Re: [PATCH] riscv: Add semihosting support [v4]

[Palmer Dabbelt]

On Thu, 30 Jan 2020 10:54:37 GMT (+0000), Peter Maydell wrote:
> On Wed, 29 Jan 2020 at 16:45, Keith Packard <address@hidden> wrote:
> > Peter Maydell <address@hidden> writes:
> >
> > > True but irrelevant. You need to refer to a proper
> > > risc-v specification for your semihosting.
> >
> > The RISC-V Foundation defined semihosting as relative to the existing
> > ARM specification, so using a link to that is appropriate here.
> >
> > Here's the current specification of the unprivileged ISA, which includes
> > the definition of semihosting
> >
> >         https://riscv.org/specifications/
> >
> > While it may be nice in some abstract sense to create a "better"
> > semihosting spec, that's not what the RISC-V foundation has decided to
> > do.
>
> We've gone round this several times. You can't just
> say "it's the arm spec", because you're not arm and
> your architecture is different. You need to actually
> document what the architecture-specific parts are,
> even if you want to mostly say "and we follow the
> Arm spec most of the time". If you really really want
> to say "32 bit RISC-V is gratuitously different from
> 64-bit RISC-V in these areas because we just blindly
> copied the way Arm happened to have historically
> developed" then you can do that if you like, but you
> need to actually write it down in a document somewhere.
>
> You're trying to implement an ABI which has multiple
> different implementations of both consumers and
> producers; it's not a good idea to shortcut the
> process of actually writing down what the requirements
> between the two ends are.

I think, more concretely, the issues here are:

* The only mention of "semihosting" in the RISC-V specification is in the
 commentary, which is not considered a normative part of the specification.
* The interface that commentary contains is defined as maybe being obsolete
 already, but no new concrete interface is defined.
* The semihosting comment doesn't define the semihosting call numbers, just the
 sequence to get to a call.  That said, we haven't written down the Linux ABI
 either -- though there's a much larger breadth of software out there that
 implements it and won't break ABI compatibility, so maybe that's considered
 sufficient in Linux land where it's not for semihosting.

It seems to me like the right way to go forward is to put together a
semihosting ABI specification as a RISC-V specification.  That would give
everyone who cares about their pre-specification implementations a chance to
get together to figure out which parts they want to keep, and give us a path to
move forwards.  There's more work that needs to be done here to put together a
system that will actually work (for example, how to we insert breakpoints if
we're using ebreak to indicate semihosting and haven't reserved NOP
sequences?), so this does warrant a bit of thought.

I know it's a bit of a headache in the short term to get a specification
together, but we've had so many headaches of the "there's no spec" sort that at
least it'll be a change of pace :)

> thanks
> -- PMM