Request for reviews (S): 6836054: java/util/Arrays/CopyMethods.java fails on solaris-sparc with IllegalArgumentException

[Vladimir Kozlov]

Sorry for long output. Here is the generated code which
loads the allocation's actual type dynamically:

  195    ConL    ===  0  [[ 196  211 ]]  #long:24
  190    CastPP  ===  182  11  [[ 176  196  196  206  242 ]]  #java/lang/Class:NotNull:exact *  Oop:java/lang/Class:NotNull:exact * !jvms: Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 CopyMethods::fullTests @ bci:72
  196    AddP    === _  190  190  195  [[ 197 ]]   Oop:java/lang/Class:NotNull:exact+24 * !jvms: Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 CopyMethods::fullTests @ bci:72
  197    LoadKlass       === _  7  196  [[ 198  203 ]]  @rawptr:BotPTR, idx=Raw; # *  Klass: * !jvms: Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 CopyMethods::fullTests @ bci:72
  203    CastPP  ===  202  197  [[ 211  211  234 ]]  #klass java/lang/Object: 0x00000000004bcd58 *  Klass:klass java/lang/Object: 0x00000000004bcd58 * !jvms: Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 CopyMethods::fullTests @ bci:72
  234    AllocateArray   ===  202  91  92  8  1 ( 232  203  24  24  10  11  68  130  24  1  1  1  1  1  1  24  11  1  1  1  1 ) [[ 235  236  237  244  245  246 ]]  rawptr:NotNull ( int:>=0, java/lang/Object:NotNull *, bool, int ) Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 
CopyMethods::fullTests @ bci:72 !jvms: Array::newInstance @ bci:2 CopyMethods::makeArray @ bci:2 CopyMethods::fullTests @ bci:72

Vladimir

Vladimir Kozlov wrote:
> We don't know statically during a method compilation the actual type
> because it is loaded dynamically for reflaction allocation.
> Escape Analysis requires to know the type statically to pass it to debug 
> info:
> 
>   const Type *t = local->bottom_type();
>   // Is it a safepoint scalar object node?
>   if (local->is_SafePointScalarObject()) {
>     SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject();
> 
>     ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx);
>     if (sv == NULL) {
>       ciKlass* cik = t->is_oopptr()->klass();
>       assert(cik->is_instance_klass() ||
>              cik->is_array_klass(), "Not supported allocation.");
>       sv = new ObjectValue(spobj->_idx,
>                            new ConstantOopWriteValue(cik->encoding()));
> 
> Vladimir
> 
> Tom Rodriguez wrote:
>> I have to admit I don't really follow this logic in the code.  How is 
>> it that we can allocate something but not know it's actual type?  
>> Isn't that required to emit the allocation in the first place?
>>
>> tom
>>
>> On May 7, 2009, at 8:29 PM, Vladimir Kozlov wrote:
>>
>>>
>>> http://cr.openjdk.java.net/~kvn/6836054/webrev.00
>>>
>>> Fixed 6836054: java/util/Arrays/CopyMethods.java fails on 
>>> solaris-sparc with IllegalArgumentException
>>>
>>> Problem:
>>> Escape Analysis requires that allocations should have
>>> exact type (by using CheckCastPP node) for restoring
>>> objects during deoptimization. But reflection allocation
>>> set type to j.l.Object assuming that caller will cast
>>> to exact type. In the failing case there is no cast
>>> to array type so debug info contains incorrect type
>>> (j.l.Object) for scalar replaced array allocation.
>>> It causes the Exception after deoptimization.
>>>
>>> Solution:
>>> Do not mark an allocation as scalar replaceable if
>>> it does not have exact type.
>>>
>>> Reviewed by:
>>>
>>> Fix verified (y/n): y, bugs case
>>>
>>> Other testing:
>>> JPRT, CTW
>>>
>>