Re: [RFC][PATCH 0/2] Add RISC-V cpu description and sim

[Ed Jones <ed dot jones at embecosm dot com>]

The generated files in the patches put them over the 400kb limit so the
patches bounced. I'll try and figure out how to get them posted.

Our hope with using CGEN was that it would be a format consistent with
other simulator ports, and that it could be iterated on quickly to add new
instructions and extensions. I generally use QEMU for testing, but there's
a certain convenience to having a simulator built along with GDB and usable
out of the box (whether that's the existing port or this cgen-based one).

We have changes to base the assembler and disassembler around CGEN, but
those changes aren't included in these patches. The assembler/disassembler
files are generated but they just sit alongside the existing riscv-opc.c
and riscv-dis.c without being used by gas/binutils. One of the motivations
for using cgen was that we would be able to base all of the tools around a
single familiar representation. That said, we also discovered that it's
very hard to get the cgen-based assembler+disassembler to match the
behaviour of the existing tools, and unless we can reliably match the
output in all of the important edge-cases it's hard to make an argument to
switch. For now, these patches just add the simulator.

Regards,
Ed


On Thu, 12 Sep 2019 at 01:44, Jim Wilson <jimw@sifive.com> wrote:

> On Wed, Sep 11, 2019 at 9:51 AM Edward Jones <ed.jones@embecosm.com>
> wrote:
> > We're looking to add a cpu description for the riscv architecture
> > plus a simulator port. The simulator has support for rv{32,64}imacfd
> > and is capable of running small programs compiled against newlib.
>
> I don't see the other two parts of the patch, but maybe they will turn
> up tomorrow.
>
> Incidentally, we have an existing gdb sim port originally written by
> Mike Frysinger (vapier), with additions from Kito Cheng and myself,
> which was implemented the traditional way, by hand.  I haven't pushed
> to upstream this on the assumption that the cgen sim port from
> Embecosm might be more useful and more maintainable.  The existing sim
> port is known to work well enough to run the gcc testsuite, but
> toolchain uses qemu by default, so I don't think that anyone other
> than Kito and myself are using it, and we don't use it very much.
> https://github.com/riscv/riscv-binutils-gdb/tree/fsf-gdb-with-sim/sim/riscv
>
> I am concerned that any attempt to replace the current assembler is
> likely to cause trouble though, as a lot of work has gone into the
> existing one to get a lot of features and edge cases right.  There is
> nothing important depending on the simulator though, so replacing that
> should not cause trouble.  The disassembler is kind of in the middle,
> there are a few things depending on it, and a few interesting features
> in it people rely on, but probably few people will notice if it
> changes.
>
> I'll need to see the other two parts of the patch series, and time to
> review them, to comment further.
>
> Jim
>