8134883: C1 hard crash in range check elimination in Nashorn test262parallel
Summary: C1's range check elimination breaks with a non-natural loop that has an exception handler as one entry
Reviewed-by: iveresov
/*
* Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved.
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
* This code is free software; you can redistribute it and/or modify it
* under the terms of the GNU General Public License version 2 only, as
* published by the Free Software Foundation.
*
* This code is distributed in the hope that it will be useful, but WITHOUT
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
* version 2 for more details (a copy is included in the LICENSE file that
* accompanied this code).
*
* You should have received a copy of the GNU General Public License version
* 2 along with this work; if not, write to the Free Software Foundation,
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
*
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
* or visit www.oracle.com if you need additional information or have any
* questions.
*
*/
#include "precompiled.hpp"
#include "c1/c1_Compilation.hpp"
#include "c1/c1_FrameMap.hpp"
#include "c1/c1_GraphBuilder.hpp"
#include "c1/c1_IR.hpp"
#include "c1/c1_InstructionPrinter.hpp"
#include "c1/c1_Optimizer.hpp"
#include "utilities/bitMap.inline.hpp"
// Implementation of XHandlers
//
// Note: This code could eventually go away if we are
// just using the ciExceptionHandlerStream.
XHandlers::XHandlers(ciMethod* method) : _list(method->exception_table_length()) {
ciExceptionHandlerStream s(method);
while (!s.is_done()) {
_list.append(new XHandler(s.handler()));
s.next();
}
assert(s.count() == method->exception_table_length(), "exception table lengths inconsistent");
}
// deep copy of all XHandler contained in list
XHandlers::XHandlers(XHandlers* other) :
_list(other->length())
{
for (int i = 0; i < other->length(); i++) {
_list.append(new XHandler(other->handler_at(i)));
}
}
// Returns whether a particular exception type can be caught. Also
// returns true if klass is unloaded or any exception handler
// classes are unloaded. type_is_exact indicates whether the throw
// is known to be exactly that class or it might throw a subtype.
bool XHandlers::could_catch(ciInstanceKlass* klass, bool type_is_exact) const {
// the type is unknown so be conservative
if (!klass->is_loaded()) {
return true;
}
for (int i = 0; i < length(); i++) {
XHandler* handler = handler_at(i);
if (handler->is_catch_all()) {
// catch of ANY
return true;
}
ciInstanceKlass* handler_klass = handler->catch_klass();
// if it's unknown it might be catchable
if (!handler_klass->is_loaded()) {
return true;
}
// if the throw type is definitely a subtype of the catch type
// then it can be caught.
if (klass->is_subtype_of(handler_klass)) {
return true;
}
if (!type_is_exact) {
// If the type isn't exactly known then it can also be caught by
// catch statements where the inexact type is a subtype of the
// catch type.
// given: foo extends bar extends Exception
// throw bar can be caught by catch foo, catch bar, and catch
// Exception, however it can't be caught by any handlers without
// bar in its type hierarchy.
if (handler_klass->is_subtype_of(klass)) {
return true;
}
}
}
return false;
}
bool XHandlers::equals(XHandlers* others) const {
if (others == NULL) return false;
if (length() != others->length()) return false;
for (int i = 0; i < length(); i++) {
if (!handler_at(i)->equals(others->handler_at(i))) return false;
}
return true;
}
bool XHandler::equals(XHandler* other) const {
assert(entry_pco() != -1 && other->entry_pco() != -1, "must have entry_pco");
if (entry_pco() != other->entry_pco()) return false;
if (scope_count() != other->scope_count()) return false;
if (_desc != other->_desc) return false;
assert(entry_block() == other->entry_block(), "entry_block must be equal when entry_pco is equal");
return true;
}
// Implementation of IRScope
BlockBegin* IRScope::build_graph(Compilation* compilation, int osr_bci) {
GraphBuilder gm(compilation, this);
NOT_PRODUCT(if (PrintValueNumbering && Verbose) gm.print_stats());
if (compilation->bailed_out()) return NULL;
return gm.start();
}
IRScope::IRScope(Compilation* compilation, IRScope* caller, int caller_bci, ciMethod* method, int osr_bci, bool create_graph)
: _callees(2)
, _compilation(compilation)
, _requires_phi_function(method->max_locals())
{
_caller = caller;
_level = caller == NULL ? 0 : caller->level() + 1;
_method = method;
_xhandlers = new XHandlers(method);
_number_of_locks = 0;
_monitor_pairing_ok = method->has_balanced_monitors();
_wrote_final = false;
_wrote_fields = false;
_wrote_volatile = false;
_start = NULL;
if (osr_bci == -1) {
_requires_phi_function.clear();
} else {
// selective creation of phi functions is not possibel in osr-methods
_requires_phi_function.set_range(0, method->max_locals());
}
assert(method->holder()->is_loaded() , "method holder must be loaded");
// build graph if monitor pairing is ok
if (create_graph && monitor_pairing_ok()) _start = build_graph(compilation, osr_bci);
}
int IRScope::max_stack() const {
int my_max = method()->max_stack();
int callee_max = 0;
for (int i = 0; i < number_of_callees(); i++) {
callee_max = MAX2(callee_max, callee_no(i)->max_stack());
}
return my_max + callee_max;
}
bool IRScopeDebugInfo::should_reexecute() {
ciMethod* cur_method = scope()->method();
int cur_bci = bci();
if (cur_method != NULL && cur_bci != SynchronizationEntryBCI) {
Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
return Interpreter::bytecode_should_reexecute(code);
} else
return false;
}
// Implementation of CodeEmitInfo
// Stack must be NON-null
CodeEmitInfo::CodeEmitInfo(ValueStack* stack, XHandlers* exception_handlers, bool deoptimize_on_exception)
: _scope(stack->scope())
, _scope_debug_info(NULL)
, _oop_map(NULL)
, _stack(stack)
, _exception_handlers(exception_handlers)
, _is_method_handle_invoke(false)
, _deoptimize_on_exception(deoptimize_on_exception) {
assert(_stack != NULL, "must be non null");
}
CodeEmitInfo::CodeEmitInfo(CodeEmitInfo* info, ValueStack* stack)
: _scope(info->_scope)
, _exception_handlers(NULL)
, _scope_debug_info(NULL)
, _oop_map(NULL)
, _stack(stack == NULL ? info->_stack : stack)
, _is_method_handle_invoke(info->_is_method_handle_invoke)
, _deoptimize_on_exception(info->_deoptimize_on_exception) {
// deep copy of exception handlers
if (info->_exception_handlers != NULL) {
_exception_handlers = new XHandlers(info->_exception_handlers);
}
}
void CodeEmitInfo::record_debug_info(DebugInformationRecorder* recorder, int pc_offset) {
// record the safepoint before recording the debug info for enclosing scopes
recorder->add_safepoint(pc_offset, _oop_map->deep_copy());
_scope_debug_info->record_debug_info(recorder, pc_offset, true/*topmost*/, _is_method_handle_invoke);
recorder->end_safepoint(pc_offset);
}
void CodeEmitInfo::add_register_oop(LIR_Opr opr) {
assert(_oop_map != NULL, "oop map must already exist");
assert(opr->is_single_cpu(), "should not call otherwise");
VMReg name = frame_map()->regname(opr);
_oop_map->set_oop(name);
}
// Mirror the stack size calculation in the deopt code
// How much stack space would we need at this point in the program in
// case of deoptimization?
