[cfe-dev] Transform (CUDA)-Attributes in the AST

[Simeon Ehrig via cfe-dev]

Hi Aaron,

thanks for the explanation. I understand why the attribute should not be 
part of the AST. Changing this behavior for my case also sounds like a 
bad idea, since I have great dependencies on third-party tools.

Now I try to understand what handleConstantAttr() does to hopefully find 
a solution. But I think this will take some time.

Thank you very much,
Simeon

On 7/12/19 1:54 PM, Aaron Ballman wrote:
> On Fri, Jul 12, 2019 at 3:06 AM Simeon Ehrig via cfe-dev
> <cfe-dev at lists.llvm.org> wrote:
>> Dear all,
>>
>> I try to transform a variable declaration with the CUDA attribute
>> __constant__ in a function. The complete code, which I want to transform
>> is the following:
>>
>>       void __cling_Un1Qu31(void* vpClingValue) {
>>         __constant__ int array[1024];
>>         ;
>>       }
>>
>> I have to transform the AST before the transformation to a semantic
>> Attribute take place, because the attribute __constant__ has the
>> semantic property that it cannot be declared as a locale variable.
>> That's also the reason why I want to transform the AST. Unfortunately
>> the function Decl::hasAttr<T>() doesn't work without semantic
>> transformation either.
>>
>> I have already started writing an ASTTransformer:
>>
>>       ASTTransformer::Result
>> CUDAConstantMemoryTransformer::Transform(clang::Decl *D) {
>>         clang::FunctionDecl *F = llvm::dyn_cast<clang::FunctionDecl>(D);
>>         if (F && F->hasBody() &&
>> (F->getNameInfo().getName().getAsString().rfind("__cling_", 0) == 0)) {
>>           //F->dump();
>>         }
>>       }
>>
>> At this point I don't know how to check if  a statement or declaration
>> has the attribute __constant__ . When I run F->dump(), I get the
>> following output:
>>
>>       FunctionDecl 0x5fb1cf0 <input_line_3:1:1, line:4:1> line:1:6
>> __cling_Un1Qu30 'void (void *)'
>>       |-ParmVarDecl 0x5fb1c50 <col:22, col:28> col:28 vpClingValue 'void *'
>>       `-CompoundStmt 0x5fb1eb8 <col:42, line:4:1>
>>         |-DeclStmt 0x5fb1e90
>> </usr/local/cuda/include/host_defines.h:86:9, input_line_3:2:30>
>>         | `-VarDecl 0x5fb1e30
>> </usr/local/cuda/include/host_defines.h:86:9, input_line_3:2:29> col:19
>> array 'int [1024]'
>>         `-NullStmt 0x5fb1ea8 <line:3:1>
>>
>> I tried different functions on different AST nodes (in special DeclStmt
>> and VarDecl), but I didn't found a solution to detect the __constant__
>> attribute.
>>
>> Can anyone help me to detect and change the __constant__ attribute in
>> the AST?
> Unfortunately, you won't be able to detect that attribute in the AST
> because it won't exist in the AST. When we ignore an attribute that
> we've parsed (because it's unknown, doesn't apply to that construct,
> etc), we do not create an AST node for it in the AST so that the
> faulty attribute doesn't cause issues with invariants elsewhere in the
> compiler or confuse third-party tools.
>
> One possible solution is to locally relax the restriction on the
> attribute so that it no longer is ignored, but that would likely not
> be acceptable for upstreaming because we want to keep that restriction
> (I presume; I know very little about CUDA specifically). See
> handleConstantAttr() in SemaDeclAttr.cpp for where this particular
> constraint is enforced.
>
> HTH!
>
> ~Aaron
>
>> Best regards,
>> Simeon
>>
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev