6680594: Load + Load isn't canonicalized leading to missed GVN opportunities
Reviewed-by: kvn, jrose
/*
* Copyright 1997-2007 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.
*
*/
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
#include "incls/_precompiled.incl"
#include "incls/_gcm.cpp.incl"
//----------------------------schedule_node_into_block-------------------------
// Insert node n into block b. Look for projections of n and make sure they
// are in b also.
void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) {
// Set basic block of n, Add n to b,
_bbs.map(n->_idx, b);
b->add_inst(n);
// After Matching, nearly any old Node may have projections trailing it.
// These are usually machine-dependent flags. In any case, they might
// float to another block below this one. Move them up.
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
Node* use = n->fast_out(i);
if (use->is_Proj()) {
Block* buse = _bbs[use->_idx];
if (buse != b) { // In wrong block?
if (buse != NULL)
buse->find_remove(use); // Remove from wrong block
_bbs.map(use->_idx, b); // Re-insert in this block
b->add_inst(use);
}
}
}
}
//------------------------------schedule_pinned_nodes--------------------------
// Set the basic block for Nodes pinned into blocks
void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) {
// Allocate node stack of size C->unique()+8 to avoid frequent realloc
GrowableArray <Node *> spstack(C->unique()+8);
spstack.push(_root);
while ( spstack.is_nonempty() ) {
Node *n = spstack.pop();
if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited
if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down!
Node *input = n->in(0);
assert( input, "pinned Node must have Control" );
while( !input->is_block_start() )
input = input->in(0);
Block *b = _bbs[input->_idx]; // Basic block of controlling input
schedule_node_into_block(n, b);
}
for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs
if( n->in(i) != NULL )
spstack.push(n->in(i));
}
}
}
}
#ifdef ASSERT
// Assert that new input b2 is dominated by all previous inputs.
// Check this by by seeing that it is dominated by b1, the deepest
// input observed until b2.
static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) {
if (b1 == NULL) return;
assert(b1->_dom_depth < b2->_dom_depth, "sanity");
Block* tmp = b2;
while (tmp != b1 && tmp != NULL) {
tmp = tmp->_idom;
}
if (tmp != b1) {
// Detected an unschedulable graph. Print some nice stuff and die.
tty->print_cr("!!! Unschedulable graph !!!");
for (uint j=0; j<n->len(); j++) { // For all inputs
Node* inn = n->in(j); // Get input
if (inn == NULL) continue; // Ignore NULL, missing inputs
Block* inb = bbs[inn->_idx];
tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order,
inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth);
inn->dump();
}
tty->print("Failing node: ");
n->dump();
assert(false, "unscheduable graph");
}
}
#endif
static Block* find_deepest_input(Node* n, Block_Array &bbs) {
// Find the last input dominated by all other inputs.
Block* deepb = NULL; // Deepest block so far
int deepb_dom_depth = 0;
for (uint k = 0; k < n->len(); k++) { // For all inputs
Node* inn = n->in(k); // Get input
if (inn == NULL) continue; // Ignore NULL, missing inputs
Block* inb = bbs[inn->_idx];
assert(inb != NULL, "must already have scheduled this input");
if (deepb_dom_depth < (int) inb->_dom_depth) {
// The new inb must be dominated by the previous deepb.
// The various inputs must be linearly ordered in the dom
// tree, or else there will not be a unique deepest block.
DEBUG_ONLY(assert_dom(deepb, inb, n, bbs));
deepb = inb; // Save deepest block
deepb_dom_depth = deepb->_dom_depth;
}
}
assert(deepb != NULL, "must be at least one input to n");
return deepb;
}
//------------------------------schedule_early---------------------------------
// Find the earliest Block any instruction can be placed in. Some instructions
// are pinned into Blocks. Unpinned instructions can appear in last block in
// which all their inputs occur.
bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) {
// Allocate stack with enough space to avoid frequent realloc
Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats
// roots.push(_root); _root will be processed among C->top() inputs
roots.push(C->top());
visited.set(C->top()->_idx);
while (roots.size() != 0) {
// Use local variables nstack_top_n & nstack_top_i to cache values
// on stack's top.
Node *nstack_top_n = roots.pop();
uint nstack_top_i = 0;
//while_nstack_nonempty:
while (true) {
// Get parent node and next input's index from stack's top.
Node *n = nstack_top_n;
uint i = nstack_top_i;
if (i == 0) {
// Special control input processing.
// While I am here, go ahead and look for Nodes which are taking control
// from a is_block_proj Node. After I inserted RegionNodes to make proper
// blocks, the control at a is_block_proj more properly comes from the
// Region being controlled by the block_proj Node.
const Node *in0 = n->in(0);
if (in0 != NULL) { // Control-dependent?
const Node *p = in0->is_block_proj();
if (p != NULL && p != n) { // Control from a block projection?
// Find trailing Region
Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block
uint j = 0;
if (pb->_num_succs != 1) { // More then 1 successor?
// Search for successor
uint max = pb->_nodes.size();
assert( max > 1, "" );
uint start = max - pb->_num_succs;
// Find which output path belongs to projection
for (j = start; j < max; j++) {
if( pb->_nodes[j] == in0 )
break;
}
assert( j < max, "must find" );
// Change control to match head of successor basic block
j -= start;
}
n->set_req(0, pb->_succs[j]->head());
}
} else { // n->in(0) == NULL
if (n->req() == 1) { // This guy is a constant with NO inputs?
n->set_req(0, _root);
}
}
}
// First, visit all inputs and force them to get a block. If an
// input is already in a block we quit following inputs (to avoid
// cycles). Instead we put that Node on a worklist to be handled
// later (since IT'S inputs may not have a block yet).
bool done = true; // Assume all n's inputs will be processed
while (i < n->len()) { // For all inputs
Node *in = n->in(i); // Get input
++i;
if (in == NULL) continue; // Ignore NULL, missing inputs
int is_visited = visited.test_set(in->_idx);
if (!_bbs.lookup(in->_idx)) { // Missing block selection?
if (is_visited) {
// assert( !visited.test(in->_idx), "did not schedule early" );
return false;
}
nstack.push(n, i); // Save parent node and next input's index.
nstack_top_n = in; // Process current input now.
nstack_top_i = 0;
done = false; // Not all n's inputs processed.
break; // continue while_nstack_nonempty;
} else if (!is_visited) { // Input not yet visited?
roots.push(in); // Visit this guy later, using worklist
}
}
if (done) {
// All of n's inputs have been processed, complete post-processing.
// Some instructions are pinned into a block. These include Region,
// Phi, Start, Return, and other control-dependent instructions and
// any projections which depend on them.
if (!n->pinned()) {
// Set earliest legal block.
_bbs.map(n->_idx, find_deepest_input(n, _bbs));
}
if (nstack.is_empty()) {
// Finished all nodes on stack.
// Process next node on the worklist 'roots'.
break;
}
// Get saved parent node and next input's index.
nstack_top_n = nstack.node();
nstack_top_i = nstack.index();
nstack.pop();
} // if (done)
} // while (true)
} // while (roots.size() != 0)
return true;
}
//------------------------------dom_lca----------------------------------------
// Find least common ancestor in dominator tree
// LCA is a current notion of LCA, to be raised above 'this'.
// As a convenient boundary condition, return 'this' if LCA is NULL.
// Find the LCA of those two nodes.
Block* Block::dom_lca(Block* LCA) {
if (LCA == NULL || LCA == this) return this;
Block* anc = this;
while (anc->_dom_depth > LCA->_dom_depth)
anc = anc->_idom; // Walk up till anc is as high as LCA
while (LCA->_dom_depth > anc->_dom_depth)
LCA = LCA->_idom; // Walk up till LCA is as high as anc
while (LCA != anc) { // Walk both up till they are the same
LCA = LCA->_idom;
anc = anc->_idom;
}
return LCA;
}
//--------------------------raise_LCA_above_use--------------------------------
// We are placing a definition, and have been given a def->use edge.
// The definition must dominate the use, so move the LCA upward in the
// dominator tree to dominate the use. If the use is a phi, adjust
// the LCA only with the phi input paths which actually use this def.
static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) {
Block* buse = bbs[use->_idx];
if (buse == NULL) return LCA; // Unused killing Projs have no use block
if (!use->is_Phi()) return buse->dom_lca(LCA);
uint pmax = use->req(); // Number of Phi inputs
// Why does not this loop just break after finding the matching input to
// the Phi? Well...it's like this. I do not have true def-use/use-def
// chains. Means I cannot distinguish, from the def-use direction, which
// of many use-defs lead from the same use to the same def. That is, this
// Phi might have several uses of the same def. Each use appears in a
// different predecessor block. But when I enter here, I cannot distinguish
// which use-def edge I should find the predecessor block for. So I find
// them all. Means I do a little extra work if a Phi uses the same value
// more than once.
for (uint j=1; j<pmax; j++) { // For all inputs
if (use->in(j) == def) { // Found matching input?
Block* pred = bbs[buse->pred(j)->_idx];
LCA = pred->dom_lca(LCA);
}
}
return LCA;
}
//----------------------------raise_LCA_above_marks----------------------------
// Return a new LCA that dominates LCA and any of its marked predecessors.
// Search all my parents up to 'early' (exclusive), looking for predecessors
// which are marked with the given index. Return the LCA (in the dom tree)
// of all marked blocks. If there are none marked, return the original
// LCA.
static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark,
Block* early, Block_Array &bbs) {
Block_List worklist;
worklist.push(LCA);
while (worklist.size() > 0) {
Block* mid = worklist.pop();
if (mid == early) continue; // stop searching here
// Test and set the visited bit.
if (mid->raise_LCA_visited() == mark) continue; // already visited
mid->set_raise_LCA_visited(mark);
// Don't process the current LCA, otherwise the search may terminate early
if (mid != LCA && mid->raise_LCA_mark() == mark) {
// Raise the LCA.
LCA = mid->dom_lca(LCA);
if (LCA == early) break; // stop searching everywhere
assert(early->dominates(LCA), "early is high enough");
// Resume searching at that point, skipping intermediate levels.
worklist.push(LCA);
} else {
// Keep searching through this block's predecessors.
for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) {
Block* mid_parent = bbs[ mid->pred(j)->_idx ];
worklist.push(mid_parent);
}
}
}
return LCA;
}
//--------------------------memory_early_block--------------------------------
// This is a variation of find_deepest_input, the heart of schedule_early.
// Find the "early" block for a load, if we considered only memory and
// address inputs, that is, if other data inputs were ignored.
//
// Because a subset of edges are considered, the resulting block will
// be earlier (at a shallower dom_depth) than the true schedule_early
// point of the node. We compute this earlier block as a more permissive
// site for anti-dependency insertion, but only if subsume_loads is enabled.
static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) {
Node* base;
Node* index;
Node* store = load->in(MemNode::Memory);
load->as_Mach()->memory_inputs(base, index);
assert(base != NodeSentinel && index != NodeSentinel,
"unexpected base/index inputs");
Node* mem_inputs[4];
int mem_inputs_length = 0;
if (base != NULL) mem_inputs[mem_inputs_length++] = base;
if (index != NULL) mem_inputs[mem_inputs_length++] = index;
if (store != NULL) mem_inputs[mem_inputs_length++] = store;
// In the comparision below, add one to account for the control input,
// which may be null, but always takes up a spot in the in array.
if (mem_inputs_length + 1 < (int) load->req()) {
// This "load" has more inputs than just the memory, base and index inputs.
// For purposes of checking anti-dependences, we need to start
// from the early block of only the address portion of the instruction,
// and ignore other blocks that may have factored into the wider
// schedule_early calculation.
if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0);
Block* deepb = NULL; // Deepest block so far
int deepb_dom_depth = 0;
for (int i = 0; i < mem_inputs_length; i++) {
Block* inb = bbs[mem_inputs[i]->_idx];
if (deepb_dom_depth < (int) inb->_dom_depth) {
// The new inb must be dominated by the previous deepb.
// The various inputs must be linearly ordered in the dom
// tree, or else there will not be a unique deepest block.
DEBUG_ONLY(assert_dom(deepb, inb, load, bbs));
deepb = inb; // Save deepest block
deepb_dom_depth = deepb->_dom_depth;
}
}
early = deepb;
}
return early;
}
//--------------------------insert_anti_dependences---------------------------
// A load may need to witness memory that nearby stores can overwrite.
// For each nearby store, either insert an "anti-dependence" edge
// from the load to the store, or else move LCA upward to force the
// load to (eventually) be scheduled in a block above the store.
//
// Do not add edges to stores on distinct control-flow paths;
// only add edges to stores which might interfere.
//
// Return the (updated) LCA. There will not be any possibly interfering
// store between the load's "early block" and the updated LCA.
// Any stores in the updated LCA will have new precedence edges
// back to the load. The caller is expected to schedule the load
// in the LCA, in which case the precedence edges will make LCM
// preserve anti-dependences. The caller may also hoist the load
// above the LCA, if it is not the early block.
Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) {
assert(load->needs_anti_dependence_check(), "must be a load of some sort");
assert(LCA != NULL, "");
DEBUG_ONLY(Block* LCA_orig = LCA);
// Compute the alias index. Loads and stores with different alias indices
// do not need anti-dependence edges.
uint load_alias_idx = C->get_alias_index(load->adr_type());
#ifdef ASSERT
if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 &&
(PrintOpto || VerifyAliases ||
PrintMiscellaneous && (WizardMode || Verbose))) {
// Load nodes should not consume all of memory.
// Reporting a bottom type indicates a bug in adlc.