int CodeEmitInfo::interpreter_frame_size() const {
ValueStack* state = _stack;
int size = 0;
int callee_parameters = 0;
int callee_locals = 0;
int extra_args = state->scope()->method()->max_stack() - state->stack_size();
while (state != NULL) {
int locks = state->locks_size();
int temps = state->stack_size();
bool is_top_frame = (state == _stack);
ciMethod* method = state->scope()->method();
int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
temps + callee_parameters,
extra_args,
locks,
callee_parameters,
callee_locals,
is_top_frame);
size += frame_size;
callee_parameters = method->size_of_parameters();
callee_locals = method->max_locals();
extra_args = 0;
state = state->caller_state();
}
return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
}
// Implementation of IR
IR::IR(Compilation* compilation, ciMethod* method, int osr_bci) :
_num_loops(0) {
// setup IR fields
_compilation = compilation;
_top_scope = new IRScope(compilation, NULL, -1, method, osr_bci, true);
_code = NULL;
}
void IR::optimize_blocks() {
Optimizer opt(this);
if (!compilation()->profile_branches()) {
if (DoCEE) {
opt.eliminate_conditional_expressions();
#ifndef PRODUCT
if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after CEE"); print(true); }
if (PrintIR || PrintIR1 ) { tty->print_cr("IR after CEE"); print(false); }
#endif
}
if (EliminateBlocks) {
opt.eliminate_blocks();
#ifndef PRODUCT
if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after block elimination"); print(true); }
if (PrintIR || PrintIR1 ) { tty->print_cr("IR after block elimination"); print(false); }
#endif
}
}
}
void IR::eliminate_null_checks() {
Optimizer opt(this);
if (EliminateNullChecks) {
opt.eliminate_null_checks();
#ifndef PRODUCT
if (PrintCFG || PrintCFG1) { tty->print_cr("CFG after null check elimination"); print(true); }
if (PrintIR || PrintIR1 ) { tty->print_cr("IR after null check elimination"); print(false); }
#endif
}
}
static int sort_pairs(BlockPair** a, BlockPair** b) {
if ((*a)->from() == (*b)->from()) {
return (*a)->to()->block_id() - (*b)->to()->block_id();
} else {
return (*a)->from()->block_id() - (*b)->from()->block_id();
}
}
class CriticalEdgeFinder: public BlockClosure {
BlockPairList blocks;
IR* _ir;
public:
CriticalEdgeFinder(IR* ir): _ir(ir) {}
void block_do(BlockBegin* bb) {
BlockEnd* be = bb->end();
int nos = be->number_of_sux();
if (nos >= 2) {
for (int i = 0; i < nos; i++) {
BlockBegin* sux = be->sux_at(i);
if (sux->number_of_preds() >= 2) {
blocks.append(new BlockPair(bb, sux));
}
}
}
}
void split_edges() {
BlockPair* last_pair = NULL;
blocks.sort(sort_pairs);
for (int i = 0; i < blocks.length(); i++) {
BlockPair* pair = blocks.at(i);
if (last_pair != NULL && pair->is_same(last_pair)) continue;
BlockBegin* from = pair->from();
BlockBegin* to = pair->to();
BlockBegin* split = from->insert_block_between(to);
#ifndef PRODUCT
if ((PrintIR || PrintIR1) && Verbose) {
tty->print_cr("Split critical edge B%d -> B%d (new block B%d)",
from->block_id(), to->block_id(), split->block_id());
}
#endif
last_pair = pair;
}
}
};
void IR::split_critical_edges() {
CriticalEdgeFinder cef(this);
iterate_preorder(&cef);
cef.split_edges();
}
class UseCountComputer: public ValueVisitor, BlockClosure {
private:
void visit(Value* n) {
// Local instructions and Phis for expression stack values at the
// start of basic blocks are not added to the instruction list
if (!(*n)->is_linked() && (*n)->can_be_linked()) {
assert(false, "a node was not appended to the graph");
Compilation::current()->bailout("a node was not appended to the graph");
}
// use n's input if not visited before
if (!(*n)->is_pinned() && !(*n)->has_uses()) {
// note: a) if the instruction is pinned, it will be handled by compute_use_count
// b) if the instruction has uses, it was touched before
// => in both cases we don't need to update n's values
uses_do(n);
}
// use n
(*n)->_use_count++;
}
Values* worklist;
int depth;
enum {
max_recurse_depth = 20
};
void uses_do(Value* n) {
depth++;
if (depth > max_recurse_depth) {
// don't allow the traversal to recurse too deeply
worklist->push(*n);
} else {
(*n)->input_values_do(this);
// special handling for some instructions
if ((*n)->as_BlockEnd() != NULL) {
// note on BlockEnd:
// must 'use' the stack only if the method doesn't
// terminate, however, in those cases stack is empty
(*n)->state_values_do(this);
}
}
depth--;
}
void block_do(BlockBegin* b) {
depth = 0;
// process all pinned nodes as the roots of expression trees
for (Instruction* n = b; n != NULL; n = n->next()) {
if (n->is_pinned()) uses_do(&n);
}
assert(depth == 0, "should have counted back down");
// now process any unpinned nodes which recursed too deeply
while (worklist->length() > 0) {
Value t = worklist->pop();
if (!t->is_pinned()) {
// compute the use count
uses_do(&t);
// pin the instruction so that LIRGenerator doesn't recurse
// too deeply during it's evaluation.
