There is a lot of documentation about using QEMU for simulating a system of particular architecture (a "platform").
For example, x86, ARM or RISCV system.
The first step is to configure QEMU target-list, for example ./configure --target-list=riscv32-softmmu.
It's also possible to provide multiple targets in the target-list, but apparently that builds an independent simulation for each specified platform.
My goal, however, is to simulate a system with mixed targets: an x86 machine which also hosts a RISCV embedded processor over PCI.
Obviously I need to implement a QEMU PCI device which would host the RISCV device on the x86 platform, and I have a good idea how to implement a generic PCI device. However, I'm not sure about the best approach to simulate both x86 and RISCV together on the same QEMU simulation.
One approach is to run two instances of QEMU (as two separate processes) and use some sort of IPC for communicating between the x86 and the RISCV simulation.
Another possible (?) approach could be to build RISCV QEMU as a loadable library and load it from x86 QEMU.
Perhaps it's even possible to have a single QEMU application that simulates both x86 and RISCV?
Yet another approach is not to use QEMU for simulating the RISCV device. I could implement a QEMU PCI device that completely encapsulates a RISCV simulation such as tiny-emu, but I would rather use QEMU for both x86 and RISCV.
https://lists.gnu.org/archive/html/qemu-devel/2021-12/msg01969.html
QEMU does not support running multiple target architectures in the same QEMU process. (This is something we would in theory like to be able to do, but it would require a lot of reworking of core parts of QEMU which assume that the target architecture is known at compile time. So far nobody has felt it important enough to put in the significant development effort needed.)
So if you want to do this you'll need to somehow stitch together a QEMU process for the primary architecture with some other process to do the secondary architecture (QEMU or otherwise). This has been done (for instance Xilinx have an out-of-tree QEMU-based system that does this kind of thing with multiple QEMU processes) but I'm not aware of any easy off-the-shelf frameworks or setups to do it. I suspect that figuring out how time/clocks interact between the two simulations is one of the tricky aspects.