Re: [PATCH v10 6/7] hw/pci-host: Add emulation of Marvell MV64361 PPC system controller

[David Gibson]

On Wed, Mar 24, 2021 at 12:11:50PM +0100, BALATON Zoltan wrote:
> On Wed, 24 Mar 2021, David Gibson wrote:
> > On Tue, Mar 23, 2021 at 02:31:07PM +0100, BALATON Zoltan wrote:
> > > On Tue, 23 Mar 2021, David Gibson wrote:
> > > > On Wed, Mar 17, 2021 at 02:17:51AM +0100, BALATON Zoltan wrote:
> > [snip]
> > > > > +static void setup_mem_windows(MV64361State *s, uint32_t val)
> > > > > +{
> > > > > +    MV64361PCIState *p;
> > > > > +    MemoryRegion *mr;
> > > > > +    uint32_t mask;
> > > > > +    int i;
> > > > > +
> > > > > +    val &= 0x1fffff;
> > > > > +    for (mask = 1, i = 0; i < 21; i++, mask <<= 1) {
> > > > 
> > > > Having a loop, where nearly the entire body is a switch over the loop
> > > > variable seems a rather odd choice to me, compared to just unrolling
> > > > the whole thing.  Or alternatively, maybe more can be be factored out
> > > > of the switch into common body code.
> > > 
> > > The loop is really over the bits in val that say which memory regions to
> > 
> > I see that, but it doesn't change the point.
> > 
> > > enable so depending on this we need to touch different mem regions. For
> > > selecting those memory regions and what to do with them a switch seems to 
> > > be
> > > obvious choice. I probably can't factor out anything as these lines in
> > > switch cases are similar but all differ in some little details (like which
> > > PCI bus, name of the region, etc.). Unrolling could be possible but would
> > > just add lines between cases that's now in the loop head so I really don't
> > 
> > I see only 2 common lines, which basically balances about the case and
> > break lines in every switchcase.
> 
> Not sure what you mean by that. To me that means that these cannot be
> factored out as they are in the middle so can't put them neither before nor
> after the switch (without adding more local variables that would just make
> the result longer than it is now).

I'm saying that I don't think unrolling would actually make things
longer, because the small amount of duplication would balanced by
removing the switch boilerplate.

> Does that mean you prefer this to be unrolled then (written without a for
> just repeating the if ((val & mask) != (s->base_addr_enable & mask)) at
> every case)? That would also be longer by about 20 lines as we also log
> regions that are not in the switch that would need their enable bits checked
> separately if it's unrolled. Basically the trace is the common part of the
> loop and handling of the individual regions are branching out from the
> switch as they are different enough that factoring out the common parts is
> not shorter than this way due to then needing local variables to hold the
> different parts (name, address, size, base) the assigning of which are as
> many or more lines than the map_pci_region call that could be factored out
> that way.
> 
> I don't see why it is a problem to have a switch in a loop. If you insist I
> can try to turn the switch into if else but I don't see how would that be
> any better. Please explain a bit more what would you prefer here as I'm not
> sure what to do with this. To me this is the shortest and simplest way to
> write this.

Hmm... you know, it still seems kinda weird to me, but I am getting
too hung up on details here.  It's good enough, go with it.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature