6912521: System.arraycopy works slower than the simple loop for little lengths

Convert small array copies to series of loads and stores

Reviewed-by: kvn, vlivanov
This commit is contained in:
Roland Westrelin 2015-02-18 18:14:07 +01:00
parent c71eca0cbf
commit 88848bd9e8
11 changed files with 1440 additions and 327 deletions

View file

@ -1660,22 +1660,7 @@ Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
// must be correct type for alignment purposes
Node* base = basic_plus_adr(ary, header);
#ifdef _LP64
// The scaled index operand to AddP must be a clean 64-bit value.
// Java allows a 32-bit int to be incremented to a negative
// value, which appears in a 64-bit register as a large
// positive number. Using that large positive number as an
// operand in pointer arithmetic has bad consequences.
// On the other hand, 32-bit overflow is rare, and the possibility
// can often be excluded, if we annotate the ConvI2L node with
// a type assertion that its value is known to be a small positive
// number. (The prior range check has ensured this.)
// This assertion is used by ConvI2LNode::Ideal.
int index_max = max_jint - 1; // array size is max_jint, index is one less
if (sizetype != NULL) index_max = sizetype->_hi - 1;
const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
idx = _gvn.transform( new ConvI2LNode(idx, lidxtype) );
#endif
idx = Compile::conv_I2X_index(&_gvn, idx, sizetype);
Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
return basic_plus_adr(ary, base, scale);
}