/* * Copyright (c) 1997, 2022, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. *
*/
// A data structure that holds all the information needed to find dominators. struct Tarjan {
Block *_block; // Basic block for this info
uint _semi; // Semi-dominators
uint _size; // Used for faster LINK and EVAL
Tarjan *_parent; // Parent in DFS
Tarjan *_label; // Used for LINK and EVAL
Tarjan *_ancestor; // Used for LINK and EVAL
Tarjan *_child; // Used for faster LINK and EVAL
Tarjan *_dom; // Parent in dominator tree (immediate dom)
Tarjan *_bucket; // Set of vertices with given semidominator
Tarjan *_dom_child; // Child in dominator tree
Tarjan *_dom_next; // Next in dominator tree
// Fast union-find work void COMPRESS();
Tarjan *EVAL(void); void LINK( Tarjan *w, Tarjan *tarjan0 );
void setdepth( uint size );
};
// Compute the dominator tree of the CFG. The CFG must already have been // constructed. This is the Lengauer & Tarjan O(E-alpha(E,V)) algorithm. void PhaseCFG::build_dominator_tree() { // Pre-grow the blocks array, prior to the ResourceMark kicking in
_blocks.map(number_of_blocks(), 0);
ResourceMark rm; // Setup mappings from my Graph to Tarjan's stuff and back // Note: Tarjan uses 1-based arrays
Tarjan* tarjan = NEW_RESOURCE_ARRAY(Tarjan, number_of_blocks() + 1);
// Tarjan's algorithm, almost verbatim: // Step 1:
uint dfsnum = do_DFS(tarjan, number_of_blocks()); if (dfsnum - 1 != number_of_blocks()) { // Check for unreachable loops! // If the returned dfsnum does not match the number of blocks, then we // must have some unreachable loops. These can be made at any time by // IterGVN. They are cleaned up by CCP or the loop opts, but the last // IterGVN can always make more that are not cleaned up. Highly unlikely // except in ZKM.jar, where endless irreducible loops cause the loop opts // to not get run. // // Having found unreachable loops, we have made a bad RPO _block layout. // We can re-run the above DFS pass with the correct number of blocks, // and hack the Tarjan algorithm below to be robust in the presence of // such dead loops (as was done for the NTarjan code farther below). // Since this situation is so unlikely, instead I've decided to bail out. // CNC 7/24/2001
C->record_method_not_compilable("unreachable loop"); return;
}
_blocks._cnt = number_of_blocks();
// Tarjan is using 1-based arrays, so these are some initialize flags
tarjan[0]._size = tarjan[0]._semi = 0;
tarjan[0]._label = &tarjan[0];
for (uint i = number_of_blocks(); i >= 2; i--) { // For all vertices in DFS order
Tarjan *w = &tarjan[i]; // Get vertex from DFS
// w is added to a bucket here, and only here. // Thus w is in at most one bucket and the sum of all bucket sizes is O(n). // Thus bucket can be a linked list. // Thus we do not need a small integer name for each Block.
w->_bucket = tarjan[w->_semi]._bucket;
tarjan[w->_semi]._bucket = w;
// Step 4: for (uint i = 2; i <= number_of_blocks(); i++) {
Tarjan *w = &tarjan[i]; if( w->_dom != &tarjan[w->_semi] )
w->_dom = w->_dom->_dom;
w->_dom_next = w->_dom_child = NULL; // Initialize for building tree later
} // No immediate dominator for the root
Tarjan *w = &tarjan[get_root_block()->_pre_order];
w->_dom = NULL;
w->_dom_next = w->_dom_child = NULL; // Initialize for building tree later
// Convert the dominator tree array into my kind of graph for(uint i = 1; i <= number_of_blocks(); i++){ // For all Tarjan vertices
Tarjan *t = &tarjan[i]; // Handy access
Tarjan *tdom = t->_dom; // Handy access to immediate dominator if( tdom ) { // Root has no immediate dominator
t->_block->_idom = tdom->_block; // Set immediate dominator
t->_dom_next = tdom->_dom_child; // Make me a sibling of parent's child
tdom->_dom_child = t; // Make me a child of my parent
} else
t->_block->_idom = NULL; // Root
}
w->setdepth(number_of_blocks() + 1); // Set depth in dominator tree
}
class Block_Stack { private: struct Block_Descr {
Block *block; // Block int index; // Index of block's successor pushed on stack int freq_idx; // Index of block's most frequent successor
};
Block_Descr *_stack_top;
Block_Descr *_stack_max;
Block_Descr *_stack;
Tarjan *_tarjan;
uint most_frequent_successor( Block *b ); public:
Block_Stack(Tarjan *tarjan, int size) : _tarjan(tarjan) {
_stack = NEW_RESOURCE_ARRAY(Block_Descr, size);
_stack_max = _stack + size;
_stack_top = _stack - 1; // stack is empty
} void push(uint pre_order, Block *b) {
Tarjan *t = &_tarjan[pre_order]; // Fast local access
b->_pre_order = pre_order; // Flag as visited
t->_block = b; // Save actual block
t->_semi = pre_order; // Block to DFS map
t->_label = t; // DFS to vertex map
t->_ancestor = NULL; // Fast LINK & EVAL setup
t->_child = &_tarjan[0]; // Sentenial
t->_size = 1;
t->_bucket = NULL; if (pre_order == 1)
t->_parent = NULL; // first block doesn't have parent else { // Save parent (current top block on stack) in DFS
t->_parent = &_tarjan[_stack_top->block->_pre_order];
} // Now put this block on stack
++_stack_top;
assert(_stack_top < _stack_max, ""); // assert if stack have to grow
_stack_top->block = b;
_stack_top->index = -1; // Find the index into b->succs[] array of the most frequent successor.
_stack_top->freq_idx = most_frequent_successor(b); // freq_idx >= 0
}
Block* pop() { Block* b = _stack_top->block; _stack_top--; return b; } bool is_nonempty() { return (_stack_top >= _stack); } bool last_successor() { return (_stack_top->index == _stack_top->freq_idx); }
Block* next_successor() { int i = _stack_top->index;
i++; if (i == _stack_top->freq_idx) i++; if (i >= (int)(_stack_top->block->_num_succs)) {
i = _stack_top->freq_idx; // process most frequent successor last
}
_stack_top->index = i; return _stack_top->block->_succs[ i ];
}
};
// Find the index into the b->succs[] array of the most frequent successor.
uint Block_Stack::most_frequent_successor( Block *b ) {
uint freq_idx = 0; int eidx = b->end_idx();
Node *n = b->get_node(eidx); int op = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : n->Opcode(); switch( op ) { case Op_CountedLoopEnd: case Op_If: { // Split frequency amongst children float prob = n->as_MachIf()->_prob; // Is succ[0] the TRUE branch or the FALSE branch? if( b->get_node(eidx+1)->Opcode() == Op_IfFalse )
prob = 1.0f - prob;
freq_idx = prob < PROB_FAIR; // freq=1 for succ[0] < 0.5 prob break;
} case Op_Catch: // Split frequency amongst children for( freq_idx = 0; freq_idx < b->_num_succs; freq_idx++ ) if( b->get_node(eidx+1+freq_idx)->as_CatchProj()->_con == CatchProjNode::fall_through_index ) break; // Handle case of no fall-thru (e.g., check-cast MUST throw an exception) if( freq_idx == b->_num_succs ) freq_idx = 0; break; // Currently there is no support for finding out the most // frequent successor for jumps, so lets just make it the first one case Op_Jump: case Op_Root: case Op_Goto: case Op_NeverBranch:
freq_idx = 0; // fall thru break; case Op_TailCall: case Op_TailJump: case Op_Return: case Op_Halt: case Op_Rethrow: break; default:
ShouldNotReachHere();
} return freq_idx;
}
// Perform DFS search. Setup 'vertex' as DFS to vertex mapping. Setup // 'semi' as vertex to DFS mapping. Set 'parent' to DFS parent.
uint PhaseCFG::do_DFS(Tarjan *tarjan, uint rpo_counter) {
Block* root_block = get_root_block();
uint pre_order = 1; // Allocate stack of size number_of_blocks() + 1 to avoid frequent realloc
Block_Stack bstack(tarjan, number_of_blocks() + 1);
// Push on stack the state for the first block
bstack.push(pre_order, root_block);
++pre_order;
while (bstack.is_nonempty()) { if (!bstack.last_successor()) { // Walk over all successors in pre-order (DFS).
