[PATCH v2 1/4] riscv: Add multiarch scaffolding for str*/mem* routines

[Jeff Law]

On 10/13/25 9:27 AM, Yao Zihong wrote:
>> For the RVV implementation, please split the symbols with one per patch. Also
>> for __memcmpeq I would just alias to memcmp, the optimization is only a couple
>> of instructions on the found patch; but it is up to you if you really think it
>> is worth it (and for memcmp, can’t you synthesize the results from the vector
>> register instead of reading the input?).
> 
>> The same applies to the string RVV routines [7]. It puzzles me why you are
>> pushing for arch-optimized strcat/strncat when strcpy/strlen/strnlen/memcpy
>> are optimized. Again, it is up to the RISC-V maintainer to see if this is valuable.
> 
> Understood - the current RVV work was mainly to intended to continue existing
> community efforts.
> 
> For memcmp/__memcmpeq I can fold or alias as suggested, and if the maintainers feel
> the str{n}cat side isn’t worth the extra complexity, I would drop or alias those as well.
I recommend we get the basics in, then do some performance comparisons 
for the more esoteric routines like memcmpeq or uarch specific routines 
and let the data guide us.


Jeff