Re: [V3] powerpc: Optimized strncmp for power10

[Peter Bergner <bergner@linux.ibm.com>]

On 5/6/24 8:04 AM, Adhemerval Zanella Netto wrote:
> If I recall correctly it was to not tie the implementation to an specific
> page size, since the ABI still allows 4k page sizes.  I think both branches
> will highly unlikely to be taken, so branch prediction will most likely
> get a high frequency hit.

Even though the hardware supports 4K pages, I thought we never built it
that way and the major distros to build it with 64K pages, but I learned
there are some minor distros the use 4K pages, so I agree we should use
that here.


> We can also try to make it dynamically if you think these checks are really
> costly, this will mean to add two extra loads and possible an extra cache
> like hit (one for GLRO struct, another for dl_pagesize).  I don't think this
> is worth.

I don't know that they costly, I just though that if they're useless because
we always use 64K pages, then it seems dumb to check the 4K boundary.
Since can/might have 4K pages, then the patch code is correct as is.


> Another question is whether this tests still make sense for POWER10, is it
> still that costly for cross page-page reads as for POWER8?

I'm not 100% sure and it would be something we'd need to test, but I suspect
it probably hasn't changed too much???

Peter