> In http://reviews.llvm.org/D10753#195190, @ABataev wrote:
> > > 1. Why you don't want to reuse an existing clang infrastructure for TLS > > support? I think it would be much easier and portable solution to reuse an > > existing code, rather than inventing a new one. > > > I am afraid I don't understand exactly what you mean by TLS infrastructure. > In my understanding, creating a TLS global variable will make each thread to > have different storage for the variable but will not cause the Ctors/Dtors to > run when a thread is spawn by the OpenMP runtime. Therefore the OpenMP > runtime has to be aware of which Ctors/Dtors to use to a given variable even > if it is declared as TLS so it can run them when it creates a thread. So from > a codegen viewpoint, the step of registering the Dtors and Ctors is still > required, only the cache creation and look-up can be skipped. That's the > reason why I am reusing the existing code to do the registration. > > It is possible, however, that I am missing some feature in clang that would > enable the execution of the Cotrs/Dtors when threads are spawn without > intervention of any runtime library. If so, please give me a pointer and I > will update the patch to use that instead. Samuel, look at file Decl.cpp, VarDecl::getTLSKind(). I think it is enough to modify this code a little bit: http://reviews.llvm.org/D10753 EMAIL PREFERENCES http://reviews.llvm.org/settings/panel/emailpreferences/ _______________________________________________ cfe-commits mailing list cfe-commits@cs.uiuc.edu http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits