Re: [PATCH V2 0/2] Split tdesc_(amd64|i386)_mpx into tdesc(amd64|i386)_mpx_* and tdesc(amd64|i386)_avx_mpx_*

[Pedro Alves <palves at redhat dot com>]

On 04/14/2016 11:29 AM, Yao Qi wrote:
> Walfred Tedeschi <walfred.tedeschi@intel.com> writes:
> 
>> CPU features can occur in any combination. The current assumption that 
>> feature "A" implies in feature "B" does not necessarily hold.
>>  
>> This patch series construct an additional combination of the Intel(R) 
>> Memory Protection Extensions (MPX) with Intel(R) Advanced Vector 
>> Extensions (AVX).
> 
> First of all, I am not against your patches.  Just think a little more
> after reading them...
> 
> This reveals a problem in gdb target description.  It doesn't scale very
> well if processors have multiple different features, and features can be
> combined differently.  A processor family has three features A, B, and
> C, and each processor implementation may have one, two or three of these
> features.  In gdb target description, we need to have many *.xml and *.c
> files, for these combinations like, A, B, C, AB, AC, BC, and ABC.
> 
> The root cause is that target description are static and pre-generated.
> If the target description can be generated dynamically according to the
> cpuid or AT_HWCAP, that would be simpler.  In this way, we only have to
> define target descriptions for feature A, B, and C, and GDB/GDBserver
> combine them together in the runtime.

I agree.  This is not the first time this is suggested.  If someone were
to do it, I'd be in favor too.

Thanks,
Pedro Alves