> No. This is a reset ordering problem. The CPU reset happens before the
> 'rom blob loader' reset, so at this point the rom data (usually an ELF file
> segment) has not been written into ram, and doing an
> address_space_read() will just read zeroes. This is also why the aliasing
> issue happens at all -- the rom blob is at a particular address, but if the
> address we use here to try to read the data is an aliased variant of it
> then rom_ptr() does the wrong thing.
>
> My preference for fixing this properly is:
> * get Damien's three-phase-reset patchset into master
> * make the ROM blob loader write its data into ram
> in phase 2 ('hold')
> * make the arm CPU reset read the data in phase 3 ('exit')
>
> This last matches better what the hardware does -- the M-profile CPU
> reads the vector table in the first couple of clock cycles when it *leaves*
> reset, not while the CPU is being *held* in reset. This kind of thing is
> always awkward to model in an emulator, especially if you were hoping
> to also handle allowing the PC to be set from an ELF file entrypoint or
> by the user in the gdbstub on startup...
Makes perfect sense. Feel free to drop the patch.
Thanks,
Jh
