#ifdef ASSERT
if( n->is_Mem() ) {
int alias_idx = get_alias_index(n->as_Mem()->adr_type());
assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
// oop will be recorded in oop map if load crosses safepoint
n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
LoadNode::is_immutable_value(n->in(MemNode::Address))),
"raw memory operations should have control edge");
}
if (n->is_MemBar()) {
MemBarNode* mb = n->as_MemBar();
if (mb->trailing_store() || mb->trailing_load_store()) {
assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
Node* mem = mb->in(MemBarNode::Precedent);
assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
(mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
} else if (mb->leading()) {
assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
}
}
#endif
// Count FPU ops and common calls, implements item (3)
switch( nop ) {
// Count all float operations that may use FPU
case Op_AddF:
case Op_SubF:
case Op_MulF:
case Op_DivF:
case Op_NegF:
case Op_ModF:
case Op_ConvI2F:
case Op_ConF:
case Op_CmpF:
case Op_CmpF3:
// case Op_ConvL2F: // longs are split into 32-bit halves
frc.inc_float_count();
break;
case Op_ConvF2D:
case Op_ConvD2F:
frc.inc_float_count();
frc.inc_double_count();
break;
// Count all double operations that may use FPU
case Op_AddD:
case Op_SubD:
case Op_MulD:
case Op_DivD:
case Op_NegD:
case Op_ModD:
case Op_ConvI2D:
case Op_ConvD2I:
// case Op_ConvL2D: // handled by leaf call
// case Op_ConvD2L: // handled by leaf call
case Op_ConD:
case Op_CmpD:
case Op_CmpD3:
frc.inc_double_count();
break;
case Op_Opaque1: // Remove Opaque Nodes before matching
case Op_Opaque2: // Remove Opaque Nodes before matching
case Op_Opaque3:
n->subsume_by(n->in(1), this);
break;
case Op_CallStaticJava:
case Op_CallJava:
case Op_CallDynamicJava:
frc.inc_java_call_count(); // Count java call site;
case Op_CallRuntime:
case Op_CallLeaf:
case Op_CallLeafNoFP: {
assert( n->is_Call(), "" );
CallNode *call = n->as_Call();
// Count call sites where the FP mode bit would have to be flipped.
// Do not count uncommon runtime calls:
// uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
// _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
frc.inc_call_count(); // Count the call site
} else { // See if uncommon argument is shared
Node *n = call->in(TypeFunc::Parms);
int nop = n->Opcode();
// Clone shared simple arguments to uncommon calls, item (1).
if( n->outcnt() > 1 &&
!n->is_Proj() &&
nop != Op_CreateEx &&
nop != Op_CheckCastPP &&
nop != Op_DecodeN &&
nop != Op_DecodeNKlass &&
!n->is_Mem() ) {
Node *x = n->clone();
call->set_req( TypeFunc::Parms, x );
}
}
break;
}
case Op_StoreD:
case Op_LoadD:
case Op_LoadD_unaligned:
frc.inc_double_count();
goto handle_mem;
case Op_StoreF:
case Op_LoadF:
frc.inc_float_count();
goto handle_mem;
case Op_StoreCM:
{
// Convert OopStore dependence into precedence edge
Node* prec = n->in(MemNode::OopStore);
n->del_req(MemNode::OopStore);
n->add_prec(prec);
eliminate_redundant_card_marks(n);
}
// fall through
case Op_StoreB:
case Op_StoreC:
case Op_StorePConditional:
case Op_StoreI:
case Op_StoreL:
case Op_StoreIConditional:
case Op_StoreLConditional:
case Op_CompareAndSwapI:
case Op_CompareAndSwapL:
case Op_CompareAndSwapP:
case Op_CompareAndSwapN:
case Op_GetAndAddI:
case Op_GetAndAddL:
case Op_GetAndSetI:
case Op_GetAndSetL:
case Op_GetAndSetP:
case Op_GetAndSetN:
case Op_StoreP:
case Op_StoreN:
case Op_StoreNKlass:
case Op_LoadB:
case Op_LoadUB:
case Op_LoadUS:
case Op_LoadI:
case Op_LoadKlass:
case Op_LoadNKlass:
case Op_LoadL:
case Op_LoadL_unaligned:
case Op_LoadPLocked:
case Op_LoadP:
case Op_LoadN:
case Op_LoadRange:
case Op_LoadS: {
handle_mem:
#ifdef ASSERT
if( VerifyOptoOopOffsets ) {
assert( n->is_Mem(), "" );
MemNode *mem = (MemNode*)n;
// Check to see if address types have grounded out somehow.
const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
assert( !tp || oop_offset_is_sane(tp), "" );
}
#endif
break;
}
case Op_AddP: { // Assert sane base pointers
Node *addp = n->in(AddPNode::Address);
assert( !addp->is_AddP() ||
addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
addp->in(AddPNode::Base) == n->in(AddPNode::Base),
"Base pointers must match" );
#ifdef _LP64
if ((UseCompressedOops || UseCompressedClassPointers) &&
addp->Opcode() == Op_ConP &&
addp == n->in(AddPNode::Base) &&
n->in(AddPNode::Offset)->is_Con()) {
// Use addressing with narrow klass to load with offset on x86.
// On sparc loading 32-bits constant and decoding it have less
// instructions (4) then load 64-bits constant (7).
// Do this transformation here since IGVN will convert ConN back to ConP.
const Type* t = addp->bottom_type();
if (t->isa_oopptr() || t->isa_klassptr()) {
Node* nn = NULL;
int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
// Look for existing ConN node of the same exact type.
Node* r = root();
uint cnt = r->outcnt();
for (uint i = 0; i < cnt; i++) {
Node* m = r->raw_out(i);
if (m!= NULL && m->Opcode() == op &&
m->bottom_type()->make_ptr() == t) {
nn = m;
break;
}
}
if (nn != NULL) {
// Decode a narrow oop to match address
// [R12 + narrow_oop_reg<<3 + offset]
if (t->isa_oopptr()) {
nn = new (this) DecodeNNode(nn, t);
} else {
nn = new (this) DecodeNKlassNode(nn, t);
}
n->set_req(AddPNode::Base, nn);
n->set_req(AddPNode::Address, nn);
if (addp->outcnt() == 0) {
addp->disconnect_inputs(NULL, this);
}
}
}
}
#endif
break;
}
#ifdef _LP64
case Op_CastPP:
if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
Node* in1 = n->in(1);
const Type* t = n->bottom_type();
Node* new_in1 = in1->clone();
new_in1->as_DecodeN()->set_type(t);
if (!Matcher::narrow_oop_use_complex_address()) {
//
// x86, ARM and friends can handle 2 adds in addressing mode
// and Matcher can fold a DecodeN node into address by using
// a narrow oop directly and do implicit NULL check in address:
//
// [R12 + narrow_oop_reg<<3 + offset]
// NullCheck narrow_oop_reg
//
// On other platforms (Sparc) we have to keep new DecodeN node and
// use it to do implicit NULL check in address:
//
// decode_not_null narrow_oop_reg, base_reg
// [base_reg + offset]
// NullCheck base_reg
//
// Pin the new DecodeN node to non-null path on these platform (Sparc)
// to keep the information to which NULL check the new DecodeN node
// corresponds to use it as value in implicit_null_check().
//
new_in1->set_req(0, n->in(0));
}
n->subsume_by(new_in1, this);
if (in1->outcnt() == 0) {
in1->disconnect_inputs(NULL, this);
}
}
break;
case Op_CmpP:
// Do this transformation here to preserve CmpPNode::sub() and
// other TypePtr related Ideal optimizations (for example, ptr nullness).
if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
Node* in1 = n->in(1);
Node* in2 = n->in(2);
if (!in1->is_DecodeNarrowPtr()) {
in2 = in1;
in1 = n->in(2);
}
assert(in1->is_DecodeNarrowPtr(), "sanity");
Node* new_in2 = NULL;
if (in2->is_DecodeNarrowPtr()) {
assert(in2->Opcode() == in1->Opcode(), "must be same node type");
new_in2 = in2->in(1);
} else if (in2->Opcode() == Op_ConP) {
const Type* t = in2->bottom_type();
if (t == TypePtr::NULL_PTR) {
assert(in1->is_DecodeN(), "compare klass to null?");
// Don't convert CmpP null check into CmpN if compressed
// oops implicit null check is not generated.
// This will allow to generate normal oop implicit null check.
if (Matcher::gen_narrow_oop_implicit_null_checks())
new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
//
// This transformation together with CastPP transformation above
// will generated code for implicit NULL checks for compressed oops.
//
// The original code after Optimize()
//
// LoadN memory, narrow_oop_reg
// decode narrow_oop_reg, base_reg
// CmpP base_reg, NULL
// CastPP base_reg // NotNull
// Load [base_reg + offset], val_reg
//
// after these transformations will be
//
// LoadN memory, narrow_oop_reg
// CmpN narrow_oop_reg, NULL
// decode_not_null narrow_oop_reg, base_reg
// Load [base_reg + offset], val_reg
//
// and the uncommon path (== NULL) will use narrow_oop_reg directly
// since narrow oops can be used in debug info now (see the code in
// final_graph_reshaping_walk()).
//
// At the end the code will be matched to
// on x86:
//
// Load_narrow_oop memory, narrow_oop_reg
// Load [R12 + narrow_oop_reg<<3 + offset], val_reg
// NullCheck narrow_oop_reg
//
// and on sparc:
//
// Load_narrow_oop memory, narrow_oop_reg
// decode_not_null narrow_oop_reg, base_reg
// Load [base_reg + offset], val_reg
// NullCheck base_reg
//
} else if (t->isa_oopptr()) {
new_in2 = ConNode::make(this, t->make_narrowoop());
} else if (t->isa_klassptr()) {
new_in2 = ConNode::make(this, t->make_narrowklass());
}
}
if (new_in2 != NULL) {
Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
n->subsume_by(cmpN, this);
if (in1->outcnt() == 0) {
in1->disconnect_inputs(NULL, this);
}
if (in2->outcnt() == 0) {
in2->disconnect_inputs(NULL, this);
}
}
}
break;
case Op_DecodeN:
case Op_DecodeNKlass:
assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
// DecodeN could be pinned when it can't be fold into
// an address expression, see the code for Op_CastPP above.
assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
break;
case Op_EncodeP:
case Op_EncodePKlass: {
Node* in1 = n->in(1);
if (in1->is_DecodeNarrowPtr()) {
n->subsume_by(in1->in(1), this);
} else if (in1->Opcode() == Op_ConP) {
const Type* t = in1->bottom_type();
if (t == TypePtr::NULL_PTR) {
assert(t->isa_oopptr(), "null klass?");
n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
} else if (t->isa_oopptr()) {
n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
} else if (t->isa_klassptr()) {
n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
}
}
if (in1->outcnt() == 0) {
in1->disconnect_inputs(NULL, this);
}
break;
}
case Op_Proj: {
if (OptimizeStringConcat) {
ProjNode* p = n->as_Proj();
if (p->_is_io_use) {
// Separate projections were used for the exception path which
// are normally removed by a late inline. If it wasn't inlined
// then they will hang around and should just be replaced with
// the original one.
Node* proj = NULL;
// Replace with just one
for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
Node *use = i.get();
if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
proj = use;
break;
}
}
assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
if (proj != NULL) {
p->subsume_by(proj, this);
}
}
}
break;
}
case Op_Phi:
if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
// The EncodeP optimization may create Phi with the same edges
// for all paths. It is not handled well by Register Allocator.
Node* unique_in = n->in(1);
assert(unique_in != NULL, "");
uint cnt = n->req();
for (uint i = 2; i < cnt; i++) {
Node* m = n->in(i);
assert(m != NULL, "");
if (unique_in != m)
unique_in = NULL;
}
if (unique_in != NULL) {
n->subsume_by(unique_in, this);
}
}
break;
#endif
#ifdef ASSERT
case Op_CastII:
// Verify that all range check dependent CastII nodes were removed.
if (n->isa_CastII()->has_range_check()) {
n->dump(3);
assert(false, "Range check dependent CastII node was not removed");
}
break;
#endif
case Op_ModI:
if (UseDivMod) {
// Check if a%b and a/b both exist
Node* d = n->find_similar(Op_DivI);
if (d) {
// Replace them with a fused divmod if supported
if (Matcher::has_match_rule(Op_DivModI)) {
DivModINode* divmod = DivModINode::make(this, n);
d->subsume_by(divmod->div_proj(), this);
n->subsume_by(divmod->mod_proj(), this);
} else {
// replace a%b with a-((a/b)*b)
Node* mult = new (this) MulINode(d, d->in(2));
Node* sub = new (this) SubINode(d->in(1), mult);
n->subsume_by(sub, this);
}
}
}
break;
case Op_ModL:
if (UseDivMod) {
// Check if a%b and a/b both exist
Node* d = n->find_similar(Op_DivL);
if (d) {
// Replace them with a fused divmod if supported
if (Matcher::has_match_rule(Op_DivModL)) {
DivModLNode* divmod = DivModLNode::make(this, n);
d->subsume_by(divmod->div_proj(), this);
n->subsume_by(divmod->mod_proj(), this);
} else {
// replace a%b with a-((a/b)*b)
Node* mult = new (this) MulLNode(d, d->in(2));
Node* sub = new (this) SubLNode(d->in(1), mult);
n->subsume_by(sub, this);
}
}
}
break;
case Op_LoadVector:
case Op_StoreVector:
break;
case Op_PackB:
case Op_PackS:
case Op_PackI:
case Op_PackF:
case Op_PackL:
case Op_PackD:
if (n->req()-1 > 2) {
// Replace many operand PackNodes with a binary tree for matching
PackNode* p = (PackNode*) n;
Node* btp = p->binary_tree_pack(this, 1, n->req());
n->subsume_by(btp, this);
}
break;
case Op_Loop:
case Op_CountedLoop:
if (n->as_Loop()->is_inner_loop()) {
frc.inc_inner_loop_count();
}
break;
case Op_LShiftI:
case Op_RShiftI:
case Op_URShiftI:
case Op_LShiftL:
case Op_RShiftL:
case Op_URShiftL:
if (Matcher::need_masked_shift_count) {
// The cpu's shift instructions don't restrict the count to the
// lower 5/6 bits. We need to do the masking ourselves.
Node* in2 = n->in(2);
juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
const TypeInt* t = in2->find_int_type();
if (t != NULL && t->is_con()) {
juint shift = t->get_con();
if (shift > mask) { // Unsigned cmp
n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
}
} else {
if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
n->set_req(2, shift);
}
}
if (in2->outcnt() == 0) { // Remove dead node
in2->disconnect_inputs(NULL, this);
}
}
break;
case Op_MemBarStoreStore:
case Op_MemBarRelease:
// Break the link with AllocateNode: it is no longer useful and
// confuses register allocation.
if (n->req() > MemBarNode::Precedent) {
n->set_req(MemBarNode::Precedent, top());
}
break;
default:
assert( !n->is_Call(), "" );
assert( !n->is_Mem(), "" );
assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
break;
}
// Collect CFG split points
if (n->is_MultiBranch())
frc._tests.push(n);
}
//------------------------------final_graph_reshaping_walk---------------------
// Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
// requires that the walk visits a node's inputs before visiting the node.
void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
ResourceArea *area = Thread::current()->resource_area();
Unique_Node_List sfpt(area);
frc._visited.set(root->_idx); // first, mark node as visited
uint cnt = root->req();
Node *n = root;
uint i = 0;
while (true) {
if (i < cnt) {
// Place all non-visited non-null inputs onto stack
Node* m = n->in(i);
++i;
if (m != NULL && !frc._visited.test_set(m->_idx)) {
if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
// compute worst case interpreter size in case of a deoptimization
update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
sfpt.push(m);
}
cnt = m->req();
nstack.push(n, i); // put on stack parent and next input's index
n = m;
i = 0;
}
} else {
// Now do post-visit work
final_graph_reshaping_impl( n, frc );
if (nstack.is_empty())
break; // finished
n = nstack.node(); // Get node from stack
cnt = n->req();
i = nstack.index();
nstack.pop(); // Shift to the next node on stack
}
}
// Skip next transformation if compressed oops are not used.
if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
(!UseCompressedOops && !UseCompressedClassPointers))
return;
// Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
// It could be done for an uncommon traps or any safepoints/calls
// if the DecodeN/DecodeNKlass node is referenced only in a debug info.
while (sfpt.size() > 0) {
n = sfpt.pop();
JVMState *jvms = n->as_SafePoint()->jvms();
assert(jvms != NULL, "sanity");
int start = jvms->debug_start();
int end = n->req();
bool is_uncommon = (n->is_CallStaticJava() &&
n->as_CallStaticJava()->uncommon_trap_request() != 0);
for (int j = start; j < end; j++) {
Node* in = n->in(j);
if (in->is_DecodeNarrowPtr()) {
bool safe_to_skip = true;
if (!is_uncommon ) {
// Is it safe to skip?
for (uint i = 0; i < in->outcnt(); i++) {
Node* u = in->raw_out(i);
if (!u->is_SafePoint() ||
u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
safe_to_skip = false;
}
}
}
if (safe_to_skip) {
n->set_req(j, in->in(1));
}
if (in->outcnt() == 0) {
in->disconnect_inputs(NULL, this);
}
}
}
}
}
//------------------------------final_graph_reshaping--------------------------
// Final Graph Reshaping.
//
// (1) Clone simple inputs to uncommon calls, so they can be scheduled late
// and not commoned up and forced early. Must come after regular
// optimizations to avoid GVN undoing the cloning. Clone constant
// inputs to Loop Phis; these will be split by the allocator anyways.
// Remove Opaque nodes.
// (2) Move last-uses by commutative operations to the left input to encourage
// Intel update-in-place two-address operations and better register usage
// on RISCs. Must come after regular optimizations to avoid GVN Ideal
// calls canonicalizing them back.
// (3) Count the number of double-precision FP ops, single-precision FP ops
// and call sites. On Intel, we can get correct rounding either by
// forcing singles to memory (requires extra stores and loads after each
// FP bytecode) or we can set a rounding mode bit (requires setting and
// clearing the mode bit around call sites). The mode bit is only used
// if the relative frequency of single FP ops to calls is low enough.
// This is a key transform for SPEC mpeg_audio.
// (4) Detect infinite loops; blobs of code reachable from above but not
// below. Several of the Code_Gen algorithms fail on such code shapes,
// so we simply bail out. Happens a lot in ZKM.jar, but also happens
// from time to time in other codes (such as -Xcomp finalizer loops, etc).
// Detection is by looking for IfNodes where only 1 projection is
// reachable from below or CatchNodes missing some targets.
// (5) Assert for insane oop offsets in debug mode.
bool Compile::final_graph_reshaping() {
// an infinite loop may have been eliminated by the optimizer,
// in which case the graph will be empty.
if (root()->req() == 1) {
record_method_not_compilable("trivial infinite loop");
return true;
}
// Expensive nodes have their control input set to prevent the GVN
// from freely commoning them. There's no GVN beyond this point so
// no need to keep the control input. We want the expensive nodes to
// be freely moved to the least frequent code path by gcm.
assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
for (int i = 0; i < expensive_count(); i++) {
_expensive_nodes->at(i)->set_req(0, NULL);
}
Final_Reshape_Counts frc;
// Visit everybody reachable!
// Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
Node_Stack nstack(live_nodes() >> 1);
final_graph_reshaping_walk(nstack, root(), frc);
// Check for unreachable (from below) code (i.e., infinite loops).
for( uint i = 0; i < frc._tests.size(); i++ ) {
MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
// Get number of CFG targets.
// Note that PCTables include exception targets after calls.
uint required_outcnt = n->required_outcnt();
if (n->outcnt() != required_outcnt) {
// Check for a few special cases. Rethrow Nodes never take the
// 'fall-thru' path, so expected kids is 1 less.
if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
if (n->in(0)->in(0)->is_Call()) {
CallNode *call = n->in(0)->in(0)->as_Call();
if (call->entry_point() == OptoRuntime::rethrow_stub()) {
required_outcnt--; // Rethrow always has 1 less kid
} else if (call->req() > TypeFunc::Parms &&
call->is_CallDynamicJava()) {
// Check for null receiver. In such case, the optimizer has
// detected that the virtual call will always result in a null
// pointer exception. The fall-through projection of this CatchNode
// will not be populated.
Node *arg0 = call->in(TypeFunc::Parms);
if (arg0->is_Type() &&
arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
required_outcnt--;
}
} else if (call->entry_point() == OptoRuntime::new_array_Java() &&
call->req() > TypeFunc::Parms+1 &&
call->is_CallStaticJava()) {
// Check for negative array length. In such case, the optimizer has
// detected that the allocation attempt will always result in an
// exception. There is no fall-through projection of this CatchNode .
Node *arg1 = call->in(TypeFunc::Parms+1);
if (arg1->is_Type() &&
arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
required_outcnt--;
}
}
}
}
// Recheck with a better notion of 'required_outcnt'
if (n->outcnt() != required_outcnt) {
record_method_not_compilable("malformed control flow");
return true; // Not all targets reachable!
}
}
// Check that I actually visited all kids. Unreached kids
// must be infinite loops.
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
if (!frc._visited.test(n->fast_out(j)->_idx)) {
record_method_not_compilable("infinite loop");
return true; // Found unvisited kid; must be unreach
}
}
// If original bytecodes contained a mixture of floats and doubles
// check if the optimizer has made it homogenous, item (3).
if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
frc.get_float_count() > 32 &&
frc.get_double_count() == 0 &&
(10 * frc.get_call_count() < frc.get_float_count()) ) {
set_24_bit_selection_and_mode( false, true );
}
set_java_calls(frc.get_java_call_count());
set_inner_loops(frc.get_inner_loop_count());
// No infinite loops, no reason to bail out.
return false;
}
//-----------------------------too_many_traps----------------------------------
// Report if there are too many traps at the current method and bci.
// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
bool Compile::too_many_traps(ciMethod* method,
int bci,
Deoptimization::DeoptReason reason) {
ciMethodData* md = method->method_data();
if (md->is_empty()) {
// Assume the trap has not occurred, or that it occurred only
// because of a transient condition during start-up in the interpreter.
return false;
}
ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
if (md->has_trap_at(bci, m, reason) != 0) {
// Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
// Also, if there are multiple reasons, or if there is no per-BCI record,
// assume the worst.
if (log())
log()->elem("observe trap='%s' count='%d'",
Deoptimization::trap_reason_name(reason),
md->trap_count(reason));
return true;
} else {
// Ignore method/bci and see if there have been too many globally.
return too_many_traps(reason, md);
}
}
// Less-accurate variant which does not require a method and bci.
bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
ciMethodData* logmd) {
if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
// Too many traps globally.
// Note that we use cumulative trap_count, not just md->trap_count.
if (log()) {
int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
Deoptimization::trap_reason_name(reason),
mcount, trap_count(reason));
}
return true;
} else {
// The coast is clear.
return false;
}
}
//--------------------------too_many_recompiles--------------------------------
// Report if there are too many recompiles at the current method and bci.
// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
// Is not eager to return true, since this will cause the compiler to use
// Action_none for a trap point, to avoid too many recompilations.
bool Compile::too_many_recompiles(ciMethod* method,
int bci,
Deoptimization::DeoptReason reason) {
ciMethodData* md = method->method_data();
if (md->is_empty()) {
// Assume the trap has not occurred, or that it occurred only
// because of a transient condition during start-up in the interpreter.
return false;
}
// Pick a cutoff point well within PerBytecodeRecompilationCutoff.
uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
Deoptimization::DeoptReason per_bc_reason
= Deoptimization::reason_recorded_per_bytecode_if_any(reason);
ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
if ((per_bc_reason == Deoptimization::Reason_none
|| md->has_trap_at(bci, m, reason) != 0)
// The trap frequency measure we care about is the recompile count:
&& md->trap_recompiled_at(bci, m)
&& md->overflow_recompile_count() >= bc_cutoff) {
// Do not emit a trap here if it has already caused recompilations.
// Also, if there are multiple reasons, or if there is no per-BCI record,
// assume the worst.
if (log())
log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
Deoptimization::trap_reason_name(reason),
md->trap_count(reason),
md->overflow_recompile_count());
return true;
} else if (trap_count(reason) != 0
&& decompile_count() >= m_cutoff) {
// Too many recompiles globally, and we have seen this sort of trap.
// Use cumulative decompile_count, not just md->decompile_count.
if (log())
log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
Deoptimization::trap_reason_name(reason),
md->trap_count(reason), trap_count(reason),
md->decompile_count(), decompile_count());
return true;
} else {
// The coast is clear.
return false;
}
}
// Compute when not to trap. Used by matching trap based nodes and
// NullCheck optimization.
void Compile::set_allowed_deopt_reasons() {
_allowed_reasons = 0;
if (is_method_compilation()) {
for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
assert(rs < BitsPerInt, "recode bit map");
if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
_allowed_reasons |= nth_bit(rs);
}
}
}
}
#ifndef PRODUCT
//------------------------------verify_graph_edges---------------------------
// Walk the Graph and verify that there is a one-to-one correspondence
// between Use-Def edges and Def-Use edges in the graph.
void Compile::verify_graph_edges(bool no_dead_code) {
if (VerifyGraphEdges) {
ResourceArea *area = Thread::current()->resource_area();
Unique_Node_List visited(area);
// Call recursive graph walk to check edges
_root->verify_edges(visited);
if (no_dead_code) {
// Now make sure that no visited node is used by an unvisited node.
bool dead_nodes = false;
Unique_Node_List checked(area);
while (visited.size() > 0) {
Node* n = visited.pop();
checked.push(n);
for (uint i = 0; i < n->outcnt(); i++) {
Node* use = n->raw_out(i);
if (checked.member(use)) continue; // already checked
if (visited.member(use)) continue; // already in the graph
if (use->is_Con()) continue; // a dead ConNode is OK
// At this point, we have found a dead node which is DU-reachable.
if (!dead_nodes) {
tty->print_cr("*** Dead nodes reachable via DU edges:");
dead_nodes = true;
}
use->dump(2);
tty->print_cr("---");
checked.push(use); // No repeats; pretend it is now checked.
}
}
assert(!dead_nodes, "using nodes must be reachable from root");
}
}
}
// Verify GC barriers consistency
// Currently supported:
// - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
void Compile::verify_barriers() {
if (UseG1GC) {
// Verify G1 pre-barriers
const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
ResourceArea *area = Thread::current()->resource_area();
Unique_Node_List visited(area);
Node_List worklist(area);
// We're going to walk control flow backwards starting from the Root
worklist.push(_root);
while (worklist.size() > 0) {
Node* x = worklist.pop();
if (x == NULL || x == top()) continue;
if (visited.member(x)) {
continue;
} else {
visited.push(x);
}
if (x->is_Region()) {
for (uint i = 1; i < x->req(); i++) {
worklist.push(x->in(i));
}
} else {
worklist.push(x->in(0));
// We are looking for the pattern:
// /->ThreadLocal
// If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
// \->ConI(0)
// We want to verify that the If and the LoadB have the same control
// See GraphKit::g1_write_barrier_pre()
if (x->is_If()) {
IfNode *iff = x->as_If();
if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
&& cmp->in(1)->is_Load()) {
LoadNode* load = cmp->in(1)->as_Load();
if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
&& load->in(2)->in(3)->is_Con()
&& load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
Node* if_ctrl = iff->in(0);
Node* load_ctrl = load->in(0);
if (if_ctrl != load_ctrl) {
// Skip possible CProj->NeverBranch in infinite loops
if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
&& (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
if_ctrl = if_ctrl->in(0)->in(0);
}
}
assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
}
}
}
}
}
}
}
}
#endif
// The Compile object keeps track of failure reasons separately from the ciEnv.
// This is required because there is not quite a 1-1 relation between the
// ciEnv and its compilation task and the Compile object. Note that one
// ciEnv might use two Compile objects, if C2Compiler::compile_method decides
// to backtrack and retry without subsuming loads. Other than this backtracking
// behavior, the Compile's failure reason is quietly copied up to the ciEnv
// by the logic in C2Compiler.
void Compile::record_failure(const char* reason) {
if (log() != NULL) {
log()->elem("failure reason='%s' phase='compile'", reason);
}
if (_failure_reason == NULL) {
// Record the first failure reason.
_failure_reason = reason;
}
if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
C->print_method(PHASE_FAILURE);
}
_root = NULL; // flush the graph, too
}
Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
: TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
_phase_name(name), _dolog(dolog)
{
if (dolog) {
C = Compile::current();
_log = C->log();
} else {
C = NULL;
_log = NULL;
}
if (_log != NULL) {
_log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
_log->stamp();
_log->end_head();
}
}
Compile::TracePhase::~TracePhase() {
C = Compile::current();
if (_dolog) {
_log = C->log();
} else {
_log = NULL;
}
#ifdef ASSERT
if (PrintIdealNodeCount) {
tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
_phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
}
if (VerifyIdealNodeCount) {
Compile::current()->print_missing_nodes();
}
#endif
if (_log != NULL) {
_log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
}
}
//=============================================================================
// Two Constant's are equal when the type and the value are equal.
bool Compile::Constant::operator==(const Constant& other) {
if (type() != other.type() ) return false;
if (can_be_reused() != other.can_be_reused()) return false;
// For floating point values we compare the bit pattern.
switch (type()) {
case T_FLOAT: return (_v._value.i == other._v._value.i);
case T_LONG:
case T_DOUBLE: return (_v._value.j == other._v._value.j);
case T_OBJECT:
case T_ADDRESS: return (_v._value.l == other._v._value.l);
case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
case T_METADATA: return (_v._metadata == other._v._metadata);
default: ShouldNotReachHere();
}
return false;
}
static int type_to_size_in_bytes(BasicType t) {
switch (t) {
case T_LONG: return sizeof(jlong );
case T_FLOAT: return sizeof(jfloat );
case T_DOUBLE: return sizeof(jdouble);
case T_METADATA: return sizeof(Metadata*);
// We use T_VOID as marker for jump-table entries (labels) which
// need an internal word relocation.
case T_VOID:
case T_ADDRESS:
case T_OBJECT: return sizeof(jobject);
}
ShouldNotReachHere();
return -1;
}
int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
// sort descending
if (a->freq() > b->freq()) return -1;
if (a->freq() < b->freq()) return 1;
return 0;
}
void Compile::ConstantTable::calculate_offsets_and_size() {
// First, sort the array by frequencies.
_constants.sort(qsort_comparator);
#ifdef ASSERT
// Make sure all jump-table entries were sorted to the end of the
// array (they have a negative frequency).
bool found_void = false;
for (int i = 0; i < _constants.length(); i++) {
Constant con = _constants.at(i);
if (con.type() == T_VOID)
found_void = true; // jump-tables
else
assert(!found_void, "wrong sorting");
}
#endif
int offset = 0;
for (int i = 0; i < _constants.length(); i++) {
Constant* con = _constants.adr_at(i);
// Align offset for type.
int typesize = type_to_size_in_bytes(con->type());
offset = align_size_up(offset, typesize);
con->set_offset(offset); // set constant's offset
if (con->type() == T_VOID) {
MachConstantNode* n = (MachConstantNode*) con->get_jobject();
offset = offset + typesize * n->outcnt(); // expand jump-table
} else {
offset = offset + typesize;
}
}
// Align size up to the next section start (which is insts; see
// CodeBuffer::align_at_start).
assert(_size == -1, "already set?");
_size = align_size_up(offset, CodeEntryAlignment);
}
void Compile::ConstantTable::emit(CodeBuffer& cb) {
MacroAssembler _masm(&cb);
for (int i = 0; i < _constants.length(); i++) {
Constant con = _constants.at(i);
address constant_addr = NULL;
switch (con.type()) {
case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
case T_OBJECT: {
jobject obj = con.get_jobject();
int oop_index = _masm.oop_recorder()->find_index(obj);
constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
break;
}
case T_ADDRESS: {
address addr = (address) con.get_jobject();
constant_addr = _masm.address_constant(addr);
break;
}
// We use T_VOID as marker for jump-table entries (labels) which
// need an internal word relocation.
case T_VOID: {
MachConstantNode* n = (MachConstantNode*) con.get_jobject();
// Fill the jump-table with a dummy word. The real value is
// filled in later in fill_jump_table.
address dummy = (address) n;
constant_addr = _masm.address_constant(dummy);
// Expand jump-table
for (uint i = 1; i < n->outcnt(); i++) {
address temp_addr = _masm.address_constant(dummy + i);
assert(temp_addr, "consts section too small");
}
break;
}
case T_METADATA: {
Metadata* obj = con.get_metadata();
int metadata_index = _masm.oop_recorder()->find_index(obj);
constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
break;
}
default: ShouldNotReachHere();
}
assert(constant_addr, "consts section too small");
assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
}
}
int Compile::ConstantTable::find_offset(Constant& con) const {
int idx = _constants.find(con);
assert(idx != -1, "constant must be in constant table");
int offset = _constants.at(idx).offset();
assert(offset != -1, "constant table not emitted yet?");
return offset;
}
void Compile::ConstantTable::add(Constant& con) {
if (con.can_be_reused()) {
int idx = _constants.find(con);
if (idx != -1 && _constants.at(idx).can_be_reused()) {
_constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
return;
}
}
(void) _constants.append(con);
}
Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
Block* b = Compile::current()->cfg()->get_block_for_node(n);
Constant con(type, value, b->_freq);
add(con);
return con;
}
Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
Constant con(metadata);
add(con);
return con;
}
Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
jvalue value;
BasicType type = oper->type()->basic_type();
switch (type) {
case T_LONG: value.j = oper->constantL(); break;
case T_FLOAT: value.f = oper->constantF(); break;
case T_DOUBLE: value.d = oper->constantD(); break;
case T_OBJECT:
case T_ADDRESS: value.l = (jobject) oper->constant(); break;
case T_METADATA: return add((Metadata*)oper->constant()); break;
default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
}
return add(n, type, value);
}
Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
jvalue value;
// We can use the node pointer here to identify the right jump-table
// as this method is called from Compile::Fill_buffer right before
// the MachNodes are emitted and the jump-table is filled (means the
// MachNode pointers do not change anymore).
value.l = (jobject) n;
Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
add(con);
return con;
}
void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
// If called from Compile::scratch_emit_size do nothing.
if (Compile::current()->in_scratch_emit_size()) return;
assert(labels.is_nonempty(), "must be");
assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
// Since MachConstantNode::constant_offset() also contains
// table_base_offset() we need to subtract the table_base_offset()
// to get the plain offset into the constant table.
int offset = n->constant_offset() - table_base_offset();
MacroAssembler _masm(&cb);
address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
for (uint i = 0; i < n->outcnt(); i++) {
address* constant_addr = &jump_table_base[i];
assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
*constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
}
}
void Compile::dump_inlining() {
if (print_inlining() || print_intrinsics()) {
// Print inlining message for candidates that we couldn't inline
// for lack of space or non constant receiver
for (int i = 0; i < _late_inlines.length(); i++) {
CallGenerator* cg = _late_inlines.at(i);
cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
}
Unique_Node_List useful;
useful.push(root());
for (uint next = 0; next < useful.size(); ++next) {
Node* n = useful.at(next);
if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
CallNode* call = n->as_Call();
CallGenerator* cg = call->generator();
cg->print_inlining_late("receiver not constant");
}
uint max = n->len();
for ( uint i = 0; i < max; ++i ) {
Node *m = n->in(i);
if ( m == NULL ) continue;
useful.push(m);
}
}
for (int i = 0; i < _print_inlining_list->length(); i++) {
tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
}
}
}
// Dump inlining replay data to the stream.
// Don't change thread state and acquire any locks.
void Compile::dump_inline_data(outputStream* out) {
InlineTree* inl_tree = ilt();
if (inl_tree != NULL) {
out->print(" inline %d", inl_tree->count());
inl_tree->dump_replay_data(out);
}
}
int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
if (n1->Opcode() < n2->Opcode()) return -1;
else if (n1->Opcode() > n2->Opcode()) return 1;
assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
for (uint i = 1; i < n1->req(); i++) {
if (n1->in(i) < n2->in(i)) return -1;
else if (n1->in(i) > n2->in(i)) return 1;
}
return 0;
}
int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
Node* n1 = *n1p;
Node* n2 = *n2p;
return cmp_expensive_nodes(n1, n2);
}
void Compile::sort_expensive_nodes() {
if (!expensive_nodes_sorted()) {
_expensive_nodes->sort(cmp_expensive_nodes);
}
}
bool Compile::expensive_nodes_sorted() const {
for (int i = 1; i < _expensive_nodes->length(); i++) {
if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
return false;
}
}
return true;
}
bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
if (_expensive_nodes->length() == 0) {
return false;
}
assert(OptimizeExpensiveOps, "optimization off?");
// Take this opportunity to remove dead nodes from the list
int j = 0;
for (int i = 0; i < _expensive_nodes->length(); i++) {
Node* n = _expensive_nodes->at(i);
if (!n->is_unreachable(igvn)) {
assert(n->is_expensive(), "should be expensive");
_expensive_nodes->at_put(j, n);
j++;
}
}
_expensive_nodes->trunc_to(j);
// Then sort the list so that similar nodes are next to each other
// and check for at least two nodes of identical kind with same data
// inputs.
sort_expensive_nodes();
for (int i = 0; i < _expensive_nodes->length()-1; i++) {
if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
return true;
}
}
return false;
}
void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
if (_expensive_nodes->length() == 0) {
return;
}
assert(OptimizeExpensiveOps, "optimization off?");
// Sort to bring similar nodes next to each other and clear the
// control input of nodes for which there's only a single copy.
sort_expensive_nodes();
int j = 0;
int identical = 0;
int i = 0;
for (; i < _expensive_nodes->length()-1; i++) {
assert(j <= i, "can't write beyond current index");
if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
identical++;
_expensive_nodes->at_put(j++, _expensive_nodes->at(i));
continue;
}
if (identical > 0) {
_expensive_nodes->at_put(j++, _expensive_nodes->at(i));
identical = 0;
} else {
Node* n = _expensive_nodes->at(i);
igvn.hash_delete(n);
n->set_req(0, NULL);
igvn.hash_insert(n);
}
}
if (identical > 0) {
_expensive_nodes->at_put(j++, _expensive_nodes->at(i));
} else if (_expensive_nodes->length() >= 1) {
Node* n = _expensive_nodes->at(i);
igvn.hash_delete(n);
n->set_req(0, NULL);
igvn.hash_insert(n);
}
_expensive_nodes->trunc_to(j);
}
void Compile::add_expensive_node(Node * n) {
assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
assert(n->is_expensive(), "expensive nodes with non-null control here only");
assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
if (OptimizeExpensiveOps) {
_expensive_nodes->append(n);
} else {
// Clear control input and let IGVN optimize expensive nodes if
// OptimizeExpensiveOps is off.
n->set_req(0, NULL);
}
}
/**
* Remove the speculative part of types and clean up the graph
*/
void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
if (UseTypeSpeculation) {
Unique_Node_List worklist;
worklist.push(root());
int modified = 0;
// Go over all type nodes that carry a speculative type, drop the
// speculative part of the type and enqueue the node for an igvn
// which may optimize it out.
for (uint next = 0; next < worklist.size(); ++next) {
Node *n = worklist.at(next);
if (n->is_Type()) {
TypeNode* tn = n->as_Type();
const Type* t = tn->type();
const Type* t_no_spec = t->remove_speculative();
if (t_no_spec != t) {
bool in_hash = igvn.hash_delete(n);
assert(in_hash, "node should be in igvn hash table");
tn->set_type(t_no_spec);
igvn.hash_insert(n);
igvn._worklist.push(n); // give it a chance to go away
modified++;
}
}
uint max = n->len();
for( uint i = 0; i < max; ++i ) {
Node *m = n->in(i);
if (not_a_node(m)) continue;
worklist.push(m);
}
}
// Drop the speculative part of all types in the igvn's type table
igvn.remove_speculative_types();
if (modified > 0) {
igvn.optimize();
}
#ifdef ASSERT
// Verify that after the IGVN is over no speculative type has resurfaced
worklist.clear();
worklist.push(root());
for (uint next = 0; next < worklist.size(); ++next) {
Node *n = worklist.at(next);
const Type* t = igvn.type_or_null(n);
assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
if (n->is_Type()) {
t = n->as_Type()->type();
assert(t == t->remove_speculative(), "no more speculative types");
}
uint max = n->len();
for( uint i = 0; i < max; ++i ) {
Node *m = n->in(i);
if (not_a_node(m)) continue;
worklist.push(m);
}
}
igvn.check_no_speculative_types();
#endif
}
}
// Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
if (ctrl != NULL) {
// Express control dependency by a CastII node with a narrow type.
value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
// Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
// node from floating above the range check during loop optimizations. Otherwise, the
// ConvI2L node may be eliminated independently of the range check, causing the data path
// to become TOP while the control path is still there (although it's unreachable).
value->set_req(0, ctrl);
// Save CastII node to remove it after loop optimizations.
phase->C->add_range_check_cast(value);
value = phase->transform(value);
}
const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
}
// Auxiliary method to support randomized stressing/fuzzing.
//
// This method can be called the arbitrary number of times, with current count
// as the argument. The logic allows selecting a single candidate from the
// running list of candidates as follows:
// int count = 0;
// Cand* selected = null;
// while(cand = cand->next()) {
// if (randomized_select(++count)) {
// selected = cand;
// }
// }
//
// Including count equalizes the chances any candidate is "selected".
// This is useful when we don't have the complete list of candidates to choose
// from uniformly. In this case, we need to adjust the randomicity of the
// selection, or else we will end up biasing the selection towards the latter
// candidates.
//
// Quick back-envelope calculation shows that for the list of n candidates
// the equal probability for the candidate to persist as "best" can be
// achieved by replacing it with "next" k-th candidate with the probability
// of 1/k. It can be easily shown that by the end of the run, the
// probability for any candidate is converged to 1/n, thus giving the
// uniform distribution among all the candidates.
//
// We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
#define RANDOMIZED_DOMAIN_POW 29
#define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
#define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
bool Compile::randomized_select(int count) {
assert(count > 0, "only positive");
return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
}
C:\hotspot-69087d08d473\src\share\vm/opto/compile.hpp
/*
* Copyright (c) 1997, 2014, 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.
*
*/
#ifndef SHARE_VM_OPTO_COMPILE_HPP
#define SHARE_VM_OPTO_COMPILE_HPP
#include "asm/codeBuffer.hpp"
#include "ci/compilerInterface.hpp"
#include "code/debugInfoRec.hpp"
#include "code/exceptionHandlerTable.hpp"
#include "compiler/compilerOracle.hpp"
#include "compiler/compileBroker.hpp"
#include "jfr/jfrEvents.hpp"
#include "libadt/dict.hpp"
#include "libadt/port.hpp"
#include "libadt/vectset.hpp"
#include "memory/resourceArea.hpp"
#include "opto/idealGraphPrinter.hpp"
#include "opto/phasetype.hpp"
#include "opto/phase.hpp"
#include "opto/regmask.hpp"
#include "runtime/deoptimization.hpp"
#include "runtime/vmThread.hpp"
#include "utilities/ticks.hpp"
class Block;
class Bundle;
class C2Compiler;
class CallGenerator;
class ConnectionGraph;
class InlineTree;
class Int_Array;
class Matcher;
class MachConstantNode;
class MachConstantBaseNode;
class MachNode;
class MachOper;
class MachSafePointNode;
class Node;
class Node_Array;
class Node_Notes;
class OptoReg;
class PhaseCFG;
class PhaseGVN;
class PhaseIterGVN;
class PhaseRegAlloc;
class PhaseCCP;
class PhaseCCP_DCE;
class RootNode;
class relocInfo;
class Scope;
class StartNode;
class SafePointNode;
class JVMState;
class Type;
class TypeData;
class TypeInt;
class TypePtr;
class TypeOopPtr;
class TypeFunc;
class Unique_Node_List;
class nmethod;
class WarmCallInfo;
class Node_Stack;
struct Final_Reshape_Counts;
//------------------------------Compile----------------------------------------
// This class defines a top-level Compiler invocation.
class Compile : public Phase {
friend class VMStructs;
public:
// Fixed alias indexes. (See also MergeMemNode.)
enum {
AliasIdxTop = 1, // pseudo-index, aliases to nothing (used as sentinel value)
AliasIdxBot = 2, // pseudo-index, aliases to everything
AliasIdxRaw = 3 // hard-wired index for TypeRawPtr::BOTTOM
};
// Variant of TraceTime(NULL, &_t_accumulator, TimeCompiler);
// Integrated with logging. If logging is turned on, and dolog is true,
// then brackets are put into the log, with time stamps and node counts.
// (The time collection itself is always conditionalized on TimeCompiler.)
class TracePhase : public TraceTime {
private:
Compile* C;
CompileLog* _log;
const char* _phase_name;
bool _dolog;
public:
TracePhase(const char* name, elapsedTimer* accumulator, bool dolog);
~TracePhase();
};
// Information per category of alias (memory slice)
class AliasType {
private:
friend class Compile;
int _index; // unique index, used with MergeMemNode
const TypePtr* _adr_type; // normalized address type
ciField* _field; // relevant instance field, or null if none
const Type* _element; // relevant array element type, or null if none
bool _is_rewritable; // false if the memory is write-once only
int _general_index; // if this is type is an instance, the general
// type that this is an instance of
void Init(int i, const TypePtr* at);
public:
int index() const { return _index; }
const TypePtr* adr_type() const { return _adr_type; }
ciField* field() const { return _field; }
const Type* element() const { return _element; }
bool is_rewritable() const { return _is_rewritable; }
bool is_volatile() const { return (_field ? _field->is_volatile() : false); }
int general_index() const { return (_general_index != 0) ? _general_index : _index; }
void set_rewritable(bool z) { _is_rewritable = z; }
void set_field(ciField* f) {
assert(!_field,"");
_field = f;
if (f->is_final() || f->is_stable()) {
// In the case of @Stable, multiple writes are possible but may be assumed to be no-ops.
_is_rewritable = false;
}
}
void set_element(const Type* e) {
assert(_element == NULL, "");
_element = e;
}
BasicType basic_type() const;
void print_on(outputStream* st) PRODUCT_RETURN;
};
enum {
logAliasCacheSize = 6,
AliasCacheSize = (1<<logAliasCacheSize)
};
struct AliasCacheEntry { const TypePtr* _adr_type; int _index; }; // simple duple type
enum {
trapHistLength = MethodData::_trap_hist_limit
};
// Constant entry of the constant table.
class Constant {
private:
BasicType _type;
union {
jvalue _value;
Metadata* _metadata;
} _v;
int _offset; // offset of this constant (in bytes) relative to the constant table base.
float _freq;
bool _can_be_reused; // true (default) if the value can be shared with other users.
public:
Constant() : _type(T_ILLEGAL), _offset(-1), _freq(0.0f), _can_be_reused(true) { _v._value.l = 0; }
Constant(BasicType type, jvalue value, float freq = 0.0f, bool can_be_reused = true) :
_type(type),
_offset(-1),
_freq(freq),
_can_be_reused(can_be_reused)
{
assert(type != T_METADATA, "wrong constructor");
_v._value = value;
}
Constant(Metadata* metadata, bool can_be_reused = true) :
_type(T_METADATA),
_offset(-1),
_freq(0.0f),
_can_be_reused(can_be_reused)
{
_v._metadata = metadata;
}
bool operator==(const Constant& other);
BasicType type() const { return _type; }
jlong get_jlong() const { return _v._value.j; }
jfloat get_jfloat() const { return _v._value.f; }
jdouble get_jdouble() const { return _v._value.d; }
jobject get_jobject() const { return _v._value.l; }
Metadata* get_metadata() const { return _v._metadata; }
int offset() const { return _offset; }
void set_offset(int offset) { _offset = offset; }
float freq() const { return _freq; }
void inc_freq(float freq) { _freq += freq; }
bool can_be_reused() const { return _can_be_reused; }
};
// Constant table.
class ConstantTable {
private:
GrowableArray<Constant> _constants; // Constants of this table.
int _size; // Size in bytes the emitted constant table takes (including padding).
int _table_base_offset; // Offset of the table base that gets added to the constant offsets.
int _nof_jump_tables; // Number of jump-tables in this constant table.
static int qsort_comparator(Constant* a, Constant* b);
// We use negative frequencies to keep the order of the
// jump-tables in which they were added. Otherwise we get into
// trouble with relocation.
float next_jump_table_freq() { return -1.0f * (++_nof_jump_tables); }
public:
ConstantTable() :
_size(-1),
_table_base_offset(-1), // We can use -1 here since the constant table is always bigger than 2 bytes (-(size / 2), see MachConstantBaseNode::emit).
_nof_jump_tables(0)
{}
int size() const { assert(_size != -1, "not calculated yet"); return _size; }
int calculate_table_base_offset() const; // AD specific
void set_table_base_offset(int x) { assert(_table_base_offset == -1 || x == _table_base_offset, "can't change"); _table_base_offset = x; }
int table_base_offset() const { assert(_table_base_offset != -1, "not set yet"); return _table_base_offset; }
void emit(CodeBuffer& cb);
// Returns the offset of the last entry (the top) of the constant table.
int top_offset() const { assert(_constants.top().offset() != -1, "not bound yet"); return _constants.top().offset(); }
void calculate_offsets_and_size();
int find_offset(Constant& con) const;
void add(Constant& con);
Constant add(MachConstantNode* n, BasicType type, jvalue value);
Constant add(Metadata* metadata);
Constant add(MachConstantNode* n, MachOper* oper);
Constant add(MachConstantNode* n, jfloat f) {
jvalue value; value.f = f;
return add(n, T_FLOAT, value);
}
Constant add(MachConstantNode* n, jdouble d) {
jvalue value; value.d = d;
return add(n, T_DOUBLE, value);
}
// Jump-table
Constant add_jump_table(MachConstantNode* n);
void fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const;
};
private:
// Fixed parameters to this compilation.
const int _compile_id;
const bool _save_argument_registers; // save/restore arg regs for trampolines
const bool _subsume_loads; // Load can be matched as part of a larger op.
const bool _do_escape_analysis; // Do escape analysis.
const bool _eliminate_boxing; // Do boxing elimination.
ciMethod* _method; // The method being compiled.
int _entry_bci; // entry bci for osr methods.
const TypeFunc* _tf; // My kind of signature
InlineTree* _ilt; // Ditto (temporary).
address _stub_function; // VM entry for stub being compiled, or NULL
const char* _stub_name; // Name of stub or adapter being compiled, or NULL
address _stub_entry_point; // Compile code entry for generated stub, or NULL
// Control of this compilation.
int _num_loop_opts; // Number of iterations for doing loop optimiztions
int _max_inline_size; // Max inline size for this compilation
int _freq_inline_size; // Max hot method inline size for this compilation
int _fixed_slots; // count of frame slots not allocated by the register
// allocator i.e. locks, original deopt pc, etc.
uintx _max_node_limit; // Max unique node count during a single compilation.
// For deopt
int _orig_pc_slot;
int _orig_pc_slot_offset_in_bytes;
int _major_progress; // Count of something big happening
bool _inlining_progress; // progress doing incremental inlining?
bool _inlining_incrementally;// Are we doing incremental inlining (post parse)
bool _has_loops; // True if the method _may_ have some loops
bool _has_split_ifs; // True if the method _may_ have some split-if
bool _has_unsafe_access; // True if the method _may_ produce faults in unsafe loads or stores.
bool _has_stringbuilder; // True StringBuffers or StringBuilders are allocated
bool _has_boxed_value; // True if a boxed object is allocated
int _max_vector_size; // Maximum size of generated vectors
uint _trap_hist[trapHistLength]; // Cumulative traps
bool _trap_can_recompile; // Have we emitted a recompiling trap?
uint _decompile_count; // Cumulative decompilation counts.
bool _do_inlining; // True if we intend to do inlining
bool _do_scheduling; // True if we intend to do scheduling
bool _do_freq_based_layout; // True if we intend to do frequency based block layout
bool _do_count_invocations; // True if we generate code to count invocations
bool _do_method_data_update; // True if we generate code to update MethodData*s
int _AliasLevel; // Locally-adjusted version of AliasLevel flag.
bool _print_assembly; // True if we should dump assembly code for this compilation
bool _print_inlining; // True if we should print inlining for this compilation
bool _print_intrinsics; // True if we should print intrinsics for this compilation
#ifndef PRODUCT
bool _trace_opto_output;
bool _parsed_irreducible_loop; // True if ciTypeFlow detected irreducible loops during parsing
#endif
bool _has_irreducible_loop; // Found irreducible loops
// JSR 292
bool _has_method_handle_invokes; // True if this method has MethodHandle invokes.
RTMState _rtm_state; // State of Restricted Transactional Memory usage
// Compilation environment.
Arena _comp_arena; // Arena with lifetime equivalent to Compile
ciEnv* _env; // CI interface
CompileLog* _log; // from CompilerThread
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.
GrowableArray<Node*>* _expensive_nodes; // List of nodes that are expensive to compute and that we'd better not let the GVN freely common
GrowableArray<Node*>* _range_check_casts; // List of CastII nodes with a range check dependency
ConnectionGraph* _congraph;
#ifndef PRODUCT
IdealGraphPrinter* _printer;
#endif
// Node management
uint _unique; // Counter for unique Node indices
VectorSet _dead_node_list; // Set of dead nodes
uint _dead_node_count; // Number of dead nodes; VectorSet::Size() is O(N).
// So use this to keep count and make the call O(1).
debug_only(static int _debug_idx;) // Monotonic counter (not reset), use -XX:BreakAtNode=<idx>
Arena _node_arena; // Arena for new-space Nodes
Arena _old_arena; // Arena for old-space Nodes, lifetime during xform
RootNode* _root; // Unique root of compilation, or NULL after bail-out.
Node* _top; // Unique top node. (Reset by various phases.)
Node* _immutable_memory; // Initial memory state
Node* _recent_alloc_obj;
Node* _recent_alloc_ctl;
// Constant table
ConstantTable _constant_table; // The constant table for this compile.
MachConstantBaseNode* _mach_constant_base_node; // Constant table base node singleton.
// Blocked array of debugging and profiling information,
// tracked per node.
enum { _log2_node_notes_block_size = 8,
_node_notes_block_size = (1<<_log2_node_notes_block_size)
};
GrowableArray<Node_Notes*>* _node_note_array;
Node_Notes* _default_node_notes; // default notes for new nodes
// After parsing and every bulk phase we hang onto the Root instruction.
// The RootNode instruction is where the whole program begins. It produces
// the initial Control and BOTTOM for everybody else.
// Type management
Arena _Compile_types; // Arena for all types
Arena* _type_arena; // Alias for _Compile_types except in Initialize_shared()
Dict* _type_dict; // Intern table
void* _type_hwm; // Last allocation (see Type::operator new/delete)
size_t _type_last_size; // Last allocation size (see Type::operator new/delete)
ciMethod* _last_tf_m; // Cache for
const TypeFunc* _last_tf; // TypeFunc::make
AliasType** _alias_types; // List of alias types seen so far.
int _num_alias_types; // Logical length of _alias_types
int _max_alias_types; // Physical length of _alias_types
AliasCacheEntry _alias_cache[AliasCacheSize]; // Gets aliases w/o data structure walking
// Parsing, optimization
PhaseGVN* _initial_gvn; // Results of parse-time PhaseGVN
Unique_Node_List* _for_igvn; // Initial work-list for next round of Iterative GVN
WarmCallInfo* _warm_calls; // Sorted work-list for heat-based inlining.
GrowableArray<CallGenerator*> _late_inlines; // List of CallGenerators to be revisited after
// main parsing has finished.
GrowableArray<CallGenerator*> _string_late_inlines; // same but for string operations
GrowableArray<CallGenerator*> _boxing_late_inlines; // same but for boxing operations
int _late_inlines_pos; // Where in the queue should the next late inlining candidate go (emulate depth first inlining)
uint _number_of_mh_late_inlines; // number of method handle late inlining still pending
// Inlining may not happen in parse order which would make
// PrintInlining output confusing. Keep track of PrintInlining
// pieces in order.
class PrintInliningBuffer : public ResourceObj {
private:
CallGenerator* _cg;
stringStream* _ss;
public:
PrintInliningBuffer()
: _cg(NULL) { _ss = new stringStream(); }
stringStream* ss() const { return _ss; }
CallGenerator* cg() const { return _cg; }
void set_cg(CallGenerator* cg) { _cg = cg; }
};
GrowableArray<PrintInliningBuffer>* _print_inlining_list;
int _print_inlining_idx;
// Only keep nodes in the expensive node list that need to be optimized
void cleanup_expensive_nodes(PhaseIterGVN &igvn);
// Use for sorting expensive nodes to bring similar nodes together
static int cmp_expensive_nodes(Node** n1, Node** n2);
// Expensive nodes list already sorted?
bool expensive_nodes_sorted() const;
// Remove the speculative part of types and clean up the graph
void remove_speculative_types(PhaseIterGVN &igvn);
void* _replay_inline_data; // Pointer to data loaded from file
public:
outputStream* print_inlining_stream() const {
return _print_inlining_list->adr_at(_print_inlining_idx)->ss();
}
void print_inlining_skip(CallGenerator* cg) {
if (_print_inlining) {
_print_inlining_list->adr_at(_print_inlining_idx)->set_cg(cg);
_print_inlining_idx++;
_print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
}
}
void print_inlining_insert(CallGenerator* cg) {
if (_print_inlining) {
for (int i = 0; i < _print_inlining_list->length(); i++) {
if (_print_inlining_list->adr_at(i)->cg() == cg) {
_print_inlining_list->insert_before(i+1, PrintInliningBuffer());
_print_inlining_idx = i+1;
_print_inlining_list->adr_at(i)->set_cg(NULL);
return;
}
}
ShouldNotReachHere();
}
}
void print_inlining(ciMethod* method, int inline_level, int bci, const char* msg = NULL) {
stringStream ss;
CompileTask::print_inlining(&ss, method, inline_level, bci, msg);
print_inlining_stream()->print("%s", ss.as_string());
}
void* replay_inline_data() const { return _replay_inline_data; }
// Dump inlining replay data to the stream.
void dump_inline_data(outputStream* out);
private:
// Matching, CFG layout, allocation, code generation
PhaseCFG* _cfg; // Results of CFG finding
bool _select_24_bit_instr; // We selected an instruction with a 24-bit result
bool _in_24_bit_fp_mode; // We are emitting instructions with 24-bit results
int _java_calls; // Number of java calls in the method
int _inner_loops; // Number of inner loops in the method
Matcher* _matcher; // Engine to map ideal to machine instructions
PhaseRegAlloc* _regalloc; // Results of register allocation.
int _frame_slots; // Size of total frame in stack slots
CodeOffsets _code_offsets; // Offsets into the code for various interesting entries
RegMask _FIRST_STACK_mask; // All stack slots usable for spills (depends on frame layout)
Arena* _indexSet_arena; // control IndexSet allocation within PhaseChaitin
void* _indexSet_free_block_list; // free list of IndexSet bit blocks
int _interpreter_frame_size;
uint _node_bundling_limit;
Bundle* _node_bundling_base; // Information for instruction bundling
// Instruction bits passed off to the VM
int _method_size; // Size of nmethod code segment in bytes
CodeBuffer _code_buffer; // Where the code is assembled
int _first_block_size; // Size of unvalidated entry point code / OSR poison code
ExceptionHandlerTable _handler_table; // Table of native-code exception handlers
ImplicitExceptionTable _inc_table; // Table of implicit null checks in native code
OopMapSet* _oop_map_set; // Table of oop maps (one for each safepoint location)
static int _CompiledZap_count; // counter compared against CompileZap[First/Last]
BufferBlob* _scratch_buffer_blob; // For temporary code buffers.
relocInfo* _scratch_locs_memory; // For temporary code buffers.
int _scratch_const_size; // For temporary code buffers.
bool _in_scratch_emit_size; // true when in scratch_emit_size.
public:
// Accessors
// The Compile instance currently active in this (compiler) thread.
static Compile* current() {
return (Compile*) ciEnv::current()->compiler_data();
}
// ID for this compilation. Useful for setting breakpoints in the debugger.
int compile_id() const { return _compile_id; }
// Does this compilation allow instructions to subsume loads? User
// instructions that subsume a load may result in an unschedulable
// instruction sequence.
bool subsume_loads() const { return _subsume_loads; }
/** Do escape analysis. */
bool do_escape_analysis() const { return _do_escape_analysis; }
/** Do boxing elimination. */
bool eliminate_boxing() const { return _eliminate_boxing; }
/** Do aggressive boxing elimination. */
bool aggressive_unboxing() const { return _eliminate_boxing && AggressiveUnboxing; }
bool save_argument_registers() const { return _save_argument_registers; }
// Other fixed compilation parameters.
ciMethod* method() const { return _method; }
int entry_bci() const { return _entry_bci; }
bool is_osr_compilation() const { return _entry_bci != InvocationEntryBci; }
bool is_method_compilation() const { return (_method != NULL && !_method->flags().is_native()); }
const TypeFunc* tf() const { assert(_tf!=NULL, ""); return _tf; }
void init_tf(const TypeFunc* tf) { assert(_tf==NULL, ""); _tf = tf; }
InlineTree* ilt() const { return _ilt; }
address stub_function() const { return _stub_function; }
const char* stub_name() const { return _stub_name; }
address stub_entry_point() const { return _stub_entry_point; }
// Control of this compilation.
int fixed_slots() const { assert(_fixed_slots >= 0, ""); return _fixed_slots; }
void set_fixed_slots(int n) { _fixed_slots = n; }
int major_progress() const { return _major_progress; }
void set_inlining_progress(bool z) { _inlining_progress = z; }
int inlining_progress() const { return _inlining_progress; }
void set_inlining_incrementally(bool z) { _inlining_incrementally = z; }
int inlining_incrementally() const { return _inlining_incrementally; }
void set_major_progress() { _major_progress++; }
void clear_major_progress() { _major_progress = 0; }
int num_loop_opts() const { return _num_loop_opts; }
void set_num_loop_opts(int n) { _num_loop_opts = n; }
int max_inline_size() const { return _max_inline_size; }
void set_freq_inline_size(int n) { _freq_inline_size = n; }
int freq_inline_size() const { return _freq_inline_size; }
void set_max_inline_size(int n) { _max_inline_size = n; }
bool has_loops() const { return _has_loops; }
void set_has_loops(bool z) { _has_loops = z; }
bool has_split_ifs() const { return _has_split_ifs; }
void set_has_split_ifs(bool z) { _has_split_ifs = z; }
bool has_unsafe_access() const { return _has_unsafe_access; }
void set_has_unsafe_access(bool z) { _has_unsafe_access = z; }
bool has_stringbuilder() const { return _has_stringbuilder; }
void set_has_stringbuilder(bool z) { _has_stringbuilder = z; }
bool has_boxed_value() const { return _has_boxed_value; }
void set_has_boxed_value(bool z) { _has_boxed_value = z; }
int max_vector_size() const { return _max_vector_size; }
void set_max_vector_size(int s) { _max_vector_size = s; }
void set_trap_count(uint r, uint c) { assert(r < trapHistLength, "oob"); _trap_hist[r] = c; }
uint trap_count(uint r) const { assert(r < trapHistLength, "oob"); return _trap_hist[r]; }
bool trap_can_recompile() const { return _trap_can_recompile; }
void set_trap_can_recompile(bool z) { _trap_can_recompile = z; }
uint decompile_count() const { return _decompile_count; }
void set_decompile_count(uint c) { _decompile_count = c; }
bool allow_range_check_smearing() const;
bool do_inlining() const { return _do_inlining; }
void set_do_inlining(bool z) { _do_inlining = z; }
bool do_scheduling() const { return _do_scheduling; }
void set_do_scheduling(bool z) { _do_scheduling = z; }
bool do_freq_based_layout() const{ return _do_freq_based_layout; }
void set_do_freq_based_layout(bool z){ _do_freq_based_layout = z; }
bool do_count_invocations() const{ return _do_count_invocations; }
void set_do_count_invocations(bool z){ _do_count_invocations = z; }
bool do_method_data_update() const { return _do_method_data_update; }
void set_do_method_data_update(bool z) { _do_method_data_update = z; }
int AliasLevel() const { return _AliasLevel; }
bool print_assembly() const { return _print_assembly; }
void set_print_assembly(bool z) { _print_assembly = z; }
bool print_inlining() const { return _print_inlining; }
void set_print_inlining(bool z) { _print_inlining = z; }
bool print_intrinsics() const { return _print_intrinsics; }
void set_print_intrinsics(bool z) { _print_intrinsics = z; }
RTMState rtm_state() const { return _rtm_state; }
void set_rtm_state(RTMState s) { _rtm_state = s; }
bool use_rtm() const { return (_rtm_state & NoRTM) == 0; }
bool profile_rtm() const { return _rtm_state == ProfileRTM; }
uint max_node_limit() const { return (uint)_max_node_limit; }
void set_max_node_limit(uint n) { _max_node_limit = n; }
// check the CompilerOracle for special behaviours for this compile
bool method_has_option(const char * option) {
return method() != NULL && method()->has_option(option);
}
template<typename T>
bool method_has_option_value(const char * option, T& value) {
return method() != NULL && method()->has_option_value(option, value);
}
#ifndef PRODUCT
bool trace_opto_output() const { return _trace_opto_output; }
bool parsed_irreducible_loop() const { return _parsed_irreducible_loop; }
void set_parsed_irreducible_loop(bool z) { _parsed_irreducible_loop = z; }
int _in_dump_cnt; // Required for dumping ir nodes.
#endif
bool has_irreducible_loop() const { return _has_irreducible_loop; }
void set_has_irreducible_loop(bool z) { _has_irreducible_loop = z; }
// JSR 292
bool has_method_handle_invokes() const { return _has_method_handle_invokes; }
void set_has_method_handle_invokes(bool z) { _has_method_handle_invokes = z; }
Ticks _latest_stage_start_counter;
void begin_method() {
#ifndef PRODUCT
if (_printer) _printer->begin_method(this);
#endif
C->_latest_stage_start_counter.stamp();
}
void print_method(CompilerPhaseType cpt, int level = 1) {
EventCompilerPhase event;
if (event.should_commit()) {
event.set_starttime(C->_latest_stage_start_counter);
event.set_phase((u1) cpt);
event.set_compileId(C->_compile_id);
event.set_phaseLevel(level);
event.commit();
}
#ifndef PRODUCT
if (_printer) _printer->print_method(this, CompilerPhaseTypeHelper::to_string(cpt), level);
#endif
C->_latest_stage_start_counter.stamp();
}
void end_method(int level = 1) {
EventCompilerPhase event;
if (event.should_commit()) {
event.set_starttime(C->_latest_stage_start_counter);
event.set_phase((u1) PHASE_END);
event.set_compileId(C->_compile_id);
event.set_phaseLevel(level);
event.commit();
}
#ifndef PRODUCT
if (_printer) _printer->end_method();
#endif
}
int macro_count() const { return _macro_nodes->length(); }
int predicate_count() const { return _predicate_opaqs->length();}
int expensive_count() const { return _expensive_nodes->length(); }
Node* macro_node(int idx) const { return _macro_nodes->at(idx); }
Node* predicate_opaque1_node(int idx) const { return _predicate_opaqs->at(idx);}
Node* expensive_node(int idx) const { return _expensive_nodes->at(idx); }
ConnectionGraph* congraph() { return _congraph;}
void set_congraph(ConnectionGraph* congraph) { _congraph = congraph;}
void add_macro_node(Node * n) {
//assert(n->is_macro(), "must be a macro node");
assert(!_macro_nodes->contains(n), "duplicate entry in expand list");
_macro_nodes->append(n);
}
void remove_macro_node(Node * n) {
// this function may be called twice for a node so check
// 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_expensive_node(Node * n);
void remove_expensive_node(Node * n) {
if (_expensive_nodes->contains(n)) {
_expensive_nodes->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);
}
// Range check dependent CastII nodes that can be removed after loop optimizations
void add_range_check_cast(Node* n);
void remove_range_check_cast(Node* n) {
if (_range_check_casts->contains(n)) {
_range_check_casts->remove(n);
}
}
Node* range_check_cast_node(int idx) const { return _range_check_casts->at(idx); }
int range_check_cast_count() const { return _range_check_casts->length(); }
// Remove all range check dependent CastIINodes.
void remove_range_check_casts(PhaseIterGVN &igvn);
// 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);
bool is_predicate_opaq(Node * n) {
return _predicate_opaqs->contains(n);
}
// Are there candidate expensive nodes for optimization?
bool should_optimize_expensive_nodes(PhaseIterGVN &igvn);
// Check whether n1 and n2 are similar
static int cmp_expensive_nodes(Node* n1, Node* n2);
// Sort expensive nodes to locate similar expensive nodes
void sort_expensive_nodes();
// Compilation environment.
Arena* comp_arena() { return &_comp_arena; }
ciEnv* env() const { return _env; }
CompileLog* log() const { return _log; }
bool failing() const { return _env->failing() || _failure_reason != NULL; }
const char* failure_reason() { return _failure_reason; }
bool failure_reason_is(const char* r) { return (r==_failure_reason) || (r!=NULL && _failure_reason!=NULL && strcmp(r, _failure_reason)==0); }
void record_failure(const char* reason);
void record_method_not_compilable(const char* reason, bool all_tiers = false) {
// All bailouts cover "all_tiers" when TieredCompilation is off.
if (!TieredCompilation) all_tiers = true;
env()->record_method_not_compilable(reason, all_tiers);
// Record failure reason.
record_failure(reason);
}
void record_method_not_compilable_all_tiers(const char* reason) {
record_method_not_compilable(reason, true);
}
bool check_node_count(uint margin, const char* reason) {
if (live_nodes() + margin > max_node_limit()) {
record_method_not_compilable(reason);
return true;
} else {
return false;
}
}
// Node management
uint unique() const { return _unique; }
uint next_unique() { return _unique++; }
void set_unique(uint i) { _unique = i; }
static int debug_idx() { return debug_only(_debug_idx)+0; }
static void set_debug_idx(int i) { debug_only(_debug_idx = i); }
Arena* node_arena() { return &_node_arena; }
Arena* old_arena() { return &_old_arena; }
RootNode* root() const { return _root; }
void set_root(RootNode* r) { _root = r; }
StartNode* start() const; // (Derived from root.)
void init_start(StartNode* s);
Node* immutable_memory();
Node* recent_alloc_ctl() const { return _recent_alloc_ctl; }
Node* recent_alloc_obj() const { return _recent_alloc_obj; }
void set_recent_alloc(Node* ctl, Node* obj) {
_recent_alloc_ctl = ctl;
_recent_alloc_obj = obj;
}
void record_dead_node(uint idx) { if (_dead_node_list.test_set(idx)) return;
_dead_node_count++;
}
bool is_dead_node(uint idx) { return _dead_node_list.test(idx) != 0; }
uint dead_node_count() { return _dead_node_count; }
void reset_dead_node_list() { _dead_node_list.Reset();
_dead_node_count = 0;
}
uint live_nodes() const {
int val = _unique - _dead_node_count;
assert (val >= 0, err_msg_res("number of tracked dead nodes %d more than created nodes %d", _unique, _dead_node_count));
return (uint) val;
}
#ifdef ASSERT
uint count_live_nodes_by_graph_walk();
void print_missing_nodes();
#endif
// Constant table
ConstantTable& constant_table() { return _constant_table; }
MachConstantBaseNode* mach_constant_base_node();
bool has_mach_constant_base_node() const { return _mach_constant_base_node != NULL; }
// Generated by adlc, true if CallNode requires MachConstantBase.
bool needs_clone_jvms();
// Handy undefined Node
Node* top() const { return _top; }
// these are used by guys who need to know about creation and transformation of top:
Node* cached_top_node() { return _top; }
void set_cached_top_node(Node* tn);
GrowableArray<Node_Notes*>* node_note_array() const { return _node_note_array; }
void set_node_note_array(GrowableArray<Node_Notes*>* arr) { _node_note_array = arr; }
Node_Notes* default_node_notes() const { return _default_node_notes; }
void set_default_node_notes(Node_Notes* n) { _default_node_notes = n; }
Node_Notes* node_notes_at(int idx) {
return locate_node_notes(_node_note_array, idx, false);
}
inline bool set_node_notes_at(int idx, Node_Notes* value);
// Copy notes from source to dest, if they exist.
// Overwrite dest only if source provides something.
// Return true if information was moved.
bool copy_node_notes_to(Node* dest, Node* source);
// Workhorse function to sort out the blocked Node_Notes array:
inline Node_Notes* locate_node_notes(GrowableArray<Node_Notes*>* arr,
int idx, bool can_grow = false);
void grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by);
// Type management
Arena* type_arena() { return _type_arena; }
Dict* type_dict() { return _type_dict; }
void* type_hwm() { return _type_hwm; }
size_t type_last_size() { return _type_last_size; }
int num_alias_types() { return _num_alias_types; }
void init_type_arena() { _type_arena = &_Compile_types; }
void set_type_arena(Arena* a) { _type_arena = a; }
void set_type_dict(Dict* d) { _type_dict = d; }
void set_type_hwm(void* p) { _type_hwm = p; }
void set_type_last_size(size_t sz) { _type_last_size = sz; }
const TypeFunc* last_tf(ciMethod* m) {
return (m == _last_tf_m) ? _last_tf : NULL;
}
void set_last_tf(ciMethod* m, const TypeFunc* tf) {
assert(m != NULL || tf == NULL, "");
_last_tf_m = m;
_last_tf = tf;
}
AliasType* alias_type(int idx) { assert(idx < num_alias_types(), "oob"); return _alias_types[idx]; }
AliasType* alias_type(const TypePtr* adr_type, ciField* field = NULL) { return find_alias_type(adr_type, false, field); }
bool have_alias_type(const TypePtr* adr_type);
AliasType* alias_type(ciField* field);
int get_alias_index(const TypePtr* at) { return alias_type(at)->index(); }
const TypePtr* get_adr_type(uint aidx) { return alias_type(aidx)->adr_type(); }
int get_general_index(uint aidx) { return alias_type(aidx)->general_index(); }
// Building nodes
void rethrow_exceptions(JVMState* jvms);
void return_values(JVMState* jvms);
JVMState* build_start_state(StartNode* start, const TypeFunc* tf);
// Decide how to build a call.
// The profile factor is a discount to apply to this site's interp. profile.
CallGenerator* call_generator(ciMethod* call_method, int vtable_index, bool call_does_dispatch,
JVMState* jvms, bool allow_inline, float profile_factor, ciKlass* speculative_receiver_type = NULL,
bool allow_intrinsics = true, bool delayed_forbidden = false);
bool should_delay_inlining(ciMethod* call_method, JVMState* jvms) {
return should_delay_string_inlining(call_method, jvms) ||
should_delay_boxing_inlining(call_method, jvms);
}
bool should_delay_string_inlining(ciMethod* call_method, JVMState* jvms);
bool should_delay_boxing_inlining(ciMethod* call_method, JVMState* jvms);
// Helper functions to identify inlining potential at call-site
ciMethod* optimize_virtual_call(ciMethod* caller, int bci, ciInstanceKlass* klass,
ciKlass* holder, ciMethod* callee,
const TypeOopPtr* receiver_type, bool is_virtual,
bool &call_does_dispatch, int &vtable_index,
bool check_access = true);
ciMethod* optimize_inlining(ciMethod* caller, int bci, ciInstanceKlass* klass,
ciMethod* callee, const TypeOopPtr* receiver_type,
bool check_access = true);
// Report if there were too many traps at a current method and bci.
// Report if a trap was recorded, and/or PerMethodTrapLimit was exceeded.
// If there is no MDO at all, report no trap unless told to assume it.
bool too_many_traps(ciMethod* method, int bci, Deoptimization::DeoptReason reason);
// This version, unspecific to a particular bci, asks if
// PerMethodTrapLimit was exceeded for all inlined methods seen so far.
bool too_many_traps(Deoptimization::DeoptReason reason,
// Privately used parameter for logging:
ciMethodData* logmd = NULL);
// Report if there were too many recompiles at a method and bci.
bool too_many_recompiles(ciMethod* method, int bci, Deoptimization::DeoptReason reason);
// Return a bitset with the reasons where deoptimization is allowed,
// i.e., where there were not too many uncommon traps.
int _allowed_reasons;
int allowed_deopt_reasons() { return _allowed_reasons; }
void set_allowed_deopt_reasons();
// Parsing, optimization
PhaseGVN* initial_gvn() { return _initial_gvn; }
Unique_Node_List* for_igvn() { return _for_igvn; }
inline void record_for_igvn(Node* n); // Body is after class Unique_Node_List.
void set_initial_gvn(PhaseGVN *gvn) { _initial_gvn = gvn; }
void set_for_igvn(Unique_Node_List *for_igvn) { _for_igvn = for_igvn; }
// Replace n by nn using initial_gvn, calling hash_delete and
// record_for_igvn as needed.
void gvn_replace_by(Node* n, Node* nn);
void identify_useful_nodes(Unique_Node_List &useful);
void update_dead_node_list(Unique_Node_List &useful);
void remove_useless_nodes (Unique_Node_List &useful);
WarmCallInfo* warm_calls() const { return _warm_calls; }
void set_warm_calls(WarmCallInfo* l) { _warm_calls = l; }
WarmCallInfo* pop_warm_call();
// Record this CallGenerator for inlining at the end of parsing.
void add_late_inline(CallGenerator* cg) {
_late_inlines.insert_before(_late_inlines_pos, cg);
_late_inlines_pos++;
}
void prepend_late_inline(CallGenerator* cg) {
_late_inlines.insert_before(0, cg);
}
void add_string_late_inline(CallGenerator* cg) {
_string_late_inlines.push(cg);
}
void add_boxing_late_inline(CallGenerator* cg) {
_boxing_late_inlines.push(cg);
}
void remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful);
void dump_inlining();
bool over_inlining_cutoff() const {
if (!inlining_incrementally()) {
return unique() > (uint)NodeCountInliningCutoff;
} else {
return live_nodes() > (uint)LiveNodeCountInliningCutoff;
}
}
void inc_number_of_mh_late_inlines() { _number_of_mh_late_inlines++; }
void dec_number_of_mh_late_inlines() { assert(_number_of_mh_late_inlines > 0, "_number_of_mh_late_inlines < 0 !"); _number_of_mh_late_inlines--; }
bool has_mh_late_inlines() const { return _number_of_mh_late_inlines > 0; }
void inline_incrementally_one(PhaseIterGVN& igvn);
void inline_incrementally(PhaseIterGVN& igvn);
void inline_string_calls(bool parse_time);
void inline_boxing_calls(PhaseIterGVN& igvn);
void remove_root_to_sfpts_edges(PhaseIterGVN& igvn);
// Matching, CFG layout, allocation, code generation
PhaseCFG* cfg() { return _cfg; }
bool select_24_bit_instr() const { return _select_24_bit_instr; }
bool in_24_bit_fp_mode() const { return _in_24_bit_fp_mode; }
bool has_java_calls() const { return _java_calls > 0; }
int java_calls() const { return _java_calls; }
int inner_loops() const { return _inner_loops; }
Matcher* matcher() { return _matcher; }
PhaseRegAlloc* regalloc() { return _regalloc; }
int frame_slots() const { return _frame_slots; }
int frame_size_in_words() const; // frame_slots in units of the polymorphic 'words'
int frame_size_in_bytes() const { return _frame_slots << LogBytesPerInt; }
RegMask& FIRST_STACK_mask() { return _FIRST_STACK_mask; }
Arena* indexSet_arena() { return _indexSet_arena; }
void* indexSet_free_block_list() { return _indexSet_free_block_list; }
uint node_bundling_limit() { return _node_bundling_limit; }
Bundle* node_bundling_base() { return _node_bundling_base; }
void set_node_bundling_limit(uint n) { _node_bundling_limit = n; }
void set_node_bundling_base(Bundle* b) { _node_bundling_base = b; }
bool starts_bundle(const Node *n) const;
bool need_stack_bang(int frame_size_in_bytes) const;
bool need_register_stack_bang() const;
void update_interpreter_frame_size(int size) {
if (_interpreter_frame_size < size) {
_interpreter_frame_size = size;
}
}
int bang_size_in_bytes() const;
void set_matcher(Matcher* m) { _matcher = m; }
//void set_regalloc(PhaseRegAlloc* ra) { _regalloc = ra; }
void set_indexSet_arena(Arena* a) { _indexSet_arena = a; }
void set_indexSet_free_block_list(void* p) { _indexSet_free_block_list = p; }
// Remember if this compilation changes hardware mode to 24-bit precision
void set_24_bit_selection_and_mode(bool selection, bool mode) {
_select_24_bit_instr = selection;
_in_24_bit_fp_mode = mode;
}
void set_java_calls(int z) { _java_calls = z; }
void set_inner_loops(int z) { _inner_loops = z; }
// Instruction bits passed off to the VM
int code_size() { return _method_size; }
CodeBuffer* code_buffer() { return &_code_buffer; }
int first_block_size() { return _first_block_size; }
void set_frame_complete(int off) { _code_offsets.set_value(CodeOffsets::Frame_Complete, off); }
ExceptionHandlerTable* handler_table() { return &_handler_table; }
ImplicitExceptionTable* inc_table() { return &_inc_table; }
OopMapSet* oop_map_set() { return _oop_map_set; }
DebugInformationRecorder* debug_info() { return env()->debug_info(); }
Dependencies* dependencies() { return env()->dependencies(); }
static int CompiledZap_count() { return _CompiledZap_count; }
BufferBlob* scratch_buffer_blob() { return _scratch_buffer_blob; }
void init_scratch_buffer_blob(int const_size);
void clear_scratch_buffer_blob();
void set_scratch_buffer_blob(BufferBlob* b) { _scratch_buffer_blob = b; }
relocInfo* scratch_locs_memory() { return _scratch_locs_memory; }
void set_scratch_locs_memory(relocInfo* b) { _scratch_locs_memory = b; }
// emit to scratch blob, report resulting size
uint scratch_emit_size(const Node* n);
void set_in_scratch_emit_size(bool x) { _in_scratch_emit_size = x; }
bool in_scratch_emit_size() const { return _in_scratch_emit_size; }
enum ScratchBufferBlob {
MAX_inst_size = 1024,
MAX_locs_size = 128, // number of relocInfo elements
MAX_const_size = 128,
MAX_stubs_size = 128
};
// Major entry point. Given a Scope, compile the associated method.
// For normal compilations, entry_bci is InvocationEntryBci. For on stack
// replacement, entry_bci indicates the bytecode for which to compile a
// continuation.
Compile(ciEnv* ci_env, C2Compiler* compiler, ciMethod* target,
int entry_bci, bool subsume_loads, bool do_escape_analysis,
bool eliminate_boxing);
// Second major entry point. From the TypeFunc signature, generate code
// to pass arguments from the Java calling convention to the C calling
// convention.
Compile(ciEnv* ci_env, const TypeFunc *(*gen)(),
address stub_function, const char *stub_name,
int is_fancy_jump, bool pass_tls,
bool save_arg_registers, bool return_pc);
// From the TypeFunc signature, generate code to pass arguments
// from Compiled calling convention to Interpreter's calling convention
void Generate_Compiled_To_Interpreter_Graph(const TypeFunc *tf, address interpreter_entry);
// From the TypeFunc signature, generate code to pass arguments
// from Interpreter's calling convention to Compiler's calling convention
void Generate_Interpreter_To_Compiled_Graph(const TypeFunc *tf);
// Are we compiling a method?
bool has_method() { return method() != NULL; }
// Maybe print some information about this compile.
void print_compile_messages();
// Final graph reshaping, a post-pass after the regular optimizer is done.
bool final_graph_reshaping();
// returns true if adr is completely contained in the given alias category
bool must_alias(const TypePtr* adr, int alias_idx);
// returns true if adr overlaps with the given alias category
bool can_alias(const TypePtr* adr, int alias_idx);
// Driver for converting compiler's IR into machine code bits
void Output();
// Accessors for node bundling info.
Bundle* node_bundling(const Node *n);
bool valid_bundle_info(const Node *n);
// Schedule and Bundle the instructions
void ScheduleAndBundle();
// Build OopMaps for each GC point
void BuildOopMaps();
// Append debug info for the node "local" at safepoint node "sfpt" to the
// "array", May also consult and add to "objs", which describes the
// scalar-replaced objects.
void FillLocArray( int idx, MachSafePointNode* sfpt,
Node *local, GrowableArray<ScopeValue*> *array,
GrowableArray<ScopeValue*> *objs );
// If "objs" contains an ObjectValue whose id is "id", returns it, else NULL.
static ObjectValue* sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id);
// Requres that "objs" does not contains an ObjectValue whose id matches
// that of "sv. Appends "sv".
static void set_sv_for_object_node(GrowableArray<ScopeValue*> *objs,
ObjectValue* sv );
// Process an OopMap Element while emitting nodes
void Process_OopMap_Node(MachNode *mach, int code_offset);
// Initialize code buffer
CodeBuffer* init_buffer(uint* blk_starts);
// Write out basic block data to code buffer
void fill_buffer(CodeBuffer* cb, uint* blk_starts);
// Determine which variable sized branches can be shortened
void shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size);
// Compute the size of first NumberOfLoopInstrToAlign instructions
// at the head of a loop.
void compute_loop_first_inst_sizes();
// Compute the information for the exception tables
void FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels);
// Stack slots that may be unused by the calling convention but must
// otherwise be preserved. On Intel this includes the return address.
// On PowerPC it includes the 4 words holding the old TOC & LR glue.
uint in_preserve_stack_slots();
// "Top of Stack" slots that may be unused by the calling convention but must
// otherwise be preserved.
// On Intel these are not necessary and the value can be zero.
// On Sparc this describes the words reserved for storing a register window
// when an interrupt occurs.
static uint out_preserve_stack_slots();
// Number of outgoing stack slots killed above the out_preserve_stack_slots
// for calls to C. Supports the var-args backing area for register parms.
uint varargs_C_out_slots_killed() const;
// Number of Stack Slots consumed by a synchronization entry
int sync_stack_slots() const;
// Compute the name of old_SP. See <arch>.ad for frame layout.
OptoReg::Name compute_old_SP();
#ifdef ENABLE_ZAP_DEAD_LOCALS
static bool is_node_getting_a_safepoint(Node*);
void Insert_zap_nodes();
Node* call_zap_node(MachSafePointNode* n, int block_no);
#endif
private:
// Phase control:
void Init(int aliaslevel); // Prepare for a single compilation
int Inline_Warm(); // Find more inlining work.
void Finish_Warm(); // Give up on further inlines.
void Optimize(); // Given a graph, optimize it
void Code_Gen(); // Generate code from a graph
// Management of the AliasType table.
void grow_alias_types();
AliasCacheEntry* probe_alias_cache(const TypePtr* adr_type);
const TypePtr *flatten_alias_type(const TypePtr* adr_type) const;
AliasType* find_alias_type(const TypePtr* adr_type, bool no_create, ciField* field);
void verify_top(Node*) const PRODUCT_RETURN;
// Intrinsic setup.
void register_library_intrinsics(); // initializer
CallGenerator* make_vm_intrinsic(ciMethod* m, bool is_virtual); // constructor
int intrinsic_insertion_index(ciMethod* m, bool is_virtual); // helper
CallGenerator* find_intrinsic(ciMethod* m, bool is_virtual); // query fn
void register_intrinsic(CallGenerator* cg); // update fn
#ifndef PRODUCT
static juint _intrinsic_hist_count[vmIntrinsics::ID_LIMIT];
static jubyte _intrinsic_hist_flags[vmIntrinsics::ID_LIMIT];
#endif
// Function calls made by the public function final_graph_reshaping.
// No need to be made public as they are not called elsewhere.
void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc);
void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc );
void eliminate_redundant_card_marks(Node* n);
public:
// Note: Histogram array size is about 1 Kb.
enum { // flag bits:
_intrinsic_worked = 1, // succeeded at least once
_intrinsic_failed = 2, // tried it but it failed
_intrinsic_disabled = 4, // was requested but disabled (e.g., -XX:-InlineUnsafeOps)
_intrinsic_virtual = 8, // was seen in the virtual form (rare)
_intrinsic_both = 16 // was seen in the non-virtual form (usual)
};
// Update histogram. Return boolean if this is a first-time occurrence.
static bool gather_intrinsic_statistics(vmIntrinsics::ID id,
bool is_virtual, int flags) PRODUCT_RETURN0;
static void print_intrinsic_statistics() PRODUCT_RETURN;
// Graph verification code
// Walk the node list, verifying that there is a one-to-one
// correspondence between Use-Def edges and Def-Use edges
// The option no_dead_code enables stronger checks that the
// graph is strongly connected from root in both directions.
void verify_graph_edges(bool no_dead_code = false) PRODUCT_RETURN;
// Verify GC barrier patterns
void verify_barriers() PRODUCT_RETURN;
// End-of-run dumps.
static void print_statistics() PRODUCT_RETURN;
// Dump formatted assembly
void dump_asm(int *pcs = NULL, uint pc_limit = 0) PRODUCT_RETURN;
void dump_pc(int *pcs, int pc_limit, Node *n);
// Verify ADLC assumptions during startup
static void adlc_verification() PRODUCT_RETURN;
// Definitions of pd methods
static void pd_compiler2_init();
// Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
static Node* constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl);
// Auxiliary method for randomized fuzzing/stressing
static bool randomized_select(int count);
#ifdef ASSERT
bool _type_verify_symmetry;
#endif
};
#endif // SHARE_VM_OPTO_COMPILE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/connode.cpp
/*
* Copyright (c) 1997, 2019, 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 "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/compile.hpp"
#include "opto/connode.hpp"
#include "opto/machnode.hpp"
#include "opto/matcher.hpp"
#include "opto/memnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
#include "runtime/sharedRuntime.hpp"
// Optimization - Graph Style
//=============================================================================
//------------------------------hash-------------------------------------------
uint ConNode::hash() const {
return (uintptr_t)in(TypeFunc::Control) + _type->hash();
}
//------------------------------make-------------------------------------------
ConNode *ConNode::make( Compile* C, const Type *t ) {
switch( t->basic_type() ) {
case T_INT: return new (C) ConINode( t->is_int() );
case T_LONG: return new (C) ConLNode( t->is_long() );
case T_FLOAT: return new (C) ConFNode( t->is_float_constant() );
case T_DOUBLE: return new (C) ConDNode( t->is_double_constant() );
case T_VOID: return new (C) ConNode ( Type::TOP );
case T_OBJECT: return new (C) ConPNode( t->is_ptr() );
case T_ARRAY: return new (C) ConPNode( t->is_aryptr() );
case T_ADDRESS: return new (C) ConPNode( t->is_ptr() );
case T_NARROWOOP: return new (C) ConNNode( t->is_narrowoop() );
case T_NARROWKLASS: return new (C) ConNKlassNode( t->is_narrowklass() );
case T_METADATA: return new (C) ConPNode( t->is_ptr() );
// Expected cases: TypePtr::NULL_PTR, any is_rawptr()
// Also seen: AnyPtr(TopPTR *+top); from command line:
// r -XX:+PrintOpto -XX:CIStart=285 -XX:+CompileTheWorld -XX:CompileTheWorldStartAt=660
// %%%% Stop using TypePtr::NULL_PTR to represent nulls: use either TypeRawPtr::NULL_PTR
// or else TypeOopPtr::NULL_PTR. Then set Type::_basic_type[AnyPtr] = T_ILLEGAL
}
ShouldNotReachHere();
return NULL;
}
//=============================================================================
/*
The major change is for CMoveP and StrComp. They have related but slightly
different problems. They both take in TWO oops which are both null-checked
independently before the using Node. After CCP removes the CastPP's they need
to pick up the guarding test edge - in this case TWO control edges. I tried
various solutions, all have problems:
(1) Do nothing. This leads to a bug where we hoist a Load from a CMoveP or a
StrComp above a guarding null check. I've seen both cases in normal -Xcomp
testing.
(2) Plug the control edge from 1 of the 2 oops in. Apparent problem here is
to figure out which test post-dominates. The real problem is that it doesn't
matter which one you pick. After you pick up, the dominating-test elider in
IGVN can remove the test and allow you to hoist up to the dominating test on
the chosen oop bypassing the test on the not-chosen oop. Seen in testing.
Oops.
(3) Leave the CastPP's in. This makes the graph more accurate in some sense;
we get to keep around the knowledge that an oop is not-null after some test.
Alas, the CastPP's interfere with GVN (some values are the regular oop, some
are the CastPP of the oop, all merge at Phi's which cannot collapse, etc).
This cost us 10% on SpecJVM, even when I removed some of the more trivial
cases in the optimizer. Removing more useless Phi's started allowing Loads to
illegally float above null checks. I gave up on this approach.
(4) Add BOTH control edges to both tests. Alas, too much code knows that
control edges are in slot-zero ONLY. Many quick asserts fail; no way to do
this one. Note that I really want to allow the CMoveP to float and add both
control edges to the dependent Load op - meaning I can select early but I
cannot Load until I pass both tests.
(5) Do not hoist CMoveP and StrComp. To this end I added the v-call
depends_only_on_test(). No obvious performance loss on Spec, but we are
clearly conservative on CMoveP (also so on StrComp but that's unlikely to
matter ever).
*/
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Move constants to the right.
Node *CMoveNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if( in(0) && remove_dead_region(phase, can_reshape) ) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
assert( !phase->eqv(in(Condition), this) &&
!phase->eqv(in(IfFalse), this) &&
!phase->eqv(in(IfTrue), this), "dead loop in CMoveNode::Ideal" );
if( phase->type(in(Condition)) == Type::TOP )
return NULL; // return NULL when Condition is dead
if( in(IfFalse)->is_Con() && !in(IfTrue)->is_Con() ) {
if( in(Condition)->is_Bool() ) {
BoolNode* b = in(Condition)->as_Bool();
BoolNode* b2 = b->negate(phase);
return make( phase->C, in(Control), phase->transform(b2), in(IfTrue), in(IfFalse), _type );
}
}
return NULL;
}
//------------------------------is_cmove_id------------------------------------
// Helper function to check for CMOVE identity. Shared with PhiNode::Identity
Node *CMoveNode::is_cmove_id( PhaseTransform *phase, Node *cmp, Node *t, Node *f, BoolNode *b ) {
// Check for Cmp'ing and CMove'ing same values
if( (phase->eqv(cmp->in(1),f) &&
phase->eqv(cmp->in(2),t)) ||
// Swapped Cmp is OK
(phase->eqv(cmp->in(2),f) &&
phase->eqv(cmp->in(1),t)) ) {
// Give up this identity check for floating points because it may choose incorrect
// value around 0.0 and -0.0
if ( cmp->Opcode()==Op_CmpF || cmp->Opcode()==Op_CmpD )
return NULL;
// Check for "(t==f)?t:f;" and replace with "f"
if( b->_test._test == BoolTest::eq )
return f;
// Allow the inverted case as well
// Check for "(t!=f)?t:f;" and replace with "t"
if( b->_test._test == BoolTest::ne )
return t;
}
return NULL;
}
//------------------------------Identity---------------------------------------
// Conditional-move is an identity if both inputs are the same, or the test
// true or false.
Node *CMoveNode::Identity( PhaseTransform *phase ) {
if( phase->eqv(in(IfFalse),in(IfTrue)) ) // C-moving identical inputs?
return in(IfFalse); // Then it doesn't matter
if( phase->type(in(Condition)) == TypeInt::ZERO )
return in(IfFalse); // Always pick left(false) input
if( phase->type(in(Condition)) == TypeInt::ONE )
return in(IfTrue); // Always pick right(true) input
// Check for CMove'ing a constant after comparing against the constant.
// Happens all the time now, since if we compare equality vs a constant in
// the parser, we "know" the variable is constant on one path and we force
// it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a
// conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more
// general in that we don't need constants.
if( in(Condition)->is_Bool() ) {
BoolNode *b = in(Condition)->as_Bool();
Node *cmp = b->in(1);
if( cmp->is_Cmp() ) {
Node *id = is_cmove_id( phase, cmp, in(IfTrue), in(IfFalse), b );
if( id ) return id;
}
}
return this;
}
//------------------------------Value------------------------------------------
// Result is the meet of inputs
const Type *CMoveNode::Value( PhaseTransform *phase ) const {
if( phase->type(in(Condition)) == Type::TOP )
return Type::TOP;
return phase->type(in(IfFalse))->meet_speculative(phase->type(in(IfTrue)));
}
//------------------------------make-------------------------------------------
// Make a correctly-flavored CMove. Since _type is directly determined
// from the inputs we do not need to specify it here.
CMoveNode *CMoveNode::make( Compile *C, Node *c, Node *bol, Node *left, Node *right, const Type *t ) {
switch( t->basic_type() ) {
case T_INT: return new (C) CMoveINode( bol, left, right, t->is_int() );
case T_FLOAT: return new (C) CMoveFNode( bol, left, right, t );
case T_DOUBLE: return new (C) CMoveDNode( bol, left, right, t );
case T_LONG: return new (C) CMoveLNode( bol, left, right, t->is_long() );
case T_OBJECT: return new (C) CMovePNode( c, bol, left, right, t->is_oopptr() );
case T_ADDRESS: return new (C) CMovePNode( c, bol, left, right, t->is_ptr() );
case T_NARROWOOP: return new (C) CMoveNNode( c, bol, left, right, t );
default:
ShouldNotReachHere();
return NULL;
}
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Check for conversions to boolean
Node *CMoveINode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Try generic ideal's first
Node *x = CMoveNode::Ideal(phase, can_reshape);
if( x ) return x;
// If zero is on the left (false-case, no-move-case) it must mean another
// constant is on the right (otherwise the shared CMove::Ideal code would
// have moved the constant to the right). This situation is bad for Intel
// and a don't-care for Sparc. It's bad for Intel because the zero has to
// be manifested in a register with a XOR which kills flags, which are live
// on input to the CMoveI, leading to a situation which causes excessive
// spilling on Intel. For Sparc, if the zero in on the left the Sparc will
// zero a register via G0 and conditionally-move the other constant. If the
// zero is on the right, the Sparc will load the first constant with a
// 13-bit set-lo and conditionally move G0. See bug 4677505.
if( phase->type(in(IfFalse)) == TypeInt::ZERO && !(phase->type(in(IfTrue)) == TypeInt::ZERO) ) {
if( in(Condition)->is_Bool() ) {
BoolNode* b = in(Condition)->as_Bool();
BoolNode* b2 = b->negate(phase);
return make( phase->C, in(Control), phase->transform(b2), in(IfTrue), in(IfFalse), _type );
}
}
// Now check for booleans
int flip = 0;
// Check for picking from zero/one
if( phase->type(in(IfFalse)) == TypeInt::ZERO && phase->type(in(IfTrue)) == TypeInt::ONE ) {
flip = 1 - flip;
} else if( phase->type(in(IfFalse)) == TypeInt::ONE && phase->type(in(IfTrue)) == TypeInt::ZERO ) {
} else return NULL;
// Check for eq/ne test
if( !in(1)->is_Bool() ) return NULL;
BoolNode *bol = in(1)->as_Bool();
if( bol->_test._test == BoolTest::eq ) {
} else if( bol->_test._test == BoolTest::ne ) {
flip = 1-flip;
} else return NULL;
// Check for vs 0 or 1
if( !bol->in(1)->is_Cmp() ) return NULL;
const CmpNode *cmp = bol->in(1)->as_Cmp();
if( phase->type(cmp->in(2)) == TypeInt::ZERO ) {
} else if( phase->type(cmp->in(2)) == TypeInt::ONE ) {
// Allow cmp-vs-1 if the other input is bounded by 0-1
if( phase->type(cmp->in(1)) != TypeInt::BOOL )
return NULL;
flip = 1 - flip;
} else return NULL;
// Convert to a bool (flipped)
// Build int->bool conversion
#ifndef PRODUCT
if( PrintOpto ) tty->print_cr("CMOV to I2B");
#endif
Node *n = new (phase->C) Conv2BNode( cmp->in(1) );
if( flip )
n = new (phase->C) XorINode( phase->transform(n), phase->intcon(1) );
return n;
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Check for absolute value
Node *CMoveFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Try generic ideal's first
Node *x = CMoveNode::Ideal(phase, can_reshape);
if( x ) return x;
int cmp_zero_idx = 0; // Index of compare input where to look for zero
int phi_x_idx = 0; // Index of phi input where to find naked x
// Find the Bool
if( !in(1)->is_Bool() ) return NULL;
BoolNode *bol = in(1)->as_Bool();
// Check bool sense
switch( bol->_test._test ) {
case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = IfTrue; break;
case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = IfFalse; break;
case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = IfTrue; break;
case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = IfFalse; break;
default: return NULL; break;
}
// Find zero input of CmpF; the other input is being abs'd
Node *cmpf = bol->in(1);
if( cmpf->Opcode() != Op_CmpF ) return NULL;
Node *X = NULL;
bool flip = false;
if( phase->type(cmpf->in(cmp_zero_idx)) == TypeF::ZERO ) {
X = cmpf->in(3 - cmp_zero_idx);
} else if (phase->type(cmpf->in(3 - cmp_zero_idx)) == TypeF::ZERO) {
// The test is inverted, we should invert the result...
X = cmpf->in(cmp_zero_idx);
flip = true;
} else {
return NULL;
}
// If X is found on the appropriate phi input, find the subtract on the other
if( X != in(phi_x_idx) ) return NULL;
int phi_sub_idx = phi_x_idx == IfTrue ? IfFalse : IfTrue;
Node *sub = in(phi_sub_idx);
// Allow only SubF(0,X) and fail out for all others; NegF is not OK
if( sub->Opcode() != Op_SubF ||
sub->in(2) != X ||
phase->type(sub->in(1)) != TypeF::ZERO ) return NULL;
Node *abs = new (phase->C) AbsFNode( X );
if( flip )
abs = new (phase->C) SubFNode(sub->in(1), phase->transform(abs));
return abs;
}
//=============================================================================
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Check for absolute value
Node *CMoveDNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Try generic ideal's first
Node *x = CMoveNode::Ideal(phase, can_reshape);
if( x ) return x;
int cmp_zero_idx = 0; // Index of compare input where to look for zero
int phi_x_idx = 0; // Index of phi input where to find naked x
// Find the Bool
if( !in(1)->is_Bool() ) return NULL;
BoolNode *bol = in(1)->as_Bool();
// Check bool sense
switch( bol->_test._test ) {
case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = IfTrue; break;
case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = IfFalse; break;
case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = IfTrue; break;
case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = IfFalse; break;
default: return NULL; break;
}
// Find zero input of CmpD; the other input is being abs'd
Node *cmpd = bol->in(1);
if( cmpd->Opcode() != Op_CmpD ) return NULL;
Node *X = NULL;
bool flip = false;
if( phase->type(cmpd->in(cmp_zero_idx)) == TypeD::ZERO ) {
X = cmpd->in(3 - cmp_zero_idx);
} else if (phase->type(cmpd->in(3 - cmp_zero_idx)) == TypeD::ZERO) {
// The test is inverted, we should invert the result...
X = cmpd->in(cmp_zero_idx);
flip = true;
} else {
return NULL;
}
// If X is found on the appropriate phi input, find the subtract on the other
if( X != in(phi_x_idx) ) return NULL;
int phi_sub_idx = phi_x_idx == IfTrue ? IfFalse : IfTrue;
Node *sub = in(phi_sub_idx);
// Allow only SubD(0,X) and fail out for all others; NegD is not OK
if( sub->Opcode() != Op_SubD ||
sub->in(2) != X ||
phase->type(sub->in(1)) != TypeD::ZERO ) return NULL;
Node *abs = new (phase->C) AbsDNode( X );
if( flip )
abs = new (phase->C) SubDNode(sub->in(1), phase->transform(abs));
return abs;
}
//=============================================================================
// If input is already higher or equal to cast type, then this is an identity.
Node *ConstraintCastNode::Identity( PhaseTransform *phase ) {
return phase->type(in(1))->higher_equal_speculative(_type) ? in(1) : this;
}
//------------------------------Value------------------------------------------
// Take 'join' of input and cast-up type
const Type *ConstraintCastNode::Value( PhaseTransform *phase ) const {
if( in(0) && phase->type(in(0)) == Type::TOP ) return Type::TOP;
const Type* ft = phase->type(in(1))->filter_speculative(_type);
#ifdef ASSERT
// Previous versions of this function had some special case logic,
// which is no longer necessary. Make sure of the required effects.
switch (Opcode()) {
case Op_CastII:
{
const Type* t1 = phase->type(in(1));
if( t1 == Type::TOP ) assert(ft == Type::TOP, "special case #1");
const Type* rt = t1->join_speculative(_type);
if (rt->empty()) assert(ft == Type::TOP, "special case #2");
break;
}
case Op_CastPP:
if (phase->type(in(1)) == TypePtr::NULL_PTR &&
_type->isa_ptr() && _type->is_ptr()->_ptr == TypePtr::NotNull)
assert(ft == Type::TOP, "special case #3");
break;
}
#endif //ASSERT
return ft;
}
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node. Strip out
// control copies
Node *ConstraintCastNode::Ideal(PhaseGVN *phase, bool can_reshape){
return (in(0) && remove_dead_region(phase, can_reshape)) ? this : NULL;
}
//------------------------------Ideal_DU_postCCP-------------------------------
// Throw away cast after constant propagation
Node *ConstraintCastNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
const Type *t = ccp->type(in(1));
ccp->hash_delete(this);
set_type(t); // Turn into ID function
ccp->hash_insert(this);
return this;
}
uint CastIINode::size_of() const {
return sizeof(*this);
}
uint CastIINode::cmp(const Node &n) const {
return TypeNode::cmp(n) &&
((CastIINode&)n)._carry_dependency == _carry_dependency &&
((CastIINode&)n)._range_check_dependency == _range_check_dependency;
}
Node *CastIINode::Identity(PhaseTransform *phase) {
if (_carry_dependency) {
return this;
}
return ConstraintCastNode::Identity(phase);
}
const Type *CastIINode::Value(PhaseTransform *phase) const {
const Type *res = ConstraintCastNode::Value(phase);
// Try to improve the type of the CastII if we recognize a CmpI/If
// pattern.
if (_carry_dependency) {
if (in(0) != NULL && in(0)->in(0) != NULL && in(0)->in(0)->is_If()) {
assert(in(0)->is_IfFalse() || in(0)->is_IfTrue(), "should be If proj");
Node* proj = in(0);
if (proj->in(0)->in(1)->is_Bool()) {
Node* b = proj->in(0)->in(1);
if (b->in(1)->Opcode() == Op_CmpI) {
Node* cmp = b->in(1);
if (cmp->in(1) == in(1) && phase->type(cmp->in(2))->isa_int()) {
const TypeInt* in2_t = phase->type(cmp->in(2))->is_int();
const Type* t = TypeInt::INT;
BoolTest test = b->as_Bool()->_test;
if (proj->is_IfFalse()) {
test = test.negate();
}
BoolTest::mask m = test._test;
jlong lo_long = min_jint;
jlong hi_long = max_jint;
if (m == BoolTest::le || m == BoolTest::lt) {
hi_long = in2_t->_hi;
if (m == BoolTest::lt) {
hi_long -= 1;
}
} else if (m == BoolTest::ge || m == BoolTest::gt) {
lo_long = in2_t->_lo;
if (m == BoolTest::gt) {
lo_long += 1;
}
} else if (m == BoolTest::eq) {
lo_long = in2_t->_lo;
hi_long = in2_t->_hi;
} else if (m == BoolTest::ne) {
// can't do any better
} else {
stringStream ss;
test.dump_on(&ss);
fatal(err_msg_res("unexpected comparison %s", ss.as_string()));
}
int lo_int = (int)lo_long;
int hi_int = (int)hi_long;
if (lo_long != (jlong)lo_int) {
lo_int = min_jint;
}
if (hi_long != (jlong)hi_int) {
hi_int = max_jint;
}
t = TypeInt::make(lo_int, hi_int, Type::WidenMax);
res = res->filter_speculative(t);
return res;
}
}
}
}
}
return res;
}
Node *CastIINode::Ideal_DU_postCCP(PhaseCCP *ccp) {
if (_carry_dependency || _range_check_dependency) {
return NULL;
}
return ConstraintCastNode::Ideal_DU_postCCP(ccp);
}
#ifndef PRODUCT
void CastIINode::dump_spec(outputStream *st) const {
TypeNode::dump_spec(st);
if (_carry_dependency) {
st->print(" carry dependency");
}
if (_range_check_dependency) {
st->print(" range check dependency");
}
}
#endif
//=============================================================================
//------------------------------Ideal_DU_postCCP-------------------------------
// If not converting int->oop, throw away cast after constant propagation
Node *CastPPNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
const Type *t = ccp->type(in(1));
if (!t->isa_oop_ptr() || ((in(1)->is_DecodeN()) && Matcher::gen_narrow_oop_implicit_null_checks())) {
return NULL; // do not transform raw pointers or narrow oops
}
return ConstraintCastNode::Ideal_DU_postCCP(ccp);
}
//=============================================================================
//------------------------------Identity---------------------------------------
// If input is already higher or equal to cast type, then this is an identity.
Node *CheckCastPPNode::Identity( PhaseTransform *phase ) {
// Toned down to rescue meeting at a Phi 3 different oops all implementing
// the same interface. CompileTheWorld starting at 502, kd12rc1.zip.
return (phase->type(in(1)) == phase->type(this)) ? in(1) : this;
}
//------------------------------Value------------------------------------------
// Take 'join' of input and cast-up type, unless working with an Interface
const Type *CheckCastPPNode::Value( PhaseTransform *phase ) const {
if( in(0) && phase->type(in(0)) == Type::TOP ) return Type::TOP;
const Type *inn = phase->type(in(1));
if( inn == Type::TOP ) return Type::TOP; // No information yet
const TypePtr *in_type = inn->isa_ptr();
const TypePtr *my_type = _type->isa_ptr();
const Type *result = _type;
if( in_type != NULL && my_type != NULL ) {
TypePtr::PTR in_ptr = in_type->ptr();
if( in_ptr == TypePtr::Null ) {
result = in_type;
} else if( in_ptr == TypePtr::Constant ) {
// Casting a constant oop to an interface?
// (i.e., a String to a Comparable?)
// Then return the interface.
const TypeOopPtr *jptr = my_type->isa_oopptr();
assert( jptr, "" );
result = (jptr->klass()->is_interface() || !in_type->higher_equal(_type))
? my_type->cast_to_ptr_type( TypePtr::NotNull )
: in_type;
} else {
result = my_type->cast_to_ptr_type( my_type->join_ptr(in_ptr) );
}
}
return result;
// JOIN NOT DONE HERE BECAUSE OF INTERFACE ISSUES.
// FIX THIS (DO THE JOIN) WHEN UNION TYPES APPEAR!
//
// Remove this code after overnight run indicates no performance
// loss from not performing JOIN at CheckCastPPNode
//
// const TypeInstPtr *in_oop = in->isa_instptr();
// const TypeInstPtr *my_oop = _type->isa_instptr();
// // If either input is an 'interface', return destination type
// assert (in_oop == NULL || in_oop->klass() != NULL, "");
// assert (my_oop == NULL || my_oop->klass() != NULL, "");
// if( (in_oop && in_oop->klass()->is_interface())
// ||(my_oop && my_oop->klass()->is_interface()) ) {
// TypePtr::PTR in_ptr = in->isa_ptr() ? in->is_ptr()->_ptr : TypePtr::BotPTR;
// // Preserve cast away nullness for interfaces
// if( in_ptr == TypePtr::NotNull && my_oop && my_oop->_ptr == TypePtr::BotPTR ) {
// return my_oop->cast_to_ptr_type(TypePtr::NotNull);
// }
// return _type;
// }
//
// // Neither the input nor the destination type is an interface,
//
// // history: JOIN used to cause weird corner case bugs
// // return (in == TypeOopPtr::NULL_PTR) ? in : _type;
// // JOIN picks up NotNull in common instance-of/check-cast idioms, both oops.
// // JOIN does not preserve NotNull in other cases, e.g. RawPtr vs InstPtr
// const Type *join = in->join(_type);
// // Check if join preserved NotNull'ness for pointers
// if( join->isa_ptr() && _type->isa_ptr() ) {
// TypePtr::PTR join_ptr = join->is_ptr()->_ptr;
// TypePtr::PTR type_ptr = _type->is_ptr()->_ptr;
// // If there isn't any NotNull'ness to preserve
// // OR if join preserved NotNull'ness then return it
// if( type_ptr == TypePtr::BotPTR || type_ptr == TypePtr::Null ||
// join_ptr == TypePtr::NotNull || join_ptr == TypePtr::Constant ) {
// return join;
// }
// // ELSE return same old type as before
// return _type;
// }
// // Not joining two pointers
// return join;
}
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node. Strip out
// control copies
Node *CheckCastPPNode::Ideal(PhaseGVN *phase, bool can_reshape){
return (in(0) && remove_dead_region(phase, can_reshape)) ? this : NULL;
}
Node* DecodeNNode::Identity(PhaseTransform* phase) {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return in(1);
if (in(1)->is_EncodeP()) {
// (DecodeN (EncodeP p)) -> p
return in(1)->in(1);
}
return this;
}
const Type *DecodeNNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if (t == Type::TOP) return Type::TOP;
if (t == TypeNarrowOop::NULL_PTR) return TypePtr::NULL_PTR;
assert(t->isa_narrowoop(), "only narrowoop here");
return t->make_ptr();
}
Node* EncodePNode::Identity(PhaseTransform* phase) {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return in(1);
if (in(1)->is_DecodeN()) {
// (EncodeP (DecodeN p)) -> p
return in(1)->in(1);
}
return this;
}
const Type *EncodePNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if (t == Type::TOP) return Type::TOP;
if (t == TypePtr::NULL_PTR) return TypeNarrowOop::NULL_PTR;
assert(t->isa_oop_ptr(), "only oopptr here");
return t->make_narrowoop();
}
Node *EncodeNarrowPtrNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
return MemNode::Ideal_common_DU_postCCP(ccp, this, in(1));
}
Node* DecodeNKlassNode::Identity(PhaseTransform* phase) {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return in(1);
if (in(1)->is_EncodePKlass()) {
// (DecodeNKlass (EncodePKlass p)) -> p
return in(1)->in(1);
}
return this;
}
const Type *DecodeNKlassNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if (t == Type::TOP) return Type::TOP;
assert(t != TypeNarrowKlass::NULL_PTR, "null klass?");
assert(t->isa_narrowklass(), "only narrow klass ptr here");
return t->make_ptr();
}
Node* EncodePKlassNode::Identity(PhaseTransform* phase) {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return in(1);
if (in(1)->is_DecodeNKlass()) {
// (EncodePKlass (DecodeNKlass p)) -> p
return in(1)->in(1);
}
return this;
}
const Type *EncodePKlassNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if (t == Type::TOP) return Type::TOP;
assert (t != TypePtr::NULL_PTR, "null klass?");
assert(UseCompressedClassPointers && t->isa_klassptr(), "only klass ptr here");
return t->make_narrowklass();
}
//=============================================================================
//------------------------------Identity---------------------------------------
Node *Conv2BNode::Identity( PhaseTransform *phase ) {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return in(1);
if( t == TypeInt::ZERO ) return in(1);
if( t == TypeInt::ONE ) return in(1);
if( t == TypeInt::BOOL ) return in(1);
return this;
}
//------------------------------Value------------------------------------------
const Type *Conv2BNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == TypeInt::ZERO ) return TypeInt::ZERO;
if( t == TypePtr::NULL_PTR ) return TypeInt::ZERO;
const TypePtr *tp = t->isa_ptr();
if( tp != NULL ) {
if( tp->ptr() == TypePtr::AnyNull ) return Type::TOP;
if( tp->ptr() == TypePtr::Constant) return TypeInt::ONE;
if (tp->ptr() == TypePtr::NotNull) return TypeInt::ONE;
return TypeInt::BOOL;
}
if (t->base() != Type::Int) return TypeInt::BOOL;
const TypeInt *ti = t->is_int();
if( ti->_hi < 0 || ti->_lo > 0 ) return TypeInt::ONE;
return TypeInt::BOOL;
}
// The conversions operations are all Alpha sorted. Please keep it that way!
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvD2FNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::DOUBLE ) return Type::FLOAT;
const TypeD *td = t->is_double_constant();
return TypeF::make( (float)td->getd() );
}
//------------------------------Identity---------------------------------------
// Float's can be converted to doubles with no loss of bits. Hence
// converting a float to a double and back to a float is a NOP.
Node *ConvD2FNode::Identity(PhaseTransform *phase) {
return (in(1)->Opcode() == Op_ConvF2D) ? in(1)->in(1) : this;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvD2INode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::DOUBLE ) return TypeInt::INT;
const TypeD *td = t->is_double_constant();
return TypeInt::make( SharedRuntime::d2i( td->getd() ) );
}
//------------------------------Ideal------------------------------------------
// If converting to an int type, skip any rounding nodes
Node *ConvD2INode::Ideal(PhaseGVN *phase, bool can_reshape) {
if( in(1)->Opcode() == Op_RoundDouble )
set_req(1,in(1)->in(1));
return NULL;
}
//------------------------------Identity---------------------------------------
// Int's can be converted to doubles with no loss of bits. Hence
// converting an integer to a double and back to an integer is a NOP.
