[vector api] RFR: Fine-tune test cases for reduction operations

[Vladimir Ivanov]

> The following is an example on AArch64 platform. Without the patch, the test is passed . But it is failed with this patch.
> 
> I make a little change to this instruction:
> instruct reduce_max8S(iRegINoSp dst, iRegIorL2I src1, vecX src2, vecX tmp, rFlagsReg cr) %{
>    format %{ "smaxv  $tmp, T8H, $src2\n\t"
>              "umov  $dst, $tmp, H, 0\n\t"----------------> smov should be used. because umov will recognize a negative as a great positive.
>              "cmpw  $dst, $src1\n\t"
>              "cselw $dst, $dst, $src1 gt\t max reduction8S" %}
> 
> %}
> smov: Signed Move vector element to general-purpose register.
> umov: Move vector element to general-purpose register.
> 
> The test is passed with original test. The error is missed, because src1 is always Short.MIN_VALUE, reduce_max8S could give the correct result of one vector.
> test Short128VectorTests.maxAllShort128VectorTests(short[i * 5]): success

Can you, please, elaborate why there's a difference on IR level w.r.t. 
src1? (Sorry, haven't had time to explore that myself.)

bool LibraryCallKit::inline_vector_reduction() {
...
   Node* init = ReductionNode::make_reduction_input(_gvn, opc, elem_bt);
   Node* rn = gvn().transform(ReductionNode::make(opc, NULL, init, opd, 
elem_bt));
...

What I'm seeing in library_call.cpp is it always initializes the value 
with the same constant. Are there any transformations happening after 
that? If yes, then where do they happen?

Best regards,
Vladimir Ivanov

> 
> The test is failed with new test. The error is caught.
> test Short128VectorTests.maxAllShort128VectorTests(short[i * 5]): failure
> java.lang.AssertionError: Final result is incorrect! expected [32765] but found [-25541]
> 
> So that I think checking both the final and intermediate result is better.
> I update the patch a little. Please check it.
> http://cr.openjdk.java.net/~yzhang/vectorapi.test/webrev.01/
> 
> Regards,
> Yang
> 
> -----Original Message-----
> From: Vladimir Ivanov <vladimir.x.ivanov at oracle.com>
> Sent: Tuesday, February 12, 2019 9:37 AM
> To: Yang Zhang (Arm Technology China) <Yang.Zhang at arm.com>; 'panama-dev at openjdk.java.net' <panama-dev at openjdk.java.net>
> Subject: Re: [vector api] RFR: Fine-tune test cases for reduction operations
> 
> Though I like the changes you propose, I don't see how they improve test coverage. It implies C2 does the following transformation, but I don't think that's the case (yet):
> 
>     AddI i (AddReductionVI (ConI 0) v) => AddReductionVI i v
> 
> Am I missing something?
> 
> Best regards,
> Vladimir Ivanov
> 
> On 11/02/2019 00:25, Yang Zhang (Arm Technology China) wrote:
>> Hi
>>
>> Currently, test cases of reduction operations only check the intermediate result (the result of one vector), not final result ( the result of one array). So that for following examples, src1 is always initial value without this patch.
>> instruct rsadd2I_reduction_reg(rRegI dst, rRegI src1, vecD src2, vecD
>> tmp, vecD tmp2) %{  ----------------------X86 instruct
>> reduce_add2I(iRegINoSp dst, iRegIorL2I src1, vecD src2, iRegINoSp tmp, iRegINoSp tmp2) ------------------AArch64 To cover more cases, I add test cases that check the final result of reduction operations. Some indentation errors are also fixed.
>> With this patch, all Vector API tests are passed.
>>
>> Could you please help to review this patch?
>> http://cr.openjdk.java.net/~yzhang/vectorapi.test/webrev.00/
>>
>> Regards,
>> Yang
>>