// If some particular type of node validly consumes all of memory,
// sharpen the preceding "if" to exclude it, so we can catch bugs here.
tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory.");
load->dump(2);
if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, "");
}
#endif
assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp),
"String compare is only known 'load' that does not conflict with any stores");
if (!C->alias_type(load_alias_idx)->is_rewritable()) {
// It is impossible to spoil this load by putting stores before it,
// because we know that the stores will never update the value
// which 'load' must witness.
return LCA;
}
node_idx_t load_index = load->_idx;
// Note the earliest legal placement of 'load', as determined by
// by the unique point in the dom tree where all memory effects
// and other inputs are first available. (Computed by schedule_early.)
// For normal loads, 'early' is the shallowest place (dom graph wise)
// to look for anti-deps between this load and any store.
Block* early = _bbs[load_index];
// If we are subsuming loads, compute an "early" block that only considers
// memory or address inputs. This block may be different than the
// schedule_early block in that it could be at an even shallower depth in the
// dominator tree, and allow for a broader discovery of anti-dependences.
if (C->subsume_loads()) {
early = memory_early_block(load, early, _bbs);
}
ResourceArea *area = Thread::current()->resource_area();
Node_List worklist_mem(area); // prior memory state to store
Node_List worklist_store(area); // possible-def to explore
Node_List worklist_visited(area); // visited mergemem nodes
Node_List non_early_stores(area); // all relevant stores outside of early
bool must_raise_LCA = false;
#ifdef TRACK_PHI_INPUTS
// %%% This extra checking fails because MergeMem nodes are not GVNed.
// Provide "phi_inputs" to check if every input to a PhiNode is from the
// original memory state. This indicates a PhiNode for which should not
// prevent the load from sinking. For such a block, set_raise_LCA_mark
// may be overly conservative.
// Mechanism: count inputs seen for each Phi encountered in worklist_store.
DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0));
#endif
// 'load' uses some memory state; look for users of the same state.
// Recurse through MergeMem nodes to the stores that use them.
// Each of these stores is a possible definition of memory
// that 'load' needs to use. We need to force 'load'
// to occur before each such store. When the store is in
// the same block as 'load', we insert an anti-dependence
// edge load->store.
// The relevant stores "nearby" the load consist of a tree rooted
// at initial_mem, with internal nodes of type MergeMem.
// Therefore, the branches visited by the worklist are of this form:
// initial_mem -> (MergeMem ->)* store
// The anti-dependence constraints apply only to the fringe of this tree.
Node* initial_mem = load->in(MemNode::Memory);
worklist_store.push(initial_mem);
worklist_visited.push(initial_mem);
worklist_mem.push(NULL);
while (worklist_store.size() > 0) {
// Examine a nearby store to see if it might interfere with our load.
Node* mem = worklist_mem.pop();
Node* store = worklist_store.pop();
uint op = store->Opcode();
// MergeMems do not directly have anti-deps.
// Treat them as internal nodes in a forward tree of memory states,
// the leaves of which are each a 'possible-def'.
if (store == initial_mem // root (exclusive) of tree we are searching
|| op == Op_MergeMem // internal node of tree we are searching
) {
mem = store; // It's not a possibly interfering store.
if (store == initial_mem)
initial_mem = NULL; // only process initial memory once
for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
store = mem->fast_out(i);
if (store->is_MergeMem()) {
// Be sure we don't get into combinatorial problems.
// (Allow phis to be repeated; they can merge two relevant states.)
uint j = worklist_visited.size();
for (; j > 0; j--) {
if (worklist_visited.at(j-1) == store) break;
}
if (j > 0) continue; // already on work list; do not repeat
worklist_visited.push(store);
}
worklist_mem.push(mem);
worklist_store.push(store);
}
continue;
}
if (op == Op_MachProj || op == Op_Catch) continue;
if (store->needs_anti_dependence_check()) continue; // not really a store
// Compute the alias index. Loads and stores with different alias
// indices do not need anti-dependence edges. Wide MemBar's are
// anti-dependent on everything (except immutable memories).
const TypePtr* adr_type = store->adr_type();
if (!C->can_alias(adr_type, load_alias_idx)) continue;
// Most slow-path runtime calls do NOT modify Java memory, but
// they can block and so write Raw memory.
if (store->is_Mach()) {
MachNode* mstore = store->as_Mach();
if (load_alias_idx != Compile::AliasIdxRaw) {
// Check for call into the runtime using the Java calling
// convention (and from there into a wrapper); it has no
// _method. Can't do this optimization for Native calls because
// they CAN write to Java memory.
if (mstore->ideal_Opcode() == Op_CallStaticJava) {
assert(mstore->is_MachSafePoint(), "");
MachSafePointNode* ms = (MachSafePointNode*) mstore;
assert(ms->is_MachCallJava(), "");
MachCallJavaNode* mcj = (MachCallJavaNode*) ms;
if (mcj->_method == NULL) {
// These runtime calls do not write to Java visible memory
// (other than Raw) and so do not require anti-dependence edges.
continue;
}
}
// Same for SafePoints: they read/write Raw but only read otherwise.
// This is basically a workaround for SafePoints only defining control
// instead of control + memory.
if (mstore->ideal_Opcode() == Op_SafePoint)
continue;
} else {
// Some raw memory, such as the load of "top" at an allocation,
// can be control dependent on the previous safepoint. See
// comments in GraphKit::allocate_heap() about control input.
// Inserting an anti-dep between such a safepoint and a use
// creates a cycle, and will cause a subsequent failure in
// local scheduling. (BugId 4919904)
// (%%% How can a control input be a safepoint and not a projection??)
if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore)
continue;
}
}
// Identify a block that the current load must be above,
// or else observe that 'store' is all the way up in the
// earliest legal block for 'load'. In the latter case,
// immediately insert an anti-dependence edge.
Block* store_block = _bbs[store->_idx];
assert(store_block != NULL, "unused killing projections skipped above");
if (store->is_Phi()) {
// 'load' uses memory which is one (or more) of the Phi's inputs.
// It must be scheduled not before the Phi, but rather before
// each of the relevant Phi inputs.
//
// Instead of finding the LCA of all inputs to a Phi that match 'mem',
// we mark each corresponding predecessor block and do a combined
// hoisting operation later (raise_LCA_above_marks).
//
// Do not assert(store_block != early, "Phi merging memory after access")
// PhiNode may be at start of block 'early' with backedge to 'early'
DEBUG_ONLY(bool found_match = false);
for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) {
if (store->in(j) == mem) { // Found matching input?
