Re: [PATCH ARM]: Architectural Extensions (Matthew Gretton-Dann or Richard Earnshaw)

[David Fernandez <david dot 00 dot fernandez at googlemail dot com>]

El 19/12/2012 9:31, Richard Earnshaw escribió:
> On 18/12/12 18:29, David Fernandez wrote:
>> On Tue, Dec 18, 2012 at 5:38 PM, Richard Earnshaw <rearnsha@arm.com>
>> wrote:
>>> On 18/12/12 17:19, David Fernandez wrote:
>>>>
>>>> On Tue, Dec 18, 2012 at 4:08 PM, Richard Earnshaw <rearnsha@arm.com>
>>>> wrote:
>>>>>
>>>>> On 18/12/12 15:30, David Fernandez wrote:
>>>>>>
>>>>>>
>>>>>> Hi Richard,
>>>>>>
>>>>>> Great, thanks for the info. That will keep me going. I'll look into
>>>>>> .inst, as I was using just a .word to define it for now.
>>>>>>
>>>>>> Just as a comment to the whole thing, AFAIK both Cortex-Rn and
>>>>>> Cortex-An have in essence the same architecture, i.e. same
>>>>>> instruction
>>>>>> set, same possible extensions, etc. The difference is that while
>>>>>> extensions like virt and sec are standard in Cortex-An, they are
>>>>>> optional in Cortex-Rn.
>>>>>>
>>>>>> So I would suggest, that both armv7-r and arm-v7a should have access
>>>>>> to the same extensions, just that armv7-r will have most disabled by
>>>>>> default, while armv7-a would have some enabled by default.
>>>>>>
>>>>>> Also, the banked registers instructions are not part of any
>>>>>> extension,
>>>>>> and should be present all time. It is a bit weird that you don't have
>>>>>> them unless you enable the virt extension.
>>>>>>
>>>>>
>>>>> I'm not sure I understand what else you think is missing.
>>>>>
>>>>> The file
>>>>>
>>>>>           .arch armv7-r
>>>>>           .thumb
>>>>>           .syntax unified
>>>>>           srsia   sp!, #4
>>>>>           rfefd   sp!
>>>>>
>>>>> assembles without error for me.
>>>>>
>>>>> Perhaps you can provide a test-case.
>>>>>
>>>>> R.
>>>>>
>>>>>
>>>>
>>>> Something like:
>>>>
>>>>                   .arch   armv7-r
>>>>                   .syntax unified
>>>>                   .thumb
>>>>                   mrs     r0, sp_usr
>>>>                   bx      lr
>>>>
>>>> Complains, saying that this architecture does not support banked
>>>> registers.
>>>>
>>>> I have to use .arch armv7-a and .arch_extension virt to make it work,
>>>> but it should work in armv7-r and without the need for virtualization
>>>> extensions.
>>>>
>>>> R
>>>>
>>>
>>> I'll look into it.  However, Rev C of the v7 ARM ARM is quite
>>> explicit in
>>> saying that this instruction is only available with the virtualization
>>> extension, so on that basis, I think GAS is right.  Do you have some
>>> document that appears to contradict this?
>>>
>>> R.
>>>
>>
>> I've read the architecture RevC, and the only reason that I can see
>> for you to say that is that any banked register tied to hypervisor or
>> monitor modes won't be available without those modes, which require
>> virtualization extensions for hypervisor (i.e. sp_hyp, spsr_hyp, or
>> elr_hyp), and security extensions for monitor mode (i.e.
>> sp_mon,spsr_mon). So depending on the architecture, and the extensions
>> enabled, gas might fail some of those instructions depending on the
>> arguments.
>>
>> But the banked registers for the normal processor modes like
>> supervisor, abort, undef, irq, and fiq, have to be there, as those
>> modes are present in all arm processors since arm7tdmi, and the
>> sections B9.3.11 & 12, does not indicate that virtualization
>> extensions are required, as it does for instructions like SMC (to
>> enter monitor mode through exception), which in section B9.3.14 says
>> "Security Extensions" on top of each encoding, or for HVC, which have
>> the virtualization extensions conditions in the instruction
>> pseudocode.
>>
>> May be I'm missing something, but the even the technical reference
>> manual shows that the normal banked register and modes are there, like
>> sp_svc, or lr_irq, ro_usr, etc.
>>
>> David.
>>
> 
> 
> Ah, I think I see where you're coming from now...
> 
> Yes, banked registers exist in both the R and A profiles and have done
> since the very first ARM devices.  However, that still doesn't mean that
> these instructions exist unless you have the virtualization extension.
> Instead, on cores that don't have the extension you have to jump through
> a few hoops to get at the banked registers.  Essentially, to read
> registers in mode X, you have to (in pseudo-code):
> 
>     Switch_to_mode (X)
>     Load/Save_banked_regs ()
>     Switch_back_to_original_mode ()
> 
> The only exception to this is when X is user mode and you're in a system
> mode (since the final step would not be possible).  In this case, there
> are variants of the LDM and STM that can access the user-mode registers
> direfctly and save them to a memory location.  You can then access those
> values from memory.
> 
> A standard book on ARM system programming will probably explain this in
> much better detail than I can in a quick email, and probably provide
> most of the code you need as well.
> 
> HTH,
> 
> R.
> 
> 

Richard,

You might be right, but the pseudocode that appears in the architecture
manual does not imply the need for virtualization extensions at all,
while the pseudo code for the instructions that I've mentioned above does.

Anyway, I'll check these instructions on the real IP from ARM and let
you know if they actually work.

Regards
David F.