Block* next_block = bstack.next_successor(); if (next_block->_pre_order == 0) { // Check for no-pre-order, not-visited // Push on stack the state of successor
bstack.push(pre_order, next_block);
++pre_order;
}
} else { // Build a reverse post-order in the CFG _blocks array
Block *stack_top = bstack.pop();
stack_top->_rpo = --rpo_counter;
_blocks.map(stack_top->_rpo, stack_top);
}
} return pre_order;
}
void Tarjan::setdepth( uint stack_size ) {
Tarjan **top = NEW_RESOURCE_ARRAY(Tarjan*, stack_size);
Tarjan **next = top;
Tarjan **last;
uint depth = 0;
*top = this;
++top; do { // next level
++depth;
last = top; do { // Set current depth for all tarjans on this level
Tarjan *t = *next; // next tarjan from stack
++next; do {
t->_block->_dom_depth = depth; // Set depth in dominator tree
Tarjan *dom_child = t->_dom_child;
t = t->_dom_next; // next tarjan if (dom_child != NULL) {
*top = dom_child; // save child on stack
++top;
}
} while (t != NULL);
} while (next < last);
} while (last < top);
}
// Compute dominators on the Sea of Nodes form // A data structure that holds all the information needed to find dominators. struct NTarjan {
Node *_control; // Control node associated with this info
uint _semi; // Semi-dominators
uint _size; // Used for faster LINK and EVAL
NTarjan *_parent; // Parent in DFS
NTarjan *_label; // Used for LINK and EVAL
NTarjan *_ancestor; // Used for LINK and EVAL
NTarjan *_child; // Used for faster LINK and EVAL
NTarjan *_dom; // Parent in dominator tree (immediate dom)
NTarjan *_bucket; // Set of vertices with given semidominator
NTarjan *_dom_child; // Child in dominator tree
NTarjan *_dom_next; // Next in dominator tree
// Perform DFS search. // Setup 'vertex' as DFS to vertex mapping. // Setup 'semi' as vertex to DFS mapping. // Set 'parent' to DFS parent. staticint DFS( NTarjan *ntarjan, VectorSet &visited, PhaseIdealLoop *pil, uint *dfsorder ); void setdepth( uint size, uint *dom_depth );
// Compute the dominator tree of the sea of nodes. This version walks all CFG // nodes (using the is_CFG() call) and places them in a dominator tree. Thus, // it needs a count of the CFG nodes for the mapping table. This is the // Lengauer & Tarjan O(E-alpha(E,V)) algorithm. void PhaseIdealLoop::Dominators() {
ResourceMark rm; // Setup mappings from my Graph to Tarjan's stuff and back // Note: Tarjan uses 1-based arrays
NTarjan *ntarjan = NEW_RESOURCE_ARRAY(NTarjan,C->unique()+1); // Initialize _control field for fast reference int i; for( i= C->unique()-1; i>=0; i-- )
ntarjan[i]._control = NULL;
// Store the DFS order for the main loop const uint fill_value = max_juint;
uint *dfsorder = NEW_RESOURCE_ARRAY(uint,C->unique()+1);
memset(dfsorder, fill_value, (C->unique()+1) * sizeof(uint));
// Tarjan's algorithm, almost verbatim: // Step 1:
VectorSet visited; int dfsnum = NTarjan::DFS( ntarjan, visited, this, dfsorder);
// Tarjan is using 1-based arrays, so these are some initialize flags
ntarjan[0]._size = ntarjan[0]._semi = 0;
ntarjan[0]._label = &ntarjan[0];
for( i = dfsnum-1; i>1; i-- ) { // For all nodes in reverse DFS order
NTarjan *w = &ntarjan[i]; // Get Node from DFS
assert(w->_control != NULL,"bad DFS walk");
// Step 2:
Node *whead = w->_control; for( uint j=0; j < whead->req(); j++ ) { // For each predecessor if( whead->in(j) == NULL || !whead->in(j)->is_CFG() ) continue; // Only process control nodes
uint b = dfsorder[whead->in(j)->_idx]; if(b == fill_value) continue;
NTarjan *vx = &ntarjan[b];
NTarjan *u = vx->EVAL(); if( u->_semi < w->_semi )
w->_semi = u->_semi;
}
// w is added to a bucket here, and only here. // Thus w is in at most one bucket and the sum of all bucket sizes is O(n). // Thus bucket can be a linked list.
w->_bucket = ntarjan[w->_semi]._bucket;
ntarjan[w->_semi]._bucket = w;
// Cleanup any unreachable loops now. Unreachable loops are loops that // flow into the main graph (and hence into ROOT) but are not reachable // from above. Such code is dead, but requires a global pass to detect // it; this global pass was the 'build_loop_tree' pass run just prior. if( !_verify_only && whead->is_Region() ) { for( uint i = 1; i < whead->req(); i++ ) { if (!has_node(whead->in(i))) { // Kill dead input path
assert( !visited.test(whead->in(i)->_idx), "input with no loop must be dead" );
_igvn.delete_input_of(whead, i); for (DUIterator_Fast jmax, j = whead->fast_outs(jmax); j < jmax; j++) {
Node* p = whead->fast_out(j); if( p->is_Phi() ) {
_igvn.delete_input_of(p, i);
}
}
i--; // Rerun same iteration
