[aarch64-port-dev ] RFR(S): 8204473 - AARCH64: register post-index addressing mode is not supported directly

[Dmitrij Pochepko]

On 06.06.2018 20:46, Dmitrij Pochepko wrote:
>
>
> On 06.06.2018 20:26, Dmitrij Pochepko wrote:
>>
>>
>> On 06.06.2018 20:16, Andrew Haley wrote:
>>> On 06/06/2018 05:59 PM, Dmitrij Pochepko wrote:
>>>> Ahh, sorry. Missed single existing [ab]usage of such addressing in
>>>> fastdebug mode:
>>> Mmm.  It's always a really, really good idea to run tests with
>>> assertions enabled.
>> Yes. My bad.
>>>> http://hg.openjdk.java.net/jdk/jdk/file/tip/src/hotspot/cpu/aarch64/assembler_aarch64.cpp#l1179 
>>>>
>>>>
>>>> However, there is a confusing comment there: "// BEGIN Generated code
>>>> -- do not edit // Generated by aarch64-asmtest.py"
>>>> What should we do about it? I don't see aarch64-asmtest.py to edit.
>>> Oh yes.  It must not have been included when we merged into mainline.
>>> Oops.
>>>
>>>> Do we need to edit this smoke test directly despite this comment?
>>> Yes, please.  The code you need to edit is *after* the line
>>> END  Generated code -- do not edit
>> Thanks. Running full test cycle now.
>>>
>>> I did this:
>>>
>>>    address PC = __ pc();
>>>    __ ld1(v0, __ T16B, Address(r16)); // No offset
>>>    __ ld1(v0, __ T16B, __ post(r16, 8)); // Post-index
>>>    __ ld1(v0, __ T16B, __ post(r16, r17)); //
>>>
>>> which at least checks that we can generate the instructions, for which
>>> I get
>>>
>>>    0x000003ffa1000764: ld1    {v0.16b}, [x16]
>>>    0x000003ffa1000768: ld1    {v0.16b}, [x16], #16
>>>    0x000003ffa100076c: ld1    {v0.16b}, [x16], x17
>>>
>>> That "#16" looks odd.  Presumably there's some scaling going on there.
>>> Do you know?
>>>
>> Yes. I also investigated this some time ago. It is also related to 
>> ld/st addressing modes.
>> For immediate post-index of ld/st only one immediate value per type 
>> is supported (which is size of loaded/stored data) and the 
>> instruction doesn't allow any other options for immediates (see 
>> specification, C7.2.152
>> LD1 (multiple structures)).
>> So, hotspot just checks if immediate is not zero and silently 
>> generates the instruction.
>>
>> Thanks,
>> Dmitrij
>
> Please take a look to webrev.02: 
> http://cr.openjdk.java.net/~dpochepk/8204473/webrev.02/
> Simple sanity with fastdebug build passed. Currently waiting for large 
> test cycle to be finished.
>
> Thanks,
> Dmitrij
large test cycle passed. no new failures(although lots of JDK-8204331 
failures).