6894779: Loop Predication for Loop Optimizer in C2
Summary: Loop predication implementation
Reviewed-by: never, kvn
--- a/hotspot/src/share/vm/includeDB_compiler2 Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/includeDB_compiler2 Tue Jan 12 14:37:35 2010 -0800
@@ -601,6 +601,7 @@
loopTransform.cpp addnode.hpp
loopTransform.cpp allocation.inline.hpp
+loopTransform.cpp callnode.hpp
loopTransform.cpp connode.hpp
loopTransform.cpp compileLog.hpp
loopTransform.cpp divnode.hpp
--- a/hotspot/src/share/vm/opto/c2_globals.hpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/c2_globals.hpp Tue Jan 12 14:37:35 2010 -0800
@@ -154,6 +154,12 @@
notproduct(bool, TraceProfileTripCount, false, \
"Trace profile loop trip count information") \
\
+ product(bool, UseLoopPredicate, true, \
+ "Generate a predicate to select fast/slow loop versions") \
+ \
+ develop(bool, TraceLoopPredicate, false, \
+ "Trace generation of loop predicates") \
+ \
develop(bool, OptoCoalesce, true, \
"Use Conservative Copy Coalescing in the Register Allocator") \
\
--- a/hotspot/src/share/vm/opto/compile.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/compile.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -932,6 +932,7 @@
_intrinsics = NULL;
_macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
+ _predicate_opaqs = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
register_library_intrinsics();
}
@@ -1553,6 +1554,19 @@
}
}
+//---------------------cleanup_loop_predicates-----------------------
+// Remove the opaque nodes that protect the predicates so that all unused
+// checks and uncommon_traps will be eliminated from the ideal graph
+void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
+ if (predicate_count()==0) return;
+ for (int i = predicate_count(); i > 0; i--) {
+ Node * n = predicate_opaque1_node(i-1);
+ assert(n->Opcode() == Op_Opaque1, "must be");
+ igvn.replace_node(n, n->in(1));
+ }
+ assert(predicate_count()==0, "should be clean!");
+ igvn.optimize();
+}
//------------------------------Optimize---------------------------------------
// Given a graph, optimize it.
@@ -1594,7 +1608,7 @@
if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
{
TracePhase t2("idealLoop", &_t_idealLoop, true);
- PhaseIdealLoop ideal_loop( igvn, true );
+ PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate);
loop_opts_cnt--;
if (major_progress()) print_method("PhaseIdealLoop 1", 2);
if (failing()) return;
@@ -1602,7 +1616,7 @@
// Loop opts pass if partial peeling occurred in previous pass
if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
TracePhase t3("idealLoop", &_t_idealLoop, true);
- PhaseIdealLoop ideal_loop( igvn, false );
+ PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
loop_opts_cnt--;
if (major_progress()) print_method("PhaseIdealLoop 2", 2);
if (failing()) return;
@@ -1610,7 +1624,7 @@
// Loop opts pass for loop-unrolling before CCP
if(major_progress() && (loop_opts_cnt > 0)) {
TracePhase t4("idealLoop", &_t_idealLoop, true);
- PhaseIdealLoop ideal_loop( igvn, false );
+ PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate);
loop_opts_cnt--;
if (major_progress()) print_method("PhaseIdealLoop 3", 2);
}
@@ -1648,13 +1662,21 @@
// peeling, unrolling, etc.
if(loop_opts_cnt > 0) {
debug_only( int cnt = 0; );
+ bool loop_predication = UseLoopPredicate;
while(major_progress() && (loop_opts_cnt > 0)) {
TracePhase t2("idealLoop", &_t_idealLoop, true);
assert( cnt++ < 40, "infinite cycle in loop optimization" );
- PhaseIdealLoop ideal_loop( igvn, true );
+ PhaseIdealLoop ideal_loop( igvn, true, loop_predication);
loop_opts_cnt--;
if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
if (failing()) return;
+ // Perform loop predication optimization during first iteration after CCP.
+ // After that switch it off and cleanup unused loop predicates.
+ if (loop_predication) {
+ loop_predication = false;
+ cleanup_loop_predicates(igvn);
+ if (failing()) return;
+ }
}
}
--- a/hotspot/src/share/vm/opto/compile.hpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/compile.hpp Tue Jan 12 14:37:35 2010 -0800
@@ -38,6 +38,7 @@
class OptoReg;
class PhaseCFG;
class PhaseGVN;
+class PhaseIterGVN;
class PhaseRegAlloc;
class PhaseCCP;
class PhaseCCP_DCE;
@@ -172,6 +173,7 @@
const char* _failure_reason; // for record_failure/failing pattern
GrowableArray<CallGenerator*>* _intrinsics; // List of intrinsics.
GrowableArray<Node*>* _macro_nodes; // List of nodes which need to be expanded before matching.
+ GrowableArray<Node*>* _predicate_opaqs; // List of Opaque1 nodes for the loop predicates.
ConnectionGraph* _congraph;
#ifndef PRODUCT
IdealGraphPrinter* _printer;
@@ -351,7 +353,9 @@
}
int macro_count() { return _macro_nodes->length(); }
+ int predicate_count() { return _predicate_opaqs->length();}
Node* macro_node(int idx) { return _macro_nodes->at(idx); }
+ Node* predicate_opaque1_node(int idx) { return _predicate_opaqs->at(idx);}
ConnectionGraph* congraph() { return _congraph;}
void add_macro_node(Node * n) {
//assert(n->is_macro(), "must be a macro node");
@@ -363,7 +367,19 @@
// that the node is in the array before attempting to remove it
if (_macro_nodes->contains(n))
_macro_nodes->remove(n);
+ // remove from _predicate_opaqs list also if it is there
+ if (predicate_count() > 0 && _predicate_opaqs->contains(n)){
+ _predicate_opaqs->remove(n);
+ }
}
+ void add_predicate_opaq(Node * n) {
+ assert(!_predicate_opaqs->contains(n), " duplicate entry in predicate opaque1");
+ assert(_macro_nodes->contains(n), "should have already been in macro list");
+ _predicate_opaqs->append(n);
+ }
+ // remove the opaque nodes that protect the predicates so that the unused checks and
+ // uncommon traps will be eliminated from the graph.
+ void cleanup_loop_predicates(PhaseIterGVN &igvn);
// Compilation environment.
Arena* comp_arena() { return &_comp_arena; }
--- a/hotspot/src/share/vm/opto/loopTransform.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/loopTransform.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -549,6 +549,10 @@
// Comparing trip+off vs limit
Node *bol = iff->in(1);
if( bol->req() != 2 ) continue; // dead constant test
+ if (!bol->is_Bool()) {
+ assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
+ continue;
+ }
Node *cmp = bol->in(1);
Node *rc_exp = cmp->in(1);
@@ -875,7 +879,7 @@
//------------------------------is_invariant-----------------------------
// Return true if n is invariant
bool IdealLoopTree::is_invariant(Node* n) const {
- Node *n_c = _phase->get_ctrl(n);
+ Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
if (n_c->is_top()) return false;
return !is_member(_phase->get_loop(n_c));
}
@@ -1594,7 +1598,7 @@
bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
// Check and remove empty loops (spam micro-benchmarks)
if( policy_do_remove_empty_loop(phase) )
- return true; // Here we removed an empty loop
+ return true; // Here we removed an empty loop
bool should_peel = policy_peeling(phase); // Should we peel?
@@ -1688,8 +1692,8 @@
// an even number of trips). If we are peeling, we might enable some RCE
// and we'd rather unroll the post-RCE'd loop SO... do not unroll if
// peeling.
- if( should_unroll && !should_peel )
- phase->do_unroll(this,old_new, true);
+ if( should_unroll && !should_peel )
+ phase->do_unroll(this,old_new, true);
// Adjust the pre-loop limits to align the main body
// iterations.
@@ -1731,9 +1735,9 @@
_allow_optimizations &&
!tail()->is_top() ) { // Also ignore the occasional dead backedge
if (!_has_call) {
- if (!iteration_split_impl( phase, old_new )) {
- return false;
- }
+ if (!iteration_split_impl( phase, old_new )) {
+ return false;
+ }
} else if (policy_unswitching(phase)) {
phase->do_unswitching(this, old_new);
}
@@ -1746,3 +1750,576 @@
return false;
return true;
}
+
+//-------------------------------is_uncommon_trap_proj----------------------------
+// Return true if proj is the form of "proj->[region->..]call_uct"
+bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate) {
+ int path_limit = 10;
+ assert(proj, "invalid argument");
+ Node* out = proj;
+ for (int ct = 0; ct < path_limit; ct++) {
+ out = out->unique_ctrl_out();
+ if (out == NULL || out->is_Root() || out->is_Start())
+ return false;
+ if (out->is_CallStaticJava()) {
+ int req = out->as_CallStaticJava()->uncommon_trap_request();
+ if (req != 0) {
+ Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(req);
+ if (!must_reason_predicate || reason == Deoptimization::Reason_predicate){
+ return true;
+ }
+ }
+ return false; // don't do further after call
+ }
+ }
+ return false;
+}
+
+//-------------------------------is_uncommon_trap_if_pattern-------------------------
+// Return true for "if(test)-> proj -> ...
+// |
+// V
+// other_proj->[region->..]call_uct"
+//
+// "must_reason_predicate" means the uct reason must be Reason_predicate
+bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, bool must_reason_predicate) {
+ Node *in0 = proj->in(0);
+ if (!in0->is_If()) return false;
+ IfNode* iff = in0->as_If();
+
+ // we need "If(Conv2B(Opaque1(...)))" pattern for must_reason_predicate
+ if (must_reason_predicate) {
+ if (iff->in(1)->Opcode() != Op_Conv2B ||
+ iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
+ return false;
+ }
+ }
+
+ ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
+ return is_uncommon_trap_proj(other_proj, must_reason_predicate);
+}
+
+//------------------------------create_new_if_for_predicate------------------------
+// create a new if above the uct_if_pattern for the predicate to be promoted.
+//
+// before after
+// ---------- ----------
+// ctrl ctrl
+// | |
+// | |
+// v v
+// iff new_iff
+// / \ / \
+// / \ / \
+// v v v v
+// uncommon_proj cont_proj if_uct if_cont
+// \ | | | |
+// \ | | | |
+// v v v | v
+// rgn loop | iff
+// | | / \
+// | | / \
+// v | v v
+// uncommon_trap | uncommon_proj cont_proj
+// \ \ | |
+// \ \ | |
+// v v v v
+// rgn loop
+// |
+// |
+// v
+// uncommon_trap
+//
+//
+// We will create a region to guard the uct call if there is no one there.
+// The true projecttion (if_cont) of the new_iff is returned.
+ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj) {
+ assert(is_uncommon_trap_if_pattern(cont_proj, true), "must be a uct if pattern!");
+ IfNode* iff = cont_proj->in(0)->as_If();
+
+ ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
+ Node *rgn = uncommon_proj->unique_ctrl_out();
+ assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
+
+ if (!rgn->is_Region()) { // create a region to guard the call
+ assert(rgn->is_Call(), "must be call uct");
+ CallNode* call = rgn->as_Call();
+ rgn = new (C, 1) RegionNode(1);
+ _igvn.set_type(rgn, rgn->bottom_type());
+ rgn->add_req(uncommon_proj);
+ set_idom(rgn, idom(uncommon_proj), dom_depth(uncommon_proj)+1);
+ _igvn.hash_delete(call);
+ call->set_req(0, rgn);
+ }
+
+ // Create new_iff
+ uint iffdd = dom_depth(iff);
+ IdealLoopTree* lp = get_loop(iff);
+ IfNode *new_iff = new (C, 2) IfNode(iff->in(0), NULL, iff->_prob, iff->_fcnt);
+ register_node(new_iff, lp, idom(iff), iffdd);
+ Node *if_cont = new (C, 1) IfTrueNode(new_iff);
+ Node *if_uct = new (C, 1) IfFalseNode(new_iff);
+ if (cont_proj->is_IfFalse()) {
+ // Swap
+ Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
+ }
+ register_node(if_cont, lp, new_iff, iffdd);
+ register_node(if_uct, get_loop(rgn), new_iff, iffdd);
+
+ // if_cont to iff
+ _igvn.hash_delete(iff);
+ iff->set_req(0, if_cont);
+ set_idom(iff, if_cont, dom_depth(iff));
+
+ // if_uct to rgn
+ _igvn.hash_delete(rgn);
+ rgn->add_req(if_uct);
+ Node* ridom = idom(rgn);
+ Node* nrdom = dom_lca(ridom, new_iff);
+ set_idom(rgn, nrdom, dom_depth(rgn));
+
+ // rgn must have no phis
+ assert(!rgn->as_Region()->has_phi(), "region must have no phis");
+
+ return if_cont->as_Proj();
+}
+
+//------------------------------find_predicate_insertion_point--------------------------
+// Find a good location to insert a predicate
+ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c) {
+ if (start_c == C->root() || !start_c->is_Proj())
+ return NULL;
+ if (is_uncommon_trap_if_pattern(start_c->as_Proj(), true/*Reason_Predicate*/)) {
+ return start_c->as_Proj();
+ }
+ return NULL;
+}
+
+//------------------------------Invariance-----------------------------------
+// Helper class for loop_predication_impl to compute invariance on the fly and
+// clone invariants.
+class Invariance : public StackObj {
+ VectorSet _visited, _invariant;
+ Node_Stack _stack;
+ VectorSet _clone_visited;
+ Node_List _old_new; // map of old to new (clone)
+ IdealLoopTree* _lpt;
+ PhaseIdealLoop* _phase;
+
+ // Helper function to set up the invariance for invariance computation
+ // If n is a known invariant, set up directly. Otherwise, look up the
+ // the possibility to push n onto the stack for further processing.
+ void visit(Node* use, Node* n) {
+ if (_lpt->is_invariant(n)) { // known invariant
+ _invariant.set(n->_idx);
+ } else if (!n->is_CFG()) {
+ Node *n_ctrl = _phase->ctrl_or_self(n);
+ Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
+ if (_phase->is_dominator(n_ctrl, u_ctrl)) {
+ _stack.push(n, n->in(0) == NULL ? 1 : 0);
+ }
+ }
+ }
+
+ // Compute invariance for "the_node" and (possibly) all its inputs recursively
+ // on the fly
+ void compute_invariance(Node* n) {
+ assert(_visited.test(n->_idx), "must be");
+ visit(n, n);
+ while (_stack.is_nonempty()) {
+ Node* n = _stack.node();
+ uint idx = _stack.index();
+ if (idx == n->req()) { // all inputs are processed
+ _stack.pop();
+ // n is invariant if it's inputs are all invariant
+ bool all_inputs_invariant = true;
+ for (uint i = 0; i < n->req(); i++) {
+ Node* in = n->in(i);
+ if (in == NULL) continue;
+ assert(_visited.test(in->_idx), "must have visited input");
+ if (!_invariant.test(in->_idx)) { // bad guy
+ all_inputs_invariant = false;
+ break;
+ }
+ }
+ if (all_inputs_invariant) {
+ _invariant.set(n->_idx); // I am a invariant too
+ }
+ } else { // process next input
+ _stack.set_index(idx + 1);
+ Node* m = n->in(idx);
+ if (m != NULL && !_visited.test_set(m->_idx)) {
+ visit(n, m);
+ }
+ }
+ }
+ }
+
+ // Helper function to set up _old_new map for clone_nodes.
+ // If n is a known invariant, set up directly ("clone" of n == n).
+ // Otherwise, push n onto the stack for real cloning.
+ void clone_visit(Node* n) {
+ assert(_invariant.test(n->_idx), "must be invariant");
+ if (_lpt->is_invariant(n)) { // known invariant
+ _old_new.map(n->_idx, n);
+ } else{ // to be cloned
+ assert (!n->is_CFG(), "should not see CFG here");
+ _stack.push(n, n->in(0) == NULL ? 1 : 0);
+ }
+ }
+
+ // Clone "n" and (possibly) all its inputs recursively
+ void clone_nodes(Node* n, Node* ctrl) {
+ clone_visit(n);
+ while (_stack.is_nonempty()) {
+ Node* n = _stack.node();
+ uint idx = _stack.index();
+ if (idx == n->req()) { // all inputs processed, clone n!
