archive/jdk
1
0
Fork 0
mirror of https://github.com/openjdk/jdk.git synced 2025-09-21 03:24:38 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

6747051: Improve code and implicit null check generation for compressed oops

Browse source 
Push DecodeN node below the Null check to the non-null path to use the mach node without 0 test.

Reviewed-by: rasbold, never
This commit is contained in:
Vladimir Kozlov 2008-09-23 12:29:06 -07:00
parent 4b1e242299
commit b15796424e
6 changed files with 209 additions and 32 deletions
Download patch file Download diff file Expand all files Collapse all files

137
hotspot/src/share/vm/opto/compile.cpp
View file

@ -2075,6 +2075,44 @@ static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
}
#ifdef _LP64
case Op_CastPP:
if (n->in(1)->is_DecodeN() && UseImplicitNullCheckForNarrowOop) {
Compile* C = Compile::current();
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::clone_shift_expressions) {
//
// 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 patforms (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);
if (in1->outcnt() == 0) {
in1->disconnect_inputs(NULL);
}
}
break;
case Op_CmpP:
// Do this transformation here to preserve CmpPNode::sub() and
// other TypePtr related Ideal optimizations (for example, ptr nullness).
@ -2094,24 +2132,44 @@ static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
} else if (in2->Opcode() == Op_ConP) {
const Type* t = in2->bottom_type();
if (t == TypePtr::NULL_PTR && UseImplicitNullCheckForNarrowOop) {
if (Matcher::clone_shift_expressions) {
// x86, ARM and friends can handle 2 adds in addressing mode.
// Decode a narrow oop and do implicit NULL check in address
// [R12 + narrow_oop_reg<<3 + offset]
new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
} else {
// Don't replace CmpP(o ,null) if 'o' is used in AddP
// to generate implicit NULL check on Sparc where
// narrow oops can't be used in address.
uint i = 0;
for (; i < in1->outcnt(); i++) {
if (in1->raw_out(i)->is_AddP())
break;
}
if (i >= in1->outcnt()) {
new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
}
}
new_in2 = ConNode::make(C, 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(C, t->make_narrowoop());
}
@ -2128,6 +2186,49 @@ static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) {
}
}
break;
case Op_DecodeN:
assert(!n->in(1)->is_EncodeP(), "should be optimized out");
break;
case Op_EncodeP: {
Node* in1 = n->in(1);
if (in1->is_DecodeN()) {
n->subsume_by(in1->in(1));
} else if (in1->Opcode() == Op_ConP) {
Compile* C = Compile::current();
const Type* t = in1->bottom_type();
if (t == TypePtr::NULL_PTR) {
n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
} else if (t->isa_oopptr()) {
n->subsume_by(ConNode::make(C, t->make_narrowoop()));
}
}
if (in1->outcnt() == 0) {
in1->disconnect_inputs(NULL);
}
break;
}
case Op_Phi:
if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
// 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);
}
}
break;
#endif
case Op_ModI: