Re: [PATCH][BZ #16214] Fix TLS access on S390 with -march=z10

[Andreas Krebbel <krebbel at linux dot vnet dot ibm dot com>]

Hi Carlos,

On 27/11/13 07:03, Carlos O'Donell wrote:
> Yes, _dl_sym and _dl_vsym use this e.g. ->do_sym->_dl_tls_symaddr.
> 
> An asm with clobbers is the only way of making this work reliably.
> 
> The asm could set r12 to GOT, call __tls_get_offset, restore r12
> (or list it clobbered along with all the register __tls_get_offset
> touches, and __tls_get_addr), save result, then the rest can be
> in C and does the math required to compute the result.
> 
> You have to avoid the compiler using r12 for something else in the
> middle and the asm is the only way to avoid that.

When building with -fpic r12 is fixed and cannot be used by the compiler freely. So this is not
supposed to happen.

...
>>>  #ifdef PIC
>>>  # define TLS_IE(x) \
>>> -  ({ unsigned long __offset;						      \
>>> +  ({ unsigned long __offset, __save;					      \
>>>       asm ("bras %0,1f\n"						      \
>>>  	  "0:\t.quad " #x "@gotntpoff\n"				      \
>>> -	  "1:\tlg %0,0(%0)\n\t"						      \
>>> -	  "lg %0,0(%0,%%r12):tls_load:" #x				      \
>>> -	  : "=&a" (__offset) : : "cc" );				      \
>>> +	  "1:\tlgr %1,%%r12\n\t"					      \
>>> +	  "larl %%r12,_GLOBAL_OFFSET_TABLE_\n\t"			      \
>>> +	  "lg %0,0(%0)\n\t"						      \
>>> +	  "lg %0,0(%0,%%r12):tls_load:" #x "\n\t"			      \
>>> +	  "lgr %%r12,%1"						      \
>>> +	  : "=&a" (__offset), "=&a" (__save) : : "cc" );		      \
>>>       (int *) (__builtin_thread_pointer() + __offset); })
>>>  #else
>>>  # define TLS_IE(x) \
>>>
>>
>> For this code to work the GOT pointer is not required to be in r12. Can't you just set it up in a
>> compiler chosen reg? Something like this:
> 
> Why? This code attempts to emulate exactly what the compiler is going to do
> and should be representative of the instruction sequences you'd see being
> generated for TLS accesses. These macros are only ever used in testing and
> never anywhere else so their speed is not important.
> 
> I suggest leaving the longer sequences that Siddhesh has here that save/restore
> r12 and follow the ABI.

All this happens inside the macro so I don't think the ABI is relevant here. It is only needed to
add the GOT pointer to an offset.  Since we cannot do the add with a single instruction we need a
scratch register. This can be r12 or any other GPR (except r0 of course).  I would prefer the
shorter variant because well ... it's shorter :)

Bye,

-Andreas-