Re: [PATCH net-next 1/6] tools: ynl-gen: bitshift the flag values in the generated code

[Jakub Kicinski <kuba@kernel.org>]

On Tue, 14 Oct 2025 16:49:22 +0000 Asbjørn Sloth Tønnesen wrote:
> On 10/14/25 12:53 AM, Jakub Kicinski wrote:
> > On Mon, 13 Oct 2025 16:49:58 +0000 Asbjørn Sloth Tønnesen wrote:  
> >> Instead of pre-computing the flag values within the code generator,
> >> then move the bitshift operation into the generated code.
> >>
> >> This IMHO makes the generated code read more like handwritten code.  
> > 
> > I like it the way it is. The values are irrelevant.  
> 
> Bit-shifting seams like the preferred way across the uAPI headers.
> 
> Would you be open to hexadecimal notation, if not bit-shifting?
> 
> Currently NLA_POLICY_MASK() is generated with a hexadecimal mask, and
> with these patches, if render-max is not set. If using literal values
> then we should properly consistently generate them as either decimal
> or hexadecimal. I prefer hexadecimal over decimal.

Hm, hex could do. For the bit/1 << x i really don't like that the values
are not aligned to columns, so they visually mix in with the names. 
But aligning them would be more LoC than it's worth.

hex could be a reasonable compromise, but I make no promise that I will
like it once I see the result :)

> > And returning a string from user_value() is quite ugly.  
> It only returns a string, when as_c is set, I am happy to duplicate
> some code instead, and add a dedicated method always returning a string,
> but can we please agree on the generated output, before implementation?

nlspec.py was supposed to be a library that abstracts away things like
default values not being present, and simplifies indexing. So having a
"give me a format for C as result" arg is not great for layering.
That kind of logic belongs in the caller. 

Regarding LoC - great code is concise, but that doesn't mean that
making code shorter always makes it better.