WIP: Adding intrinsic implementations for LSX/LASX targets.

[jinboson]

https://github.com/google/highway/blob/aba14629c2aaff4f842d7ec94c964b9a1f4aff39/hwy/ops/x86_256-inl.h#L267C12-L267C21

It seems that VFromD has not been defined yet around line 267, so why can it be used here ? Or I missed something ?

[jan-wassenberg]

Hi, VFromD is actually defined in x86_128-inl.h, which is included from the x86_256-inl.h and thus available before that point.

FYI we do sometimes have bugs where one compiler is more tolerant of "use before definition", which we fix by reordering. This is why some ops have their dependencies in parentheses, like "// ------------------------------ LoadMaskBits (TestBit)", so we can more easily see this.

HTH?

[jinboson]

Hi, VFromD is actually defined in x86_128-inl.h, which is included from the x86_256-inl.h and thus available before that point.

Yes, I find it in x86_128-inl.h，but line 267 seems need a Vec256<TFromD> return type right ? There seems a mismatch problem in the size of return type. That' my confunsed point.

[jan-wassenberg]

Ah, I see. This seems permissible due to the two-phase lookup of templates: the Zero overload is only required to be present at the point of instantiation.
But there is no harm in moving the 256-bit Zero overloads before their usage, we can do that. Thanks for pointing this out.

[jinboson]

Really appreciate your help in understanding this. Thanks.