DEBUG_ONLY(found_match = true);
Block* pred_block = _bbs[store_block->pred(j)->_idx];
if (pred_block != early) {
// If any predecessor of the Phi matches the load's "early block",
// we do not need a precedence edge between the Phi and 'load'
// since the load will be forced into a block preceeding the Phi.
pred_block->set_raise_LCA_mark(load_index);
assert(!LCA_orig->dominates(pred_block) ||
early->dominates(pred_block), "early is high enough");
must_raise_LCA = true;
}
}
}
assert(found_match, "no worklist bug");
#ifdef TRACK_PHI_INPUTS
#ifdef ASSERT
// This assert asks about correct handling of PhiNodes, which may not
// have all input edges directly from 'mem'. See BugId 4621264
int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1;
// Increment by exactly one even if there are multiple copies of 'mem'
// coming into the phi, because we will run this block several times
// if there are several copies of 'mem'. (That's how DU iterators work.)
phi_inputs.at_put(store->_idx, num_mem_inputs);
assert(PhiNode::Input + num_mem_inputs < store->req(),
"Expect at least one phi input will not be from original memory state");
#endif //ASSERT
#endif //TRACK_PHI_INPUTS
} else if (store_block != early) {
// 'store' is between the current LCA and earliest possible block.
// Label its block, and decide later on how to raise the LCA
// to include the effect on LCA of this store.
// If this store's block gets chosen as the raised LCA, we
// will find him on the non_early_stores list and stick him
// with a precedence edge.
// (But, don't bother if LCA is already raised all the way.)
if (LCA != early) {
store_block->set_raise_LCA_mark(load_index);
must_raise_LCA = true;
non_early_stores.push(store);
}
} else {
// Found a possibly-interfering store in the load's 'early' block.
// This means 'load' cannot sink at all in the dominator tree.
// Add an anti-dep edge, and squeeze 'load' into the highest block.
assert(store != load->in(0), "dependence cycle found");
if (verify) {
assert(store->find_edge(load) != -1, "missing precedence edge");
} else {
store->add_prec(load);
}
LCA = early;
// This turns off the process of gathering non_early_stores.
}
}
// (Worklist is now empty; all nearby stores have been visited.)
// Finished if 'load' must be scheduled in its 'early' block.
// If we found any stores there, they have already been given
// precedence edges.
if (LCA == early) return LCA;
// We get here only if there are no possibly-interfering stores
// in the load's 'early' block. Move LCA up above all predecessors
// which contain stores we have noted.
//
// The raised LCA block can be a home to such interfering stores,
// but its predecessors must not contain any such stores.
//
// The raised LCA will be a lower bound for placing the load,
// preventing the load from sinking past any block containing
// a store that may invalidate the memory state required by 'load'.
if (must_raise_LCA)
LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs);
if (LCA == early) return LCA;
// Insert anti-dependence edges from 'load' to each store
// in the non-early LCA block.
// Mine the non_early_stores list for such stores.
if (LCA->raise_LCA_mark() == load_index) {
while (non_early_stores.size() > 0) {
Node* store = non_early_stores.pop();
Block* store_block = _bbs[store->_idx];
if (store_block == LCA) {
// add anti_dependence from store to load in its own block
assert(store != load->in(0), "dependence cycle found");
if (verify) {
assert(store->find_edge(load) != -1, "missing precedence edge");
} else {
store->add_prec(load);
}
} else {
assert(store_block->raise_LCA_mark() == load_index, "block was marked");
// Any other stores we found must be either inside the new LCA
// or else outside the original LCA. In the latter case, they
// did not interfere with any use of 'load'.
assert(LCA->dominates(store_block)
|| !LCA_orig->dominates(store_block), "no stray stores");
}
}
}
// Return the highest block containing stores; any stores
// within that block have been given anti-dependence edges.
return LCA;
}
// This class is used to iterate backwards over the nodes in the graph.
class Node_Backward_Iterator {
private:
Node_Backward_Iterator();
public:
// Constructor for the iterator
Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs);
// Postincrement operator to iterate over the nodes
Node *next();
private:
VectorSet &_visited;
Node_List &_stack;
Block_Array &_bbs;
};
// Constructor for the Node_Backward_Iterator
Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs )
: _visited(visited), _stack(stack), _bbs(bbs) {
// The stack should contain exactly the root
stack.clear();
stack.push(root);
// Clear the visited bits
visited.Clear();
}
// Iterator for the Node_Backward_Iterator
Node *Node_Backward_Iterator::next() {
// If the _stack is empty, then just return NULL: finished.
if ( !_stack.size() )
return NULL;
// '_stack' is emulating a real _stack. The 'visit-all-users' loop has been
// made stateless, so I do not need to record the index 'i' on my _stack.
// Instead I visit all users each time, scanning for unvisited users.
// I visit unvisited not-anti-dependence users first, then anti-dependent
// children next.
Node *self = _stack.pop();
// I cycle here when I am entering a deeper level of recursion.
// The key variable 'self' was set prior to jumping here.
while( 1 ) {
_visited.set(self->_idx);
// Now schedule all uses as late as possible.
uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx;
uint src_rpo = _bbs[src]->_rpo;
// Schedule all nodes in a post-order visit
Node *unvisited = NULL; // Unvisited anti-dependent Node, if any
// Scan for unvisited nodes
for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
// For all uses, schedule late
Node* n = self->fast_out(i); // Use
// Skip already visited children
if ( _visited.test(n->_idx) )
continue;
// do not traverse backward control edges
Node *use = n->is_Proj() ? n->in(0) : n;
uint use_rpo = _bbs[use->_idx]->_rpo;
if ( use_rpo < src_rpo )
continue;
// Phi nodes always precede uses in a basic block
if ( use_rpo == src_rpo && use->is_Phi() )
continue;
unvisited = n; // Found unvisited
// Check for possible-anti-dependent
if( !n->needs_anti_dependence_check() )
break; // Not visited, not anti-dep; schedule it NOW
}
// Did I find an unvisited not-anti-dependent Node?
if ( !unvisited )
break; // All done with children; post-visit 'self'
// Visit the unvisited Node. Contains the obvious push to
// indicate I'm entering a deeper level of recursion. I push the
// old state onto the _stack and set a new state and loop (recurse).
