[PATCH] D15305: [CUDA] Do not allow dynamic initialization of global device side variables.

Richard Smith via cfe-commits cfe-commits at lists.llvm.org
Fri Jan 15 17:32:21 PST 2016

On Fri, Jan 15, 2016 at 5:29 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Fri, Jan 15, 2016 at 4:22 PM, Artem Belevich <tra at google.com> wrote:
>> tra added inline comments.
>> ================
>> Comment at: lib/CodeGen/CodeGenModule.cpp:2334
>> @@ -2339,1 +2333,3 @@
>> +      D->hasAttr<CUDASharedAttr>())
>>      Init = llvm::UndefValue::get(getTypes().ConvertType(ASTTy));
>> +  else if (!InitExpr) {
>> ----------------
>> rsmith wrote:
>>> As this is a global variable, it should presumably still be statically zero-initialized.
>> There is no way to initialize __shared__ variables. They are rough equivalent of local variables, only in this case CUDA allocates them per kernel invocation from a shared buffer with no guarantees regarding its contents.
>> They used to be zero-initialized by compiler, but that was intentionally changed to undef in r245786 / http://reviews.llvm.org/D12241
> That doesn't seem right. C++ guarantees zero-initialization for all
> globals, prior to performing any other initialization.

It looks like the problem being fixed by D12241 was probably caused by
the __shared__ variables having the wrong linkage.

More information about the cfe-commits mailing list