t->pin();
}
}
assert(depth == 0, "should have counted back down");
}
UseCountComputer() {
worklist = new Values();
depth = 0;
}
public:
static void compute(BlockList* blocks) {
UseCountComputer ucc;
blocks->iterate_backward(&ucc);
}
};
// helper macro for short definition of trace-output inside code
#ifndef PRODUCT
#define TRACE_LINEAR_SCAN(level, code) \
if (TraceLinearScanLevel >= level) { \
code; \
}
#else
#define TRACE_LINEAR_SCAN(level, code)
#endif
class ComputeLinearScanOrder : public StackObj {
private:
int _max_block_id; // the highest block_id of a block
int _num_blocks; // total number of blocks (smaller than _max_block_id)
int _num_loops; // total number of loops
bool _iterative_dominators;// method requires iterative computation of dominatiors
BlockList* _linear_scan_order; // the resulting list of blocks in correct order
BitMap _visited_blocks; // used for recursive processing of blocks
BitMap _active_blocks; // used for recursive processing of blocks
BitMap _dominator_blocks; // temproary BitMap used for computation of dominator
intArray _forward_branches; // number of incoming forward branches for each block
BlockList _loop_end_blocks; // list of all loop end blocks collected during count_edges
BitMap2D _loop_map; // two-dimensional bit set: a bit is set if a block is contained in a loop
BlockList _work_list; // temporary list (used in mark_loops and compute_order)
BlockList _loop_headers;
Compilation* _compilation;
// accessors for _visited_blocks and _active_blocks
void init_visited() { _active_blocks.clear(); _visited_blocks.clear(); }
bool is_visited(BlockBegin* b) const { return _visited_blocks.at(b->block_id()); }
bool is_active(BlockBegin* b) const { return _active_blocks.at(b->block_id()); }
void set_visited(BlockBegin* b) { assert(!is_visited(b), "already set"); _visited_blocks.set_bit(b->block_id()); }
void set_active(BlockBegin* b) { assert(!is_active(b), "already set"); _active_blocks.set_bit(b->block_id()); }
void clear_active(BlockBegin* b) { assert(is_active(b), "not already"); _active_blocks.clear_bit(b->block_id()); }
// accessors for _forward_branches
void inc_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) + 1); }
int dec_forward_branches(BlockBegin* b) { _forward_branches.at_put(b->block_id(), _forward_branches.at(b->block_id()) - 1); return _forward_branches.at(b->block_id()); }
// accessors for _loop_map
bool is_block_in_loop (int loop_idx, BlockBegin* b) const { return _loop_map.at(loop_idx, b->block_id()); }
void set_block_in_loop (int loop_idx, BlockBegin* b) { _loop_map.set_bit(loop_idx, b->block_id()); }
void clear_block_in_loop(int loop_idx, int block_id) { _loop_map.clear_bit(loop_idx, block_id); }
// count edges between blocks
void count_edges(BlockBegin* cur, BlockBegin* parent);
// loop detection
void mark_loops();
void clear_non_natural_loops(BlockBegin* start_block);
void assign_loop_depth(BlockBegin* start_block);
// computation of final block order
BlockBegin* common_dominator(BlockBegin* a, BlockBegin* b);
void compute_dominator(BlockBegin* cur, BlockBegin* parent);
int compute_weight(BlockBegin* cur);
bool ready_for_processing(BlockBegin* cur);
void sort_into_work_list(BlockBegin* b);
void append_block(BlockBegin* cur);
void compute_order(BlockBegin* start_block);
// fixup of dominators for non-natural loops
bool compute_dominators_iter();
void compute_dominators();
// debug functions
NOT_PRODUCT(void print_blocks();)
DEBUG_ONLY(void verify();)
Compilation* compilation() const { return _compilation; }
public:
ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block);
// accessors for final result
BlockList* linear_scan_order() const { return _linear_scan_order; }
int num_loops() const { return _num_loops; }
};
ComputeLinearScanOrder::ComputeLinearScanOrder(Compilation* c, BlockBegin* start_block) :
_max_block_id(BlockBegin::number_of_blocks()),
_num_blocks(0),
_num_loops(0),
_iterative_dominators(false),
_visited_blocks(_max_block_id),
_active_blocks(_max_block_id),
_dominator_blocks(_max_block_id),
_forward_branches(_max_block_id, 0),
_loop_end_blocks(8),
_work_list(8),
_linear_scan_order(NULL), // initialized later with correct size
_loop_map(0, 0), // initialized later with correct size
_compilation(c)
{
TRACE_LINEAR_SCAN(2, tty->print_cr("***** computing linear-scan block order"));
init_visited();
count_edges(start_block, NULL);
if (compilation()->is_profiling()) {
ciMethod *method = compilation()->method();
if (!method->is_accessor()) {
ciMethodData* md = method->method_data_or_null();
assert(md != NULL, "Sanity");
md->set_compilation_stats(_num_loops, _num_blocks);
}
}
if (_num_loops > 0) {
mark_loops();
clear_non_natural_loops(start_block);
assign_loop_depth(start_block);
}
compute_order(start_block);
compute_dominators();
NOT_PRODUCT(print_blocks());
DEBUG_ONLY(verify());
}
// Traverse the CFG:
// * count total number of blocks
// * count all incoming edges and backward incoming edges
// * number loop header blocks
// * create a list with all loop end blocks
void ComputeLinearScanOrder::count_edges(BlockBegin* cur, BlockBegin* parent) {
TRACE_LINEAR_SCAN(3, tty->print_cr("Enter count_edges for block B%d coming from B%d", cur->block_id(), parent != NULL ? parent->block_id() : -1));
assert(cur->dominator() == NULL, "dominator already initialized");
if (is_active(cur)) {
TRACE_LINEAR_SCAN(3, tty->print_cr("backward branch"));
assert(is_visited(cur), "block must be visisted when block is active");
assert(parent != NULL, "must have parent");
cur->set(BlockBegin::backward_branch_target_flag);
// When a loop header is also the start of an exception handler, then the backward branch is
// an exception edge. Because such edges are usually critical edges which cannot be split, the
// loop must be excluded here from processing.
if (cur->is_set(BlockBegin::exception_entry_flag)) {
// Make sure that dominators are correct in this weird situation
_iterative_dominators = true;
return;
}
cur->set(BlockBegin::linear_scan_loop_header_flag);
parent->set(BlockBegin::linear_scan_loop_end_flag);
assert(parent->number_of_sux() == 1 && parent->sux_at(0) == cur,
"loop end blocks must have one successor (critical edges are split)");
_loop_end_blocks.append(parent);
return;
}
// increment number of incoming forward branches
inc_forward_branches(cur);
if (is_visited(cur)) {
TRACE_LINEAR_SCAN(3, tty->print_cr("block already visited"));
return;
}
_num_blocks++;
set_visited(cur);
set_active(cur);
// recursive call for all successors
int i;
for (i = cur->number_of_sux() - 1; i >= 0; i--) {
count_edges(cur->sux_at(i), cur);
}
for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
count_edges(cur->exception_handler_at(i), cur);
}
clear_active(cur);
// Each loop has a unique number.
// When multiple loops are nested, assign_loop_depth assumes that the
// innermost loop has the lowest number. This is guaranteed by setting
// the loop number after the recursive calls for the successors above
// have returned.
if (cur->is_set(BlockBegin::linear_scan_loop_header_flag)) {
assert(cur->loop_index() == -1, "cannot set loop-index twice");
TRACE_LINEAR_SCAN(3, tty->print_cr("Block B%d is loop header of loop %d", cur->block_id(), _num_loops));
cur->set_loop_index(_num_loops);
_loop_headers.append(cur);
_num_loops++;
}
TRACE_LINEAR_SCAN(3, tty->print_cr("Finished count_edges for block B%d", cur->block_id()));
}
void ComputeLinearScanOrder::mark_loops() {
TRACE_LINEAR_SCAN(3, tty->print_cr("----- marking loops"));
_loop_map = BitMap2D(_num_loops, _max_block_id);
_loop_map.clear();
for (int i = _loop_end_blocks.length() - 1; i >= 0; i--) {
BlockBegin* loop_end = _loop_end_blocks.at(i);
BlockBegin* loop_start = loop_end->sux_at(0);
int loop_idx = loop_start->loop_index();
TRACE_LINEAR_SCAN(3, tty->print_cr("Processing loop from B%d to B%d (loop %d):", loop_start->block_id(), loop_end->block_id(), loop_idx));
assert(loop_end->is_set(BlockBegin::linear_scan_loop_end_flag), "loop end flag must be set");
assert(loop_end->number_of_sux() == 1, "incorrect number of successors");
assert(loop_start->is_set(BlockBegin::linear_scan_loop_header_flag), "loop header flag must be set");
assert(loop_idx >= 0 && loop_idx < _num_loops, "loop index not set");
assert(_work_list.is_empty(), "work list must be empty before processing");
// add the end-block of the loop to the working list
_work_list.push(loop_end);
set_block_in_loop(loop_idx, loop_end);
do {
BlockBegin* cur = _work_list.pop();
TRACE_LINEAR_SCAN(3, tty->print_cr(" processing B%d", cur->block_id()));
assert(is_block_in_loop(loop_idx, cur), "bit in loop map must be set when block is in work list");
// recursive processing of all predecessors ends when start block of loop is reached
if (cur != loop_start && !cur->is_set(BlockBegin::osr_entry_flag)) {
for (int j = cur->number_of_preds() - 1; j >= 0; j--) {
BlockBegin* pred = cur->pred_at(j);
if (!is_block_in_loop(loop_idx, pred) /*&& !pred->is_set(BlockBeginosr_entry_flag)*/) {
// this predecessor has not been processed yet, so add it to work list
TRACE_LINEAR_SCAN(3, tty->print_cr(" pushing B%d", pred->block_id()));
_work_list.push(pred);
set_block_in_loop(loop_idx, pred);
}
}
}
} while (!_work_list.is_empty());
}
}
// check for non-natural loops (loops where the loop header does not dominate
// all other loop blocks = loops with mulitple entries).