+ _stack.pop();
+ // clone invariant node
+ Node* n_cl = n->clone();
+ _old_new.map(n->_idx, n_cl);
+ _phase->register_new_node(n_cl, ctrl);
+ for (uint i = 0; i < n->req(); i++) {
+ Node* in = n_cl->in(i);
+ if (in == NULL) continue;
+ n_cl->set_req(i, _old_new[in->_idx]);
+ }
+ } else { // process next input
+ _stack.set_index(idx + 1);
+ Node* m = n->in(idx);
+ if (m != NULL && !_clone_visited.test_set(m->_idx)) {
+ clone_visit(m); // visit the input
+ }
+ }
+ }
+ }
+
+ public:
+ Invariance(Arena* area, IdealLoopTree* lpt) :
+ _lpt(lpt), _phase(lpt->_phase),
+ _visited(area), _invariant(area), _stack(area, 10 /* guess */),
+ _clone_visited(area), _old_new(area)
+ {}
+
+ // Map old to n for invariance computation and clone
+ void map_ctrl(Node* old, Node* n) {
+ assert(old->is_CFG() && n->is_CFG(), "must be");
+ _old_new.map(old->_idx, n); // "clone" of old is n
+ _invariant.set(old->_idx); // old is invariant
+ _clone_visited.set(old->_idx);
+ }
+
+ // Driver function to compute invariance
+ bool is_invariant(Node* n) {
+ if (!_visited.test_set(n->_idx))
+ compute_invariance(n);
+ return (_invariant.test(n->_idx) != 0);
+ }
+
+ // Driver function to clone invariant
+ Node* clone(Node* n, Node* ctrl) {
+ assert(ctrl->is_CFG(), "must be");
+ assert(_invariant.test(n->_idx), "must be an invariant");
+ if (!_clone_visited.test(n->_idx))
+ clone_nodes(n, ctrl);
+ return _old_new[n->_idx];
+ }
+};
+
+//------------------------------is_range_check_if -----------------------------------
+// Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
+// Note: this function is particularly designed for loop predication. We require load_range
+// and offset to be loop invariant computed on the fly by "invar"
+bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
+ if (!is_loop_exit(iff)) {
+ return false;
+ }
+ if (!iff->in(1)->is_Bool()) {
+ return false;
+ }
+ const BoolNode *bol = iff->in(1)->as_Bool();
+ if (bol->_test._test != BoolTest::lt) {
+ return false;
+ }
+ if (!bol->in(1)->is_Cmp()) {
+ return false;
+ }
+ const CmpNode *cmp = bol->in(1)->as_Cmp();
+ if (cmp->Opcode() != Op_CmpU ) {
+ return false;
+ }
+ if (cmp->in(2)->Opcode() != Op_LoadRange) {
+ return false;
+ }
+ LoadRangeNode* lr = (LoadRangeNode*)cmp->in(2);
+ if (!invar.is_invariant(lr)) { // loadRange must be invariant
+ return false;
+ }
+ Node *iv = _head->as_CountedLoop()->phi();
+ int scale = 0;
+ Node *offset = NULL;
+ if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
+ return false;
+ }
+ if(offset && !invar.is_invariant(offset)) { // offset must be invariant
+ return false;
+ }
+ return true;
+}
+
+//------------------------------rc_predicate-----------------------------------
+// Create a range check predicate
+//
+// for (i = init; i < limit; i += stride) {
+// a[scale*i+offset]
+// }
+//
+// Compute max(scale*i + offset) for init <= i < limit and build the predicate
+// as "max(scale*i + offset) u< a.length".
+//
+// There are two cases for max(scale*i + offset):
+// (1) stride*scale > 0
+// max(scale*i + offset) = scale*(limit-stride) + offset
+// (2) stride*scale < 0
+// max(scale*i + offset) = scale*init + offset
+BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl,
+ int scale, Node* offset,
+ Node* init, Node* limit, Node* stride,
+ Node* range) {
+ Node* max_idx_expr = init;
+ int stride_con = stride->get_int();
+ if ((stride_con > 0) == (scale > 0)) {
+ max_idx_expr = new (C, 3) SubINode(limit, stride);
+ register_new_node(max_idx_expr, ctrl);
+ }
+
+ if (scale != 1) {
+ ConNode* con_scale = _igvn.intcon(scale);
+ max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
+ register_new_node(max_idx_expr, ctrl);
+ }
+
+ if (offset && (!offset->is_Con() || offset->get_int() != 0)){
+ max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
+ register_new_node(max_idx_expr, ctrl);
+ }
+
+ CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
+ register_new_node(cmp, ctrl);
+ BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
+ register_new_node(bol, ctrl);
+ return bol;
+}
+
+//------------------------------ loop_predication_impl--------------------------
+// Insert loop predicates for null checks and range checks
+bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
+ if (!UseLoopPredicate) return false;
+
+ // Too many traps seen?
+ bool tmt = C->too_many_traps(C->method(), 0, Deoptimization::Reason_predicate);
+ int tc = C->trap_count(Deoptimization::Reason_predicate);
+ if (tmt || tc > 0) {
+ if (TraceLoopPredicate) {
+ tty->print_cr("too many predicate traps: %d", tc);
+ C->method()->print(); // which method has too many predicate traps
+ tty->print_cr("");
+ }
+ return false;
+ }
+
+ CountedLoopNode *cl = NULL;
+ if (loop->_head->is_CountedLoop()) {
+ cl = loop->_head->as_CountedLoop();
+ // do nothing for iteration-splitted loops
+ if(!cl->is_normal_loop()) return false;
+ }
+
+ LoopNode *lpn = loop->_head->as_Loop();
+ Node* entry = lpn->in(LoopNode::EntryControl);
+
+ ProjNode *predicate_proj = find_predicate_insertion_point(entry);
+ if (!predicate_proj){
+#ifndef PRODUCT
+ if (TraceLoopPredicate) {
+ tty->print("missing predicate:");
+ loop->dump_head();
+ }
+#endif
+ return false;
+ }
+
+ ConNode* zero = _igvn.intcon(0);
+ set_ctrl(zero, C->root());
+ Node *cond_false = new (C, 2) Conv2BNode(zero);
+ register_new_node(cond_false, C->root());
+ ConNode* one = _igvn.intcon(1);
+ set_ctrl(one, C->root());
+ Node *cond_true = new (C, 2) Conv2BNode(one);
+ register_new_node(cond_true, C->root());
+
+ ResourceArea *area = Thread::current()->resource_area();
+ Invariance invar(area, loop);
+
+ // Create list of if-projs such that a newer proj dominates all older
+ // projs in the list, and they all dominate loop->tail()
+ Node_List if_proj_list(area);
+ LoopNode *head = loop->_head->as_Loop();
+ Node *current_proj = loop->tail(); //start from tail
+ while ( current_proj != head ) {
+ if (loop == get_loop(current_proj) && // still in the loop ?
+ current_proj->is_Proj() && // is a projection ?
+ current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
+ if_proj_list.push(current_proj);
+ }
+ current_proj = idom(current_proj);
+ }
+
+ bool hoisted = false; // true if at least one proj is promoted
+ while (if_proj_list.size() > 0) {
+ // Following are changed to nonnull when a predicate can be hoisted
+ ProjNode* new_predicate_proj = NULL;
+ BoolNode* new_predicate_bol = NULL;
+
+ ProjNode* proj = if_proj_list.pop()->as_Proj();
+ IfNode* iff = proj->in(0)->as_If();
+
+ if (!is_uncommon_trap_if_pattern(proj)) {
+ if (loop->is_loop_exit(iff)) {
+ // stop processing the remaining projs in the list because the execution of them
+ // depends on the condition of "iff" (iff->in(1)).
+ break;
+ } else {
+ // Both arms are inside the loop. There are two cases:
+ // (1) there is one backward branch. In this case, any remaining proj
+ // in the if_proj list post-dominates "iff". So, the condition of "iff"
+ // does not determine the execution the remining projs directly, and we
+ // can safely continue.
+ // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
+ // does not dominate loop->tail(), so it can not be in the if_proj list.
+ continue;
+ }
+ }
+
+ Node* test = iff->in(1);
+ if (!test->is_Bool()){ //Conv2B, ...
+ continue;
+ }
+ BoolNode* bol = test->as_Bool();
+ if (invar.is_invariant(bol)) {
+ // Invariant test
+ new_predicate_proj = create_new_if_for_predicate(predicate_proj);
+ Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
+ new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
+ if (TraceLoopPredicate) tty->print("invariant");
+ } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
+ // Range check (only for counted loops)
+ new_predicate_proj = create_new_if_for_predicate(predicate_proj);
+ Node *ctrl = new_predicate_proj->in(0)->as_If()->in(0);
+ const Node* cmp = bol->in(1)->as_Cmp();
+ Node* idx = cmp->in(1);
+ assert(!invar.is_invariant(idx), "index is variant");
+ assert(cmp->in(2)->Opcode() == Op_LoadRange, "must be");
+ LoadRangeNode* ld_rng = (LoadRangeNode*)cmp->in(2); // LoadRangeNode
+ assert(invar.is_invariant(ld_rng), "load range must be invariant");
+ ld_rng = (LoadRangeNode*)invar.clone(ld_rng, ctrl);
+ int scale = 1;
+ Node* offset = zero;
+ bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
+ assert(ok, "must be index expression");
+ if (offset && offset != zero) {
+ assert(invar.is_invariant(offset), "offset must be loop invariant");
+ offset = invar.clone(offset, ctrl);
+ }
+ Node* init = cl->init_trip();
+ Node* limit = cl->limit();
+ Node* stride = cl->stride();
+ new_predicate_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, ld_rng);
+ if (TraceLoopPredicate) tty->print("range check");
+ }
+
+ if (new_predicate_proj == NULL) {
+ // The other proj of the "iff" is a uncommon trap projection, and we can assume
+ // the other proj will not be executed ("executed" means uct raised).
+ continue;
+ } else {
+ // Success - attach condition (new_predicate_bol) to predicate if
+ invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
+ IfNode* new_iff = new_predicate_proj->in(0)->as_If();
+
+ // Negate test if necessary
+ if (proj->_con != predicate_proj->_con) {
+ new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
+ register_new_node(new_predicate_bol, new_iff->in(0));
+ if (TraceLoopPredicate) tty->print_cr(" if negated: %d", iff->_idx);
+ } else {
+ if (TraceLoopPredicate) tty->print_cr(" if: %d", iff->_idx);
+ }
+
+ _igvn.hash_delete(new_iff);
+ new_iff->set_req(1, new_predicate_bol);
+
+ _igvn.hash_delete(iff);
+ iff->set_req(1, proj->is_IfFalse() ? cond_false : cond_true);
+
+ Node* ctrl = new_predicate_proj; // new control
+ ProjNode* dp = proj; // old control
+ assert(get_loop(dp) == loop, "guarenteed at the time of collecting proj");
+ // Find nodes (depends only on the test) off the surviving projection;
+ // move them outside the loop with the control of proj_clone
+ for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
+ Node* cd = dp->fast_out(i); // Control-dependent node
+ if (cd->depends_only_on_test()) {
+ assert(cd->in(0) == dp, "");
+ _igvn.hash_delete(cd);
+ cd->set_req(0, ctrl); // ctrl, not NULL
+ set_early_ctrl(cd);
+ _igvn._worklist.push(cd);
+ IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
+ if (new_loop != loop) {
+ if (!loop->_child) loop->_body.yank(cd);
+ if (!new_loop->_child ) new_loop->_body.push(cd);
+ }
+ --i;
+ --imax;
+ }
+ }
+
+ hoisted = true;
+ C->set_major_progress();
+ }
+ } // end while
+
+#ifndef PRODUCT
+ // report that the loop predication has been actually performed
+ // for this loop
+ if (TraceLoopPredicate && hoisted) {
+ tty->print("Loop Predication Performed:");
+ loop->dump_head();
+ }
+#endif
+
+ return hoisted;
+}
+
+//------------------------------loop_predication--------------------------------
+// driver routine for loop predication optimization
+bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
+ bool hoisted = false;
+ // Recursively promote predicates
+ if ( _child ) {
+ hoisted = _child->loop_predication( phase);
+ }
+
+ // self
+ if (!_irreducible && !tail()->is_top()) {
+ hoisted |= phase->loop_predication_impl(this);
+ }
+
+ if ( _next ) { //sibling
+ hoisted |= _next->loop_predication( phase);
+ }
+
+ return hoisted;
+}
--- a/hotspot/src/share/vm/opto/loopnode.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/loopnode.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -1420,11 +1420,57 @@
}
}
+//---------------------collect_potentially_useful_predicates-----------------------
+// Helper function to collect potentially useful predicates to prevent them from
+// being eliminated by PhaseIdealLoop::eliminate_useless_predicates
+void PhaseIdealLoop::collect_potentially_useful_predicates(
+ IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
+ if (loop->_child) { // child
+ collect_potentially_useful_predicates(loop->_child, useful_predicates);
+ }
+
+ // self (only loops that we can apply loop predication may use their predicates)
+ if (loop->_head->is_Loop() &&
+ !loop->_irreducible &&
+ !loop->tail()->is_top()) {
+ LoopNode *lpn = loop->_head->as_Loop();
+ Node* entry = lpn->in(LoopNode::EntryControl);
+ ProjNode *predicate_proj = find_predicate_insertion_point(entry);
+ if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
+ assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
+ useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
+ }
+ }
+
+ if ( loop->_next ) { // sibling
+ collect_potentially_useful_predicates(loop->_next, useful_predicates);
+ }
+}
+
+//------------------------eliminate_useless_predicates-----------------------------
+// Eliminate all inserted predicates if they could not be used by loop predication.
+void PhaseIdealLoop::eliminate_useless_predicates() {
+ if (C->predicate_count() == 0) return; // no predicate left
+
+ Unique_Node_List useful_predicates; // to store useful predicates
+ if (C->has_loops()) {
+ collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
+ }
+
+ for (int i = C->predicate_count(); i > 0; i--) {
+ Node * n = C->predicate_opaque1_node(i-1);
+ assert(n->Opcode() == Op_Opaque1, "must be");
+ if (!useful_predicates.member(n)) { // not in the useful list
+ _igvn.replace_node(n, n->in(1));
+ }
+ }
+}
+
//=============================================================================
//----------------------------build_and_optimize-------------------------------
// Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
// its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
-void PhaseIdealLoop::build_and_optimize(bool do_split_ifs) {
+void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool do_loop_pred) {
int old_progress = C->major_progress();
// Reset major-progress flag for the driver's heuristics
@@ -1577,6 +1623,12 @@
return;
}
+ // some parser-inserted loop predicates could never be used by loop
+ // predication. Eliminate them before loop optimization
+ if (UseLoopPredicate) {
+ eliminate_useless_predicates();
+ }
+
// clear out the dead code
while(_deadlist.size()) {
_igvn.remove_globally_dead_node(_deadlist.pop());
@@ -1603,7 +1655,7 @@
// Because RCE opportunities can be masked by split_thru_phi,
// look for RCE candidates and inhibit split_thru_phi
// on just their loop-phi's for this pass of loop opts
- if( SplitIfBlocks && do_split_ifs ) {
+ if (SplitIfBlocks && do_split_ifs) {
if (lpt->policy_range_check(this)) {
lpt->_rce_candidate = 1; // = true
}
@@ -1619,12 +1671,17 @@
NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
}
+ // Perform loop predication before iteration splitting
+ if (do_loop_pred && C->has_loops() && !C->major_progress()) {
+ _ltree_root->_child->loop_predication(this);
+ }
+
// Perform iteration-splitting on inner loops. Split iterations to avoid
// range checks or one-shot null checks.
// If split-if's didn't hack the graph too bad (no CFG changes)
// then do loop opts.
- if( C->has_loops() && !C->major_progress() ) {
+ if (C->has_loops() && !C->major_progress()) {
memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
_ltree_root->_child->iteration_split( this, worklist );
// No verify after peeling! GCM has hoisted code out of the loop.
@@ -1636,7 +1693,7 @@
// Do verify graph edges in any case
NOT_PRODUCT( C->verify_graph_edges(); );
- if( !do_split_ifs ) {
+ if (!do_split_ifs) {
// We saw major progress in Split-If to get here. We forced a
// pass with unrolling and not split-if, however more split-if's
// might make progress. If the unrolling didn't make progress
@@ -2763,6 +2820,22 @@
Node *legal = LCA; // Walk 'legal' up the IDOM chain
Node *least = legal; // Best legal position so far
while( early != legal ) { // While not at earliest legal
+#ifdef ASSERT
+ if (legal->is_Start() && !early->is_Root()) {
+ // Bad graph. Print idom path and fail.
+ tty->print_cr( "Bad graph detected in build_loop_late");
+ tty->print("n: ");n->dump(); tty->cr();
+ tty->print("early: ");early->dump(); tty->cr();
+ int ct = 0;
+ Node *dbg_legal = LCA;
+ while(!dbg_legal->is_Start() && ct < 100) {
+ tty->print("idom[%d] ",ct); dbg_legal->dump(); tty->cr();
+ ct++;
+ dbg_legal = idom(dbg_legal);
+ }
+ assert(false, "Bad graph detected in build_loop_late");
+ }
+#endif
// Find least loop nesting depth
legal = idom(legal); // Bump up the IDOM tree
// Check for lower nesting depth
--- a/hotspot/src/share/vm/opto/loopnode.hpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/loopnode.hpp Tue Jan 12 14:37:35 2010 -0800
@@ -30,6 +30,7 @@
class Node;
class PhaseIdealLoop;
class VectorSet;
+class Invariance;
struct small_cache;
//
@@ -325,6 +326,10 @@
// Returns TRUE if loop tree is structurally changed.
bool beautify_loops( PhaseIdealLoop *phase );
+ // Perform optimization to use the loop predicates for null checks and range checks.
+ // Applies to any loop level (not just the innermost one)
+ bool loop_predication( PhaseIdealLoop *phase);
+
// Perform iteration-splitting on inner loops. Split iterations to
// avoid range checks or one-shot null checks. Returns false if the
// current round of loop opts should stop.
@@ -395,6 +400,9 @@
// into longer memory ops, we may want to increase alignment.
bool policy_align( PhaseIdealLoop *phase ) const;
+ // Return TRUE if "iff" is a range check.
+ bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
+
// Compute loop trip count from profile data
void compute_profile_trip_cnt( PhaseIdealLoop *phase );
@@ -521,9 +529,6 @@
}
Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
- // true if CFG node d dominates CFG node n
- bool is_dominator(Node *d, Node *n);
-
// Helper function for directing control inputs away from CFG split
// points.
Node *find_non_split_ctrl( Node *ctrl ) const {
@@ -572,6 +577,17 @@
assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
return n;
}
+ // true if CFG node d dominates CFG node n
+ bool is_dominator(Node *d, Node *n);
+ // return get_ctrl for a data node and self(n) for a CFG node
+ Node* ctrl_or_self(Node* n) {
+ if (has_ctrl(n))
+ return get_ctrl(n);
+ else {
+ assert (n->is_CFG(), "must be a CFG node");
+ return n;
+ }
+ }
private:
Node *get_ctrl_no_update( Node *i ) const {
@@ -600,7 +616,7 @@
// Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
// the 'old_node' with 'new_node'. Kill old-node. Add a reference
// from old_node to new_node to support the lazy update. Reference
- // replaces loop reference, since that is not neede for dead node.
+ // replaces loop reference, since that is not needed for dead node.
public:
void lazy_update( Node *old_node, Node *new_node ) {
assert( old_node != new_node, "no cycles please" );
@@ -679,11 +695,11 @@
_dom_lca_tags(C->comp_arena()),
_verify_me(NULL),
_verify_only(true) {
- build_and_optimize(false);
+ build_and_optimize(false, false);
}
// build the loop tree and perform any requested optimizations
- void build_and_optimize(bool do_split_if);
+ void build_and_optimize(bool do_split_if, bool do_loop_pred);
public:
// Dominators for the sea of nodes
@@ -694,13 +710,13 @@
Node *dom_lca_internal( Node *n1, Node *n2 ) const;
// Compute the Ideal Node to Loop mapping
- PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs) :
+ PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) :
PhaseTransform(Ideal_Loop),
_igvn(igvn),
_dom_lca_tags(C->comp_arena()),
_verify_me(NULL),
_verify_only(false) {
- build_and_optimize(do_split_ifs);
+ build_and_optimize(do_split_ifs, do_loop_pred);
}
// Verify that verify_me made the same decisions as a fresh run.
@@ -710,7 +726,7 @@
_dom_lca_tags(C->comp_arena()),
_verify_me(verify_me),
_verify_only(false) {
- build_and_optimize(false);
+ build_and_optimize(false, false);
}
// Build and verify the loop tree without modifying the graph. This
@@ -790,6 +806,30 @@
// Return true if exp is a scaled induction var plus (or minus) constant
bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
+ // Return true if proj is for "proj->[region->..]call_uct"
+ bool is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate = false);
+ // Return true for "if(test)-> proj -> ...
+ // |
+ // V
+ // other_proj->[region->..]call_uct"
+ bool is_uncommon_trap_if_pattern(ProjNode* proj, bool must_reason_predicate = false);
+ // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
+ ProjNode* create_new_if_for_predicate(ProjNode* cont_proj);
+ // Find a good location to insert a predicate
+ ProjNode* find_predicate_insertion_point(Node* start_c);
+ // Construct a range check for a predicate if
+ BoolNode* rc_predicate(Node* ctrl,
+ int scale, Node* offset,
+ Node* init, Node* limit, Node* stride,
+ Node* range);
+
+ // Implementation of the loop predication to promote checks outside the loop
+ bool loop_predication_impl(IdealLoopTree *loop);
+
+ // Helper function to collect predicate for eliminating the useless ones
+ void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
+ void eliminate_useless_predicates();
+
// Eliminate range-checks and other trip-counter vs loop-invariant tests.
void do_range_check( IdealLoopTree *loop, Node_List &old_new );
@@ -906,7 +946,6 @@
const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
// Helper functions
- void register_new_node( Node *n, Node *blk );
Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
void handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true );
@@ -918,6 +957,7 @@
public:
void set_created_loop_node() { _created_loop_node = true; }
bool created_loop_node() { return _created_loop_node; }
+ void register_new_node( Node *n, Node *blk );
#ifndef PRODUCT
void dump( ) const;
--- a/hotspot/src/share/vm/opto/parse.hpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/parse.hpp Tue Jan 12 14:37:35 2010 -0800
@@ -430,6 +430,11 @@
}
}
+ // Return true if the parser should add a loop predicate
+ bool should_add_predicate(int target_bci);
+ // Insert a loop predicate into the graph
+ void add_predicate();
+
// Note: Intrinsic generation routines may be found in library_call.cpp.
