On 8/12/21 2:27 PM, Peter Maydell wrote: > On Thu, 12 Aug 2021 at 13:13, Alexandre IOOSS <erdn...@crans.org> wrote: >> >> On 8/12/21 11:33 AM, Peter Maydell wrote: >>> In the realize methods of the stm32f100 and stm32f205 SoC objects, we >>> call g_new() to create new MemoryRegion objjects for the sram, flash, >>> and flash_alias. This is unnecessary (and leaves open the >>> possibility of leaking the allocations if we exit from realize with >>> an error). Make these MemoryRegions member fields of the device >>> state struct instead, as stm32f405 already does. >> >> There is a typo in "objjects". >> >> This is something I had issue understanding as I was seeing both >> patterns in the codebase, thank you for making this clear. > > Basically if there's a struct that the MemoryRegion can live > in then that's the best place for it. For some board-level code > where we haven't needed to subclass MachineState there is no > convenient struct, so we just g_new(). In a few places like this > one the board-code pattern has been copied into an SoC object.
FYI since more than 2 years now Igor recommends using DEFINE_TYPES() for newer QOM style, which makes adding board-specific fields to MachineState quite easy (see hw/avr/arduino.c or hw/arm/raspi.c).
