[LLVMdev] LLC Bug x86 with thread local storage

[Patrick Marlier]

On 07/08/2010 09:37 AM, Eric Christopher wrote:
>
> On Jul 7, 2010, at 11:20 AM, Eric Christopher wrote:
>
>>
>> On Jul 7, 2010, at 4:52 AM, Patrick Marlier wrote:
>>
>>> Which one is correct ?
>>> - movl    $tm_nest_level at TPOFF, %ecx
>>> or
>>> - movq    $tm_nest_level at TPOFF, %rcx
>>> or
>>> - movl    tm_nest_level at TPOFF, %ecx
>>>
>>
>> I believe this is initial exec and so from:
>>
>> http://people.redhat.com/drepper/tls.pdf
>>
>> it would be movl tm_nest_level at TPOFF, %ecx
>
> As a correction, I don't know that any of these is correct.
> I need to look into this a bit more.  I think that it requires the
> relocation to be off of a register, i.e.
>
> 	movabs  $foo at tpoff, %rax
> 	add     $foo at tpoff, %rax
>          mov     foo at tpoff(%rbx), %eax
>
> are all legal, while:
>
> 	mov     foo at tpoff(%ebx), %eax
>      	call    foo at tpoff
>       	mov     $foo at tpoff, %eax
>       	mov     $foo at tpoff, %ax
>       	mov     $foo at tpoff, %al
>
> are all illegal.
>
> So we appear to be lowering something, and trying to make
> an illegal move out of it.
>
> Unfortunately we're really short on testcases here. :)  I've made
> an attempt that works for the current test and passes our set of tests
> in:
>
> Sending        lib/Target/X86/X86Instr64bit.td
> Sending        test/CodeGen/X86/x86-64-tls-1.ll
> Transmitting file data ..
> Committed revision 107860.
>
> Going to need some more testcases soon though if you run into
> any more problems :)

I tested on our application and now it compiles :)

Thank you so much!
Have an nice day,

Patrick Marlier.