_stack.push(self);
self = unvisited;
} // End recursion loop
return self;
}
//------------------------------ComputeLatenciesBackwards----------------------
// Compute the latency of all the instructions.
void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) {
#ifndef PRODUCT
if (trace_opto_pipelining())
tty->print("\n#---- ComputeLatenciesBackwards ----\n");
#endif
Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
Node *n;
// Walk over all the nodes from last to first
while (n = iter.next()) {
// Set the latency for the definitions of this instruction
partial_latency_of_defs(n);
}
} // end ComputeLatenciesBackwards
//------------------------------partial_latency_of_defs------------------------
// Compute the latency impact of this node on all defs. This computes
// a number that increases as we approach the beginning of the routine.
void PhaseCFG::partial_latency_of_defs(Node *n) {
// Set the latency for this instruction
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# latency_to_inputs: node_latency[%d] = %d for node",
n->_idx, _node_latency.at_grow(n->_idx));
dump();
}
#endif
if (n->is_Proj())
n = n->in(0);
if (n->is_Root())
return;
uint nlen = n->len();
uint use_latency = _node_latency.at_grow(n->_idx);
uint use_pre_order = _bbs[n->_idx]->_pre_order;
for ( uint j=0; j<nlen; j++ ) {
Node *def = n->in(j);
if (!def || def == n)
continue;
// Walk backwards thru projections
if (def->is_Proj())
def = def->in(0);
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# in(%2d): ", j);
def->dump();
}
#endif
// If the defining block is not known, assume it is ok
Block *def_block = _bbs[def->_idx];
uint def_pre_order = def_block ? def_block->_pre_order : 0;
if ( (use_pre_order < def_pre_order) ||
(use_pre_order == def_pre_order && n->is_Phi()) )
continue;
uint delta_latency = n->latency(j);
uint current_latency = delta_latency + use_latency;
if (_node_latency.at_grow(def->_idx) < current_latency) {
_node_latency.at_put_grow(def->_idx, current_latency);
}
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d",
use_latency, j, delta_latency, current_latency, def->_idx,
_node_latency.at_grow(def->_idx));
}
#endif
}
}
//------------------------------latency_from_use-------------------------------
// Compute the latency of a specific use
int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) {
// If self-reference, return no latency
if (use == n || use->is_Root())
return 0;
uint def_pre_order = _bbs[def->_idx]->_pre_order;
uint latency = 0;
// If the use is not a projection, then it is simple...
if (!use->is_Proj()) {
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# out(): ");
use->dump();
}
#endif
uint use_pre_order = _bbs[use->_idx]->_pre_order;
if (use_pre_order < def_pre_order)
return 0;
if (use_pre_order == def_pre_order && use->is_Phi())
return 0;
uint nlen = use->len();
uint nl = _node_latency.at_grow(use->_idx);
for ( uint j=0; j<nlen; j++ ) {
if (use->in(j) == n) {
// Change this if we want local latencies
uint ul = use->latency(j);
uint l = ul + nl;
if (latency < l) latency = l;
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d",
nl, j, ul, l, latency);
}
#endif
}
}
} else {
// This is a projection, just grab the latency of the use(s)
for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
uint l = latency_from_use(use, def, use->fast_out(j));
if (latency < l) latency = l;
}
}
return latency;
}
//------------------------------latency_from_uses------------------------------
// Compute the latency of this instruction relative to all of it's uses.
// This computes a number that increases as we approach the beginning of the
// routine.
void PhaseCFG::latency_from_uses(Node *n) {
// Set the latency for this instruction
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# latency_from_outputs: node_latency[%d] = %d for node",
n->_idx, _node_latency.at_grow(n->_idx));
dump();
}
#endif
uint latency=0;
const Node *def = n->is_Proj() ? n->in(0): n;
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
uint l = latency_from_use(n, def, n->fast_out(i));
if (latency < l) latency = l;
}
_node_latency.at_put_grow(n->_idx, latency);
}
//------------------------------hoist_to_cheaper_block-------------------------
// Pick a block for node self, between early and LCA, that is a cheaper
// alternative to LCA.
Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) {
const double delta = 1+PROB_UNLIKELY_MAG(4);
Block* least = LCA;
double least_freq = least->_freq;
uint target = _node_latency.at_grow(self->_idx);
uint start_latency = _node_latency.at_grow(LCA->_nodes[0]->_idx);
uint end_latency = _node_latency.at_grow(LCA->_nodes[LCA->end_idx()]->_idx);
bool in_latency = (target <= start_latency);
const Block* root_block = _bbs[_root->_idx];
// Turn off latency scheduling if scheduling is just plain off
if (!C->do_scheduling())
in_latency = true;
// Do not hoist (to cover latency) instructions which target a
// single register. Hoisting stretches the live range of the
// single register and may force spilling.
MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty())
in_latency = true;
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("# Find cheaper block for latency %d: ",
_node_latency.at_grow(self->_idx));
self->dump();
tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
LCA->_pre_order,
LCA->_nodes[0]->_idx,
start_latency,
LCA->_nodes[LCA->end_idx()]->_idx,
end_latency,
least_freq);
}
#endif
// Walk up the dominator tree from LCA (Lowest common ancestor) to
// the earliest legal location. Capture the least execution frequency.
while (LCA != early) {
LCA = LCA->_idom; // Follow up the dominator tree
if (LCA == NULL) {
// Bailout without retry
C->record_method_not_compilable("late schedule failed: LCA == NULL");
return least;
}
// Don't hoist machine instructions to the root basic block
if (mach && LCA == root_block)
break;
uint start_lat = _node_latency.at_grow(LCA->_nodes[0]->_idx);
uint end_idx = LCA->end_idx();
uint end_lat = _node_latency.at_grow(LCA->_nodes[end_idx]->_idx);
double LCA_freq = LCA->_freq;
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g",
LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq);
}
#endif
if (LCA_freq < least_freq || // Better Frequency
( !in_latency && // No block containing latency
LCA_freq < least_freq * delta && // No worse frequency
target >= end_lat && // within latency range
!self->is_iteratively_computed() ) // But don't hoist IV increments
// because they may end up above other uses of their phi forcing
// their result register to be different from their input.
) {
least = LCA; // Found cheaper block
least_freq = LCA_freq;
start_latency = start_lat;
end_latency = end_lat;
if (target <= start_lat)
in_latency = true;
}
}
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print_cr("# Choose block B%d with start latency=%d and freq=%g",
least->_pre_order, start_latency, least_freq);
}
#endif
// See if the latency needs to be updated
if (target < end_latency) {
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency);
}
#endif
_node_latency.at_put_grow(self->_idx, end_latency);
partial_latency_of_defs(self);
}
return least;
}
//------------------------------schedule_late-----------------------------------
// Now schedule all codes as LATE as possible. This is the LCA in the
// dominator tree of all USES of a value. Pick the block with the least
// loop nesting depth that is lowest in the dominator tree.
extern const char must_clone[];
void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) {
#ifndef PRODUCT
if (trace_opto_pipelining())
tty->print("\n#---- schedule_late ----\n");
#endif
Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs);
Node *self;
// Walk over all the nodes from last to first
while (self = iter.next()) {
Block* early = _bbs[self->_idx]; // Earliest legal placement
if (self->is_top()) {
// Top node goes in bb #2 with other constants.
// It must be special-cased, because it has no out edges.
early->add_inst(self);
continue;
}
// No uses, just terminate
if (self->outcnt() == 0) {
assert(self->Opcode() == Op_MachProj, "sanity");
continue; // Must be a dead machine projection
}
// If node is pinned in the block, then no scheduling can be done.
if( self->pinned() ) // Pinned in block?
continue;
MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL;
if (mach) {
switch (mach->ideal_Opcode()) {
case Op_CreateEx:
// Don't move exception creation
early->add_inst(self);
continue;
break;
case Op_CheckCastPP:
// Don't move CheckCastPP nodes away from their input, if the input
// is a rawptr (5071820).
