[RFC 0/2] Use address spaces for x86 tls

[Richard Henderson]

On 10/08/2015 08:24 PM, Florian Weimer wrote:
> On 10/08/2015 05:58 AM, Richard Henderson wrote:
>
>> Otherwise the code is practically identical.  Which, I suppose is
>> good in that glibc is currently doing a fair job with its macros.
>> So the change is perhaps only of interest in terms of (very) long term
>> maintenance (i.e. once gcc6 is the minimal require build compiler).
>
> What is the debugging experience like?  Can I still do
>
>    print *__thread_self
>
> in GDB and get useful results?

An excellent question, which somehow I hadn't even considered.  That's 
embarrassing.

I can't imagine how it could get useful results at the moment, since we're not 
telling gdb about the address space.

There's no obvious answer for how this should be represented.  There's 
DW_AT_segment, though how exactly to describe %fs or %gs as a location 
description is a bit lacking.  There's DW_AT_address_class, which has an 
integral argument, which could be defined for x86 in some reasonable way.

Either way we require a debugger change.


r~