not an oop or narrow oop
return;
}
uint i;
for (i = 1; i < n->req() ; i++) {
continue; // ignore NULL
continue; // ignore top or inputs which go back this node
break;
} else {
}
}
if (i >= n->req())
else
break;
}
case Op_Proj:
{
// we are only interested in the oop result projection from a call
// The call may not be registered yet (since not all its inputs are registered)
// if this is the projection from backbranch edge of Phi.
}
// The call's result may need to be processed later if the call
// returns it's argument and the argument is not processed yet.
}
break;
}
}
break;
}
case Op_Return:
{
// Treat Return value as LocalVar with GlobalEscape escape state.
break;
} else {
}
}
break;
}
case Op_StoreP:
case Op_StoreN:
{
if (adr_type->isa_oopptr()) {
} else {
// We are computing a raw address for a store captured
// by an Initialize compute an appropriate address type.
} else {
return;
}
}
break;
}
case Op_StorePConditional:
case Op_CompareAndSwapP:
case Op_CompareAndSwapN:
{
if (adr_type->isa_oopptr()) {
} else {
return;
}
break;
}
case Op_AryEq:
case Op_StrComp:
case Op_StrEquals:
case Op_StrIndexOf:
{
// char[] arrays passed to string intrinsics are not scalar replaceable.
break;
}
case Op_ThreadLocal:
{
break;
}
default:
;
// nothing to do
}
return;
}
// Don't set processed bit for AddP, LoadP, StoreP since
// they may need more then one pass to process.
// Also don't mark as processed Call nodes since their
// arguments may need more then one pass to process.
return; // No need to redefine node's state.
if (n->is_Call()) {
return;
}
switch (n->Opcode()) {
case Op_AddP:
{
// Create a field edge to this node from everything base could point to.
}
break;
}
case Op_CastX2P:
{
assert(false, "Op_CastX2P");
break;
}
case Op_CastPP:
case Op_CheckCastPP:
case Op_EncodeP:
case Op_DecodeN:
{
} else {
}
break;
}
case Op_ConP:
{
assert(false, "Op_ConP");
break;
}
case Op_ConN:
{
assert(false, "Op_ConN");
break;
}
case Op_CreateEx:
{
assert(false, "Op_CreateEx");
break;
}
case Op_LoadKlass:
case Op_LoadNKlass:
{
assert(false, "Op_LoadKlass");
break;
}
case Op_LoadP:
case Op_LoadN:
{
#ifdef ASSERT
assert(false, "Op_LoadP");
#endif
} else {
}
// For everything "adr_base" could point to, create a deferred edge from
// this node to each field with the same offset.
// Add field edge if it is missing.
}
}
break;
}
case Op_Parm:
{
assert(false, "Op_Parm");
break;
}
case Op_PartialSubtypeCheck:
{
assert(false, "Op_PartialSubtypeCheck");
break;
}
case Op_Phi:
{
#ifdef ASSERT
assert(false, "Op_Phi");
#endif
continue; // ignore NULL
continue; // ignore top or inputs which go back this node
} else {
}
}
break;
}
case Op_Proj:
{
// we are only interested in the oop result projection from a call
"all nodes should be registered");
break;
}
}
assert(false, "Op_Proj");
break;
}
case Op_Return:
{
#ifdef ASSERT
assert(false, "Op_Return");
}
#endif
} else {
}
break;
}
case Op_StoreP:
case Op_StoreN:
case Op_StorePConditional:
case Op_CompareAndSwapP:
case Op_CompareAndSwapN:
{
#ifdef ASSERT
if (!adr_type->isa_oopptr())
#endif
// For everything "adr_base" could point to, create a deferred edge
// to "val" from each field with the same offset.
// Add field edge if it is missing.
}
break;
}
case Op_AryEq:
case Op_StrComp:
case Op_StrEquals:
case Op_StrIndexOf:
{
// char[] arrays passed to string intrinsic do not escape but
// they are not scalar replaceable. Adjust escape state for them.
// Start from in(2) edge since in(1) is memory edge.
}
// Mark as ArgEscape everything "adr" could point to.
}
}
break;
}
case Op_ThreadLocal:
{
assert(false, "Op_ThreadLocal");
break;
}
default:
// This method should be called only for EA specific nodes.
}
}
#ifndef PRODUCT
void ConnectionGraph::dump() {
bool first = true;
continue;
if (first) {
first = false;
}
// Print all locals which reference this allocation
}
}
if (Verbose) {
// Print all fields which reference this allocation
}
}
}
}
}
#endif