[llvm-commits] [llvm] r66922 - in /llvm/trunk: lib/Target/X86/X86InstrInfo.td test/CodeGen/X86/tls13.ll test/CodeGen/X86/tls14.ll

[Dan Gohman]

On Mar 31, 2009, at 3:17 PM, Rafael Espindola wrote:

>> Instead of making it TLS specific, I'd just make that operand  
>> "specify
>> the segment".  Even though we do nothing with FS or ES or DS etc, it
>> would be nice to be able to have a clean model where the instructions
>> think about these machine level things, and isel lowering handles the
>> translation of "TLS to GS"
>
> I am having some problem along this path.
>
> I have added a X86::SegmentAddress node and TLS addresses are lowered
> into it. This
> works perfectly for the case we do a load. Just another recursive call
> in MatchAddress
> and record the segment.
>
> The problem is when we just want to compute the address as in
>
> int __thread v;
> int *f(void)
> {
> return &v;
> }
>
> In this case, Select is called on the wrapper and it must be expanded
> into two instructions. A "mov gs:0, %a_register" to get the thread  
> pointer
> and a leal/add to get the address. Is Select really the best place  
> to do
> such expansion?
>
>
> An option that looks better so far is to keep the current  
> THREAD_POINTER
> node. The idea is:
>
> 1) Lowering produces an add of and offset and the THREAD_POINTER
> as it does now.
> 2) The THREAD_POINTER is selected into a mov gs:0, a_register
> 3) If we are computing and address, we are done
> 4) If we are doing a load, the two instructions get folded post the
>   instruction selection.
>
> In summary, at the DAG level we will see a THREAD_POINTER and
> not addresses with segments. At the machine instruction level we
> will see addresses with segments.
>
> I am just not sure if the folding described in 4 will always be done.
>
> comments?

I agree with your analysis. It looks like too much complexity to handle
in Select. This is something X86ISelLowering should handle.

I think Chris' interest is to have a single place where TLS is lowered
to segment register usage, with the particular place being less
important. This can be accomplished with the approach you describe
only with THREAD_POINTER renamed to GET_GS, so that it's clear that
Lowering has fully lowered the TLS addressing.

Dan