Node *def = self->in(1);
if (def != NULL && def->bottom_type()->base() == Type::RawPtr) {
early->add_inst(self);
continue;
}
break;
}
}
// Gather LCA of all uses
Block *LCA = NULL;
{
for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) {
// For all uses, find LCA
Node* use = self->fast_out(i);
LCA = raise_LCA_above_use(LCA, use, self, _bbs);
}
} // (Hide defs of imax, i from rest of block.)
// Place temps in the block of their use. This isn't a
// requirement for correctness but it reduces useless
// interference between temps and other nodes.
if (mach != NULL && mach->is_MachTemp()) {
_bbs.map(self->_idx, LCA);
LCA->add_inst(self);
continue;
}
// Check if 'self' could be anti-dependent on memory
if (self->needs_anti_dependence_check()) {
// Hoist LCA above possible-defs and insert anti-dependences to
// defs in new LCA block.
LCA = insert_anti_dependences(LCA, self);
}
if (early->_dom_depth > LCA->_dom_depth) {
// Somehow the LCA has moved above the earliest legal point.
// (One way this can happen is via memory_early_block.)
if (C->subsume_loads() == true && !C->failing()) {
// Retry with subsume_loads == false
// If this is the first failure, the sentinel string will "stick"
// to the Compile object, and the C2Compiler will see it and retry.
C->record_failure(C2Compiler::retry_no_subsuming_loads());
} else {
// Bailout without retry when (early->_dom_depth > LCA->_dom_depth)
C->record_method_not_compilable("late schedule failed: incorrect graph");
}
return;
}
// If there is no opportunity to hoist, then we're done.
bool try_to_hoist = (LCA != early);
// Must clone guys stay next to use; no hoisting allowed.
// Also cannot hoist guys that alter memory or are otherwise not
// allocatable (hoisting can make a value live longer, leading to
// anti and output dependency problems which are normally resolved
// by the register allocator giving everyone a different register).
if (mach != NULL && must_clone[mach->ideal_Opcode()])
try_to_hoist = false;
Block* late = NULL;
if (try_to_hoist) {
// Now find the block with the least execution frequency.
// Start at the latest schedule and work up to the earliest schedule
// in the dominator tree. Thus the Node will dominate all its uses.
late = hoist_to_cheaper_block(LCA, early, self);
} else {
// Just use the LCA of the uses.
late = LCA;
}
// Put the node into target block
schedule_node_into_block(self, late);
#ifdef ASSERT
if (self->needs_anti_dependence_check()) {
// since precedence edges are only inserted when we're sure they
// are needed make sure that after placement in a block we don't
// need any new precedence edges.
verify_anti_dependences(late, self);
}
#endif
} // Loop until all nodes have been visited
} // end ScheduleLate
//------------------------------GlobalCodeMotion-------------------------------
void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) {
ResourceMark rm;
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("\n---- Start GlobalCodeMotion ----\n");
}
#endif
// Initialize the bbs.map for things on the proj_list
uint i;
for( i=0; i < proj_list.size(); i++ )
_bbs.map(proj_list[i]->_idx, NULL);
// Set the basic block for Nodes pinned into blocks
Arena *a = Thread::current()->resource_area();
VectorSet visited(a);
schedule_pinned_nodes( visited );
// Find the earliest Block any instruction can be placed in. Some
// instructions are pinned into Blocks. Unpinned instructions can
// appear in last block in which all their inputs occur.
visited.Clear();
Node_List stack(a);
stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list
if (!schedule_early(visited, stack)) {
// Bailout without retry
C->record_method_not_compilable("early schedule failed");
return;
}
// Build Def-Use edges.
proj_list.push(_root); // Add real root as another root
proj_list.pop();
// Compute the latency information (via backwards walk) for all the
// instructions in the graph
GrowableArray<uint> node_latency;
_node_latency = node_latency;
if( C->do_scheduling() )
ComputeLatenciesBackwards(visited, stack);
// Now schedule all codes as LATE as possible. This is the LCA in the
// dominator tree of all USES of a value. Pick the block with the least
// loop nesting depth that is lowest in the dominator tree.
// ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() )
schedule_late(visited, stack);
if( C->failing() ) {
// schedule_late fails only when graph is incorrect.
assert(!VerifyGraphEdges, "verification should have failed");
return;
}
unique = C->unique();
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("\n---- Detect implicit null checks ----\n");
}
#endif
// Detect implicit-null-check opportunities. Basically, find NULL checks
// with suitable memory ops nearby. Use the memory op to do the NULL check.
// I can generate a memory op if there is not one nearby.
if (C->is_method_compilation()) {
// Don't do it for natives, adapters, or runtime stubs
int allowed_reasons = 0;
// ...and don't do it when there have been too many traps, globally.
for (int reason = (int)Deoptimization::Reason_none+1;
reason < Compile::trapHistLength; reason++) {
assert(reason < BitsPerInt, "recode bit map");
if (!C->too_many_traps((Deoptimization::DeoptReason) reason))
allowed_reasons |= nth_bit(reason);
}
// By reversing the loop direction we get a very minor gain on mpegaudio.
// Feel free to revert to a forward loop for clarity.
// for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) {
for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) {
Node *proj = matcher._null_check_tests[i ];
Node *val = matcher._null_check_tests[i+1];
_bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons);
// The implicit_null_check will only perform the transformation
// if the null branch is truly uncommon, *and* it leads to an
// uncommon trap. Combined with the too_many_traps guards
// above, this prevents SEGV storms reported in 6366351,
// by recompiling offending methods without this optimization.
}
}
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("\n---- Start Local Scheduling ----\n");
}
#endif
// Schedule locally. Right now a simple topological sort.
// Later, do a real latency aware scheduler.
int *ready_cnt = NEW_RESOURCE_ARRAY(int,C->unique());
memset( ready_cnt, -1, C->unique() * sizeof(int) );
visited.Clear();
for (i = 0; i < _num_blocks; i++) {
if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) {
if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
C->record_method_not_compilable("local schedule failed");
}
return;
}
}
// If we inserted any instructions between a Call and his CatchNode,
// clone the instructions on all paths below the Catch.
for( i=0; i < _num_blocks; i++ )
_blocks[i]->call_catch_cleanup(_bbs);
#ifndef PRODUCT
if (trace_opto_pipelining()) {
tty->print("\n---- After GlobalCodeMotion ----\n");
for (uint i = 0; i < _num_blocks; i++) {
_blocks[i]->dump();
}
}
#endif
}
//------------------------------Estimate_Block_Frequency-----------------------
// Estimate block frequencies based on IfNode probabilities.
void PhaseCFG::Estimate_Block_Frequency() {
int cnts = C->method() ? C->method()->interpreter_invocation_count() : 1;
// Most of our algorithms will die horribly if frequency can become
// negative so make sure cnts is a sane value.
if( cnts <= 0 ) cnts = 1;
float f = (float)cnts/(float)FreqCountInvocations;
// Create the loop tree and calculate loop depth.
