RFR: 8333684: C2 SuperWord: multiple smaller refactorings in preparation for JDK-8332163

Christian Hagedorn chagedorn at openjdk.org
Wed Jun 12 11:57:27 UTC 2024


On Thu, 6 Jun 2024 10:34:24 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

>> In preparation for https://github.com/openjdk/jdk/pull/19261, I made some smaller refactorings / moving code around:
>> 
>> - `SuperWord::same_input` -> `PackSet::isa_unique_input_or_null`
>> - Rename print-method tags: `SUPERWORD1_BEFORE_SCHEDULE` -> `AUTO_VECTORIZATION1_BEFORE_APPLY` etc.
>> - Refactored `SuperWord::schedule / output` -> `SuperWord::schedule_and_apply`:
>>   - Reorganize so that we can separate out all methods that change the C2 graph into `SuperWord::apply`.
>>   - Move all `phase()->C->print_method` to `SuperWord::apply`.
>>   - Rename `SuperWord::schedule_reorder_memops` -> `SuperWord::apply_memops_reordering_with_schedule`.
>>   - Rename `SuperWord::output` -> `SuperWord::apply_vectorization`.
>> - Move `SuperWord::vectors_should_be_aligned` -> `VLoop::vectors_should_be_aligned`.
>> - Move `SuperWord::requires_long_to_int_conversion` -> `VectorNode::is_scalar_op_that_returns_int_but_vector_op_returns_long`, and move comments.
>> - Move `VectorNode::can_transform_shift_op` -> `VectorNode::can_use_RShiftI_instead_of_URShiftI`, and move comments.
>> - Extract out `PackSet::get_bool_test` from `SuperWord::output / apply_vectorization`.
>> - Extract opcode check to `VectorNode::is_scalar_unary_op_with_equal_input_and_output_types`.
>
> src/hotspot/share/opto/superword.cpp line 1779:
> 
>> 1777:       retValue = UseVectorCmov;
>> 1778:     } else if (VectorNode::is_scalar_op_that_returns_int_but_vector_op_returns_long(opc)) {
>> 1779:       // Requires extra vector long -> int conversion.
> 
> Note: moved comments to method definition. Chose better name.

The name is quite verbose. How about keeping it as suggested in this comment, something like `needs_vector_long_to_int_conversion()`?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/19573#discussion_r1636250038


More information about the hotspot-compiler-dev mailing list