// such loops are ignored
void ComputeLinearScanOrder::clear_non_natural_loops(BlockBegin* start_block) {
for (int i = _num_loops - 1; i >= 0; i--) {
if (is_block_in_loop(i, start_block)) {
// loop i contains the entry block of the method
// -> this is not a natural loop, so ignore it
TRACE_LINEAR_SCAN(2, tty->print_cr("Loop %d is non-natural, so it is ignored", i));
BlockBegin *loop_header = _loop_headers.at(i);
assert(loop_header->is_set(BlockBegin::linear_scan_loop_header_flag), "Must be loop header");
for (int j = 0; j < loop_header->number_of_preds(); j++) {
BlockBegin *pred = loop_header->pred_at(j);
pred->clear(BlockBegin::linear_scan_loop_end_flag);
}
loop_header->clear(BlockBegin::linear_scan_loop_header_flag);
for (int block_id = _max_block_id - 1; block_id >= 0; block_id--) {
clear_block_in_loop(i, block_id);
}
_iterative_dominators = true;
}
}
}
void ComputeLinearScanOrder::assign_loop_depth(BlockBegin* start_block) {
TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing loop-depth and weight"));
init_visited();
assert(_work_list.is_empty(), "work list must be empty before processing");
_work_list.append(start_block);
do {
BlockBegin* cur = _work_list.pop();
if (!is_visited(cur)) {
set_visited(cur);
TRACE_LINEAR_SCAN(4, tty->print_cr("Computing loop depth for block B%d", cur->block_id()));
// compute loop-depth and loop-index for the block
assert(cur->loop_depth() == 0, "cannot set loop-depth twice");
int i;
int loop_depth = 0;
int min_loop_idx = -1;
for (i = _num_loops - 1; i >= 0; i--) {
if (is_block_in_loop(i, cur)) {
loop_depth++;
min_loop_idx = i;
}
}
cur->set_loop_depth(loop_depth);
cur->set_loop_index(min_loop_idx);
// append all unvisited successors to work list
for (i = cur->number_of_sux() - 1; i >= 0; i--) {
_work_list.append(cur->sux_at(i));
}
for (i = cur->number_of_exception_handlers() - 1; i >= 0; i--) {
_work_list.append(cur->exception_handler_at(i));
}
}
} while (!_work_list.is_empty());
}
BlockBegin* ComputeLinearScanOrder::common_dominator(BlockBegin* a, BlockBegin* b) {
assert(a != NULL && b != NULL, "must have input blocks");
_dominator_blocks.clear();
while (a != NULL) {
_dominator_blocks.set_bit(a->block_id());
assert(a->dominator() != NULL || a == _linear_scan_order->at(0), "dominator must be initialized");
a = a->dominator();
}
while (b != NULL && !_dominator_blocks.at(b->block_id())) {
assert(b->dominator() != NULL || b == _linear_scan_order->at(0), "dominator must be initialized");
b = b->dominator();
}
assert(b != NULL, "could not find dominator");
return b;
}
void ComputeLinearScanOrder::compute_dominator(BlockBegin* cur, BlockBegin* parent) {
if (cur->dominator() == NULL) {
TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: initializing dominator of B%d to B%d", cur->block_id(), parent->block_id()));
cur->set_dominator(parent);
} else if (!(cur->is_set(BlockBegin::linear_scan_loop_header_flag) && parent->is_set(BlockBegin::linear_scan_loop_end_flag))) {
TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: computing dominator of B%d: common dominator of B%d and B%d is B%d", cur->block_id(), parent->block_id(), cur->dominator()->block_id(), common_dominator(cur->dominator(), parent)->block_id()));
// Does not hold for exception blocks
assert(cur->number_of_preds() > 1 || cur->is_set(BlockBegin::exception_entry_flag), "");
cur->set_dominator(common_dominator(cur->dominator(), parent));
}
// Additional edge to xhandler of all our successors
// range check elimination needs that the state at the end of a
// block be valid in every block it dominates so cur must dominate
// the exception handlers of its successors.
int num_cur_xhandler = cur->number_of_exception_handlers();
for (int j = 0; j < num_cur_xhandler; j++) {
BlockBegin* xhandler = cur->exception_handler_at(j);
compute_dominator(xhandler, parent);
}
}
int ComputeLinearScanOrder::compute_weight(BlockBegin* cur) {
BlockBegin* single_sux = NULL;
if (cur->number_of_sux() == 1) {
single_sux = cur->sux_at(0);
}
// limit loop-depth to 15 bit (only for security reason, it will never be so big)
int weight = (cur->loop_depth() & 0x7FFF) << 16;
// general macro for short definition of weight flags
// the first instance of INC_WEIGHT_IF has the highest priority
int cur_bit = 15;
#define INC_WEIGHT_IF(condition) if ((condition)) { weight |= (1 << cur_bit); } cur_bit--;
// this is necessery for the (very rare) case that two successing blocks have
// the same loop depth, but a different loop index (can happen for endless loops
// with exception handlers)
INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_header_flag));
// loop end blocks (blocks that end with a backward branch) are added
// after all other blocks of the loop.
INC_WEIGHT_IF(!cur->is_set(BlockBegin::linear_scan_loop_end_flag));
// critical edge split blocks are prefered because than they have a bigger
// proability to be completely empty
INC_WEIGHT_IF(cur->is_set(BlockBegin::critical_edge_split_flag));
// exceptions should not be thrown in normal control flow, so these blocks
// are added as late as possible
INC_WEIGHT_IF(cur->end()->as_Throw() == NULL && (single_sux == NULL || single_sux->end()->as_Throw() == NULL));
INC_WEIGHT_IF(cur->end()->as_Return() == NULL && (single_sux == NULL || single_sux->end()->as_Return() == NULL));
// exceptions handlers are added as late as possible
INC_WEIGHT_IF(!cur->is_set(BlockBegin::exception_entry_flag));
// guarantee that weight is > 0
weight |= 1;
#undef INC_WEIGHT_IF
assert(cur_bit >= 0, "too many flags");
assert(weight > 0, "weight cannot become negative");
return weight;
}
bool ComputeLinearScanOrder::ready_for_processing(BlockBegin* cur) {
// Discount the edge just traveled.
