[Patch ARM] Add support for bare-metal cpu initialization on A profile cores.

[Ramana Radhakrishnan]

On 09/23/13 04:12, Ye Joey wrote:
> Ramana,
>
> This patch breaks combined build for Cortex-M, i.e., configured with
>
> --with-mode=thumb --with-cpu=cortex-m3
> combine/src/libgloss/arm/cpu-init/rdimon-aem.S:84: Error: thumb
> conditional instruction should be in IT block -- `moveq r4,#0'
>
>
> I think when configured for M-profile, it shouldn't build rdimon-aem.S at all.

Ugggh ... sorry about the breakage.

Agreed - and something's off in the way in which we're multilibbing in 
the cpu-init directory which is why I've not seen this.

The way I'm planning on fixing this is to

* Turn rdimon-aem.S into an empty file for M class cores.
* Force everything into ARM state anyway as it doesn't matter in T32 
state whether one runs cpu-init in ARM or Thumb state.

Patch under testing.

regards
Ramana

>
> - Joey
>
> On Thu, Sep 19, 2013 at 9:34 PM, Ramana Radhakrishnan <ramrad01@arm.com> wrote:
>> Hi,
>>
>>          Increasingly we've found the need for a start-up routine required
>> for bare-metal applications that need to initialize the mmu and the caches
>> for use during hardware bring-up and/or for bare-metal testing on ARM's
>> propreitary models as well as allow for semi-hosting to happen at the same
>> time. This has been found to be useful in a number of situations and this is
>> just the initial patch set required for this to work on A profile CPUs.  I
>> would recommend that this not be used for earlier cores without testing.
>>
>> The way this is implemented is by adding a new weak function called
>> _rdimon_init_hook that deals with this initialization. I expect this to work
>> on newer hardware because our testing involves testing with the AEM models
>> that model the newer hardware but I don't have access to any such hardware
>> to test this right now.
>>
>> The SMP enable bit in the ACTLR is enabled only for A15 and A7 but not for
>> any other cores as this is needed only on those cores. We use this
>> internally for some of our GCC test runs on our fast models, we've tried it
>> in the past to run some simple applications on bare-metal A7 and A15 boards.
>>
>> This necessitates the use of a new specs file aprofile-ve.specs which I've
>> added and provided as an example. I have tested this on an AEM model but am
>> confident that this will work just fine on a bare-metal board connected via
>> D-Stream or anything that implements the standard Angel API.
>>
>> If people want to use a different base address for the text section and / or
>> for the vectors page in the cpu init code this is available by doing
>> --defsym _rdimon_vector_base=FOO or -Ttext_section=BAR . The specs file only
>> eases the use for this.
>>
>>
>> Tested with an arm-eabi bare-metal GCC multilib build across architecture
>> variants, Thumbness and fpu abi variants using the new specs file in my
>> baremetal build script and tested with --specs=aprofile-ve.specs passed by
>> default in my site.exp - tested on AEM VE model internally .
>>
>>
>> Ok to apply ?
>>
>> regards
>> Ramana
>>
>> <DATE>  Matthew Gretton-Dann  <matthew.gretton-dann@arm.com>
>>       Ramana Radhakrishnan  <ramana.radhakrishnan@arm.com>
>>       Greta Yorsh  <greta.yorsh@arm.com>
>>
>>          * arm/Makefile.in: Add support for cpu-init directory and add
>>            elf-aprofile-ve.specs.
>>          * arm/configure.in: Likewise.
>>          * arm/configure: Regenerate.
>>          * arm/cpu-init: New directory.
>>          * arm/cpu-init/Makefile.in: New file.
>>          * arm/cpu-init/rdimon-aem.S: Likewise.
>>          * arm/crt0.S: Call _rdimon_init_hook
>>          * arm/elf-aprofile-ve.specs: New file.
>