RFR: 8370481: C2 SuperWord: Long/Integer.compareUnsigned return wrong value in SLP [v4]

[Emanuel Peter]

On Thu, 23 Oct 2025 15:02:13 GMT, Hamlin Li <mli at openjdk.org> wrote:

>> Hi,
>> Can you help to review the patch? @eme64 
>> 
>> ##  Issue
>> 
>> Currently, in SLP when transform from (Bool + CmpU + CMove) to (VectorMaskCmp + VectorBlend), the unsigned-ness in CmpU is lost, then end up doing a signed instead of unsigned comparison in VectorMaskCmp.
>> For details please check code at `SuperWordVTransformBuilder::make_vector_vtnode_for_pack` and `PackSet::get_bool_test`.
>> 
>> ##  Fix
>> Currently, `BoolTest` does not support an unsigned construction (`BoolTest( mask btm ) : _test(btm) { assert((btm & unsigned_compare) == 0, "unsupported");}`), seems to me a feasible solution would be get the unsigned information from CmpU (which could be an input of Bool) and pass it to VectorMaskCmp.
>> 
>> Thanks
>> 
>> This pr could also lead to more optimizations, like: https://github.com/openjdk/jdk/pull/25336 and https://github.com/openjdk/jdk/pull/25341.
>
> Hamlin Li has updated the pull request incrementally with one additional commit since the last revision:
> 
>   catch unexpected Cmp ops

src/hotspot/share/opto/superword.cpp line 1752:

> 1750:   case Op_CmpUL:
> 1751:     // Carry unsigned-ness information from CmpUxx to VTransformBoolTest,
> 1752:     // which will be passed to e.g. VectorMaskCmp.

Suggestion:

    // When we have CmpU->Bool, the mask of the Bool has no unsigned-ness information,
    // but the mask is implicitly unsigned only because of the CmpU. Since we will replace
    // the CmpU->Bool with a single VectorMaskCmp, we need to now make the unsigned-ness
    // explicit.

This would give more of a reason why we need to do the "unsign...ing". What do you think?

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

PR Review Comment: https://git.openjdk.org/jdk/pull/27942#discussion_r2455569832