// When the number drops to zero, all forward branches were processed
if (dec_forward_branches(cur) != 0) {
return false;
}
assert(_linear_scan_order->index_of(cur) == -1, "block already processed (block can be ready only once)");
assert(_work_list.index_of(cur) == -1, "block already in work-list (block can be ready only once)");
return true;
}
void ComputeLinearScanOrder::sort_into_work_list(BlockBegin* cur) {
assert(_work_list.index_of(cur) == -1, "block already in work list");
int cur_weight = compute_weight(cur);
// the linear_scan_number is used to cache the weight of a block
cur->set_linear_scan_number(cur_weight);
#ifndef PRODUCT
if (StressLinearScan) {
_work_list.insert_before(0, cur);
return;
}
#endif
_work_list.append(NULL); // provide space for new element
int insert_idx = _work_list.length() - 1;
while (insert_idx > 0 && _work_list.at(insert_idx - 1)->linear_scan_number() > cur_weight) {
_work_list.at_put(insert_idx, _work_list.at(insert_idx - 1));
insert_idx--;
}
_work_list.at_put(insert_idx, cur);
TRACE_LINEAR_SCAN(3, tty->print_cr("Sorted B%d into worklist. new worklist:", cur->block_id()));
TRACE_LINEAR_SCAN(3, for (int i = 0; i < _work_list.length(); i++) tty->print_cr("%8d B%2d weight:%6x", i, _work_list.at(i)->block_id(), _work_list.at(i)->linear_scan_number()));
#ifdef ASSERT
for (int i = 0; i < _work_list.length(); i++) {
assert(_work_list.at(i)->linear_scan_number() > 0, "weight not set");
assert(i == 0 || _work_list.at(i - 1)->linear_scan_number() <= _work_list.at(i)->linear_scan_number(), "incorrect order in worklist");
}
#endif
}
void ComputeLinearScanOrder::append_block(BlockBegin* cur) {
TRACE_LINEAR_SCAN(3, tty->print_cr("appending block B%d (weight 0x%6x) to linear-scan order", cur->block_id(), cur->linear_scan_number()));
assert(_linear_scan_order->index_of(cur) == -1, "cannot add the same block twice");
// currently, the linear scan order and code emit order are equal.
// therefore the linear_scan_number and the weight of a block must also
// be equal.
cur->set_linear_scan_number(_linear_scan_order->length());
_linear_scan_order->append(cur);
}
void ComputeLinearScanOrder::compute_order(BlockBegin* start_block) {
TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing final block order"));
// the start block is always the first block in the linear scan order
_linear_scan_order = new BlockList(_num_blocks);
append_block(start_block);
assert(start_block->end()->as_Base() != NULL, "start block must end with Base-instruction");
BlockBegin* std_entry = ((Base*)start_block->end())->std_entry();
BlockBegin* osr_entry = ((Base*)start_block->end())->osr_entry();
BlockBegin* sux_of_osr_entry = NULL;
if (osr_entry != NULL) {
// special handling for osr entry:
// ignore the edge between the osr entry and its successor for processing
// the osr entry block is added manually below
assert(osr_entry->number_of_sux() == 1, "osr entry must have exactly one successor");
assert(osr_entry->sux_at(0)->number_of_preds() >= 2, "sucessor of osr entry must have two predecessors (otherwise it is not present in normal control flow");
sux_of_osr_entry = osr_entry->sux_at(0);
dec_forward_branches(sux_of_osr_entry);
compute_dominator(osr_entry, start_block);
_iterative_dominators = true;
}
compute_dominator(std_entry, start_block);
// start processing with standard entry block
assert(_work_list.is_empty(), "list must be empty before processing");
if (ready_for_processing(std_entry)) {
sort_into_work_list(std_entry);
} else {
assert(false, "the std_entry must be ready for processing (otherwise, the method has no start block)");
}
do {
BlockBegin* cur = _work_list.pop();
if (cur == sux_of_osr_entry) {
// the osr entry block is ignored in normal processing, it is never added to the
// work list. Instead, it is added as late as possible manually here.
append_block(osr_entry);
compute_dominator(cur, osr_entry);
}
append_block(cur);
int i;
int num_sux = cur->number_of_sux();
// changed loop order to get "intuitive" order of if- and else-blocks
for (i = 0; i < num_sux; i++) {
BlockBegin* sux = cur->sux_at(i);
compute_dominator(sux, cur);
if (ready_for_processing(sux)) {
sort_into_work_list(sux);
}
}
num_sux = cur->number_of_exception_handlers();
for (i = 0; i < num_sux; i++) {
BlockBegin* sux = cur->exception_handler_at(i);
if (ready_for_processing(sux)) {
sort_into_work_list(sux);
}
}
} while (_work_list.length() > 0);
}
bool ComputeLinearScanOrder::compute_dominators_iter() {
bool changed = false;
int num_blocks = _linear_scan_order->length();
assert(_linear_scan_order->at(0)->dominator() == NULL, "must not have dominator");
assert(_linear_scan_order->at(0)->number_of_preds() == 0, "must not have predecessors");
for (int i = 1; i < num_blocks; i++) {
BlockBegin* block = _linear_scan_order->at(i);
BlockBegin* dominator = block->pred_at(0);
int num_preds = block->number_of_preds();
TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: Processing B%d", block->block_id()));
for (int j = 0; j < num_preds; j++) {
BlockBegin *pred = block->pred_at(j);
TRACE_LINEAR_SCAN(4, tty->print_cr(" DOM: Subrocessing B%d", pred->block_id()));
if (block->is_set(BlockBegin::exception_entry_flag)) {
dominator = common_dominator(dominator, pred);
int num_pred_preds = pred->number_of_preds();
for (int k = 0; k < num_pred_preds; k++) {
dominator = common_dominator(dominator, pred->pred_at(k));
}
} else {
dominator = common_dominator(dominator, pred);
}
}
if (dominator != block->dominator()) {
TRACE_LINEAR_SCAN(4, tty->print_cr("DOM: updating dominator of B%d from B%d to B%d", block->block_id(), block->dominator()->block_id(), dominator->block_id()));
block->set_dominator(dominator);
changed = true;
}
}
return changed;
}
void ComputeLinearScanOrder::compute_dominators() {
TRACE_LINEAR_SCAN(3, tty->print_cr("----- computing dominators (iterative computation reqired: %d)", _iterative_dominators));
// iterative computation of dominators is only required for methods with non-natural loops
// and OSR-methods. For all other methods, the dominators computed when generating the
// linear scan block order are correct.
if (_iterative_dominators) {
do {
TRACE_LINEAR_SCAN(1, tty->print_cr("DOM: next iteration of fix-point calculation"));
} while (compute_dominators_iter());
}
// check that dominators are correct
assert(!compute_dominators_iter(), "fix point not reached");
// Add Blocks to dominates-Array
int num_blocks = _linear_scan_order->length();
for (int i = 0; i < num_blocks; i++) {
BlockBegin* block = _linear_scan_order->at(i);
BlockBegin *dom = block->dominator();
if (dom) {
assert(dom->dominator_depth() != -1, "Dominator must have been visited before");
dom->dominates()->append(block);
block->set_dominator_depth(dom->dominator_depth() + 1);
} else {
block->set_dominator_depth(0);
}
}
}
#ifndef PRODUCT
void ComputeLinearScanOrder::print_blocks() {
if (TraceLinearScanLevel >= 2) {
tty->print_cr("----- loop information:");
for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
BlockBegin* cur = _linear_scan_order->at(block_idx);
tty->print("%4d: B%2d: ", cur->linear_scan_number(), cur->block_id());
for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
tty->print ("%d ", is_block_in_loop(loop_idx, cur));
}
tty->print_cr(" -> loop_index: %2d, loop_depth: %2d", cur->loop_index(), cur->loop_depth());
}
}
if (TraceLinearScanLevel >= 1) {
tty->print_cr("----- linear-scan block order:");
for (int block_idx = 0; block_idx < _linear_scan_order->length(); block_idx++) {
BlockBegin* cur = _linear_scan_order->at(block_idx);
tty->print("%4d: B%2d loop: %2d depth: %2d", cur->linear_scan_number(), cur->block_id(), cur->loop_index(), cur->loop_depth());
tty->print(cur->is_set(BlockBegin::exception_entry_flag) ? " ex" : " ");
tty->print(cur->is_set(BlockBegin::critical_edge_split_flag) ? " ce" : " ");
tty->print(cur->is_set(BlockBegin::linear_scan_loop_header_flag) ? " lh" : " ");
tty->print(cur->is_set(BlockBegin::linear_scan_loop_end_flag) ? " le" : " ");
if (cur->dominator() != NULL) {
tty->print(" dom: B%d ", cur->dominator()->block_id());
} else {
tty->print(" dom: NULL ");
}
if (cur->number_of_preds() > 0) {
tty->print(" preds: ");
for (int j = 0; j < cur->number_of_preds(); j++) {
BlockBegin* pred = cur->pred_at(j);
tty->print("B%d ", pred->block_id());
}
}
if (cur->number_of_sux() > 0) {
tty->print(" sux: ");
for (int j = 0; j < cur->number_of_sux(); j++) {
BlockBegin* sux = cur->sux_at(j);
tty->print("B%d ", sux->block_id());
}
}
if (cur->number_of_exception_handlers() > 0) {
tty->print(" ex: ");
for (int j = 0; j < cur->number_of_exception_handlers(); j++) {
BlockBegin* ex = cur->exception_handler_at(j);
tty->print("B%d ", ex->block_id());
}
}
tty->cr();
}
}
}
#endif
#ifdef ASSERT
void ComputeLinearScanOrder::verify() {
assert(_linear_scan_order->length() == _num_blocks, "wrong number of blocks in list");
if (StressLinearScan) {
// blocks are scrambled when StressLinearScan is used
return;
}
// check that all successors of a block have a higher linear-scan-number
// and that all predecessors of a block have a lower linear-scan-number
// (only backward branches of loops are ignored)
int i;
for (i = 0; i < _linear_scan_order->length(); i++) {
BlockBegin* cur = _linear_scan_order->at(i);
assert(cur->linear_scan_number() == i, "incorrect linear_scan_number");
assert(cur->linear_scan_number() >= 0 && cur->linear_scan_number() == _linear_scan_order->index_of(cur), "incorrect linear_scan_number");
int j;
for (j = cur->number_of_sux() - 1; j >= 0; j--) {
BlockBegin* sux = cur->sux_at(j);
assert(sux->linear_scan_number() >= 0 && sux->linear_scan_number() == _linear_scan_order->index_of(sux), "incorrect linear_scan_number");
if (!sux->is_set(BlockBegin::backward_branch_target_flag)) {
assert(cur->linear_scan_number() < sux->linear_scan_number(), "invalid order");
}
if (cur->loop_depth() == sux->loop_depth()) {
assert(cur->loop_index() == sux->loop_index() || sux->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index");
}
}
for (j = cur->number_of_preds() - 1; j >= 0; j--) {
BlockBegin* pred = cur->pred_at(j);
assert(pred->linear_scan_number() >= 0 && pred->linear_scan_number() == _linear_scan_order->index_of(pred), "incorrect linear_scan_number");
if (!cur->is_set(BlockBegin::backward_branch_target_flag)) {
assert(cur->linear_scan_number() > pred->linear_scan_number(), "invalid order");
}
if (cur->loop_depth() == pred->loop_depth()) {
assert(cur->loop_index() == pred->loop_index() || cur->is_set(BlockBegin::linear_scan_loop_header_flag), "successing blocks with same loop depth must have same loop index");
}
assert(cur->dominator()->linear_scan_number() <= cur->pred_at(j)->linear_scan_number(), "dominator must be before predecessors");
}
// check dominator
if (i == 0) {
assert(cur->dominator() == NULL, "first block has no dominator");
} else {
assert(cur->dominator() != NULL, "all but first block must have dominator");
}
// Assertion does not hold for exception handlers
assert(cur->number_of_preds() != 1 || cur->dominator() == cur->pred_at(0) || cur->is_set(BlockBegin::exception_entry_flag), "Single predecessor must also be dominator");
}
// check that all loops are continuous
for (int loop_idx = 0; loop_idx < _num_loops; loop_idx++) {
int block_idx = 0;
assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "the first block must not be present in any loop");
// skip blocks before the loop
while (block_idx < _num_blocks && !is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
block_idx++;
}
// skip blocks of loop
while (block_idx < _num_blocks && is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx))) {
block_idx++;
}
// after the first non-loop block, there must not be another loop-block
while (block_idx < _num_blocks) {
assert(!is_block_in_loop(loop_idx, _linear_scan_order->at(block_idx)), "loop not continuous in linear-scan order");
block_idx++;
}
}
}
#endif
void IR::compute_code() {
assert(is_valid(), "IR must be valid");
ComputeLinearScanOrder compute_order(compilation(), start());
_num_loops = compute_order.num_loops();
_code = compute_order.linear_scan_order();
}
void IR::compute_use_counts() {
// make sure all values coming out of this block get evaluated.
int num_blocks = _code->length();
for (int i = 0; i < num_blocks; i++) {
_code->at(i)->end()->state()->pin_stack_for_linear_scan();
}
// compute use counts
UseCountComputer::compute(_code);
}
void IR::iterate_preorder(BlockClosure* closure) {
assert(is_valid(), "IR must be valid");
start()->iterate_preorder(closure);
}
void IR::iterate_postorder(BlockClosure* closure) {
assert(is_valid(), "IR must be valid");
start()->iterate_postorder(closure);
}
void IR::iterate_linear_scan_order(BlockClosure* closure) {
linear_scan_order()->iterate_forward(closure);
}
#ifndef PRODUCT
class BlockPrinter: public BlockClosure {
private:
InstructionPrinter* _ip;
bool _cfg_only;
bool _live_only;
public:
BlockPrinter(InstructionPrinter* ip, bool cfg_only, bool live_only = false) {
_ip = ip;
_cfg_only = cfg_only;
_live_only = live_only;
}
virtual void block_do(BlockBegin* block) {
if (_cfg_only) {
_ip->print_instr(block); tty->cr();
} else {
block->print_block(*_ip, _live_only);
}
}
};
void IR::print(BlockBegin* start, bool cfg_only, bool live_only) {
ttyLocker ttyl;
InstructionPrinter ip(!cfg_only);
BlockPrinter bp(&ip, cfg_only, live_only);
start->iterate_preorder(&bp);
tty->cr();
}
void IR::print(bool cfg_only, bool live_only) {
if (is_valid()) {
print(start(), cfg_only, live_only);
} else {
tty->print_cr("invalid IR");
}
}
define_array(BlockListArray, BlockList*)
define_stack(BlockListList, BlockListArray)
class PredecessorValidator : public BlockClosure {
private:
BlockListList* _predecessors;
BlockList* _blocks;
static int cmp(BlockBegin** a, BlockBegin** b) {
return (*a)->block_id() - (*b)->block_id();
}
public:
PredecessorValidator(IR* hir) {
ResourceMark rm;
_predecessors = new BlockListList(BlockBegin::number_of_blocks(), NULL);
_blocks = new BlockList();
int i;
hir->start()->iterate_preorder(this);
if (hir->code() != NULL) {
assert(hir->code()->length() == _blocks->length(), "must match");
for (i = 0; i < _blocks->length(); i++) {
assert(hir->code()->contains(_blocks->at(i)), "should be in both lists");
}
}
for (i = 0; i < _blocks->length(); i++) {
BlockBegin* block = _blocks->at(i);
BlockList* preds = _predecessors->at(block->block_id());
if (preds == NULL) {
assert(block->number_of_preds() == 0, "should be the same");
continue;
}
// clone the pred list so we can mutate it
BlockList* pred_copy = new BlockList();
int j;
for (j = 0; j < block->number_of_preds(); j++) {
pred_copy->append(block->pred_at(j));
}
// sort them in the same order
preds->sort(cmp);
pred_copy->sort(cmp);
int length = MIN2(preds->length(), block->number_of_preds());
for (j = 0; j < block->number_of_preds(); j++) {
assert(preds->at(j) == pred_copy->at(j), "must match");
}
assert(preds->length() == block->number_of_preds(), "should be the same");
}
}
virtual void block_do(BlockBegin* block) {
_blocks->append(block);
BlockEnd* be = block->end();
int n = be->number_of_sux();
int i;
for (i = 0; i < n; i++) {
BlockBegin* sux = be->sux_at(i);
assert(!sux->is_set(BlockBegin::exception_entry_flag), "must not be xhandler");
BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL);
if (preds == NULL) {
preds = new BlockList();
_predecessors->at_put(sux->block_id(), preds);
}
preds->append(block);
}
n = block->number_of_exception_handlers();
for (i = 0; i < n; i++) {
BlockBegin* sux = block->exception_handler_at(i);
assert(sux->is_set(BlockBegin::exception_entry_flag), "must be xhandler");
BlockList* preds = _predecessors->at_grow(sux->block_id(), NULL);
if (preds == NULL) {
preds = new BlockList();
_predecessors->at_put(sux->block_id(), preds);
}
preds->append(block);
}
}
};
class VerifyBlockBeginField : public BlockClosure {
public:
virtual void block_do(BlockBegin *block) {
for ( Instruction *cur = block; cur != NULL; cur = cur->next()) {
assert(cur->block() == block, "Block begin is not correct");
}
}
};
void IR::verify() {
#ifdef ASSERT
PredecessorValidator pv(this);
VerifyBlockBeginField verifier;
this->iterate_postorder(&verifier);
#endif
}
#endif // PRODUCT
void SubstitutionResolver::visit(Value* v) {
Value v0 = *v;
if (v0) {
Value vs = v0->subst();
if (vs != v0) {
*v = v0->subst();
}
}
}
#ifdef ASSERT
class SubstitutionChecker: public ValueVisitor {
void visit(Value* v) {
Value v0 = *v;
if (v0) {
Value vs = v0->subst();
assert(vs == v0, "missed substitution");
}
}
};
#endif
void SubstitutionResolver::block_do(BlockBegin* block) {
Instruction* last = NULL;
for (Instruction* n = block; n != NULL;) {
n->values_do(this);
// need to remove this instruction from the instruction stream
if (n->subst() != n) {
assert(last != NULL, "must have last");
last->set_next(n->next());
} else {
last = n;
}
n = last->next();
}
#ifdef ASSERT
SubstitutionChecker check_substitute;
if (block->state()) block->state()->values_do(&check_substitute);
block->block_values_do(&check_substitute);
if (block->end() && block->end()->state()) block->end()->state()->values_do(&check_substitute);
#endif
}