_root_loop = create_loop_tree();
_root_loop->compute_loop_depth(0);
// Compute block frequency of each block, relative to a single loop entry.
_root_loop->compute_freq();
// Adjust all frequencies to be relative to a single method entry
_root_loop->_freq = f * 1.0;
_root_loop->scale_freq();
// force paths ending at uncommon traps to be infrequent
Block_List worklist;
Block* root_blk = _blocks[0];
for (uint i = 0; i < root_blk->num_preds(); i++) {
Block *pb = _bbs[root_blk->pred(i)->_idx];
if (pb->has_uncommon_code()) {
worklist.push(pb);
}
}
while (worklist.size() > 0) {
Block* uct = worklist.pop();
uct->_freq = PROB_MIN;
for (uint i = 0; i < uct->num_preds(); i++) {
Block *pb = _bbs[uct->pred(i)->_idx];
if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) {
worklist.push(pb);
}
}
}
#ifndef PRODUCT
if (PrintCFGBlockFreq) {
tty->print_cr("CFG Block Frequencies");
_root_loop->dump_tree();
if (Verbose) {
tty->print_cr("PhaseCFG dump");
dump();
tty->print_cr("Node dump");
_root->dump(99999);
}
}
#endif
}
//----------------------------create_loop_tree--------------------------------
// Create a loop tree from the CFG
CFGLoop* PhaseCFG::create_loop_tree() {
#ifdef ASSERT
assert( _blocks[0] == _broot, "" );
for (uint i = 0; i < _num_blocks; i++ ) {
Block *b = _blocks[i];
// Check that _loop field are clear...we could clear them if not.
assert(b->_loop == NULL, "clear _loop expected");
// Sanity check that the RPO numbering is reflected in the _blocks array.
// It doesn't have to be for the loop tree to be built, but if it is not,
// then the blocks have been reordered since dom graph building...which
// may question the RPO numbering
assert(b->_rpo == i, "unexpected reverse post order number");
}
#endif
int idct = 0;
CFGLoop* root_loop = new CFGLoop(idct++);
Block_List worklist;
// Assign blocks to loops
for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block
Block *b = _blocks[i];
if (b->head()->is_Loop()) {
Block* loop_head = b;
assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
Node* tail_n = loop_head->pred(LoopNode::LoopBackControl);
Block* tail = _bbs[tail_n->_idx];
// Defensively filter out Loop nodes for non-single-entry loops.
// For all reasonable loops, the head occurs before the tail in RPO.
if (i <= tail->_rpo) {
// The tail and (recursive) predecessors of the tail
// are made members of a new loop.
assert(worklist.size() == 0, "nonempty worklist");
CFGLoop* nloop = new CFGLoop(idct++);
assert(loop_head->_loop == NULL, "just checking");
loop_head->_loop = nloop;
// Add to nloop so push_pred() will skip over inner loops
nloop->add_member(loop_head);
nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs);
while (worklist.size() > 0) {
Block* member = worklist.pop();
if (member != loop_head) {
for (uint j = 1; j < member->num_preds(); j++) {
nloop->push_pred(member, j, worklist, _bbs);
}
}
}
}
}
}
// Create a member list for each loop consisting
// of both blocks and (immediate child) loops.
for (uint i = 0; i < _num_blocks; i++) {
Block *b = _blocks[i];
CFGLoop* lp = b->_loop;
if (lp == NULL) {
// Not assigned to a loop. Add it to the method's pseudo loop.
b->_loop = root_loop;
lp = root_loop;
}
if (lp == root_loop || b != lp->head()) { // loop heads are already members
lp->add_member(b);
}
if (lp != root_loop) {
if (lp->parent() == NULL) {
// Not a nested loop. Make it a child of the method's pseudo loop.
root_loop->add_nested_loop(lp);
}
if (b == lp->head()) {
// Add nested loop to member list of parent loop.
lp->parent()->add_member(lp);
}
}
}
return root_loop;
}
//------------------------------push_pred--------------------------------------
void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) {
Node* pred_n = blk->pred(i);
Block* pred = node_to_blk[pred_n->_idx];
CFGLoop *pred_loop = pred->_loop;
if (pred_loop == NULL) {
// Filter out blocks for non-single-entry loops.
// For all reasonable loops, the head occurs before the tail in RPO.
if (pred->_rpo > head()->_rpo) {
pred->_loop = this;
worklist.push(pred);
}
} else if (pred_loop != this) {
// Nested loop.
while (pred_loop->_parent != NULL && pred_loop->_parent != this) {
pred_loop = pred_loop->_parent;
}
// Make pred's loop be a child
if (pred_loop->_parent == NULL) {
add_nested_loop(pred_loop);
// Continue with loop entry predecessor.
Block* pred_head = pred_loop->head();
assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors");
assert(pred_head != head(), "loop head in only one loop");
push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk);
} else {
assert(pred_loop->_parent == this && _parent == NULL, "just checking");
}
}
}
//------------------------------add_nested_loop--------------------------------
// Make cl a child of the current loop in the loop tree.
void CFGLoop::add_nested_loop(CFGLoop* cl) {
assert(_parent == NULL, "no parent yet");
assert(cl != this, "not my own parent");
cl->_parent = this;
CFGLoop* ch = _child;
if (ch == NULL) {
_child = cl;
} else {
while (ch->_sibling != NULL) { ch = ch->_sibling; }
ch->_sibling = cl;
}
}
//------------------------------compute_loop_depth-----------------------------
// Store the loop depth in each CFGLoop object.
// Recursively walk the children to do the same for them.
void CFGLoop::compute_loop_depth(int depth) {
_depth = depth;
CFGLoop* ch = _child;
while (ch != NULL) {
ch->compute_loop_depth(depth + 1);
ch = ch->_sibling;
}
}
//------------------------------compute_freq-----------------------------------
// Compute the frequency of each block and loop, relative to a single entry
// into the dominating loop head.
void CFGLoop::compute_freq() {
// Bottom up traversal of loop tree (visit inner loops first.)
// Set loop head frequency to 1.0, then transitively
// compute frequency for all successors in the loop,
// as well as for each exit edge. Inner loops are
// treated as single blocks with loop exit targets
// as the successor blocks.
// Nested loops first
CFGLoop* ch = _child;
while (ch != NULL) {
ch->compute_freq();
ch = ch->_sibling;
}
assert (_members.length() > 0, "no empty loops");
Block* hd = head();
hd->_freq = 1.0f;
for (int i = 0; i < _members.length(); i++) {
CFGElement* s = _members.at(i);
float freq = s->_freq;
if (s->is_block()) {
Block* b = s->as_Block();
for (uint j = 0; j < b->_num_succs; j++) {
Block* sb = b->_succs[j];
update_succ_freq(sb, freq * b->succ_prob(j));
}
} else {
CFGLoop* lp = s->as_CFGLoop();
assert(lp->_parent == this, "immediate child");
for (int k = 0; k < lp->_exits.length(); k++) {
Block* eb = lp->_exits.at(k).get_target();
float prob = lp->_exits.at(k).get_prob();
update_succ_freq(eb, freq * prob);
}
}
}
#if 0
// Raise frequency of the loop backedge block, in an effort
// to keep it empty. Skip the method level "loop".
if (_parent != NULL) {
CFGElement* s = _members.at(_members.length() - 1);
if (s->is_block()) {
Block* bk = s->as_Block();
if (bk->_num_succs == 1 && bk->_succs[0] == hd) {
// almost any value >= 1.0f works
// FIXME: raw constant
bk->_freq = 1.05f;
}
}
}
#endif
// For all loops other than the outer, "method" loop,
// sum and normalize the exit probability. The "method" loop
// should keep the initial exit probability of 1, so that
// inner blocks do not get erroneously scaled.
if (_depth != 0) {
// Total the exit probabilities for this loop.
float exits_sum = 0.0f;
for (int i = 0; i < _exits.length(); i++) {
exits_sum += _exits.at(i).get_prob();
}
// Normalize the exit probabilities. Until now, the
// probabilities estimate the possibility of exit per
// a single loop iteration; afterward, they estimate
// the probability of exit per loop entry.
for (int i = 0; i < _exits.length(); i++) {
Block* et = _exits.at(i).get_target();
float new_prob = _exits.at(i).get_prob() / exits_sum;
BlockProbPair bpp(et, new_prob);
_exits.at_put(i, bpp);
}
// Save the total, but guard against unreasoable probability,
// as the value is used to estimate the loop trip count.
// An infinite trip count would blur relative block
// frequencies.
if (exits_sum > 1.0f) exits_sum = 1.0;
if (exits_sum < PROB_MIN) exits_sum = PROB_MIN;
_exit_prob = exits_sum;
}
}
//------------------------------succ_prob-------------------------------------
// Determine the probability of reaching successor 'i' from the receiver block.
float Block::succ_prob(uint i) {
int eidx = end_idx();
Node *n = _nodes[eidx]; // Get ending Node
int op = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : n->Opcode();
// Switch on branch type
switch( op ) {
case Op_CountedLoopEnd:
case Op_If: {
assert (i < 2, "just checking");
// Conditionals pass on only part of their frequency
float prob = n->as_MachIf()->_prob;
assert(prob >= 0.0 && prob <= 1.0, "out of range probability");
// If succ[i] is the FALSE branch, invert path info
if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) {
return 1.0f - prob; // not taken
} else {
return prob; // taken
}
}
case Op_Jump:
// Divide the frequency between all successors evenly
return 1.0f/_num_succs;
case Op_Catch: {
const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj();
if (ci->_con == CatchProjNode::fall_through_index) {
// Fall-thru path gets the lion's share.
return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs;
} else {
// Presume exceptional paths are equally unlikely
return PROB_UNLIKELY_MAG(5);
}
}
case Op_Root:
case Op_Goto:
// Pass frequency straight thru to target
return 1.0f;
case Op_NeverBranch:
return 0.0f;
case Op_TailCall:
case Op_TailJump:
case Op_Return:
case Op_Halt:
case Op_Rethrow:
// Do not push out freq to root block
return 0.0f;
default:
ShouldNotReachHere();
}
return 0.0f;
}
//------------------------------update_succ_freq-------------------------------
// Update the appropriate frequency associated with block 'b', a succesor of
// a block in this loop.
void CFGLoop::update_succ_freq(Block* b, float freq) {
if (b->_loop == this) {
if (b == head()) {
// back branch within the loop
// Do nothing now, the loop carried frequency will be
// adjust later in scale_freq().
} else {
// simple branch within the loop
b->_freq += freq;
}
} else if (!in_loop_nest(b)) {
// branch is exit from this loop
BlockProbPair bpp(b, freq);
_exits.append(bpp);
} else {
// branch into nested loop
CFGLoop* ch = b->_loop;
ch->_freq += freq;
}
}
//------------------------------in_loop_nest-----------------------------------
// Determine if block b is in the receiver's loop nest.
bool CFGLoop::in_loop_nest(Block* b) {
int depth = _depth;
CFGLoop* b_loop = b->_loop;
int b_depth = b_loop->_depth;
if (depth == b_depth) {
return true;
}
while (b_depth > depth) {
b_loop = b_loop->_parent;
b_depth = b_loop->_depth;
}
return b_loop == this;
}
//------------------------------scale_freq-------------------------------------
// Scale frequency of loops and blocks by trip counts from outer loops
// Do a top down traversal of loop tree (visit outer loops first.)
void CFGLoop::scale_freq() {
float loop_freq = _freq * trip_count();
for (int i = 0; i < _members.length(); i++) {
CFGElement* s = _members.at(i);
s->_freq *= loop_freq;
}
CFGLoop* ch = _child;
while (ch != NULL) {
ch->scale_freq();
ch = ch->_sibling;
}
}
#ifndef PRODUCT
//------------------------------dump_tree--------------------------------------
void CFGLoop::dump_tree() const {
dump();
if (_child != NULL) _child->dump_tree();
if (_sibling != NULL) _sibling->dump_tree();
}
//------------------------------dump-------------------------------------------
void CFGLoop::dump() const {
for (int i = 0; i < _depth; i++) tty->print(" ");
tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n",
_depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq);
for (int i = 0; i < _depth; i++) tty->print(" ");
tty->print(" members:", _id);
int k = 0;
for (int i = 0; i < _members.length(); i++) {
if (k++ >= 6) {
tty->print("\n ");
for (int j = 0; j < _depth+1; j++) tty->print(" ");
k = 0;
}
CFGElement *s = _members.at(i);
if (s->is_block()) {
Block *b = s->as_Block();
tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq);
} else {
CFGLoop* lp = s->as_CFGLoop();
tty->print(" L%d(%6.3f)", lp->_id, lp->_freq);
}
}
tty->print("\n");
for (int i = 0; i < _depth; i++) tty->print(" ");
tty->print(" exits: ");
k = 0;
for (int i = 0; i < _exits.length(); i++) {
if (k++ >= 7) {
tty->print("\n ");
for (int j = 0; j < _depth+1; j++) tty->print(" ");
k = 0;
}
Block *blk = _exits.at(i).get_target();
float prob = _exits.at(i).get_prob();
tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100));
}
tty->print("\n");
}
#endif