[Mesa-dev] [PATCH] r600g-llvm: Crude fix for a race in initialization of the llvm backend

[Tom Stellard]

On Sat, Aug 18, 2012 at 11:43:45AM +0200, Mathias Fröhlich wrote:
> 
> Tom,
> 
> On Thursday, August 16, 2012 14:50:57 you wrote:
> > Actually, I just noticed that the LLVM docs say that only one thread can
> > access the target registry at a time:
> > http://llvm.org/docs/doxygen/html/structllvm_1_1TargetRegistry.html#a0b078b4
> > 68553a84ec2b9fd70e93f7b43
> > 
> > Also, I think we should be doing the target registry once per
> > context rather than once per compile.
> Yep, I think so too.
> 
> For the problem you mentioned, we can even avoid using the target registry 
> here. This registry just only calls the usual constructor as far as I could 
> see now. I have, during debugging this, now also tested just directly 
> instantiating an AMDGPUTargetMachine instance directly on the stack without 
> querying the target registry. this workes fine too.
> 
> So, when cleaning this up a little more, it might be an option to just 
> instanciate our required target machine in context initialization.
>

This will work, but I would like to avoid instantiating the
AMDGPUTargetMachine directly in order to minimize the differences in the
code when the R600 backend is built as part of the LLVM libs.

-Tom
> But for today, this does not affect the problem I observe!
> I have found a race that we trigger inside llvm's pass registry and that is 
> the reason of the problem I observe at least.
> I will post a patch for that in the next minutes.
> 
> Greetings
> 
> Mathias
> 
> 
>