Re: [PATCH 07/32] mm: Bring back vmalloc_exec

[Kent Overstreet]

On Fri, May 12, 2023 at 06:57:52PM -0700, Eric Biggers wrote:
> What I would suggest instead is preprocessing the list of 6 field lengths to
> create some information that can be used to extract all 6 fields branchlessly
> with no dependencies between different fields.  (And you clearly *can* add a
> preprocessing step, as you already have one -- the dynamic code generator.)
> 
> So, something like the following:
> 
>     const struct field_info *info = &format->fields[0];
> 
>     field0 = (in->u64s[info->word_idx] >> info->shift1) & info->mask;
>     field0 |= in->u64s[info->word_idx - 1] >> info->shift2;
> 
> ... but with the code for all 6 fields interleaved.
> 
> On modern CPUs, I think that would be faster than your current C code.
> 
> You could do better by creating variants that are specialized for specific
> common sets of parameters.  During "preprocessing", you would select a variant
> and set an enum accordingly.  During decoding, you would switch on that enum and
> call the appropriate variant.  (This could also be done with a function pointer,
> of course, but indirect calls are slow these days...)
> 
> For example, you mentioned that 8-byte packed keys is a common case.  In that
> case there is only a single u64 to decode from, so you could create a function
> that just handles that case:
> 
>     field0 = (word >> info->shift) & info->mask;
> 
> You could also create other variants, e.g.:
> 
> - 16-byte packed keys (which you mentioned are common)
> - Some specific set of fields have zero width so don't need to be extracted
>   (which it sounds like is common, or is it different fields each time?)
> - All fields having specific lengths (are there any particularly common cases?)
> 
> Have you considered any of these ideas?

I like that idea.

Gonna hack some code... :)