} // End of if dead input path
} // End of for all input paths
} // End if if whead is a Region
} // End of for all Nodes in reverse DFS order
// Step 4: for( i=2; i < dfsnum; i++ ) { // DFS order
NTarjan *w = &ntarjan[i];
assert(w->_control != NULL,"Bad DFS walk"); if( w->_dom != &ntarjan[w->_semi] )
w->_dom = w->_dom->_dom;
w->_dom_next = w->_dom_child = NULL; // Initialize for building tree later
} // No immediate dominator for the root
NTarjan *w = &ntarjan[dfsorder[C->root()->_idx]];
w->_dom = NULL;
w->_parent = NULL;
w->_dom_next = w->_dom_child = NULL; // Initialize for building tree later
// Convert the dominator tree array into my kind of graph for( i=1; i<dfsnum; i++ ) { // For all Tarjan vertices
NTarjan *t = &ntarjan[i]; // Handy access
assert(t->_control != NULL,"Bad DFS walk");
NTarjan *tdom = t->_dom; // Handy access to immediate dominator if( tdom ) { // Root has no immediate dominator
_idom[t->_control->_idx] = tdom->_control; // Set immediate dominator
t->_dom_next = tdom->_dom_child; // Make me a sibling of parent's child
tdom->_dom_child = t; // Make me a child of my parent
} else
_idom[C->root()->_idx] = NULL; // Root
}
w->setdepth( C->unique()+1, _dom_depth ); // Set depth in dominator tree // Pick up the 'top' node as well
_idom [C->top()->_idx] = C->root();
_dom_depth[C->top()->_idx] = 1;
// Debug Print of Dominator tree if( PrintDominators ) { #ifndef PRODUCT
w->dump(0); #endif
}
}
// Perform DFS search. Setup 'vertex' as DFS to vertex mapping. Setup // 'semi' as vertex to DFS mapping. Set 'parent' to DFS parent. int NTarjan::DFS( NTarjan *ntarjan, VectorSet &visited, PhaseIdealLoop *pil, uint *dfsorder) { // Allocate stack of size C->live_nodes()/8 to avoid frequent realloc
GrowableArray <Node *> dfstack(pil->C->live_nodes() >> 3);
Node *b = pil->C->root(); int dfsnum = 1;
dfsorder[b->_idx] = dfsnum; // Cache parent's dfsnum for a later use
dfstack.push(b);
while (dfstack.is_nonempty()) {
b = dfstack.pop(); if( !visited.test_set(b->_idx) ) { // Test node and flag it as visited
NTarjan *w = &ntarjan[dfsnum]; // Only fully process control nodes
w->_control = b; // Save actual node // Use parent's cached dfsnum to identify "Parent in DFS"
w->_parent = &ntarjan[dfsorder[b->_idx]];
dfsorder[b->_idx] = dfsnum; // Save DFS order info
w->_semi = dfsnum; // Node to DFS map
w->_label = w; // DFS to vertex map
w->_ancestor = NULL; // Fast LINK & EVAL setup
w->_child = &ntarjan[0]; // Sentinel
w->_size = 1;
w->_bucket = NULL;
// Need DEF-USE info for this pass for ( int i = b->outcnt(); i-- > 0; ) { // Put on stack backwards
Node* s = b->raw_out(i); // Get a use // CFG nodes only and not dead stuff if( s->is_CFG() && pil->has_node(s) && !visited.test(s->_idx) ) {
dfsorder[s->_idx] = dfsnum; // Cache parent's dfsnum for a later use
dfstack.push(s);
}
}
dfsnum++; // update after parent's dfsnum has been cached.
}
}
void NTarjan::setdepth( uint stack_size, uint *dom_depth ) {
NTarjan **top = NEW_RESOURCE_ARRAY(NTarjan*, stack_size);
NTarjan **next = top;
NTarjan **last;
uint depth = 0;
*top = this;
++top; do { // next level
++depth;
last = top; do { // Set current depth for all tarjans on this level
NTarjan *t = *next; // next tarjan from stack
++next; do {
dom_depth[t->_control->_idx] = depth; // Set depth in dominator tree
NTarjan *dom_child = t->_dom_child;
t = t->_dom_next; // next tarjan if (dom_child != NULL) {
*top = dom_child; // save child on stack
++top;
}
} while (t != NULL);
} while (next < last);
} while (last < top);
}
#ifndef PRODUCT void NTarjan::dump(int offset) const { // Dump the data from this node int i; for(i = offset; i >0; i--) // Use indenting for tree structure
tty->print(" ");
tty->print("Dominator Node: ");
_control->dump(); // Control node for this dom node
tty->print("\n"); for(i = offset; i >0; i--) // Use indenting for tree structure
tty->print(" ");
tty->print("semi:%d, size:%d\n",_semi, _size); for(i = offset; i >0; i--) // Use indenting for tree structure
tty->print(" ");
tty->print("DFS Parent: "); if(_parent != NULL)
_parent->_control->dump(); // Parent in DFS
tty->print("\n"); for(i = offset; i >0; i--) // Use indenting for tree structure
tty->print(" ");
tty->print("Dom Parent: "); if(_dom != NULL)
_dom->_control->dump(); // Parent in Dominator Tree
tty->print("\n");
// Recurse over remaining tree if( _dom_child ) _dom_child->dump(offset+2); // Children in dominator tree if( _dom_next ) _dom_next ->dump(offset ); // Siblings in dominator tree
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.