RFR: 8340079: Modify rearrange/selectFrom Vector API methods to perform wrapIndexes instead of checkIndexes

[Sandhya Viswanathan]

On Fri, 13 Sep 2024 05:30:36 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:

>> Currently the rearrange and selectFrom APIs check shuffle indices and throw IndexOutOfBoundsException if there is any exceptional source index in the shuffle. This causes the generated code to be less optimal. This PR modifies the rearrange/selectFrom Vector API methods to perform wrapIndexes instead of checkIndexes and performs optimizations to generate efficient code.
>> 
>> Summary of changes is as follows:
>>  1) The rearrange/selectFrom methods do wrapIndexes instead of checkIndexes.
>>  2) Intrinsic for wrapIndexes and selectFrom to generate efficient code
>> 
>> For the following source:
>> 
>> 
>>     public void test() {
>>         var index = ByteVector.fromArray(bspecies128, shuffles[1], 0);
>>         for (int j = 0; j < bspecies128.loopBound(size); j += bspecies128.length()) {
>>             var inpvect = ByteVector.fromArray(bspecies128, byteinp, j);
>>             index.selectFrom(inpvect).intoArray(byteres, j);
>>         }
>>     }
>> 
>> 
>> The code generated for inner main now looks as follows:
>> ;; B24: #      out( B24 B25 ) <- in( B23 B24 ) Loop( B24-B24 inner main of N173 strip mined) Freq: 4160.96
>>   0x00007f40d02274d0:   movslq %ebx,%r13
>>   0x00007f40d02274d3:   vmovdqu 0x10(%rsi,%r13,1),%xmm1
>>   0x00007f40d02274da:   vpshufb %xmm2,%xmm1,%xmm1
>>   0x00007f40d02274df:   vmovdqu %xmm1,0x10(%rax,%r13,1)
>>   0x00007f40d02274e6:   vmovdqu 0x20(%rsi,%r13,1),%xmm1
>>   0x00007f40d02274ed:   vpshufb %xmm2,%xmm1,%xmm1
>>   0x00007f40d02274f2:   vmovdqu %xmm1,0x20(%rax,%r13,1)
>>   0x00007f40d02274f9:   vmovdqu 0x30(%rsi,%r13,1),%xmm1
>>   0x00007f40d0227500:   vpshufb %xmm2,%xmm1,%xmm1
>>   0x00007f40d0227505:   vmovdqu %xmm1,0x30(%rax,%r13,1)
>>   0x00007f40d022750c:   vmovdqu 0x40(%rsi,%r13,1),%xmm1
>>   0x00007f40d0227513:   vpshufb %xmm2,%xmm1,%xmm1
>>   0x00007f40d0227518:   vmovdqu %xmm1,0x40(%rax,%r13,1)
>>   0x00007f40d022751f:   add    $0x40,%ebx
>>   0x00007f40d0227522:   cmp    %r8d,%ebx
>>   0x00007f40d0227525:   jl     0x00007f40d02274d0
>> 
>> Best Regards,
>> Sandhya
>
> src/hotspot/share/opto/vectorIntrinsics.cpp line 2206:
> 
>> 2204:   const Type * byte_bt = Type::get_const_basic_type(T_BYTE);
>> 2205:   const TypeVect * byte_vt  = TypeVect::make(byte_bt, num_elem);
>> 2206:   Node* byte_shuffle = gvn().transform(VectorCastNode::make(cast_vopc, v1, T_BYTE, num_elem));
> 
> We can be optimal here and prevent down casting and subsequent load shuffles in applicable scenarios, e.g. indexes held in integral vectors.

@jatin-bhateja If you could expand on this comment with specific cases it will be helpful. The loadShuffle generation is needed for platform specific handling of shuffles and cannot be optimized out here.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/20634#discussion_r1759296031