[9] RFR(S): 8144212: JDK 9 b93 breaks Apache Lucene due to compact strings

[Tobias Hartmann]

On 08.01.2016 20:41, Vladimir Kozlov wrote:
> On 1/8/16 2:37 AM, Tobias Hartmann wrote:
>>
>> On 07.01.2016 21:49, Vladimir Kozlov wrote:
>>> On 1/7/16 6:52 AM, Tobias Hartmann wrote:
>>>> Hi Vladimir,
>>>>
>>>> On 07.01.2016 00:58, Vladimir Kozlov wrote:
>>>>> Andrew is right.
>>>>
>>>> Yes, he's right that the membar is not needed in this case. I noticed that GraphKit::inflate_string() sets the output memory to TypeAryPtr::BYTES although inflate writes to a char[] array in this case. This caused the subsequent char load to be on a different slice allowing C2 to move the load to before the intrinsic.
>>>
>>> Right. It was the root of this bug, see below.
>>>
>>>>
>>>> I fixed this for the inflate and compress intrinsics.
>>>>
>>>>> GraphKit::inflate_string() should have SCMemProjNode as compress_string() does to prevent loads move up.
>>>>> StrInflatedCopyNode is not memory node.
>>>>
>>>> Okay, why are above changes not sufficient to prevent the load from moving up? Also, the comment for SCMemProjNode says:
>>>
>>> I did not get the question. Is it before your webrev.01 change? Or even with the change?
>>
>> I meant with webrev.01 but you answered my question below.
>>
>>>>    // This class defines a projection of the memory  state of a store conditional node.
>>>>    // These nodes return a value, but also update memory.
>>>>
>>>> But inflate does not return any value.
>>>
>>> Hmm, according to bottom type inflate produce memory:
>>>
>>> StrInflatedCopyNode::bottom_type() const { return Type::MEMORY; }
>>>
>>> So it really does not need SCMemProjNode. Sorry about that.
>>> So load was LoadUS which is char load and originally memory slice of inflate was incorrect BYTES.
>>
>> Exactly.
>>
>>> Instead of SCMemProjNode we should have to change the idx of your dst_type:
>>>
>>> set_memory(str, dst_type);
>>
>> Yes, that's what I do now in webrev.01 by passing the dst_type as an argument to inflate_string.
>>
>>> And you should rollback part of changes in escape.cpp and macro.cpp.
>>
>> Okay, I'll to that.
>>
>>>> Here is the new webrev, including the SCMemProjNode and adapting escape analysis and macro expansion accordingly:
>>>> http://cr.openjdk.java.net/~thartmann/8144212/webrev.01/
>>>
>>> In general when src & dst arrays have different type we may need to use TypeOopPtr::BOTTOM to prevent related store & loads bypass these copy nodes.
>>
>> Okay, should we then use BOTTOM for both the input and output type?
> 
> Only input. Output type corresponds to dst array type which you set correctly now.

It seems like that this is not sufficient. As Roland pointed out (off-thread), there may still be a problem in the following case:
 StoreC
 inflate_string
 LoadC

The memory graph (def->use) now looks like this:
 LoadC -> inflate_string -> ByteMem
               ... StoreC-> CharMem


The intrinsic hides the dependency between LoadC and StoreC, causing the load to read from memory not containing the result of the StoreC. I was able to write a regression test for this (see 'TestStringIntrinsicMemoryFlow::testInflate2').

Setting the input to BOTTOM, generates the following graph:
http://cr.openjdk.java.net/~thartmann/8144212/inflate_bottom.png
The 349 LoadUS does not read the result of the 96 StoreC because the StrInflateCopyNode does not capture it's memory. The test fails.

I adapted the fix to emit a MergeMemoryNode to capture the entire memory state as input to the intrinsic. The graph then looks like this:
 LoadC -> inflate_string -> MergeMem(ByteMem, StoreC(CharMem))
http://cr.openjdk.java.net/~thartmann/8144212/inflate_merge.png

Here is the new webrev:
http://cr.openjdk.java.net/~thartmann/8144212/webrev.02/

Probably, we could also only capture the byte and char slices instead of merging everything. What do you think?

Best,
Tobias

>>>> Related question:
>>>> In library_call.cpp, I now use TypeAryPtr::get_array_body_type(dst_elem) to get the correct TypeAryPtr for the destination (we support both BYTES and CHARS). For a char[] destination, it returns:
>>>>    char[int:>=0]:exact+any *
>>>>
>>>> which is equal to the type of the char load.
>>>
>>> Please, explain this. I thought string's array will always be byte[] when compressed strings are enabled. Is it used for getChars() which returns char array?
>>
>> Yes, both the compress and inflate intrinsics are used for different types of src and dst arrays. See comment in library_call.cpp:
>>
>> // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
>> //   int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
>> //   int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
>> // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
>> //   void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
>> //   void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
>>
>> I.e., the inflate intrinsic is used for inflation from byte[] to byte[]/char[].
>>
>>> Should we also be more careful in inflate_string_slow()? Is it used?
>>
>> No, inflate_string_slow() is only called from PhaseStringOpts::copy_latin1_string() where it is used to inflate from byte[] to byte[].
>>
>>>> I also tried to derive the type from the array by using dst_type->isa_aryptr(). However, this returns a more specific type:
>>>>    char[int:1]:NotNull:exact *
>>>>
>>>> Using this results in C2 assuming that the subsequent char load is independent and again moving it to before the intrinsic. I don't understand why that is. Shouldn't the second type be a "subtype" of the first type?
>>>
>>> It is indeed strange. What memory type of LoadUS? It could be bug.
>>
>> LoadUS has memory type "char[int:>=0]:exact+any *" which has alias index 4. dst_type->isa_aryptr() returns memory type "char[int:1]:NotNull:exact *" which has alias index 8.
>>
>> I will look into this again and try to understand what happens.
> 
> It could that aryptr is pointer to array and load type is pointer to array's element.
> 
> Thanks,
> Vladimir
> 
>>
>> Thanks,
>> Tobias
>>
>>>>> On 1/6/16 5:34 AM, Andrew Haley wrote:
>>>>>> On 01/06/2016 01:06 PM, Tobias Hartmann wrote:
>>>>>>
>>>>>>> The problem here is that C2 reorders memory instructions and moves
>>>>>>> an array load before an array store. The MemBarCPUOrder is now used
>>>>>>> (compiler internally) to prevent this. We do the same for normal
>>>>>>> array copys in PhaseMacroExpand::expand_arraycopy_node(). No actual
>>>>>>> code is emitted. See also the comment in memnode.hpp:
>>>>>>>
>>>>>>>     // Ordering within the same CPU.  Used to order unsafe memory references
>>>>>>>     // inside the compiler when we lack alias info.  Not needed "outside" the
>>>>>>>     // compiler because the CPU does all the ordering for us.
>>>>>>>
>>>>>>> "CPU does all the ordering for us" means that even with a relaxed
>>>>>>> memory ordering, loads are never moved before dependent stores.
>>>>>>>
>>>>>>> Or did I misunderstand your question?
>>>>>>
>>>>>> No, I don't think so.  I was just checking: I am very aware that
>>>>>> HotSpot has presented those of use with relaxed memory order machines
>>>>>> with some interesting gotchas over the years, that's all.  I'm a bit
>>>>>> surprised that C2 needs this barrier, given that there is a
>>>>>> read-after-write dependency, but never mind.
>>>>>>
>>>>>> Thanks,
>>>>>>
>>>>>> Andrew.
>>>>>>