// Helper function to setup Ideal Call nodes
@@ -491,7 +496,7 @@
void do_ifnull(BoolTest::mask btest, Node* c);
void do_if(BoolTest::mask btest, Node* c);
- void repush_if_args();
+ int repush_if_args();
void adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
Block* path, Block* other_path);
IfNode* jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask);
--- a/hotspot/src/share/vm/opto/parse1.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/parse1.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -1383,6 +1383,10 @@
set_parse_bci(iter().cur_bci());
if (bci() == block()->limit()) {
+ // insert a predicate if it falls through to a loop head block
+ if (should_add_predicate(bci())){
+ add_predicate();
+ }
// Do not walk into the next block until directed by do_all_blocks.
merge(bci());
break;
@@ -2083,6 +2087,37 @@
}
}
+//------------------------------should_add_predicate--------------------------
+bool Parse::should_add_predicate(int target_bci) {
+ if (!UseLoopPredicate) return false;
+ Block* target = successor_for_bci(target_bci);
+ if (target != NULL &&
+ target->is_loop_head() &&
+ block()->rpo() < target->rpo()) {
+ return true;
+ }
+ return false;
+}
+
+//------------------------------add_predicate---------------------------------
+void Parse::add_predicate() {
+ assert(UseLoopPredicate,"use only for loop predicate");
+ Node *cont = _gvn.intcon(1);
+ Node* opq = _gvn.transform(new (C, 2) Opaque1Node(C, cont));
+ Node *bol = _gvn.transform(new (C, 2) Conv2BNode(opq));
+ IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
+ Node* iffalse = _gvn.transform(new (C, 1) IfFalseNode(iff));
+ C->add_predicate_opaq(opq);
+ {
+ PreserveJVMState pjvms(this);
+ set_control(iffalse);
+ uncommon_trap(Deoptimization::Reason_predicate,
+ Deoptimization::Action_maybe_recompile);
+ }
+ Node* iftrue = _gvn.transform(new (C, 1) IfTrueNode(iff));
+ set_control(iftrue);
+}
+
#ifndef PRODUCT
//------------------------show_parse_info--------------------------------------
void Parse::show_parse_info() {
--- a/hotspot/src/share/vm/opto/parse2.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/parse2.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -278,6 +278,11 @@
if (len < 1) {
// If this is a backward branch, add safepoint
maybe_add_safepoint(default_dest);
+ if (should_add_predicate(default_dest)){
+ _sp += 1; // set original stack for use by uncommon_trap
+ add_predicate();
+ _sp -= 1;
+ }
merge(default_dest);
return;
}
@@ -324,6 +329,11 @@
if (len < 1) { // If this is a backward branch, add safepoint
maybe_add_safepoint(default_dest);
+ if (should_add_predicate(default_dest)){
+ _sp += 1; // set original stack for use by uncommon_trap
+ add_predicate();
+ _sp -= 1;
+ }
merge(default_dest);
return;
}
@@ -731,6 +741,9 @@
push(_gvn.makecon(ret_addr));
// Flow to the jsr.
+ if (should_add_predicate(jsr_bci)){
+ add_predicate();
+ }
merge(jsr_bci);
}
@@ -881,7 +894,7 @@
//-------------------------------repush_if_args--------------------------------
// Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
-inline void Parse::repush_if_args() {
+inline int Parse::repush_if_args() {
#ifndef PRODUCT
if (PrintOpto && WizardMode) {
tty->print("defending against excessive implicit null exceptions on %s @%d in ",
@@ -895,6 +908,7 @@
assert(argument(0) != NULL, "must exist");
assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
_sp += bc_depth;
+ return bc_depth;
}
//----------------------------------do_ifnull----------------------------------
@@ -954,8 +968,14 @@
// Update method data
profile_taken_branch(target_bci);
adjust_map_after_if(btest, c, prob, branch_block, next_block);
- if (!stopped())
+ if (!stopped()) {
+ if (should_add_predicate(target_bci)){ // add a predicate if it branches to a loop
+ int nargs = repush_if_args(); // set original stack for uncommon_trap
+ add_predicate();
+ _sp -= nargs;
+ }
merge(target_bci);
+ }
}
}
@@ -1076,8 +1096,14 @@
// Update method data
profile_taken_branch(target_bci);
adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
- if (!stopped())
+ if (!stopped()) {
+ if (should_add_predicate(target_bci)){ // add a predicate if it branches to a loop
+ int nargs = repush_if_args(); // set original stack for the uncommon_trap
+ add_predicate();
+ _sp -= nargs;
+ }
merge(target_bci);
+ }
}
}
@@ -2080,6 +2106,10 @@
// Update method data
profile_taken_branch(target_bci);
+ // Add loop predicate if it goes to a loop
+ if (should_add_predicate(target_bci)){
+ add_predicate();
+ }
// Merge the current control into the target basic block
merge(target_bci);
--- a/hotspot/src/share/vm/opto/split_if.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/opto/split_if.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -219,6 +219,7 @@
//------------------------------register_new_node------------------------------
void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) {
+ assert(!n->is_CFG(), "must be data node");
_igvn.register_new_node_with_optimizer(n);
set_ctrl(n, blk);
IdealLoopTree *loop = get_loop(blk);
--- a/hotspot/src/share/vm/runtime/deoptimization.cpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/runtime/deoptimization.cpp Tue Jan 12 14:37:35 2010 -0800
@@ -1672,7 +1672,8 @@
"unhandled",
"constraint",
"div0_check",
- "age"
+ "age",
+ "predicate"
};
const char* Deoptimization::_trap_action_name[Action_LIMIT] = {
// Note: Keep this in sync. with enum DeoptAction.
--- a/hotspot/src/share/vm/runtime/deoptimization.hpp Sat Jan 09 00:59:35 2010 -0800
+++ b/hotspot/src/share/vm/runtime/deoptimization.hpp Tue Jan 12 14:37:35 2010 -0800
@@ -46,6 +46,7 @@
Reason_constraint, // arbitrary runtime constraint violated
Reason_div0_check, // a null_check due to division by zero
Reason_age, // nmethod too old; tier threshold reached
+ Reason_predicate, // compiler generated predicate failed
Reason_LIMIT,
// Note: Keep this enum in sync. with _trap_reason_name.
Reason_RECORDED_LIMIT = Reason_unloaded // some are not recorded per bc