Node *ConvD2INode::Identity(PhaseTransform *phase) {
return (in(1)->Opcode() == Op_ConvI2D) ? in(1)->in(1) : this;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvD2LNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::DOUBLE ) return TypeLong::LONG;
const TypeD *td = t->is_double_constant();
return TypeLong::make( SharedRuntime::d2l( td->getd() ) );
}
//------------------------------Identity---------------------------------------
Node *ConvD2LNode::Identity(PhaseTransform *phase) {
// Remove ConvD2L->ConvL2D->ConvD2L sequences.
if( in(1) ->Opcode() == Op_ConvL2D &&
in(1)->in(1)->Opcode() == Op_ConvD2L )
return in(1)->in(1);
return this;
}
//------------------------------Ideal------------------------------------------
// If converting to an int type, skip any rounding nodes
Node *ConvD2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if( in(1)->Opcode() == Op_RoundDouble )
set_req(1,in(1)->in(1));
return NULL;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvF2DNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::FLOAT ) return Type::DOUBLE;
const TypeF *tf = t->is_float_constant();
return TypeD::make( (double)tf->getf() );
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvF2INode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::FLOAT ) return TypeInt::INT;
const TypeF *tf = t->is_float_constant();
return TypeInt::make( SharedRuntime::f2i( tf->getf() ) );
}
//------------------------------Identity---------------------------------------
Node *ConvF2INode::Identity(PhaseTransform *phase) {
// Remove ConvF2I->ConvI2F->ConvF2I sequences.
if( in(1) ->Opcode() == Op_ConvI2F &&
in(1)->in(1)->Opcode() == Op_ConvF2I )
return in(1)->in(1);
return this;
}
//------------------------------Ideal------------------------------------------
// If converting to an int type, skip any rounding nodes
Node *ConvF2INode::Ideal(PhaseGVN *phase, bool can_reshape) {
if( in(1)->Opcode() == Op_RoundFloat )
set_req(1,in(1)->in(1));
return NULL;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvF2LNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::FLOAT ) return TypeLong::LONG;
const TypeF *tf = t->is_float_constant();
return TypeLong::make( SharedRuntime::f2l( tf->getf() ) );
}
//------------------------------Identity---------------------------------------
Node *ConvF2LNode::Identity(PhaseTransform *phase) {
// Remove ConvF2L->ConvL2F->ConvF2L sequences.
if( in(1) ->Opcode() == Op_ConvL2F &&
in(1)->in(1)->Opcode() == Op_ConvF2L )
return in(1)->in(1);
return this;
}
//------------------------------Ideal------------------------------------------
// If converting to an int type, skip any rounding nodes
Node *ConvF2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if( in(1)->Opcode() == Op_RoundFloat )
set_req(1,in(1)->in(1));
return NULL;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvI2DNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeInt *ti = t->is_int();
if( ti->is_con() ) return TypeD::make( (double)ti->get_con() );
return bottom_type();
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvI2FNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeInt *ti = t->is_int();
if( ti->is_con() ) return TypeF::make( (float)ti->get_con() );
return bottom_type();
}
//------------------------------Identity---------------------------------------
Node *ConvI2FNode::Identity(PhaseTransform *phase) {
// Remove ConvI2F->ConvF2I->ConvI2F sequences.
if( in(1) ->Opcode() == Op_ConvF2I &&
in(1)->in(1)->Opcode() == Op_ConvI2F )
return in(1)->in(1);
return this;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvI2LNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeInt *ti = t->is_int();
const Type* tl = TypeLong::make(ti->_lo, ti->_hi, ti->_widen);
// Join my declared type against my incoming type.
tl = tl->filter(_type);
return tl;
}
#ifdef _LP64
static inline bool long_ranges_overlap(jlong lo1, jlong hi1,
jlong lo2, jlong hi2) {
// Two ranges overlap iff one range's low point falls in the other range.
return (lo2 <= lo1 && lo1 <= hi2) || (lo1 <= lo2 && lo2 <= hi1);
}
#endif
//------------------------------Ideal------------------------------------------
Node *ConvI2LNode::Ideal(PhaseGVN *phase, bool can_reshape) {
const TypeLong* this_type = this->type()->is_long();
Node* this_changed = NULL;
// If _major_progress, then more loop optimizations follow. Do NOT
// remove this node's type assertion until no more loop ops can happen.
// The progress bit is set in the major loop optimizations THEN comes the
// call to IterGVN and any chance of hitting this code. Cf. Opaque1Node.
if (can_reshape && !phase->C->major_progress()) {
const TypeInt* in_type = phase->type(in(1))->isa_int();
if (in_type != NULL && this_type != NULL &&
(in_type->_lo != this_type->_lo ||
in_type->_hi != this_type->_hi)) {
// Although this WORSENS the type, it increases GVN opportunities,
// because I2L nodes with the same input will common up, regardless
// of slightly differing type assertions. Such slight differences
// arise routinely as a result of loop unrolling, so this is a
// post-unrolling graph cleanup. Choose a type which depends only
// on my input. (Exception: Keep a range assertion of >=0 or <0.)
jlong lo1 = this_type->_lo;
jlong hi1 = this_type->_hi;
int w1 = this_type->_widen;
if (lo1 != (jint)lo1 ||
hi1 != (jint)hi1 ||
lo1 > hi1) {
// Overflow leads to wraparound, wraparound leads to range saturation.
lo1 = min_jint; hi1 = max_jint;
} else if (lo1 >= 0) {
// Keep a range assertion of >=0.
lo1 = 0; hi1 = max_jint;
} else if (hi1 < 0) {
// Keep a range assertion of <0.
lo1 = min_jint; hi1 = -1;
} else {
lo1 = min_jint; hi1 = max_jint;
}
const TypeLong* wtype = TypeLong::make(MAX2((jlong)in_type->_lo, lo1),
MIN2((jlong)in_type->_hi, hi1),
MAX2((int)in_type->_widen, w1));
if (wtype != type()) {
set_type(wtype);
// Note: this_type still has old type value, for the logic below.
this_changed = this;
}
}
}
#ifdef _LP64
// Convert ConvI2L(AddI(x, y)) to AddL(ConvI2L(x), ConvI2L(y))
// but only if x and y have subranges that cannot cause 32-bit overflow,
// under the assumption that x+y is in my own subrange this->type().
// This assumption is based on a constraint (i.e., type assertion)
// established in Parse::array_addressing or perhaps elsewhere.
// This constraint has been adjoined to the "natural" type of
// the incoming argument in(0). We know (because of runtime
// checks) - that the result value I2L(x+y) is in the joined range.
// Hence we can restrict the incoming terms (x, y) to values such
// that their sum also lands in that range.
// This optimization is useful only on 64-bit systems, where we hope
// the addition will end up subsumed in an addressing mode.
// It is necessary to do this when optimizing an unrolled array
// copy loop such as x[i++] = y[i++].
// On 32-bit systems, it's better to perform as much 32-bit math as
// possible before the I2L conversion, because 32-bit math is cheaper.
// There's no common reason to "leak" a constant offset through the I2L.
// Addressing arithmetic will not absorb it as part of a 64-bit AddL.
Node* z = in(1);
int op = z->Opcode();
if (op == Op_AddI || op == Op_SubI) {
Node* x = z->in(1);
Node* y = z->in(2);
assert (x != z && y != z, "dead loop in ConvI2LNode::Ideal");
if (phase->type(x) == Type::TOP) return this_changed;
if (phase->type(y) == Type::TOP) return this_changed;
const TypeInt* tx = phase->type(x)->is_int();
const TypeInt* ty = phase->type(y)->is_int();
const TypeLong* tz = this_type;
jlong xlo = tx->_lo;
jlong xhi = tx->_hi;
jlong ylo = ty->_lo;
jlong yhi = ty->_hi;
jlong zlo = tz->_lo;
jlong zhi = tz->_hi;
jlong vbit = CONST64(1) << BitsPerInt;
int widen = MAX2(tx->_widen, ty->_widen);
if (op == Op_SubI) {
jlong ylo0 = ylo;
ylo = -yhi;
yhi = -ylo0;
}
// See if x+y can cause positive overflow into z+2**32
if (long_ranges_overlap(xlo+ylo, xhi+yhi, zlo+vbit, zhi+vbit)) {
return this_changed;
}
// See if x+y can cause negative overflow into z-2**32
if (long_ranges_overlap(xlo+ylo, xhi+yhi, zlo-vbit, zhi-vbit)) {
return this_changed;
}
// Now it's always safe to assume x+y does not overflow.
// This is true even if some pairs x,y might cause overflow, as long
// as that overflow value cannot fall into [zlo,zhi].
// Confident that the arithmetic is "as if infinite precision",
// we can now use z's range to put constraints on those of x and y.
// The "natural" range of x [xlo,xhi] can perhaps be narrowed to a
// more "restricted" range by intersecting [xlo,xhi] with the
// range obtained by subtracting y's range from the asserted range
// of the I2L conversion. Here's the interval arithmetic algebra:
// x == z-y == [zlo,zhi]-[ylo,yhi] == [zlo,zhi]+[-yhi,-ylo]
// => x in [zlo-yhi, zhi-ylo]
// => x in [zlo-yhi, zhi-ylo] INTERSECT [xlo,xhi]
// => x in [xlo MAX zlo-yhi, xhi MIN zhi-ylo]
jlong rxlo = MAX2(xlo, zlo - yhi);
jlong rxhi = MIN2(xhi, zhi - ylo);
// And similarly, x changing place with y:
jlong rylo = MAX2(ylo, zlo - xhi);
jlong ryhi = MIN2(yhi, zhi - xlo);
if (rxlo > rxhi || rylo > ryhi) {
return this_changed; // x or y is dying; don't mess w/ it
}
if (op == Op_SubI) {
jlong rylo0 = rylo;
rylo = -ryhi;
ryhi = -rylo0;
}
assert(rxlo == (int)rxlo && rxhi == (int)rxhi, "x should not overflow");
assert(rylo == (int)rylo && ryhi == (int)ryhi, "y should not overflow");
Node* cx = phase->C->constrained_convI2L(phase, x, TypeInt::make(rxlo, rxhi, widen), NULL);
Node *hook = new (phase->C) Node(1);
hook->init_req(0, cx); // Add a use to cx to prevent him from dying
Node* cy = phase->C->constrained_convI2L(phase, y, TypeInt::make(rylo, ryhi, widen), NULL);
hook->del_req(0); // Just yank bogus edge
hook->destruct();
switch (op) {
case Op_AddI: return new (phase->C) AddLNode(cx, cy);
case Op_SubI: return new (phase->C) SubLNode(cx, cy);
default: ShouldNotReachHere();
}
}
#endif //_LP64
return this_changed;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvL2DNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeLong *tl = t->is_long();
if( tl->is_con() ) return TypeD::make( (double)tl->get_con() );
return bottom_type();
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ConvL2FNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeLong *tl = t->is_long();
if( tl->is_con() ) return TypeF::make( (float)tl->get_con() );
return bottom_type();
}
//=============================================================================
//----------------------------Identity-----------------------------------------
Node *ConvL2INode::Identity( PhaseTransform *phase ) {
// Convert L2I(I2L(x)) => x
if (in(1)->Opcode() == Op_ConvI2L) return in(1)->in(1);
return this;
}
//------------------------------Value------------------------------------------
const Type *ConvL2INode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeLong *tl = t->is_long();
if (tl->is_con())
// Easy case.
return TypeInt::make((jint)tl->get_con());
return bottom_type();
}
//------------------------------Ideal------------------------------------------
// Return a node which is more "ideal" than the current node.
// Blow off prior masking to int
Node *ConvL2INode::Ideal(PhaseGVN *phase, bool can_reshape) {
Node *andl = in(1);
uint andl_op = andl->Opcode();
if( andl_op == Op_AndL ) {
// Blow off prior masking to int
if( phase->type(andl->in(2)) == TypeLong::make( 0xFFFFFFFF ) ) {
set_req(1,andl->in(1));
return this;
}
}
// Swap with a prior add: convL2I(addL(x,y)) ==> addI(convL2I(x),convL2I(y))
// This replaces an 'AddL' with an 'AddI'.
if( andl_op == Op_AddL ) {
// Don't do this for nodes which have more than one user since
// we'll end up computing the long add anyway.
if (andl->outcnt() > 1) return NULL;
Node* x = andl->in(1);
Node* y = andl->in(2);
assert( x != andl && y != andl, "dead loop in ConvL2INode::Ideal" );
if (phase->type(x) == Type::TOP) return NULL;
if (phase->type(y) == Type::TOP) return NULL;
Node *add1 = phase->transform(new (phase->C) ConvL2INode(x));
Node *add2 = phase->transform(new (phase->C) ConvL2INode(y));
return new (phase->C) AddINode(add1,add2);
}
// Disable optimization: LoadL->ConvL2I ==> LoadI.
// It causes problems (sizes of Load and Store nodes do not match)
// in objects initialization code and Escape Analysis.
return NULL;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *CastX2PNode::Value( PhaseTransform *phase ) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
if (t->base() == Type_X && t->singleton()) {
uintptr_t bits = (uintptr_t) t->is_intptr_t()->get_con();
if (bits == 0) return TypePtr::NULL_PTR;
return TypeRawPtr::make((address) bits);
}
return CastX2PNode::bottom_type();
}
//------------------------------Idealize---------------------------------------
static inline bool fits_in_int(const Type* t, bool but_not_min_int = false) {
if (t == Type::TOP) return false;
const TypeX* tl = t->is_intptr_t();
jint lo = min_jint;
jint hi = max_jint;
if (but_not_min_int) ++lo; // caller wants to negate the value w/o overflow
return (tl->_lo >= lo) && (tl->_hi <= hi);
}
static inline Node* addP_of_X2P(PhaseGVN *phase,
Node* base,
Node* dispX,
bool negate = false) {
if (negate) {
dispX = new (phase->C) SubXNode(phase->MakeConX(0), phase->transform(dispX));
}
return new (phase->C) AddPNode(phase->C->top(),
phase->transform(new (phase->C) CastX2PNode(base)),
phase->transform(dispX));
}
Node *CastX2PNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// convert CastX2P(AddX(x, y)) to AddP(CastX2P(x), y) if y fits in an int
int op = in(1)->Opcode();
Node* x;
Node* y;
switch (op) {
case Op_SubX:
x = in(1)->in(1);
// Avoid ideal transformations ping-pong between this and AddP for raw pointers.
if (phase->find_intptr_t_con(x, -1) == 0)
break;
y = in(1)->in(2);
if (fits_in_int(phase->type(y), true)) {
return addP_of_X2P(phase, x, y, true);
}
break;
case Op_AddX:
x = in(1)->in(1);
y = in(1)->in(2);
if (fits_in_int(phase->type(y))) {
return addP_of_X2P(phase, x, y);
}
if (fits_in_int(phase->type(x))) {
return addP_of_X2P(phase, y, x);
}
break;
}
return NULL;
}
//------------------------------Identity---------------------------------------
Node *CastX2PNode::Identity( PhaseTransform *phase ) {
if (in(1)->Opcode() == Op_CastP2X) return in(1)->in(1);
return this;
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *CastP2XNode::Value( PhaseTransform *phase ) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
if (t->base() == Type::RawPtr && t->singleton()) {
uintptr_t bits = (uintptr_t) t->is_rawptr()->get_con();
return TypeX::make(bits);
}
return CastP2XNode::bottom_type();
}
Node *CastP2XNode::Ideal(PhaseGVN *phase, bool can_reshape) {
return (in(0) && remove_dead_region(phase, can_reshape)) ? this : NULL;
}
//------------------------------Identity---------------------------------------
Node *CastP2XNode::Identity( PhaseTransform *phase ) {
if (in(1)->Opcode() == Op_CastX2P) return in(1)->in(1);
return this;
}
//=============================================================================
//------------------------------Identity---------------------------------------
// Remove redundant roundings
Node *RoundFloatNode::Identity( PhaseTransform *phase ) {
assert(Matcher::strict_fp_requires_explicit_rounding, "should only generate for Intel");
// Do not round constants
if (phase->type(in(1))->base() == Type::FloatCon) return in(1);
int op = in(1)->Opcode();
// Redundant rounding
if( op == Op_RoundFloat ) return in(1);
// Already rounded
if( op == Op_Parm ) return in(1);
if( op == Op_LoadF ) return in(1);
return this;
}
//------------------------------Value------------------------------------------
const Type *RoundFloatNode::Value( PhaseTransform *phase ) const {
return phase->type( in(1) );
}
//=============================================================================
//------------------------------Identity---------------------------------------
// Remove redundant roundings. Incoming arguments are already rounded.
Node *RoundDoubleNode::Identity( PhaseTransform *phase ) {
assert(Matcher::strict_fp_requires_explicit_rounding, "should only generate for Intel");
// Do not round constants
if (phase->type(in(1))->base() == Type::DoubleCon) return in(1);
int op = in(1)->Opcode();
// Redundant rounding
if( op == Op_RoundDouble ) return in(1);
// Already rounded
if( op == Op_Parm ) return in(1);
if( op == Op_LoadD ) return in(1);
if( op == Op_ConvF2D ) return in(1);
if( op == Op_ConvI2D ) return in(1);
return this;
}
//------------------------------Value------------------------------------------
const Type *RoundDoubleNode::Value( PhaseTransform *phase ) const {
return phase->type( in(1) );
}
//=============================================================================
// Do not allow value-numbering
uint Opaque1Node::hash() const { return NO_HASH; }
uint Opaque1Node::cmp( const Node &n ) const {
return (&n == this); // Always fail except on self
}
//------------------------------Identity---------------------------------------
// If _major_progress, then more loop optimizations follow. Do NOT remove
// the opaque Node until no more loop ops can happen. Note the timing of
// _major_progress; it's set in the major loop optimizations THEN comes the
// call to IterGVN and any chance of hitting this code. Hence there's no
// phase-ordering problem with stripping Opaque1 in IGVN followed by some
// more loop optimizations that require it.
Node *Opaque1Node::Identity( PhaseTransform *phase ) {
return phase->C->major_progress() ? this : in(1);
}
//=============================================================================
// A node to prevent unwanted optimizations. Allows constant folding. Stops
// value-numbering, most Ideal calls or Identity functions. This Node is
// specifically designed to prevent the pre-increment value of a loop trip
// counter from being live out of the bottom of the loop (hence causing the
// pre- and post-increment values both being live and thus requiring an extra
// temp register and an extra move). If we "accidentally" optimize through
// this kind of a Node, we'll get slightly pessimal, but correct, code. Thus
// it's OK to be slightly sloppy on optimizations here.
// Do not allow value-numbering
uint Opaque2Node::hash() const { return NO_HASH; }
uint Opaque2Node::cmp( const Node &n ) const {
return (&n == this); // Always fail except on self
}
//=============================================================================
uint ProfileBooleanNode::hash() const { return NO_HASH; }
uint ProfileBooleanNode::cmp( const Node &n ) const {
return (&n == this);
}
Node *ProfileBooleanNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (can_reshape && _delay_removal) {
_delay_removal = false;
return this;
} else {
return NULL;
}
}
Node *ProfileBooleanNode::Identity( PhaseTransform *phase ) {
if (_delay_removal) {
return this;
} else {
assert(_consumed, "profile should be consumed before elimination");
return in(1);
}
}
//------------------------------Value------------------------------------------
const Type *MoveL2DNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeLong *tl = t->is_long();
if( !tl->is_con() ) return bottom_type();
JavaValue v;
v.set_jlong(tl->get_con());
return TypeD::make( v.get_jdouble() );
}
//------------------------------Value------------------------------------------
const Type *MoveI2FNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
const TypeInt *ti = t->is_int();
if( !ti->is_con() ) return bottom_type();
JavaValue v;
v.set_jint(ti->get_con());
return TypeF::make( v.get_jfloat() );
}
//------------------------------Value------------------------------------------
const Type *MoveF2INode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::FLOAT ) return TypeInt::INT;
const TypeF *tf = t->is_float_constant();
JavaValue v;
v.set_jfloat(tf->getf());
return TypeInt::make( v.get_jint() );
}
//------------------------------Value------------------------------------------
const Type *MoveD2LNode::Value( PhaseTransform *phase ) const {
const Type *t = phase->type( in(1) );
if( t == Type::TOP ) return Type::TOP;
if( t == Type::DOUBLE ) return TypeLong::LONG;
const TypeD *td = t->is_double_constant();
JavaValue v;
v.set_jdouble(td->getd());
return TypeLong::make( v.get_jlong() );
}
//------------------------------Value------------------------------------------
const Type* CountLeadingZerosINode::Value(PhaseTransform* phase) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
const TypeInt* ti = t->isa_int();
if (ti && ti->is_con()) {
jint i = ti->get_con();
// HD, Figure 5-6
if (i == 0)
return TypeInt::make(BitsPerInt);
int n = 1;
unsigned int x = i;
if (x >> 16 == 0) { n += 16; x <<= 16; }
if (x >> 24 == 0) { n += 8; x <<= 8; }
if (x >> 28 == 0) { n += 4; x <<= 4; }
if (x >> 30 == 0) { n += 2; x <<= 2; }
n -= x >> 31;
return TypeInt::make(n);
}
return TypeInt::INT;
}
//------------------------------Value------------------------------------------
const Type* CountLeadingZerosLNode::Value(PhaseTransform* phase) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
const TypeLong* tl = t->isa_long();
if (tl && tl->is_con()) {
jlong l = tl->get_con();
// HD, Figure 5-6
if (l == 0)
return TypeInt::make(BitsPerLong);
int n = 1;
unsigned int x = (((julong) l) >> 32);
if (x == 0) { n += 32; x = (int) l; }
if (x >> 16 == 0) { n += 16; x <<= 16; }
if (x >> 24 == 0) { n += 8; x <<= 8; }
if (x >> 28 == 0) { n += 4; x <<= 4; }
if (x >> 30 == 0) { n += 2; x <<= 2; }
n -= x >> 31;
return TypeInt::make(n);
}
return TypeInt::INT;
}
//------------------------------Value------------------------------------------
const Type* CountTrailingZerosINode::Value(PhaseTransform* phase) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
const TypeInt* ti = t->isa_int();
if (ti && ti->is_con()) {
jint i = ti->get_con();
// HD, Figure 5-14
int y;
if (i == 0)
return TypeInt::make(BitsPerInt);
int n = 31;
y = i << 16; if (y != 0) { n = n - 16; i = y; }
y = i << 8; if (y != 0) { n = n - 8; i = y; }
y = i << 4; if (y != 0) { n = n - 4; i = y; }
y = i << 2; if (y != 0) { n = n - 2; i = y; }
y = i << 1; if (y != 0) { n = n - 1; }
return TypeInt::make(n);
}
return TypeInt::INT;
}
//------------------------------Value------------------------------------------
const Type* CountTrailingZerosLNode::Value(PhaseTransform* phase) const {
const Type* t = phase->type(in(1));
if (t == Type::TOP) return Type::TOP;
const TypeLong* tl = t->isa_long();
if (tl && tl->is_con()) {
jlong l = tl->get_con();
// HD, Figure 5-14
int x, y;
if (l == 0)
return TypeInt::make(BitsPerLong);
int n = 63;
y = (int) l; if (y != 0) { n = n - 32; x = y; } else x = (((julong) l) >> 32);
y = x << 16; if (y != 0) { n = n - 16; x = y; }
y = x << 8; if (y != 0) { n = n - 8; x = y; }
y = x << 4; if (y != 0) { n = n - 4; x = y; }
y = x << 2; if (y != 0) { n = n - 2; x = y; }
y = x << 1; if (y != 0) { n = n - 1; }
return TypeInt::make(n);
}
return TypeInt::INT;
}
C:\hotspot-69087d08d473\src\share\vm/opto/connode.hpp
/*
* Copyright (c) 1997, 2012, 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.
*
*/
#ifndef SHARE_VM_OPTO_CONNODE_HPP
#define SHARE_VM_OPTO_CONNODE_HPP
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/type.hpp"
class PhaseTransform;
class MachNode;
//------------------------------ConNode----------------------------------------
// Simple constants
class ConNode : public TypeNode {
public:
ConNode( const Type *t ) : TypeNode(t->remove_speculative(),1) {
init_req(0, (Node*)Compile::current()->root());
init_flags(Flag_is_Con);
}
virtual int Opcode() const;
virtual uint hash() const;
virtual const RegMask &out_RegMask() const { return RegMask::Empty; }
virtual const RegMask &in_RegMask(uint) const { return RegMask::Empty; }
// Polymorphic factory method:
static ConNode* make( Compile* C, const Type *t );
};
//------------------------------ConINode---------------------------------------
// Simple integer constants
class ConINode : public ConNode {
public:
ConINode( const TypeInt *t ) : ConNode(t) {}
virtual int Opcode() const;
// Factory method:
static ConINode* make( Compile* C, int con ) {
return new (C) ConINode( TypeInt::make(con) );
}
};
//------------------------------ConPNode---------------------------------------
// Simple pointer constants
class ConPNode : public ConNode {
public:
ConPNode( const TypePtr *t ) : ConNode(t) {}
virtual int Opcode() const;
// Factory methods:
static ConPNode* make( Compile *C ,address con ) {
if (con == NULL)
return new (C) ConPNode( TypePtr::NULL_PTR ) ;
else
return new (C) ConPNode( TypeRawPtr::make(con) );
}
};
//------------------------------ConNNode--------------------------------------
// Simple narrow oop constants
class ConNNode : public ConNode {
public:
ConNNode( const TypeNarrowOop *t ) : ConNode(t) {}
virtual int Opcode() const;
};
//------------------------------ConNKlassNode---------------------------------
// Simple narrow klass constants
class ConNKlassNode : public ConNode {
public:
ConNKlassNode( const TypeNarrowKlass *t ) : ConNode(t) {}
virtual int Opcode() const;
};
//------------------------------ConLNode---------------------------------------
// Simple long constants
class ConLNode : public ConNode {
public:
ConLNode( const TypeLong *t ) : ConNode(t) {}
virtual int Opcode() const;
// Factory method:
static ConLNode* make( Compile *C ,jlong con ) {
return new (C) ConLNode( TypeLong::make(con) );
}
};
//------------------------------ConFNode---------------------------------------
// Simple float constants
class ConFNode : public ConNode {
public:
ConFNode( const TypeF *t ) : ConNode(t) {}
virtual int Opcode() const;
// Factory method:
static ConFNode* make( Compile *C, float con ) {
return new (C) ConFNode( TypeF::make(con) );
}
};
//------------------------------ConDNode---------------------------------------
// Simple double constants
class ConDNode : public ConNode {
public:
ConDNode( const TypeD *t ) : ConNode(t) {}
virtual int Opcode() const;
// Factory method:
static ConDNode* make( Compile *C, double con ) {
return new (C) ConDNode( TypeD::make(con) );
}
};
//------------------------------BinaryNode-------------------------------------
// Place holder for the 2 conditional inputs to a CMove. CMove needs 4
// inputs: the Bool (for the lt/gt/eq/ne bits), the flags (result of some
// compare), and the 2 values to select between. The Matcher requires a
// binary tree so we break it down like this:
// (CMove (Binary bol cmp) (Binary src1 src2))
class BinaryNode : public Node {
public:
BinaryNode( Node *n1, Node *n2 ) : Node(0,n1,n2) { }
virtual int Opcode() const;
virtual uint ideal_reg() const { return 0; }
};
//------------------------------CMoveNode--------------------------------------
// Conditional move
class CMoveNode : public TypeNode {
public:
enum { Control, // When is it safe to do this cmove?
Condition, // Condition controlling the cmove
IfFalse, // Value if condition is false
IfTrue }; // Value if condition is true
CMoveNode( Node *bol, Node *left, Node *right, const Type *t ) : TypeNode(t,4)
{
init_class_id(Class_CMove);
// all inputs are nullified in Node::Node(int)
// init_req(Control,NULL);
init_req(Condition,bol);
init_req(IfFalse,left);
init_req(IfTrue,right);
}
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
static CMoveNode *make( Compile *C, Node *c, Node *bol, Node *left, Node *right, const Type *t );
// Helper function to spot cmove graph shapes
static Node *is_cmove_id( PhaseTransform *phase, Node *cmp, Node *t, Node *f, BoolNode *b );
};
//------------------------------CMoveDNode-------------------------------------
class CMoveDNode : public CMoveNode {
public:
CMoveDNode( Node *bol, Node *left, Node *right, const Type* t) : CMoveNode(bol,left,right,t){}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
};
//------------------------------CMoveFNode-------------------------------------
class CMoveFNode : public CMoveNode {
public:
CMoveFNode( Node *bol, Node *left, Node *right, const Type* t ) : CMoveNode(bol,left,right,t) {}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
};
//------------------------------CMoveINode-------------------------------------
class CMoveINode : public CMoveNode {
public:
CMoveINode( Node *bol, Node *left, Node *right, const TypeInt *ti ) : CMoveNode(bol,left,right,ti){}
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
};
//------------------------------CMoveLNode-------------------------------------
class CMoveLNode : public CMoveNode {
public:
CMoveLNode(Node *bol, Node *left, Node *right, const TypeLong *tl ) : CMoveNode(bol,left,right,tl){}
virtual int Opcode() const;
};
//------------------------------CMovePNode-------------------------------------
class CMovePNode : public CMoveNode {
public:
CMovePNode( Node *c, Node *bol, Node *left, Node *right, const TypePtr* t ) : CMoveNode(bol,left,right,t) { init_req(Control,c); }
virtual int Opcode() const;
};
//------------------------------CMoveNNode-------------------------------------
class CMoveNNode : public CMoveNode {
public:
CMoveNNode( Node *c, Node *bol, Node *left, Node *right, const Type* t ) : CMoveNode(bol,left,right,t) { init_req(Control,c); }
virtual int Opcode() const;
};
//------------------------------ConstraintCastNode-----------------------------
// cast to a different range
class ConstraintCastNode: public TypeNode {
public:
ConstraintCastNode (Node *n, const Type *t ): TypeNode(t,2) {
init_class_id(Class_ConstraintCast);
init_req(1, n);
}
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual int Opcode() const;
virtual uint ideal_reg() const = 0;
virtual Node *Ideal_DU_postCCP( PhaseCCP * );
};
//------------------------------CastIINode-------------------------------------
// cast integer to integer (different range)
class CastIINode: public ConstraintCastNode {
private:
// Can this node be removed post CCP or does it carry a required dependency?
const bool _carry_dependency;
// Is this node dependent on a range check?
const bool _range_check_dependency;
protected:
virtual uint cmp( const Node &n ) const;
virtual uint size_of() const;
public:
CastIINode(Node *n, const Type *t, bool carry_dependency = false, bool range_check_dependency = false)
: ConstraintCastNode(n,t), _carry_dependency(carry_dependency), _range_check_dependency(range_check_dependency) {
init_class_id(Class_CastII);
}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegI; }
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal_DU_postCCP( PhaseCCP * );
const bool has_range_check() {
#ifdef _LP64
return _range_check_dependency;
#else
assert(!_range_check_dependency, "Should not have range check dependency");
return false;
#endif
}
#ifndef PRODUCT
virtual void dump_spec(outputStream *st) const;
#endif
};
//------------------------------CastPPNode-------------------------------------
// cast pointer to pointer (different type)
class CastPPNode: public ConstraintCastNode {
public:
CastPPNode (Node *n, const Type *t ): ConstraintCastNode(n, t) {}
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegP; }
virtual Node *Ideal_DU_postCCP( PhaseCCP * );
};
//------------------------------CheckCastPPNode--------------------------------
// for _checkcast, cast pointer to pointer (different type), without JOIN,
class CheckCastPPNode: public TypeNode {
public:
CheckCastPPNode( Node *c, Node *n, const Type *t ) : TypeNode(t,2) {
init_class_id(Class_CheckCastPP);
init_req(0, c);
init_req(1, n);
}
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegP; }
// No longer remove CheckCast after CCP as it gives me a place to hang
// the proper address type - which is required to compute anti-deps.
//virtual Node *Ideal_DU_postCCP( PhaseCCP * );
};
//------------------------------EncodeNarrowPtr--------------------------------
class EncodeNarrowPtrNode : public TypeNode {
protected:
EncodeNarrowPtrNode(Node* value, const Type* type):
TypeNode(type, 2) {
init_class_id(Class_EncodeNarrowPtr);
init_req(0, NULL);
init_req(1, value);
}
public:
virtual uint ideal_reg() const { return Op_RegN; }
virtual Node *Ideal_DU_postCCP( PhaseCCP *ccp );
};
//------------------------------EncodeP--------------------------------
// Encodes an oop pointers into its compressed form
// Takes an extra argument which is the real heap base as a long which
// may be useful for code generation in the backend.
class EncodePNode : public EncodeNarrowPtrNode {
public:
EncodePNode(Node* value, const Type* type):
EncodeNarrowPtrNode(value, type) {
init_class_id(Class_EncodeP);
}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
};
//------------------------------EncodePKlass--------------------------------
// Encodes a klass pointer into its compressed form
// Takes an extra argument which is the real heap base as a long which
// may be useful for code generation in the backend.
class EncodePKlassNode : public EncodeNarrowPtrNode {
public:
EncodePKlassNode(Node* value, const Type* type):
EncodeNarrowPtrNode(value, type) {
init_class_id(Class_EncodePKlass);
}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
};
//------------------------------DecodeNarrowPtr--------------------------------
class DecodeNarrowPtrNode : public TypeNode {
protected:
DecodeNarrowPtrNode(Node* value, const Type* type):
TypeNode(type, 2) {
init_class_id(Class_DecodeNarrowPtr);
init_req(0, NULL);
init_req(1, value);
}
public:
virtual uint ideal_reg() const { return Op_RegP; }
};
//------------------------------DecodeN--------------------------------
// Converts a narrow oop into a real oop ptr.
// Takes an extra argument which is the real heap base as a long which
// may be useful for code generation in the backend.
class DecodeNNode : public DecodeNarrowPtrNode {
public:
DecodeNNode(Node* value, const Type* type):
DecodeNarrowPtrNode(value, type) {
init_class_id(Class_DecodeN);
}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
};
//------------------------------DecodeNKlass--------------------------------
// Converts a narrow klass pointer into a real klass ptr.
// Takes an extra argument which is the real heap base as a long which
// may be useful for code generation in the backend.
class DecodeNKlassNode : public DecodeNarrowPtrNode {
public:
DecodeNKlassNode(Node* value, const Type* type):
DecodeNarrowPtrNode(value, type) {
init_class_id(Class_DecodeNKlass);
}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
};
//------------------------------Conv2BNode-------------------------------------
// Convert int/pointer to a Boolean. Map zero to zero, all else to 1.
class Conv2BNode : public Node {
public:
Conv2BNode( Node *i ) : Node(0,i) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::BOOL; }
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const { return Op_RegI; }
};
// The conversions operations are all Alpha sorted. Please keep it that way!
//------------------------------ConvD2FNode------------------------------------
// Convert double to float
class ConvD2FNode : public Node {
public:
ConvD2FNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------ConvD2INode------------------------------------
// Convert Double to Integer
class ConvD2INode : public Node {
public:
ConvD2INode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------ConvD2LNode------------------------------------
// Convert Double to Long
class ConvD2LNode : public Node {
public:
ConvD2LNode( Node *dbl ) : Node(0,dbl) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeLong::LONG; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------ConvF2DNode------------------------------------
// Convert Float to a Double.
class ConvF2DNode : public Node {
public:
ConvF2DNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const { return Op_RegD; }
};
//------------------------------ConvF2INode------------------------------------
// Convert float to integer
class ConvF2INode : public Node {
public:
ConvF2INode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------ConvF2LNode------------------------------------
// Convert float to long
class ConvF2LNode : public Node {
public:
ConvF2LNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeLong::LONG; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------ConvI2DNode------------------------------------
// Convert Integer to Double
class ConvI2DNode : public Node {
public:
ConvI2DNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const { return Op_RegD; }
};
//------------------------------ConvI2FNode------------------------------------
// Convert Integer to Float
class ConvI2FNode : public Node {
public:
ConvI2FNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Identity( PhaseTransform *phase );
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------ConvI2LNode------------------------------------
// Convert integer to long
class ConvI2LNode : public TypeNode {
public:
ConvI2LNode(Node *in1, const TypeLong* t = TypeLong::INT)
: TypeNode(t, 2)
{ init_req(1, in1); }
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------ConvL2DNode------------------------------------
// Convert Long to Double
class ConvL2DNode : public Node {
public:
ConvL2DNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const { return Op_RegD; }
};
//------------------------------ConvL2FNode------------------------------------
// Convert Long to Float
class ConvL2FNode : public Node {
public:
ConvL2FNode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual const Type *Value( PhaseTransform *phase ) const;
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------ConvL2INode------------------------------------
// Convert long to integer
class ConvL2INode : public Node {
public:
ConvL2INode( Node *in1 ) : Node(0,in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------CastX2PNode-------------------------------------
// convert a machine-pointer-sized integer to a raw pointer
class CastX2PNode : public Node {
public:
CastX2PNode( Node *n ) : Node(NULL, n) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual Node *Identity( PhaseTransform *phase );
virtual uint ideal_reg() const { return Op_RegP; }
virtual const Type *bottom_type() const { return TypeRawPtr::BOTTOM; }
};
//------------------------------CastP2XNode-------------------------------------
// Used in both 32-bit and 64-bit land.
// Used for card-marks and unsafe pointer math.
class CastP2XNode : public Node {
public:
CastP2XNode( Node *ctrl, Node *n ) : Node(ctrl, n) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual Node *Identity( PhaseTransform *phase );
virtual uint ideal_reg() const { return Op_RegX; }
virtual const Type *bottom_type() const { return TypeX_X; }
// Return false to keep node from moving away from an associated card mark.
virtual bool depends_only_on_test() const { return false; }
};
//------------------------------ThreadLocalNode--------------------------------
// Ideal Node which returns the base of ThreadLocalStorage.
class ThreadLocalNode : public Node {
public:
ThreadLocalNode( ) : Node((Node*)Compile::current()->root()) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeRawPtr::BOTTOM;}
virtual uint ideal_reg() const { return Op_RegP; }
};
//------------------------------LoadReturnPCNode-------------------------------
class LoadReturnPCNode: public Node {
public:
LoadReturnPCNode(Node *c) : Node(c) { }
virtual int Opcode() const;
virtual uint ideal_reg() const { return Op_RegP; }
};
//-----------------------------RoundFloatNode----------------------------------
class RoundFloatNode: public Node {
public:
RoundFloatNode(Node* c, Node *in1): Node(c, in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual uint ideal_reg() const { return Op_RegF; }
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
};
//-----------------------------RoundDoubleNode---------------------------------
class RoundDoubleNode: public Node {
public:
RoundDoubleNode(Node* c, Node *in1): Node(c, in1) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual uint ideal_reg() const { return Op_RegD; }
virtual Node *Identity( PhaseTransform *phase );
virtual const Type *Value( PhaseTransform *phase ) const;
};
//------------------------------Opaque1Node------------------------------------
// A node to prevent unwanted optimizations. Allows constant folding.
// Stops value-numbering, Ideal calls or Identity functions.
class Opaque1Node : public Node {
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const;
public:
Opaque1Node( Compile* C, Node *n ) : Node(0,n) {
// Put it on the Macro nodes list to removed during macro nodes expansion.
init_flags(Flag_is_macro);
C->add_macro_node(this);
}
// Special version for the pre-loop to hold the original loop limit
// which is consumed by range check elimination.
Opaque1Node( Compile* C, Node *n, Node* orig_limit ) : Node(0,n,orig_limit) {
// Put it on the Macro nodes list to removed during macro nodes expansion.
init_flags(Flag_is_macro);
C->add_macro_node(this);
}
Node* original_loop_limit() { return req()==3 ? in(2) : NULL; }
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual Node *Identity( PhaseTransform *phase );
};
//------------------------------Opaque2Node------------------------------------
// A node to prevent unwanted optimizations. Allows constant folding. Stops
// value-numbering, most Ideal calls or Identity functions. This Node is
// specifically designed to prevent the pre-increment value of a loop trip
// counter from being live out of the bottom of the loop (hence causing the
// pre- and post-increment values both being live and thus requiring an extra
// temp register and an extra move). If we "accidentally" optimize through
// this kind of a Node, we'll get slightly pessimal, but correct, code. Thus
// it's OK to be slightly sloppy on optimizations here.
class Opaque2Node : public Node {
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const;
public:
Opaque2Node( Compile* C, Node *n ) : Node(0,n) {
// Put it on the Macro nodes list to removed during macro nodes expansion.
init_flags(Flag_is_macro);
C->add_macro_node(this);
}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
};
//------------------------------Opaque3Node------------------------------------
// A node to prevent unwanted optimizations. Will be optimized only during
// macro nodes expansion.
class Opaque3Node : public Opaque2Node {
int _opt; // what optimization it was used for
public:
enum { RTM_OPT };
Opaque3Node(Compile* C, Node *n, int opt) : Opaque2Node(C, n), _opt(opt) {}
virtual int Opcode() const;
bool rtm_opt() const { return (_opt == RTM_OPT); }
};
//------------------------------ProfileBooleanNode-------------------------------
// A node represents value profile for a boolean during parsing.
// Once parsing is over, the node goes away (during IGVN).
// It is used to override branch frequencies from MDO (see has_injected_profile in parse2.cpp).
class ProfileBooleanNode : public Node {
uint _false_cnt;
uint _true_cnt;
bool _consumed;
bool _delay_removal;
virtual uint hash() const ; // { return NO_HASH; }
virtual uint cmp( const Node &n ) const;
public:
ProfileBooleanNode(Node *n, uint false_cnt, uint true_cnt) : Node(0, n),
_false_cnt(false_cnt), _true_cnt(true_cnt), _delay_removal(true), _consumed(false) {}
uint false_count() const { return _false_cnt; }
uint true_count() const { return _true_cnt; }
void consume() { _consumed = true; }
virtual int Opcode() const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual Node *Identity(PhaseTransform *phase);
virtual const Type *bottom_type() const { return TypeInt::BOOL; }
};
//----------------------PartialSubtypeCheckNode--------------------------------
// The 2nd slow-half of a subtype check. Scan the subklass's 2ndary superklass
// array for an instance of the superklass. Set a hidden internal cache on a
// hit (cache is checked with exposed code in gen_subtype_check()). Return
// not zero for a miss or zero for a hit.
class PartialSubtypeCheckNode : public Node {
public:
PartialSubtypeCheckNode(Node* c, Node* sub, Node* super) : Node(c,sub,super) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeRawPtr::BOTTOM; }
virtual uint ideal_reg() const { return Op_RegP; }
};
//
class MoveI2FNode : public Node {
public:
MoveI2FNode( Node *value ) : Node(0,value) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual uint ideal_reg() const { return Op_RegF; }
virtual const Type* Value( PhaseTransform *phase ) const;
};
class MoveL2DNode : public Node {
public:
MoveL2DNode( Node *value ) : Node(0,value) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual uint ideal_reg() const { return Op_RegD; }
virtual const Type* Value( PhaseTransform *phase ) const;
};
class MoveF2INode : public Node {
public:
MoveF2INode( Node *value ) : Node(0,value) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
virtual const Type* Value( PhaseTransform *phase ) const;
};
class MoveD2LNode : public Node {
public:
MoveD2LNode( Node *value ) : Node(0,value) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
virtual const Type* Value( PhaseTransform *phase ) const;
};
//---------- CountBitsNode -----------------------------------------------------
class CountBitsNode : public Node {
public:
CountBitsNode(Node* in1) : Node(0, in1) {}
const Type* bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//---------- CountLeadingZerosINode --------------------------------------------
// Count leading zeros (0-bit count starting from MSB) of an integer.
class CountLeadingZerosINode : public CountBitsNode {
public:
CountLeadingZerosINode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
virtual const Type* Value(PhaseTransform* phase) const;
};
//---------- CountLeadingZerosLNode --------------------------------------------
// Count leading zeros (0-bit count starting from MSB) of a long.
class CountLeadingZerosLNode : public CountBitsNode {
public:
CountLeadingZerosLNode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
virtual const Type* Value(PhaseTransform* phase) const;
};
//---------- CountTrailingZerosINode -------------------------------------------
// Count trailing zeros (0-bit count starting from LSB) of an integer.
class CountTrailingZerosINode : public CountBitsNode {
public:
CountTrailingZerosINode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
virtual const Type* Value(PhaseTransform* phase) const;
};
//---------- CountTrailingZerosLNode -------------------------------------------
// Count trailing zeros (0-bit count starting from LSB) of a long.
class CountTrailingZerosLNode : public CountBitsNode {
public:
CountTrailingZerosLNode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
virtual const Type* Value(PhaseTransform* phase) const;
};
//---------- PopCountINode -----------------------------------------------------
// Population count (bit count) of an integer.
class PopCountINode : public CountBitsNode {
public:
PopCountINode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
};
//---------- PopCountLNode -----------------------------------------------------
// Population count (bit count) of a long.
class PopCountLNode : public CountBitsNode {
public:
PopCountLNode(Node* in1) : CountBitsNode(in1) {}
virtual int Opcode() const;
};
#endif // SHARE_VM_OPTO_CONNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/divnode.cpp
/*
* Copyright (c) 1997, 2012, 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 "memory/allocation.inline.hpp"
#include "opto/addnode.hpp"
#include "opto/connode.hpp"
#include "opto/divnode.hpp"
#include "opto/machnode.hpp"
#include "opto/matcher.hpp"
#include "opto/mulnode.hpp"
#include "opto/phaseX.hpp"
#include "opto/subnode.hpp"
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
#include <math.h>
//----------------------magic_int_divide_constants-----------------------------
// Compute magic multiplier and shift constant for converting a 32 bit divide
// by constant into a multiply/shift/add series. Return false if calculations
// fail.
//
// Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with
// minor type name and parameter changes.
static bool magic_int_divide_constants(jint d, jint &M, jint &s) {
int32_t p;
uint32_t ad, anc, delta, q1, r1, q2, r2, t;
const uint32_t two31 = 0x80000000L; // 2**31.
ad = ABS(d);
if (d == 0 || d == 1) return false;
t = two31 + ((uint32_t)d >> 31);
anc = t - 1 - t%ad; // Absolute value of nc.
p = 31; // Init. p.
q1 = two31/anc; // Init. q1 = 2**p/|nc|.
r1 = two31 - q1*anc; // Init. r1 = rem(2**p, |nc|).
q2 = two31/ad; // Init. q2 = 2**p/|d|.
r2 = two31 - q2*ad; // Init. r2 = rem(2**p, |d|).
do {
p = p + 1;
q1 = 2*q1; // Update q1 = 2**p/|nc|.
r1 = 2*r1; // Update r1 = rem(2**p, |nc|).
if (r1 >= anc) { // (Must be an unsigned
q1 = q1 + 1; // comparison here).
r1 = r1 - anc;
}
q2 = 2*q2; // Update q2 = 2**p/|d|.
r2 = 2*r2; // Update r2 = rem(2**p, |d|).
if (r2 >= ad) { // (Must be an unsigned
q2 = q2 + 1; // comparison here).
r2 = r2 - ad;
}
delta = ad - r2;
} while (q1 < delta || (q1 == delta && r1 == 0));
M = q2 + 1;
if (d < 0) M = -M; // Magic number and
s = p - 32; // shift amount to return.
return true;
}
//--------------------------transform_int_divide-------------------------------
// Convert a division by constant divisor into an alternate Ideal graph.
// Return NULL if no transformation occurs.
static Node *transform_int_divide( PhaseGVN *phase, Node *dividend, jint divisor ) {
// Check for invalid divisors
assert( divisor != 0 && divisor != min_jint,
"bad divisor for transforming to long multiply" );
bool d_pos = divisor >= 0;
jint d = d_pos ? divisor : -divisor;
const int N = 32;
// Result
Node *q = NULL;
if (d == 1) {
// division by +/- 1
if (!d_pos) {
// Just negate the value
q = new (phase->C) SubINode(phase->intcon(0), dividend);
}
} else if ( is_power_of_2(d) ) {
// division by +/- a power of 2
// See if we can simply do a shift without rounding
bool needs_rounding = true;
const Type *dt = phase->type(dividend);
const TypeInt *dti = dt->isa_int();
if (dti && dti->_lo >= 0) {
// we don't need to round a positive dividend
needs_rounding = false;
} else if( dividend->Opcode() == Op_AndI ) {
// An AND mask of sufficient size clears the low bits and
// I can avoid rounding.
const TypeInt *andconi_t = phase->type( dividend->in(2) )->isa_int();
if( andconi_t && andconi_t->is_con() ) {
jint andconi = andconi_t->get_con();
if( andconi < 0 && is_power_of_2(-andconi) && (-andconi) >= d ) {
if( (-andconi) == d ) // Remove AND if it clears bits which will be shifted
dividend = dividend->in(1);
needs_rounding = false;
}
}
}
// Add rounding to the shift to handle the sign bit
int l = log2_jint(d-1)+1;
if (needs_rounding) {
// Divide-by-power-of-2 can be made into a shift, but you have to do
// more math for the rounding. You need to add 0 for positive
// numbers, and "i-1" for negative numbers. Example: i=4, so the
// shift is by 2. You need to add 3 to negative dividends and 0 to
// positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1,
// (-2+3)>>2 becomes 0, etc.
// Compute 0 or -1, based on sign bit
Node *sign = phase->transform(new (phase->C) RShiftINode(dividend, phase->intcon(N - 1)));
// Mask sign bit to the low sign bits
Node *round = phase->transform(new (phase->C) URShiftINode(sign, phase->intcon(N - l)));
// Round up before shifting
dividend = phase->transform(new (phase->C) AddINode(dividend, round));
}
// Shift for division
q = new (phase->C) RShiftINode(dividend, phase->intcon(l));
if (!d_pos) {
q = new (phase->C) SubINode(phase->intcon(0), phase->transform(q));
}
} else {
// Attempt the jint constant divide -> multiply transform found in
// "Division by Invariant Integers using Multiplication"
// by Granlund and Montgomery
// See also "Hacker's Delight", chapter 10 by Warren.
jint magic_const;
jint shift_const;
if (magic_int_divide_constants(d, magic_const, shift_const)) {
Node *magic = phase->longcon(magic_const);
Node *dividend_long = phase->transform(new (phase->C) ConvI2LNode(dividend));
// Compute the high half of the dividend x magic multiplication
Node *mul_hi = phase->transform(new (phase->C) MulLNode(dividend_long, magic));
if (magic_const < 0) {
mul_hi = phase->transform(new (phase->C) RShiftLNode(mul_hi, phase->intcon(N)));
mul_hi = phase->transform(new (phase->C) ConvL2INode(mul_hi));
// The magic multiplier is too large for a 32 bit constant. We've adjusted
// it down by 2^32, but have to add 1 dividend back in after the multiplication.
// This handles the "overflow" case described by Granlund and Montgomery.
mul_hi = phase->transform(new (phase->C) AddINode(dividend, mul_hi));
// Shift over the (adjusted) mulhi
if (shift_const != 0) {
mul_hi = phase->transform(new (phase->C) RShiftINode(mul_hi, phase->intcon(shift_const)));
}
} else {
// No add is required, we can merge the shifts together.
mul_hi = phase->transform(new (phase->C) RShiftLNode(mul_hi, phase->intcon(N + shift_const)));
mul_hi = phase->transform(new (phase->C) ConvL2INode(mul_hi));
}
// Get a 0 or -1 from the sign of the dividend.
Node *addend0 = mul_hi;
Node *addend1 = phase->transform(new (phase->C) RShiftINode(dividend, phase->intcon(N-1)));
// If the divisor is negative, swap the order of the input addends;
// this has the effect of negating the quotient.
if (!d_pos) {
Node *temp = addend0; addend0 = addend1; addend1 = temp;
}
// Adjust the final quotient by subtracting -1 (adding 1)
// from the mul_hi.
q = new (phase->C) SubINode(addend0, addend1);
}
}
return q;
}
//---------------------magic_long_divide_constants-----------------------------
// Compute magic multiplier and shift constant for converting a 64 bit divide
// by constant into a multiply/shift/add series. Return false if calculations
// fail.
//
// Borrowed almost verbatim from Hacker's Delight by Henry S. Warren, Jr. with
// minor type name and parameter changes. Adjusted to 64 bit word width.
static bool magic_long_divide_constants(jlong d, jlong &M, jint &s) {
int64_t p;
uint64_t ad, anc, delta, q1, r1, q2, r2, t;
const uint64_t two63 = 0x8000000000000000LL; // 2**63.
ad = ABS(d);
if (d == 0 || d == 1) return false;
t = two63 + ((uint64_t)d >> 63);
anc = t - 1 - t%ad; // Absolute value of nc.
p = 63; // Init. p.
q1 = two63/anc; // Init. q1 = 2**p/|nc|.
r1 = two63 - q1*anc; // Init. r1 = rem(2**p, |nc|).
q2 = two63/ad; // Init. q2 = 2**p/|d|.
r2 = two63 - q2*ad; // Init. r2 = rem(2**p, |d|).
do {
p = p + 1;
q1 = 2*q1; // Update q1 = 2**p/|nc|.
r1 = 2*r1; // Update r1 = rem(2**p, |nc|).
if (r1 >= anc) { // (Must be an unsigned
q1 = q1 + 1; // comparison here).
r1 = r1 - anc;
}
q2 = 2*q2; // Update q2 = 2**p/|d|.
r2 = 2*r2; // Update r2 = rem(2**p, |d|).
if (r2 >= ad) { // (Must be an unsigned
q2 = q2 + 1; // comparison here).
r2 = r2 - ad;
}
delta = ad - r2;
} while (q1 < delta || (q1 == delta && r1 == 0));
M = q2 + 1;
if (d < 0) M = -M; // Magic number and
s = p - 64; // shift amount to return.
return true;
}
//---------------------long_by_long_mulhi--------------------------------------
// Generate ideal node graph for upper half of a 64 bit x 64 bit multiplication
static Node* long_by_long_mulhi(PhaseGVN* phase, Node* dividend, jlong magic_const) {
// If the architecture supports a 64x64 mulhi, there is
// no need to synthesize it in ideal nodes.
if (Matcher::has_match_rule(Op_MulHiL)) {
Node* v = phase->longcon(magic_const);
return new (phase->C) MulHiLNode(dividend, v);
}
// Taken from Hacker's Delight, Fig. 8-2. Multiply high signed.
// (http://www.hackersdelight.org/HDcode/mulhs.c)
//
// int mulhs(int u, int v) {
// unsigned u0, v0, w0;
// int u1, v1, w1, w2, t;
//
// u0 = u & 0xFFFF; u1 = u >> 16;
// v0 = v & 0xFFFF; v1 = v >> 16;
// w0 = u0*v0;
// t = u1*v0 + (w0 >> 16);
// w1 = t & 0xFFFF;
// w2 = t >> 16;
// w1 = u0*v1 + w1;
// return u1*v1 + w2 + (w1 >> 16);
// }
//
// Note: The version above is for 32x32 multiplications, while the
// following inline comments are adapted to 64x64.
const int N = 64;
// Dummy node to keep intermediate nodes alive during construction
Node* hook = new (phase->C) Node(4);
// u0 = u & 0xFFFFFFFF; u1 = u >> 32;
Node* u0 = phase->transform(new (phase->C) AndLNode(dividend, phase->longcon(0xFFFFFFFF)));
Node* u1 = phase->transform(new (phase->C) RShiftLNode(dividend, phase->intcon(N / 2)));
hook->init_req(0, u0);
hook->init_req(1, u1);
// v0 = v & 0xFFFFFFFF; v1 = v >> 32;
Node* v0 = phase->longcon(magic_const & 0xFFFFFFFF);
Node* v1 = phase->longcon(magic_const >> (N / 2));
// w0 = u0*v0;
Node* w0 = phase->transform(new (phase->C) MulLNode(u0, v0));
// t = u1*v0 + (w0 >> 32);
Node* u1v0 = phase->transform(new (phase->C) MulLNode(u1, v0));
Node* temp = phase->transform(new (phase->C) URShiftLNode(w0, phase->intcon(N / 2)));
Node* t = phase->transform(new (phase->C) AddLNode(u1v0, temp));
hook->init_req(2, t);
// w1 = t & 0xFFFFFFFF;
Node* w1 = phase->transform(new (phase->C) AndLNode(t, phase->longcon(0xFFFFFFFF)));
hook->init_req(3, w1);
// w2 = t >> 32;
Node* w2 = phase->transform(new (phase->C) RShiftLNode(t, phase->intcon(N / 2)));
// w1 = u0*v1 + w1;
Node* u0v1 = phase->transform(new (phase->C) MulLNode(u0, v1));
w1 = phase->transform(new (phase->C) AddLNode(u0v1, w1));
// return u1*v1 + w2 + (w1 >> 32);
Node* u1v1 = phase->transform(new (phase->C) MulLNode(u1, v1));
Node* temp1 = phase->transform(new (phase->C) AddLNode(u1v1, w2));
Node* temp2 = phase->transform(new (phase->C) RShiftLNode(w1, phase->intcon(N / 2)));
// Remove the bogus extra edges used to keep things alive
PhaseIterGVN* igvn = phase->is_IterGVN();
if (igvn != NULL) {
igvn->remove_dead_node(hook);
} else {
for (int i = 0; i < 4; i++) {
hook->set_req(i, NULL);
}
}
return new (phase->C) AddLNode(temp1, temp2);
}
//--------------------------transform_long_divide------------------------------
// Convert a division by constant divisor into an alternate Ideal graph.
// Return NULL if no transformation occurs.
static Node *transform_long_divide( PhaseGVN *phase, Node *dividend, jlong divisor ) {
// Check for invalid divisors
assert( divisor != 0L && divisor != min_jlong,
"bad divisor for transforming to long multiply" );
bool d_pos = divisor >= 0;
jlong d = d_pos ? divisor : -divisor;
const int N = 64;
// Result
Node *q = NULL;
if (d == 1) {
// division by +/- 1
if (!d_pos) {
// Just negate the value
q = new (phase->C) SubLNode(phase->longcon(0), dividend);
}
} else if ( is_power_of_2_long(d) ) {
// division by +/- a power of 2
// See if we can simply do a shift without rounding
bool needs_rounding = true;
const Type *dt = phase->type(dividend);
const TypeLong *dtl = dt->isa_long();
if (dtl && dtl->_lo > 0) {
// we don't need to round a positive dividend
needs_rounding = false;
} else if( dividend->Opcode() == Op_AndL ) {
// An AND mask of sufficient size clears the low bits and
// I can avoid rounding.
const TypeLong *andconl_t = phase->type( dividend->in(2) )->isa_long();
if( andconl_t && andconl_t->is_con() ) {
jlong andconl = andconl_t->get_con();
if( andconl < 0 && is_power_of_2_long(-andconl) && (-andconl) >= d ) {
if( (-andconl) == d ) // Remove AND if it clears bits which will be shifted
dividend = dividend->in(1);
needs_rounding = false;
}
}
}
// Add rounding to the shift to handle the sign bit
int l = log2_long(d-1)+1;
if (needs_rounding) {
// Divide-by-power-of-2 can be made into a shift, but you have to do
// more math for the rounding. You need to add 0 for positive
// numbers, and "i-1" for negative numbers. Example: i=4, so the
// shift is by 2. You need to add 3 to negative dividends and 0 to
// positive ones. So (-7+3)>>2 becomes -1, (-4+3)>>2 becomes -1,
// (-2+3)>>2 becomes 0, etc.
// Compute 0 or -1, based on sign bit
Node *sign = phase->transform(new (phase->C) RShiftLNode(dividend, phase->intcon(N - 1)));
// Mask sign bit to the low sign bits
Node *round = phase->transform(new (phase->C) URShiftLNode(sign, phase->intcon(N - l)));
// Round up before shifting
dividend = phase->transform(new (phase->C) AddLNode(dividend, round));
}
// Shift for division
q = new (phase->C) RShiftLNode(dividend, phase->intcon(l));
if (!d_pos) {
q = new (phase->C) SubLNode(phase->longcon(0), phase->transform(q));
}
} else if ( !Matcher::use_asm_for_ldiv_by_con(d) ) { // Use hardware DIV instruction when
// it is faster than code generated below.
// Attempt the jlong constant divide -> multiply transform found in
// "Division by Invariant Integers using Multiplication"
// by Granlund and Montgomery
// See also "Hacker's Delight", chapter 10 by Warren.
jlong magic_const;
jint shift_const;
if (magic_long_divide_constants(d, magic_const, shift_const)) {
// Compute the high half of the dividend x magic multiplication
Node *mul_hi = phase->transform(long_by_long_mulhi(phase, dividend, magic_const));
// The high half of the 128-bit multiply is computed.
if (magic_const < 0) {
// The magic multiplier is too large for a 64 bit constant. We've adjusted
// it down by 2^64, but have to add 1 dividend back in after the multiplication.
// This handles the "overflow" case described by Granlund and Montgomery.
mul_hi = phase->transform(new (phase->C) AddLNode(dividend, mul_hi));
}
// Shift over the (adjusted) mulhi
if (shift_const != 0) {
mul_hi = phase->transform(new (phase->C) RShiftLNode(mul_hi, phase->intcon(shift_const)));
}
// Get a 0 or -1 from the sign of the dividend.
Node *addend0 = mul_hi;
Node *addend1 = phase->transform(new (phase->C) RShiftLNode(dividend, phase->intcon(N-1)));
// If the divisor is negative, swap the order of the input addends;
// this has the effect of negating the quotient.
if (!d_pos) {
Node *temp = addend0; addend0 = addend1; addend1 = temp;
}
// Adjust the final quotient by subtracting -1 (adding 1)
// from the mul_hi.
q = new (phase->C) SubLNode(addend0, addend1);
}
}
return q;
}
//=============================================================================
//------------------------------Identity---------------------------------------
// If the divisor is 1, we are an identity on the dividend.
Node *DivINode::Identity( PhaseTransform *phase ) {
return (phase->type( in(2) )->higher_equal(TypeInt::ONE)) ? in(1) : this;
}
//------------------------------Idealize---------------------------------------
// Divides can be changed to multiplies and/or shifts
Node *DivINode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (in(0) && remove_dead_region(phase, can_reshape)) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
const Type *t = phase->type( in(2) );
if( t == TypeInt::ONE ) // Identity?
return NULL; // Skip it
const TypeInt *ti = t->isa_int();
if( !ti ) return NULL;
if( !ti->is_con() ) return NULL;
jint i = ti->get_con(); // Get divisor
if (i == 0) return NULL; // Dividing by zero constant does not idealize
set_req(0,NULL); // Dividing by a not-zero constant; no faulting
// Dividing by MININT does not optimize as a power-of-2 shift.
if( i == min_jint ) return NULL;
return transform_int_divide( phase, in(1), i );
}
//------------------------------Value------------------------------------------
// A DivINode divides its inputs. The third input is a Control input, used to
// prevent hoisting the divide above an unsafe test.
const Type *DivINode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// x/x == 1 since we always generate the dynamic divisor check for 0.
if( phase->eqv( in(1), in(2) ) )
return TypeInt::ONE;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// Divide the two numbers. We approximate.
// If divisor is a constant and not zero
const TypeInt *i1 = t1->is_int();
const TypeInt *i2 = t2->is_int();
int widen = MAX2(i1->_widen, i2->_widen);
if( i2->is_con() && i2->get_con() != 0 ) {
int32 d = i2->get_con(); // Divisor
jint lo, hi;
if( d >= 0 ) {
lo = i1->_lo/d;
hi = i1->_hi/d;
} else {
if( d == -1 && i1->_lo == min_jint ) {
// 'min_jint/-1' throws arithmetic exception during compilation
lo = min_jint;
// do not support holes, 'hi' must go to either min_jint or max_jint:
// [min_jint, -10]/[-1,-1] ==> [min_jint] UNION [10,max_jint]
hi = i1->_hi == min_jint ? min_jint : max_jint;
} else {
lo = i1->_hi/d;
hi = i1->_lo/d;
}
}
return TypeInt::make(lo, hi, widen);
}
// If the dividend is a constant
if( i1->is_con() ) {
int32 d = i1->get_con();
if( d < 0 ) {
if( d == min_jint ) {
// (-min_jint) == min_jint == (min_jint / -1)
return TypeInt::make(min_jint, max_jint/2 + 1, widen);
} else {
return TypeInt::make(d, -d, widen);
}
}
return TypeInt::make(-d, d, widen);
}
// Otherwise we give up all hope
return TypeInt::INT;
}
//=============================================================================
//------------------------------Identity---------------------------------------
// If the divisor is 1, we are an identity on the dividend.
Node *DivLNode::Identity( PhaseTransform *phase ) {
return (phase->type( in(2) )->higher_equal(TypeLong::ONE)) ? in(1) : this;
}
//------------------------------Idealize---------------------------------------
// Dividing by a power of 2 is a shift.
Node *DivLNode::Ideal( PhaseGVN *phase, bool can_reshape) {
if (in(0) && remove_dead_region(phase, can_reshape)) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
const Type *t = phase->type( in(2) );
if( t == TypeLong::ONE ) // Identity?
return NULL; // Skip it
const TypeLong *tl = t->isa_long();
if( !tl ) return NULL;
if( !tl->is_con() ) return NULL;
jlong l = tl->get_con(); // Get divisor
if (l == 0) return NULL; // Dividing by zero constant does not idealize
set_req(0,NULL); // Dividing by a not-zero constant; no faulting
// Dividing by MINLONG does not optimize as a power-of-2 shift.
if( l == min_jlong ) return NULL;
return transform_long_divide( phase, in(1), l );
}
//------------------------------Value------------------------------------------
// A DivLNode divides its inputs. The third input is a Control input, used to
// prevent hoisting the divide above an unsafe test.
const Type *DivLNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// x/x == 1 since we always generate the dynamic divisor check for 0.
if( phase->eqv( in(1), in(2) ) )
return TypeLong::ONE;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// Divide the two numbers. We approximate.
// If divisor is a constant and not zero
const TypeLong *i1 = t1->is_long();
const TypeLong *i2 = t2->is_long();
int widen = MAX2(i1->_widen, i2->_widen);
if( i2->is_con() && i2->get_con() != 0 ) {
jlong d = i2->get_con(); // Divisor
jlong lo, hi;
if( d >= 0 ) {
lo = i1->_lo/d;
hi = i1->_hi/d;
} else {
if( d == CONST64(-1) && i1->_lo == min_jlong ) {
// 'min_jlong/-1' throws arithmetic exception during compilation
lo = min_jlong;
// do not support holes, 'hi' must go to either min_jlong or max_jlong:
// [min_jlong, -10]/[-1,-1] ==> [min_jlong] UNION [10,max_jlong]
hi = i1->_hi == min_jlong ? min_jlong : max_jlong;
} else {
lo = i1->_hi/d;
hi = i1->_lo/d;
}
}
return TypeLong::make(lo, hi, widen);
}
// If the dividend is a constant
if( i1->is_con() ) {
jlong d = i1->get_con();
if( d < 0 ) {
if( d == min_jlong ) {
// (-min_jlong) == min_jlong == (min_jlong / -1)
return TypeLong::make(min_jlong, max_jlong/2 + 1, widen);
} else {
return TypeLong::make(d, -d, widen);
}
}
return TypeLong::make(-d, d, widen);
}
// Otherwise we give up all hope
return TypeLong::LONG;
}
//=============================================================================
//------------------------------Value------------------------------------------
// An DivFNode divides its inputs. The third input is a Control input, used to
// prevent hoisting the divide above an unsafe test.
const Type *DivFNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// x/x == 1, we ignore 0/0.
// Note: if t1 and t2 are zero then result is NaN (JVMS page 213)
// Does not work for variables because of NaN's
if( phase->eqv( in(1), in(2) ) && t1->base() == Type::FloatCon)
if (!g_isnan(t1->getf()) && g_isfinite(t1->getf()) && t1->getf() != 0.0) // could be negative ZERO or NaN
return TypeF::ONE;
if( t2 == TypeF::ONE )
return t1;
// If divisor is a constant and not zero, divide them numbers
if( t1->base() == Type::FloatCon &&
t2->base() == Type::FloatCon &&
t2->getf() != 0.0 ) // could be negative zero
return TypeF::make( t1->getf()/t2->getf() );
// If the dividend is a constant zero
// Note: if t1 and t2 are zero then result is NaN (JVMS page 213)
// Test TypeF::ZERO is not sufficient as it could be negative zero
if( t1 == TypeF::ZERO && !g_isnan(t2->getf()) && t2->getf() != 0.0 )
return TypeF::ZERO;
// Otherwise we give up all hope
return Type::FLOAT;
}
//------------------------------isA_Copy---------------------------------------
// Dividing by self is 1.
// If the divisor is 1, we are an identity on the dividend.
Node *DivFNode::Identity( PhaseTransform *phase ) {
return (phase->type( in(2) ) == TypeF::ONE) ? in(1) : this;
}
//------------------------------Idealize---------------------------------------
Node *DivFNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (in(0) && remove_dead_region(phase, can_reshape)) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
const Type *t2 = phase->type( in(2) );
if( t2 == TypeF::ONE ) // Identity?
return NULL; // Skip it
const TypeF *tf = t2->isa_float_constant();
if( !tf ) return NULL;
if( tf->base() != Type::FloatCon ) return NULL;
// Check for out of range values
if( tf->is_nan() || !tf->is_finite() ) return NULL;
// Get the value
float f = tf->getf();
int exp;
// Only for special case of dividing by a power of 2
if( frexp((double)f, &exp) != 0.5 ) return NULL;
// Limit the range of acceptable exponents
if( exp < -126 || exp > 126 ) return NULL;
// Compute the reciprocal
float reciprocal = ((float)1.0) / f;
assert( frexp((double)reciprocal, &exp) == 0.5, "reciprocal should be power of 2" );
// return multiplication by the reciprocal
return (new (phase->C) MulFNode(in(1), phase->makecon(TypeF::make(reciprocal))));
}
//=============================================================================
//------------------------------Value------------------------------------------
// An DivDNode divides its inputs. The third input is a Control input, used to
// prevent hoisting the divide above an unsafe test.
const Type *DivDNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// x/x == 1, we ignore 0/0.
// Note: if t1 and t2 are zero then result is NaN (JVMS page 213)
// Does not work for variables because of NaN's
if( phase->eqv( in(1), in(2) ) && t1->base() == Type::DoubleCon)
if (!g_isnan(t1->getd()) && g_isfinite(t1->getd()) && t1->getd() != 0.0) // could be negative ZERO or NaN
return TypeD::ONE;
if( t2 == TypeD::ONE )
return t1;
#if defined(IA32)
if (!phase->C->method()->is_strict())
// Can't trust native compilers to properly fold strict double
// division with round-to-zero on this platform.
#endif
{
// If divisor is a constant and not zero, divide them numbers
if( t1->base() == Type::DoubleCon &&
t2->base() == Type::DoubleCon &&
t2->getd() != 0.0 ) // could be negative zero
return TypeD::make( t1->getd()/t2->getd() );
}
// If the dividend is a constant zero
// Note: if t1 and t2 are zero then result is NaN (JVMS page 213)
// Test TypeF::ZERO is not sufficient as it could be negative zero
if( t1 == TypeD::ZERO && !g_isnan(t2->getd()) && t2->getd() != 0.0 )
return TypeD::ZERO;
// Otherwise we give up all hope
return Type::DOUBLE;
}
//------------------------------isA_Copy---------------------------------------
// Dividing by self is 1.
// If the divisor is 1, we are an identity on the dividend.
Node *DivDNode::Identity( PhaseTransform *phase ) {
return (phase->type( in(2) ) == TypeD::ONE) ? in(1) : this;
}
//------------------------------Idealize---------------------------------------
Node *DivDNode::Ideal(PhaseGVN *phase, bool can_reshape) {
if (in(0) && remove_dead_region(phase, can_reshape)) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
const Type *t2 = phase->type( in(2) );
if( t2 == TypeD::ONE ) // Identity?
return NULL; // Skip it
const TypeD *td = t2->isa_double_constant();
if( !td ) return NULL;
if( td->base() != Type::DoubleCon ) return NULL;
// Check for out of range values
if( td->is_nan() || !td->is_finite() ) return NULL;
// Get the value
double d = td->getd();
int exp;
// Only for special case of dividing by a power of 2
if( frexp(d, &exp) != 0.5 ) return NULL;
// Limit the range of acceptable exponents
if( exp < -1021 || exp > 1022 ) return NULL;
// Compute the reciprocal
double reciprocal = 1.0 / d;
assert( frexp(reciprocal, &exp) == 0.5, "reciprocal should be power of 2" );
// return multiplication by the reciprocal
return (new (phase->C) MulDNode(in(1), phase->makecon(TypeD::make(reciprocal))));
}
//=============================================================================
//------------------------------Idealize---------------------------------------
Node *ModINode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Check for dead control input
if( in(0) && remove_dead_region(phase, can_reshape) ) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
// Get the modulus
const Type *t = phase->type( in(2) );
if( t == Type::TOP ) return NULL;
const TypeInt *ti = t->is_int();
// Check for useless control input
// Check for excluding mod-zero case
if( in(0) && (ti->_hi < 0 || ti->_lo > 0) ) {
set_req(0, NULL); // Yank control input
return this;
}
// See if we are MOD'ing by 2^k or 2^k-1.
if( !ti->is_con() ) return NULL;
jint con = ti->get_con();
Node *hook = new (phase->C) Node(1);
// First, special check for modulo 2^k-1
if( con >= 0 && con < max_jint && is_power_of_2(con+1) ) {
uint k = exact_log2(con+1); // Extract k
// Basic algorithm by David Detlefs. See fastmod_int.java for gory details.
static int unroll_factor[] = { 999, 999, 29, 14, 9, 7, 5, 4, 4, 3, 3, 2, 2, 2, 2, 2, 1 /*past here we assume 1 forever*/};
int trip_count = 1;
if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k];
// If the unroll factor is not too large, and if conditional moves are
// ok, then use this case
if( trip_count <= 5 && ConditionalMoveLimit != 0 ) {
Node *x = in(1); // Value being mod'd
Node *divisor = in(2); // Also is mask
hook->init_req(0, x); // Add a use to x to prevent him from dying
// Generate code to reduce X rapidly to nearly 2^k-1.
for( int i = 0; i < trip_count; i++ ) {
Node *xl = phase->transform( new (phase->C) AndINode(x,divisor) );
Node *xh = phase->transform( new (phase->C) RShiftINode(x,phase->intcon(k)) ); // Must be signed
x = phase->transform( new (phase->C) AddINode(xh,xl) );
hook->set_req(0, x);
}
// Generate sign-fixup code. Was original value positive?
// int hack_res = (i >= 0) ? divisor : 1;
Node *cmp1 = phase->transform( new (phase->C) CmpINode( in(1), phase->intcon(0) ) );
Node *bol1 = phase->transform( new (phase->C) BoolNode( cmp1, BoolTest::ge ) );
Node *cmov1= phase->transform( new (phase->C) CMoveINode(bol1, phase->intcon(1), divisor, TypeInt::POS) );
// if( x >= hack_res ) x -= divisor;
Node *sub = phase->transform( new (phase->C) SubINode( x, divisor ) );
Node *cmp2 = phase->transform( new (phase->C) CmpINode( x, cmov1 ) );
Node *bol2 = phase->transform( new (phase->C) BoolNode( cmp2, BoolTest::ge ) );
// Convention is to not transform the return value of an Ideal
// since Ideal is expected to return a modified 'this' or a new node.
Node *cmov2= new (phase->C) CMoveINode(bol2, x, sub, TypeInt::INT);
// cmov2 is now the mod
// Now remove the bogus extra edges used to keep things alive
if (can_reshape) {
phase->is_IterGVN()->remove_dead_node(hook);
} else {
hook->set_req(0, NULL); // Just yank bogus edge during Parse phase
}
return cmov2;
}
}
// Fell thru, the unroll case is not appropriate. Transform the modulo
// into a long multiply/int multiply/subtract case
// Cannot handle mod 0, and min_jint isn't handled by the transform
if( con == 0 || con == min_jint ) return NULL;
// Get the absolute value of the constant; at this point, we can use this
jint pos_con = (con >= 0) ? con : -con;
// integer Mod 1 is always 0
if( pos_con == 1 ) return new (phase->C) ConINode(TypeInt::ZERO);
int log2_con = -1;
// If this is a power of two, they maybe we can mask it
if( is_power_of_2(pos_con) ) {
log2_con = log2_intptr((intptr_t)pos_con);
const Type *dt = phase->type(in(1));
const TypeInt *dti = dt->isa_int();
// See if this can be masked, if the dividend is non-negative
if( dti && dti->_lo >= 0 )
return ( new (phase->C) AndINode( in(1), phase->intcon( pos_con-1 ) ) );
}
// Save in(1) so that it cannot be changed or deleted
hook->init_req(0, in(1));
// Divide using the transform from DivI to MulL
Node *result = transform_int_divide( phase, in(1), pos_con );
if (result != NULL) {
Node *divide = phase->transform(result);
// Re-multiply, using a shift if this is a power of two
Node *mult = NULL;
if( log2_con >= 0 )
mult = phase->transform( new (phase->C) LShiftINode( divide, phase->intcon( log2_con ) ) );
else
mult = phase->transform( new (phase->C) MulINode( divide, phase->intcon( pos_con ) ) );
// Finally, subtract the multiplied divided value from the original
result = new (phase->C) SubINode( in(1), mult );
}
// Now remove the bogus extra edges used to keep things alive
if (can_reshape) {
phase->is_IterGVN()->remove_dead_node(hook);
} else {
hook->set_req(0, NULL); // Just yank bogus edge during Parse phase
}
// return the value
return result;
}
//------------------------------Value------------------------------------------
const Type *ModINode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// We always generate the dynamic check for 0.
// 0 MOD X is 0
if( t1 == TypeInt::ZERO ) return TypeInt::ZERO;
// X MOD X is 0
if( phase->eqv( in(1), in(2) ) ) return TypeInt::ZERO;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
const TypeInt *i1 = t1->is_int();
const TypeInt *i2 = t2->is_int();
if( !i1->is_con() || !i2->is_con() ) {
if( i1->_lo >= 0 && i2->_lo >= 0 )
return TypeInt::POS;
// If both numbers are not constants, we know little.
return TypeInt::INT;
}
// Mod by zero? Throw exception at runtime!
if( !i2->get_con() ) return TypeInt::POS;
// We must be modulo'ing 2 float constants.
// Check for min_jint % '-1', result is defined to be '0'.
if( i1->get_con() == min_jint && i2->get_con() == -1 )
return TypeInt::ZERO;
return TypeInt::make( i1->get_con() % i2->get_con() );
}
//=============================================================================
//------------------------------Idealize---------------------------------------
Node *ModLNode::Ideal(PhaseGVN *phase, bool can_reshape) {
// Check for dead control input
if( in(0) && remove_dead_region(phase, can_reshape) ) return this;
// Don't bother trying to transform a dead node
if( in(0) && in(0)->is_top() ) return NULL;
// Get the modulus
const Type *t = phase->type( in(2) );
if( t == Type::TOP ) return NULL;
const TypeLong *tl = t->is_long();
// Check for useless control input
// Check for excluding mod-zero case
if( in(0) && (tl->_hi < 0 || tl->_lo > 0) ) {
set_req(0, NULL); // Yank control input
return this;
}
// See if we are MOD'ing by 2^k or 2^k-1.
if( !tl->is_con() ) return NULL;
jlong con = tl->get_con();
Node *hook = new (phase->C) Node(1);
// Expand mod
if( con >= 0 && con < max_jlong && is_power_of_2_long(con+1) ) {
uint k = exact_log2_long(con+1); // Extract k
// Basic algorithm by David Detlefs. See fastmod_long.java for gory details.
// Used to help a popular random number generator which does a long-mod
// of 2^31-1 and shows up in SpecJBB and SciMark.
static int unroll_factor[] = { 999, 999, 61, 30, 20, 15, 12, 10, 8, 7, 6, 6, 5, 5, 4, 4, 4, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1 /*past here we assume 1 forever*/};
int trip_count = 1;
if( k < ARRAY_SIZE(unroll_factor)) trip_count = unroll_factor[k];
// If the unroll factor is not too large, and if conditional moves are
// ok, then use this case
if( trip_count <= 5 && ConditionalMoveLimit != 0 ) {
Node *x = in(1); // Value being mod'd
Node *divisor = in(2); // Also is mask
hook->init_req(0, x); // Add a use to x to prevent him from dying
// Generate code to reduce X rapidly to nearly 2^k-1.
for( int i = 0; i < trip_count; i++ ) {
Node *xl = phase->transform( new (phase->C) AndLNode(x,divisor) );
Node *xh = phase->transform( new (phase->C) RShiftLNode(x,phase->intcon(k)) ); // Must be signed
x = phase->transform( new (phase->C) AddLNode(xh,xl) );
hook->set_req(0, x); // Add a use to x to prevent him from dying
}
// Generate sign-fixup code. Was original value positive?
// long hack_res = (i >= 0) ? divisor : CONST64(1);
Node *cmp1 = phase->transform( new (phase->C) CmpLNode( in(1), phase->longcon(0) ) );
Node *bol1 = phase->transform( new (phase->C) BoolNode( cmp1, BoolTest::ge ) );
Node *cmov1= phase->transform( new (phase->C) CMoveLNode(bol1, phase->longcon(1), divisor, TypeLong::LONG) );
// if( x >= hack_res ) x -= divisor;
Node *sub = phase->transform( new (phase->C) SubLNode( x, divisor ) );
Node *cmp2 = phase->transform( new (phase->C) CmpLNode( x, cmov1 ) );
Node *bol2 = phase->transform( new (phase->C) BoolNode( cmp2, BoolTest::ge ) );
// Convention is to not transform the return value of an Ideal
// since Ideal is expected to return a modified 'this' or a new node.
Node *cmov2= new (phase->C) CMoveLNode(bol2, x, sub, TypeLong::LONG);
// cmov2 is now the mod
// Now remove the bogus extra edges used to keep things alive
if (can_reshape) {
phase->is_IterGVN()->remove_dead_node(hook);
} else {
hook->set_req(0, NULL); // Just yank bogus edge during Parse phase
}
return cmov2;
}
}
// Fell thru, the unroll case is not appropriate. Transform the modulo
// into a long multiply/int multiply/subtract case
// Cannot handle mod 0, and min_jlong isn't handled by the transform
if( con == 0 || con == min_jlong ) return NULL;
// Get the absolute value of the constant; at this point, we can use this
jlong pos_con = (con >= 0) ? con : -con;
// integer Mod 1 is always 0
if( pos_con == 1 ) return new (phase->C) ConLNode(TypeLong::ZERO);
int log2_con = -1;
// If this is a power of two, then maybe we can mask it
if( is_power_of_2_long(pos_con) ) {
log2_con = exact_log2_long(pos_con);
const Type *dt = phase->type(in(1));
const TypeLong *dtl = dt->isa_long();
// See if this can be masked, if the dividend is non-negative
if( dtl && dtl->_lo >= 0 )
return ( new (phase->C) AndLNode( in(1), phase->longcon( pos_con-1 ) ) );
}
// Save in(1) so that it cannot be changed or deleted
hook->init_req(0, in(1));
// Divide using the transform from DivL to MulL
Node *result = transform_long_divide( phase, in(1), pos_con );
if (result != NULL) {
Node *divide = phase->transform(result);
// Re-multiply, using a shift if this is a power of two
Node *mult = NULL;
if( log2_con >= 0 )
mult = phase->transform( new (phase->C) LShiftLNode( divide, phase->intcon( log2_con ) ) );
else
mult = phase->transform( new (phase->C) MulLNode( divide, phase->longcon( pos_con ) ) );
// Finally, subtract the multiplied divided value from the original
result = new (phase->C) SubLNode( in(1), mult );
}
// Now remove the bogus extra edges used to keep things alive
if (can_reshape) {
phase->is_IterGVN()->remove_dead_node(hook);
} else {
hook->set_req(0, NULL); // Just yank bogus edge during Parse phase
}
// return the value
return result;
}
//------------------------------Value------------------------------------------
const Type *ModLNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// We always generate the dynamic check for 0.
// 0 MOD X is 0
if( t1 == TypeLong::ZERO ) return TypeLong::ZERO;
// X MOD X is 0
if( phase->eqv( in(1), in(2) ) ) return TypeLong::ZERO;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
const TypeLong *i1 = t1->is_long();
const TypeLong *i2 = t2->is_long();
if( !i1->is_con() || !i2->is_con() ) {
if( i1->_lo >= CONST64(0) && i2->_lo >= CONST64(0) )
return TypeLong::POS;
// If both numbers are not constants, we know little.
return TypeLong::LONG;
}
// Mod by zero? Throw exception at runtime!
if( !i2->get_con() ) return TypeLong::POS;
// We must be modulo'ing 2 float constants.
// Check for min_jint % '-1', result is defined to be '0'.
if( i1->get_con() == min_jlong && i2->get_con() == -1 )
return TypeLong::ZERO;
return TypeLong::make( i1->get_con() % i2->get_con() );
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ModFNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// If either number is not a constant, we know nothing.
if ((t1->base() != Type::FloatCon) || (t2->base() != Type::FloatCon)) {
return Type::FLOAT; // note: x%x can be either NaN or 0
}
float f1 = t1->getf();
float f2 = t2->getf();
jint x1 = jint_cast(f1); // note: *(int*)&f1, not just (int)f1
jint x2 = jint_cast(f2);
// If either is a NaN, return an input NaN
if (g_isnan(f1)) return t1;
if (g_isnan(f2)) return t2;
// If an operand is infinity or the divisor is +/- zero, punt.
if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jint)
return Type::FLOAT;
// We must be modulo'ing 2 float constants.
// Make sure that the sign of the fmod is equal to the sign of the dividend
jint xr = jint_cast(fmod(f1, f2));
if ((x1 ^ xr) < 0) {
xr ^= min_jint;
}
return TypeF::make(jfloat_cast(xr));
}
//=============================================================================
//------------------------------Value------------------------------------------
const Type *ModDNode::Value( PhaseTransform *phase ) const {
// Either input is TOP ==> the result is TOP
const Type *t1 = phase->type( in(1) );
const Type *t2 = phase->type( in(2) );
if( t1 == Type::TOP ) return Type::TOP;
if( t2 == Type::TOP ) return Type::TOP;
// Either input is BOTTOM ==> the result is the local BOTTOM
const Type *bot = bottom_type();
if( (t1 == bot) || (t2 == bot) ||
(t1 == Type::BOTTOM) || (t2 == Type::BOTTOM) )
return bot;
// If either number is not a constant, we know nothing.
if ((t1->base() != Type::DoubleCon) || (t2->base() != Type::DoubleCon)) {
return Type::DOUBLE; // note: x%x can be either NaN or 0
}
double f1 = t1->getd();
double f2 = t2->getd();
jlong x1 = jlong_cast(f1); // note: *(long*)&f1, not just (long)f1
jlong x2 = jlong_cast(f2);
// If either is a NaN, return an input NaN
if (g_isnan(f1)) return t1;
if (g_isnan(f2)) return t2;
// If an operand is infinity or the divisor is +/- zero, punt.
if (!g_isfinite(f1) || !g_isfinite(f2) || x2 == 0 || x2 == min_jlong)
return Type::DOUBLE;
// We must be modulo'ing 2 double constants.
// Make sure that the sign of the fmod is equal to the sign of the dividend
jlong xr = jlong_cast(fmod(f1, f2));
if ((x1 ^ xr) < 0) {
xr ^= min_jlong;
}
return TypeD::make(jdouble_cast(xr));
}
//=============================================================================
DivModNode::DivModNode( Node *c, Node *dividend, Node *divisor ) : MultiNode(3) {
init_req(0, c);
init_req(1, dividend);
init_req(2, divisor);
}
//------------------------------make------------------------------------------
DivModINode* DivModINode::make(Compile* C, Node* div_or_mod) {
Node* n = div_or_mod;
assert(n->Opcode() == Op_DivI || n->Opcode() == Op_ModI,
"only div or mod input pattern accepted");
DivModINode* divmod = new (C) DivModINode(n->in(0), n->in(1), n->in(2));
Node* dproj = new (C) ProjNode(divmod, DivModNode::div_proj_num);
Node* mproj = new (C) ProjNode(divmod, DivModNode::mod_proj_num);
return divmod;
}
//------------------------------make------------------------------------------
DivModLNode* DivModLNode::make(Compile* C, Node* div_or_mod) {
Node* n = div_or_mod;
assert(n->Opcode() == Op_DivL || n->Opcode() == Op_ModL,
"only div or mod input pattern accepted");
DivModLNode* divmod = new (C) DivModLNode(n->in(0), n->in(1), n->in(2));
Node* dproj = new (C) ProjNode(divmod, DivModNode::div_proj_num);
Node* mproj = new (C) ProjNode(divmod, DivModNode::mod_proj_num);
return divmod;
}
//------------------------------match------------------------------------------
// return result(s) along with their RegMask info
Node *DivModINode::match( const ProjNode *proj, const Matcher *match ) {
uint ideal_reg = proj->ideal_reg();
RegMask rm;
if (proj->_con == div_proj_num) {
rm = match->divI_proj_mask();
} else {
assert(proj->_con == mod_proj_num, "must be div or mod projection");
rm = match->modI_proj_mask();
}
return new (match->C)MachProjNode(this, proj->_con, rm, ideal_reg);
}
//------------------------------match------------------------------------------
// return result(s) along with their RegMask info
Node *DivModLNode::match( const ProjNode *proj, const Matcher *match ) {
uint ideal_reg = proj->ideal_reg();
RegMask rm;
if (proj->_con == div_proj_num) {
rm = match->divL_proj_mask();
} else {
assert(proj->_con == mod_proj_num, "must be div or mod projection");
rm = match->modL_proj_mask();
}
return new (match->C)MachProjNode(this, proj->_con, rm, ideal_reg);
}
C:\hotspot-69087d08d473\src\share\vm/opto/divnode.hpp
/*
* Copyright (c) 1997, 2010, 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.
*
*/
#ifndef SHARE_VM_OPTO_DIVNODE_HPP
#define SHARE_VM_OPTO_DIVNODE_HPP
#include "opto/multnode.hpp"
#include "opto/node.hpp"
#include "opto/opcodes.hpp"
#include "opto/type.hpp"
// Portions of code courtesy of Clifford Click
// Optimization - Graph Style
//------------------------------DivINode---------------------------------------
// Integer division
// Note: this is division as defined by JVMS, i.e., MinInt/-1 == MinInt.
