Re: [RFC v2 21/39] net: add HAS_IOPORT dependencies

["Maciej W. Rozycki" <macro@orcam.me.uk>]

On Tue, 3 May 2022, Niklas Schnelle wrote:

> >  The driver works just fine with MMIO where available, so if `inb'/`outb' 
> > do get removed, then only parts that rely on port I/O need to be disabled.  
> > In fact there's already such provision there in drivers/net/fddi/defxx.c 
> > for TURBOchannel systems (CONFIG_TC), which have no port I/O space either:
> > 
> > #if defined(CONFIG_EISA) || defined(CONFIG_PCI)
> > #define dfx_use_mmio bp->mmio
> > #else
> > #define dfx_use_mmio true
> > #endif
> > 
> > so I guess it's just the conditional that will have to be changed to:
> > 
> > #ifdef CONFIG_HAS_IOPORT
> > 
> > replacing the current explicit bus dependency list.  The compiler will 
> > then optimise away all the port I/O stuff (though I suspect dummy function 
> > declarations may be required for `inb'/`outb', etc.).
[...]
> With dfx_use_mmio changed as you propose above things compile on s390
> which previously ran into missing (now __compile_error()) inl() via
> dfx_port_read_long() -> dfx_inl() ->  inl().

 Great, thanks for checking!  And I note referring `__compile_error' is 
roughly equivalent to a dummy declaration, so you've got that part sorted.

> Looking at the other uses of dfx_use_mmio I notice however that in
> dfx_get_bars(), inb() actually gets called when dfx_use_mmio is true.
> This happens if dfx_bus_eisa is also true. Now that variable is just
> the cached result of DFX_BUS_EISA(dev) which is defined to 0 if
> CONFIG_EISA is unset. I'm not 100% sure if going through a local
> variable is still considered trivial enough dead code elimination, at
> least it works for me™. I did also check the GCC docs and they
> explicitly say that __attribute__(error) is supposed to be used when
> dead code elimination gets rid of the error paths.

 Yeah, dead code elimination is supposed to handle such cases.  The local
automatic variable is essentially a syntactic feature not to use the same 
expression inline over and over throughout a function (for clarity the 
variable should probably be declared `const', but that is not essential) 
and it is up to the compiler whether to reuse the value previously 
calculated or to re-evaluate the expression.

> I think we also need a "depends on HAS_IOPORT" for "config HAVE_EISA"
> just as I'm adding for "config ISA".

 Oh absolutely!  There's the slot-specific port I/O space that is used to 
identify EISA option cards in device discovery, so no EISA device will 
ever work without port I/O.  Have a look at `decode_eisa_sig' in 
drivers/eisa/eisa-bus.c for the very obvious code.  Note that some ISA 
cards can be configured to appear as EISA devices as well (I have a 3c509B 
Ethernet NIC set up that way).

  Maciej