// On processors which don't naturally support this special case (e.g., x86),
// the matcher or runtime system must take care of this.
class DivINode : public Node {
public:
DivINode( Node *c, Node *dividend, Node *divisor ) : Node(c, dividend, divisor ) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------DivLNode---------------------------------------
// Long division
class DivLNode : public Node {
public:
DivLNode( Node *c, Node *dividend, Node *divisor ) : Node(c, dividend, divisor ) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------DivFNode---------------------------------------
// Float division
class DivFNode : public Node {
public:
DivFNode( Node *c, Node *dividend, Node *divisor ) : Node(c, dividend, divisor) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------DivDNode---------------------------------------
// Double division
class DivDNode : public Node {
public:
DivDNode( Node *c, Node *dividend, Node *divisor ) : Node(c,dividend, divisor) {}
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase );
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual uint ideal_reg() const { return Op_RegD; }
};
//------------------------------ModINode---------------------------------------
// Integer modulus
class ModINode : public Node {
public:
ModINode( Node *c, Node *in1, Node *in2 ) : Node(c,in1, in2) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *bottom_type() const { return TypeInt::INT; }
virtual uint ideal_reg() const { return Op_RegI; }
};
//------------------------------ModLNode---------------------------------------
// Long modulus
class ModLNode : public Node {
public:
ModLNode( Node *c, Node *in1, Node *in2 ) : Node(c,in1, in2) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
virtual const Type *bottom_type() const { return TypeLong::LONG; }
virtual uint ideal_reg() const { return Op_RegL; }
};
//------------------------------ModFNode---------------------------------------
// Float Modulus
class ModFNode : public Node {
public:
ModFNode( Node *c, Node *in1, Node *in2 ) : Node(c,in1, in2) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return Type::FLOAT; }
virtual uint ideal_reg() const { return Op_RegF; }
};
//------------------------------ModDNode---------------------------------------
// Double Modulus
class ModDNode : public Node {
public:
ModDNode( Node *c, Node *in1, Node *in2 ) : Node(c, in1, in2) {}
virtual int Opcode() const;
virtual const Type *Value( PhaseTransform *phase ) const;
virtual const Type *bottom_type() const { return Type::DOUBLE; }
virtual uint ideal_reg() const { return Op_RegD; }
};
//------------------------------DivModNode---------------------------------------
// Division with remainder result.
class DivModNode : public MultiNode {
protected:
DivModNode( Node *c, Node *dividend, Node *divisor );
public:
enum {
div_proj_num = 0, // quotient
mod_proj_num = 1 // remainder
};
virtual int Opcode() const;
virtual Node *Identity( PhaseTransform *phase ) { return this; }
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape) { return NULL; }
virtual const Type *Value( PhaseTransform *phase ) const { return bottom_type(); }
virtual uint hash() const { return Node::hash(); }
virtual bool is_CFG() const { return false; }
virtual uint ideal_reg() const { return NotAMachineReg; }
ProjNode* div_proj() { return proj_out(div_proj_num); }
ProjNode* mod_proj() { return proj_out(mod_proj_num); }
};
//------------------------------DivModINode---------------------------------------
// Integer division with remainder result.
class DivModINode : public DivModNode {
public:
DivModINode( Node *c, Node *dividend, Node *divisor ) : DivModNode(c, dividend, divisor) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeTuple::INT_PAIR; }
virtual Node *match( const ProjNode *proj, const Matcher *m );
// Make a divmod and associated projections from a div or mod.
static DivModINode* make(Compile* C, Node* div_or_mod);
};
//------------------------------DivModLNode---------------------------------------
// Long division with remainder result.
class DivModLNode : public DivModNode {
public:
DivModLNode( Node *c, Node *dividend, Node *divisor ) : DivModNode(c, dividend, divisor) {}
virtual int Opcode() const;
virtual const Type *bottom_type() const { return TypeTuple::LONG_PAIR; }
virtual Node *match( const ProjNode *proj, const Matcher *m );
// Make a divmod and associated projections from a div or mod.
static DivModLNode* make(Compile* C, Node* div_or_mod);
};
#endif // SHARE_VM_OPTO_DIVNODE_HPP
C:\hotspot-69087d08d473\src\share\vm/opto/doCall.cpp
/*
* Copyright (c) 1998, 2016, 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 "ci/ciCallSite.hpp"
#include "ci/ciMethodHandle.hpp"
#include "classfile/vmSymbols.hpp"
#include "compiler/compileBroker.hpp"
#include "compiler/compileLog.hpp"
#include "interpreter/linkResolver.hpp"
#include "opto/addnode.hpp"
#include "opto/callGenerator.hpp"
#include "opto/cfgnode.hpp"
#include "opto/mulnode.hpp"
#include "opto/parse.hpp"
#include "opto/rootnode.hpp"
#include "opto/runtime.hpp"
#include "opto/subnode.hpp"
#include "prims/nativeLookup.hpp"
#include "runtime/sharedRuntime.hpp"
void trace_type_profile(Compile* C, ciMethod *method, int depth, int bci, ciMethod *prof_method, ciKlass *prof_klass, int site_count, int receiver_count) {
if (TraceTypeProfile || C->print_inlining()) {
outputStream* out = tty;
if (!C->print_inlining()) {
if (NOT_PRODUCT(!PrintOpto &&) !PrintCompilation) {
method->print_short_name();
tty->cr();
}
CompileTask::print_inlining(prof_method, depth, bci);
} else {
out = C->print_inlining_stream();
}
CompileTask::print_inline_indent(depth, out);
out->print(" \\-> TypeProfile (%d/%d counts) = ", receiver_count, site_count);
stringStream ss;
prof_klass->name()->print_symbol_on(&ss);
out->print("%s", ss.as_string());
out->cr();
}
}
CallGenerator* Compile::call_generator(ciMethod* callee, int vtable_index, bool call_does_dispatch,
JVMState* jvms, bool allow_inline,
float prof_factor, ciKlass* speculative_receiver_type,
bool allow_intrinsics, bool delayed_forbidden) {
ciMethod* caller = jvms->method();
int bci = jvms->bci();
Bytecodes::Code bytecode = caller->java_code_at_bci(bci);
guarantee(callee != NULL, "failed method resolution");
// Dtrace currently doesn't work unless all calls are vanilla
if (env()->dtrace_method_probes()) {
allow_inline = false;
}
// Note: When we get profiling during stage-1 compiles, we want to pull
// from more specific profile data which pertains to this inlining.
// Right now, ignore the information in jvms->caller(), and do method[bci].
ciCallProfile profile = caller->call_profile_at_bci(bci);
// See how many times this site has been invoked.
int site_count = profile.count();
int receiver_count = -1;
if (call_does_dispatch && UseTypeProfile && profile.has_receiver(0)) {
// Receivers in the profile structure are ordered by call counts
// so that the most called (major) receiver is profile.receiver(0).
receiver_count = profile.receiver_count(0);
}
CompileLog* log = this->log();
if (log != NULL) {
int rid = (receiver_count >= 0)? log->identify(profile.receiver(0)): -1;
int r2id = (rid != -1 && profile.has_receiver(1))? log->identify(profile.receiver(1)):-1;
log->begin_elem("call method='%d' count='%d' prof_factor='%g'",
log->identify(callee), site_count, prof_factor);
if (call_does_dispatch) log->print(" virtual='1'");
if (allow_inline) log->print(" inline='1'");
if (receiver_count >= 0) {
log->print(" receiver='%d' receiver_count='%d'", rid, receiver_count);
if (profile.has_receiver(1)) {
log->print(" receiver2='%d' receiver2_count='%d'", r2id, profile.receiver_count(1));
}
}
log->end_elem();
}
// Special case the handling of certain common, profitable library
// methods. If these methods are replaced with specialized code,
// then we return it as the inlined version of the call.
// We do this before the strict f.p. check below because the
// intrinsics handle strict f.p. correctly.
CallGenerator* cg_intrinsic = NULL;
if (allow_inline && allow_intrinsics) {
CallGenerator* cg = find_intrinsic(callee, call_does_dispatch);
if (cg != NULL) {
if (cg->is_predicated()) {
// Code without intrinsic but, hopefully, inlined.
CallGenerator* inline_cg = this->call_generator(callee,
vtable_index, call_does_dispatch, jvms, allow_inline, prof_factor, speculative_receiver_type, false);
if (inline_cg != NULL) {
cg = CallGenerator::for_predicated_intrinsic(cg, inline_cg);
}
}
// If intrinsic does the virtual dispatch, we try to use the type profile
// first, and hopefully inline it as the regular virtual call below.
// We will retry the intrinsic if nothing had claimed it afterwards.
if (cg->does_virtual_dispatch()) {
cg_intrinsic = cg;
cg = NULL;
} else {
return cg;
}
}
}
// Do method handle calls.
// NOTE: This must happen before normal inlining logic below since
// MethodHandle.invoke* are native methods which obviously don't
// have bytecodes and so normal inlining fails.
if (callee->is_method_handle_intrinsic()) {
CallGenerator* cg = CallGenerator::for_method_handle_call(jvms, caller, callee, delayed_forbidden);
assert(cg == NULL || !delayed_forbidden || !cg->is_late_inline() || cg->is_mh_late_inline(), "unexpected CallGenerator");
return cg;
}
// Do not inline strict fp into non-strict code, or the reverse
if (caller->is_strict() ^ callee->is_strict()) {
allow_inline = false;
}
// Attempt to inline...
if (allow_inline) {
// The profile data is only partly attributable to this caller,
// scale back the call site information.
float past_uses = jvms->method()->scale_count(site_count, prof_factor);
// This is the number of times we expect the call code to be used.
float expected_uses = past_uses;
// Try inlining a bytecoded method:
if (!call_does_dispatch) {
InlineTree* ilt = InlineTree::find_subtree_from_root(this->ilt(), jvms->caller(), jvms->method());
WarmCallInfo scratch_ci;
bool should_delay = false;
WarmCallInfo* ci = ilt->ok_to_inline(callee, jvms, profile, &scratch_ci, should_delay);
assert(ci != &scratch_ci, "do not let this pointer escape");
bool allow_inline = (ci != NULL && !ci->is_cold());
bool require_inline = (allow_inline && ci->is_hot());
if (allow_inline) {
CallGenerator* cg = CallGenerator::for_inline(callee, expected_uses);
if (require_inline && cg != NULL) {
// Delay the inlining of this method to give us the
// opportunity to perform some high level optimizations
// first.
if (should_delay_string_inlining(callee, jvms)) {
assert(!delayed_forbidden, "strange");
return CallGenerator::for_string_late_inline(callee, cg);
} else if (should_delay_boxing_inlining(callee, jvms)) {
assert(!delayed_forbidden, "strange");
return CallGenerator::for_boxing_late_inline(callee, cg);
} else if ((should_delay || AlwaysIncrementalInline) && !delayed_forbidden) {
return CallGenerator::for_late_inline(callee, cg);
}
}
if (cg == NULL || should_delay) {
// Fall through.
} else if (require_inline || !InlineWarmCalls) {
return cg;
} else {
CallGenerator* cold_cg = call_generator(callee, vtable_index, call_does_dispatch, jvms, false, prof_factor);
return CallGenerator::for_warm_call(ci, cold_cg, cg);
}
}
}
// Try using the type profile.
if (call_does_dispatch && site_count > 0 && receiver_count > 0) {
// The major receiver's count >= TypeProfileMajorReceiverPercent of site_count.
bool have_major_receiver = (100.*profile.receiver_prob(0) >= (float)TypeProfileMajorReceiverPercent);
ciMethod* receiver_method = NULL;
int morphism = profile.morphism();
if (speculative_receiver_type != NULL) {
if (!too_many_traps(caller, bci, Deoptimization::Reason_speculate_class_check)) {
// We have a speculative type, we should be able to resolve
// the call. We do that before looking at the profiling at
// this invoke because it may lead to bimorphic inlining which
// a speculative type should help us avoid.
receiver_method = callee->resolve_invoke(jvms->method()->holder(),
speculative_receiver_type);
if (receiver_method == NULL) {
speculative_receiver_type = NULL;
} else {
morphism = 1;
}
} else {
// speculation failed before. Use profiling at the call
// (could allow bimorphic inlining for instance).
speculative_receiver_type = NULL;
}
}
if (receiver_method == NULL &&
(have_major_receiver || morphism == 1 ||
(morphism == 2 && UseBimorphicInlining))) {
// receiver_method = profile.method();
// Profiles do not suggest methods now. Look it up in the major receiver.
receiver_method = callee->resolve_invoke(jvms->method()->holder(),
profile.receiver(0));
}
if (receiver_method != NULL) {
// The single majority receiver sufficiently outweighs the minority.
CallGenerator* hit_cg = this->call_generator(receiver_method,
vtable_index, !call_does_dispatch, jvms, allow_inline, prof_factor);
if (hit_cg != NULL) {
// Look up second receiver.
CallGenerator* next_hit_cg = NULL;
ciMethod* next_receiver_method = NULL;
if (morphism == 2 && UseBimorphicInlining) {
next_receiver_method = callee->resolve_invoke(jvms->method()->holder(),
profile.receiver(1));
if (next_receiver_method != NULL) {
next_hit_cg = this->call_generator(next_receiver_method,
vtable_index, !call_does_dispatch, jvms,
allow_inline, prof_factor);
if (next_hit_cg != NULL && !next_hit_cg->is_inline() &&
have_major_receiver && UseOnlyInlinedBimorphic) {
// Skip if we can't inline second receiver's method
next_hit_cg = NULL;
}
}
}
CallGenerator* miss_cg;
Deoptimization::DeoptReason reason = morphism == 2 ?
Deoptimization::Reason_bimorphic :
(speculative_receiver_type == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check);
if ((morphism == 1 || (morphism == 2 && next_hit_cg != NULL)) &&
!too_many_traps(caller, bci, reason)
) {
// Generate uncommon trap for class check failure path
// in case of monomorphic or bimorphic virtual call site.
miss_cg = CallGenerator::for_uncommon_trap(callee, reason,
Deoptimization::Action_maybe_recompile);
} else {
// Generate virtual call for class check failure path
// in case of polymorphic virtual call site.
miss_cg = CallGenerator::for_virtual_call(callee, vtable_index);
}
if (miss_cg != NULL) {
if (next_hit_cg != NULL) {
assert(speculative_receiver_type == NULL, "shouldn't end up here if we used speculation");
trace_type_profile(C, jvms->method(), jvms->depth() - 1, jvms->bci(), next_receiver_method, profile.receiver(1), site_count, profile.receiver_count(1));
// We don't need to record dependency on a receiver here and below.
// Whenever we inline, the dependency is added by Parse::Parse().
miss_cg = CallGenerator::for_predicted_call(profile.receiver(1), miss_cg, next_hit_cg, PROB_MAX);
}
if (miss_cg != NULL) {
trace_type_profile(C, jvms->method(), jvms->depth() - 1, jvms->bci(), receiver_method, profile.receiver(0), site_count, receiver_count);
ciKlass* k = speculative_receiver_type != NULL ? speculative_receiver_type : profile.receiver(0);
float hit_prob = speculative_receiver_type != NULL ? 1.0 : profile.receiver_prob(0);
CallGenerator* cg = CallGenerator::for_predicted_call(k, miss_cg, hit_cg, hit_prob);
if (cg != NULL) return cg;
}
}
}
}
}
}
// Nothing claimed the intrinsic, we go with straight-forward inlining
// for already discovered intrinsic.
if (allow_inline && allow_intrinsics && cg_intrinsic != NULL) {
assert(cg_intrinsic->does_virtual_dispatch(), "sanity");
return cg_intrinsic;
}
// There was no special inlining tactic, or it bailed out.
// Use a more generic tactic, like a simple call.
if (call_does_dispatch) {
return CallGenerator::for_virtual_call(callee, vtable_index);
} else {
// Class Hierarchy Analysis or Type Profile reveals a unique target,
// or it is a static or special call.
return CallGenerator::for_direct_call(callee, should_delay_inlining(callee, jvms));
}
}
// Return true for methods that shouldn't be inlined early so that
// they are easier to analyze and optimize as intrinsics.
bool Compile::should_delay_string_inlining(ciMethod* call_method, JVMState* jvms) {
if (has_stringbuilder()) {
if ((call_method->holder() == C->env()->StringBuilder_klass() ||
call_method->holder() == C->env()->StringBuffer_klass()) &&
(jvms->method()->holder() == C->env()->StringBuilder_klass() ||
jvms->method()->holder() == C->env()->StringBuffer_klass())) {
// Delay SB calls only when called from non-SB code
return false;
}
switch (call_method->intrinsic_id()) {
case vmIntrinsics::_StringBuilder_void:
case vmIntrinsics::_StringBuilder_int:
case vmIntrinsics::_StringBuilder_String:
case vmIntrinsics::_StringBuilder_append_char:
case vmIntrinsics::_StringBuilder_append_int:
case vmIntrinsics::_StringBuilder_append_String:
case vmIntrinsics::_StringBuilder_toString:
case vmIntrinsics::_StringBuffer_void:
case vmIntrinsics::_StringBuffer_int:
case vmIntrinsics::_StringBuffer_String:
case vmIntrinsics::_StringBuffer_append_char:
case vmIntrinsics::_StringBuffer_append_int:
case vmIntrinsics::_StringBuffer_append_String:
case vmIntrinsics::_StringBuffer_toString:
case vmIntrinsics::_Integer_toString:
return true;
case vmIntrinsics::_String_String:
{
Node* receiver = jvms->map()->in(jvms->argoff() + 1);
if (receiver->is_Proj() && receiver->in(0)->is_CallStaticJava()) {
CallStaticJavaNode* csj = receiver->in(0)->as_CallStaticJava();
ciMethod* m = csj->method();
if (m != NULL &&
(m->intrinsic_id() == vmIntrinsics::_StringBuffer_toString ||
m->intrinsic_id() == vmIntrinsics::_StringBuilder_toString))
// Delay String.<init>(new SB())
return true;
}
return false;
}
default:
return false;
}
}
return false;
}
bool Compile::should_delay_boxing_inlining(ciMethod* call_method, JVMState* jvms) {
if (eliminate_boxing() && call_method->is_boxing_method()) {
set_has_boxed_value(true);
return aggressive_unboxing();
}
return false;
}
// uncommon-trap call-sites where callee is unloaded, uninitialized or will not link
bool Parse::can_not_compile_call_site(ciMethod *dest_method, ciInstanceKlass* klass) {
// Additional inputs to consider...
// bc = bc()
// caller = method()
// iter().get_method_holder_index()
assert( dest_method->is_loaded(), "ciTypeFlow should not let us get here" );
// Interface classes can be loaded & linked and never get around to
// being initialized. Uncommon-trap for not-initialized static or
// v-calls. Let interface calls happen.
ciInstanceKlass* holder_klass = dest_method->holder();
if (!holder_klass->is_being_initialized() &&
!holder_klass->is_initialized() &&
!holder_klass->is_interface()) {
uncommon_trap(Deoptimization::Reason_uninitialized,
Deoptimization::Action_reinterpret,
holder_klass);
return true;
}
assert(dest_method->is_loaded(), "dest_method: typeflow responsibility");
return false;
}
//------------------------------do_call----------------------------------------
// Handle your basic call. Inline if we can & want to, else just setup call.
void Parse::do_call() {
// It's likely we are going to add debug info soon.
// Also, if we inline a guy who eventually needs debug info for this JVMS,
// our contribution to it is cleaned up right here.
kill_dead_locals();
// Set frequently used booleans
const bool is_virtual = bc() == Bytecodes::_invokevirtual;
const bool is_virtual_or_interface = is_virtual || bc() == Bytecodes::_invokeinterface;
const bool has_receiver = Bytecodes::has_receiver(bc());
// Find target being called
bool will_link;
ciSignature* declared_signature = NULL;
ciMethod* orig_callee = iter().get_method(will_link, &declared_signature); // callee in the bytecode
ciInstanceKlass* holder_klass = orig_callee->holder();
ciKlass* holder = iter().get_declared_method_holder();
ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder);
assert(declared_signature != NULL, "cannot be null");
// Bump max node limit for JSR292 users
if (bc() == Bytecodes::_invokedynamic || orig_callee->is_method_handle_intrinsic()) {
C->set_max_node_limit(3*MaxNodeLimit);
}
// uncommon-trap when callee is unloaded, uninitialized or will not link
// bailout when too many arguments for register representation
if (!will_link || can_not_compile_call_site(orig_callee, klass)) {
#ifndef PRODUCT
if (PrintOpto && (Verbose || WizardMode)) {
method()->print_name(); tty->print_cr(" can not compile call at bci %d to:", bci());
orig_callee->print_name(); tty->cr();
}
#endif
return;
}
assert(holder_klass->is_loaded(), "");
//assert((bc_callee->is_static() || is_invokedynamic) == !has_receiver , "must match bc"); // XXX invokehandle (cur_bc_raw)
// Note: this takes into account invokeinterface of methods declared in java/lang/Object,
// which should be invokevirtuals but according to the VM spec may be invokeinterfaces
assert(holder_klass->is_interface() || holder_klass->super() == NULL || (bc() != Bytecodes::_invokeinterface), "must match bc");
// Note: In the absence of miranda methods, an abstract class K can perform
// an invokevirtual directly on an interface method I.m if K implements I.
// orig_callee is the resolved callee which's signature includes the
// appendix argument.
const int nargs = orig_callee->arg_size();
const bool is_signature_polymorphic = MethodHandles::is_signature_polymorphic(orig_callee->intrinsic_id());
// Push appendix argument (MethodType, CallSite, etc.), if one.
if (iter().has_appendix()) {
ciObject* appendix_arg = iter().get_appendix();
const TypeOopPtr* appendix_arg_type = TypeOopPtr::make_from_constant(appendix_arg);
Node* appendix_arg_node = _gvn.makecon(appendix_arg_type);
push(appendix_arg_node);
}
// ---------------------
// Does Class Hierarchy Analysis reveal only a single target of a v-call?
// Then we may inline or make a static call, but become dependent on there being only 1 target.
// Does the call-site type profile reveal only one receiver?
// Then we may introduce a run-time check and inline on the path where it succeeds.
// The other path may uncommon_trap, check for another receiver, or do a v-call.
// Try to get the most accurate receiver type
ciMethod* callee = orig_callee;
int vtable_index = Method::invalid_vtable_index;
bool call_does_dispatch = false;
// Speculative type of the receiver if any
ciKlass* speculative_receiver_type = NULL;
if (is_virtual_or_interface) {
Node* receiver_node = stack(sp() - nargs);
const TypeOopPtr* receiver_type = _gvn.type(receiver_node)->isa_oopptr();
// call_does_dispatch and vtable_index are out-parameters. They might be changed.
// For arrays, klass below is Object. When vtable calls are used,
// resolving the call with Object would allow an illegal call to
// finalize() on an array. We use holder instead: illegal calls to
// finalize() won't be compiled as vtable calls (IC call
// resolution will catch the illegal call) and the few legal calls
// on array types won't be either.
callee = C->optimize_virtual_call(method(), bci(), klass, holder, orig_callee,
receiver_type, is_virtual,
call_does_dispatch, vtable_index); // out-parameters
speculative_receiver_type = receiver_type != NULL ? receiver_type->speculative_type() : NULL;
}
// invoke-super-special
if (iter().cur_bc_raw() == Bytecodes::_invokespecial && !orig_callee->is_object_initializer()) {
ciInstanceKlass* calling_klass = method()->holder();
ciInstanceKlass* sender_klass =
calling_klass->is_anonymous() ? calling_klass->host_klass() :
calling_klass;
if (sender_klass->is_interface()) {
Node* receiver_node = stack(sp() - nargs);
Node* cls_node = makecon(TypeKlassPtr::make(sender_klass));
Node* bad_type_ctrl = NULL;
Node* casted_receiver = gen_checkcast(receiver_node, cls_node, &bad_type_ctrl);
if (bad_type_ctrl != NULL) {
PreserveJVMState pjvms(this);
set_control(bad_type_ctrl);
uncommon_trap(Deoptimization::Reason_class_check,
Deoptimization::Action_none);
}
if (stopped()) {
return; // MUST uncommon-trap?
}
set_stack(sp() - nargs, casted_receiver);
}
}
// Note: It's OK to try to inline a virtual call.
// The call generator will not attempt to inline a polymorphic call
// unless it knows how to optimize the receiver dispatch.
bool try_inline = (C->do_inlining() || InlineAccessors);
// ---------------------
dec_sp(nargs); // Temporarily pop args for JVM state of call
JVMState* jvms = sync_jvms();
// ---------------------
// Decide call tactic.
// This call checks with CHA, the interpreter profile, intrinsics table, etc.
// It decides whether inlining is desirable or not.
CallGenerator* cg = C->call_generator(callee, vtable_index, call_does_dispatch, jvms, try_inline, prof_factor(), speculative_receiver_type);
// NOTE: Don't use orig_callee and callee after this point! Use cg->method() instead.
orig_callee = callee = NULL;
// ---------------------
// Round double arguments before call
round_double_arguments(cg->method());
// Feed profiling data for arguments to the type system so it can
// propagate it as speculative types
record_profiled_arguments_for_speculation(cg->method(), bc());
#ifndef PRODUCT
// bump global counters for calls
count_compiled_calls(/*at_method_entry*/ false, cg->is_inline());
// Record first part of parsing work for this call
parse_histogram()->record_change();
#endif // not PRODUCT
assert(jvms == this->jvms(), "still operating on the right JVMS");
assert(jvms_in_sync(), "jvms must carry full info into CG");
// save across call, for a subsequent cast_not_null.
Node* receiver = has_receiver ? argument(0) : NULL;
// The extra CheckCastPP for speculative types mess with PhaseStringOpts
if (receiver != NULL && !call_does_dispatch && !cg->is_string_late_inline()) {
// Feed profiling data for a single receiver to the type system so
// it can propagate it as a speculative type
receiver = record_profiled_receiver_for_speculation(receiver);
}
// Bump method data counters (We profile *before* the call is made
// because exceptions don't return to the call site.)
profile_call(receiver);
JVMState* new_jvms = cg->generate(jvms);
if (new_jvms == NULL) {
// When inlining attempt fails (e.g., too many arguments),
// it may contaminate the current compile state, making it
// impossible to pull back and try again. Once we call
// cg->generate(), we are committed. If it fails, the whole
// compilation task is compromised.
if (failing()) return;
// This can happen if a library intrinsic is available, but refuses
// the call site, perhaps because it did not match a pattern the
// intrinsic was expecting to optimize. Should always be possible to
// get a normal java call that may inline in that case
cg = C->call_generator(cg->method(), vtable_index, call_does_dispatch, jvms, try_inline, prof_factor(), speculative_receiver_type, /* allow_intrinsics= */ false);
if ((new_jvms = cg->generate(jvms)) == NULL) {
guarantee(failing(), "call failed to generate: calls should work");
return;
}
}
if (cg->is_inline()) {
// Accumulate has_loops estimate
C->set_has_loops(C->has_loops() || cg->method()->has_loops());
C->env()->notice_inlined_method(cg->method());
}
// Reset parser state from [new_]jvms, which now carries results of the call.
// Return value (if any) is already pushed on the stack by the cg.
add_exception_states_from(new_jvms);
if (new_jvms->map()->control() == top()) {
stop_and_kill_map();
} else {
assert(new_jvms->same_calls_as(jvms), "method/bci left unchanged");
set_jvms(new_jvms);
}
if (!stopped()) {
// This was some sort of virtual call, which did a null check for us.
// Now we can assert receiver-not-null, on the normal return path.
if (receiver != NULL && cg->is_virtual()) {
Node* cast = cast_not_null(receiver);
// %%% assert(receiver == cast, "should already have cast the receiver");
}
// Round double result after a call from strict to non-strict code
round_double_result(cg->method());
ciType* rtype = cg->method()->return_type();
ciType* ctype = declared_signature->return_type();
if (Bytecodes::has_optional_appendix(iter().cur_bc_raw()) || is_signature_polymorphic) {
// Be careful here with return types.
if (ctype != rtype) {
BasicType rt = rtype->basic_type();
BasicType ct = ctype->basic_type();
if (ct == T_VOID) {
// It's OK for a method to return a value that is discarded.
// The discarding does not require any special action from the caller.
// The Java code knows this, at VerifyType.isNullConversion.
pop_node(rt); // whatever it was, pop it
} else if (rt == T_INT || is_subword_type(rt)) {
// Nothing. These cases are handled in lambda form bytecode.
assert(ct == T_INT || is_subword_type(ct), err_msg_res("must match: rt=%s, ct=%s", type2name(rt), type2name(ct)));
} else if (rt == T_OBJECT || rt == T_ARRAY) {
assert(ct == T_OBJECT || ct == T_ARRAY, err_msg_res("rt=%s, ct=%s", type2name(rt), type2name(ct)));
if (ctype->is_loaded()) {
const TypeOopPtr* arg_type = TypeOopPtr::make_from_klass(rtype->as_klass());
const Type* sig_type = TypeOopPtr::make_from_klass(ctype->as_klass());
if (arg_type != NULL && !arg_type->higher_equal(sig_type)) {
Node* retnode = pop();
Node* cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(), retnode, sig_type));
push(cast_obj);
}
}
} else {
assert(rt == ct, err_msg_res("unexpected mismatch: rt=%s, ct=%s", type2name(rt), type2name(ct)));
// push a zero; it's better than getting an oop/int mismatch
pop_node(rt);
Node* retnode = zerocon(ct);
push_node(ct, retnode);
}
// Now that the value is well-behaved, continue with the call-site type.
rtype = ctype;
}
} else {
// Symbolic resolution enforces the types to be the same.
// NOTE: We must relax the assert for unloaded types because two
// different ciType instances of the same unloaded class type
// can appear to be "loaded" by different loaders (depending on
// the accessing class).
assert(!rtype->is_loaded() || !ctype->is_loaded() || rtype == ctype,
err_msg_res("mismatched return types: rtype=%s, ctype=%s", rtype->name(), ctype->name()));
}
// If the return type of the method is not loaded, assert that the
// value we got is a null. Otherwise, we need to recompile.
if (!rtype->is_loaded()) {
#ifndef PRODUCT
if (PrintOpto && (Verbose || WizardMode)) {
method()->print_name(); tty->print_cr(" asserting nullness of result at bci: %d", bci());
cg->method()->print_name(); tty->cr();
}
#endif
if (C->log() != NULL) {
C->log()->elem("assert_null reason='return' klass='%d'",
C->log()->identify(rtype));
}
// If there is going to be a trap, put it at the next bytecode:
set_bci(iter().next_bci());
null_assert(peek());
set_bci(iter().cur_bci()); // put it back
}
BasicType ct = ctype->basic_type();
if (ct == T_OBJECT || ct == T_ARRAY) {
ciKlass* better_type = method()->return_profiled_type(bci());
if (UseTypeSpeculation && better_type != NULL) {
// If profiling reports a single type for the return value,
// feed it to the type system so it can propagate it as a
// speculative type
record_profile_for_speculation(stack(sp()-1), better_type);
}
}
}
// Restart record of parsing work after possible inlining of call
#ifndef PRODUCT
parse_histogram()->set_initial_state(bc());
#endif
}
03-28
08-27
1101
1101
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
