llvm dev - Apr 2008 - [LLVMdev] Can't invoke an intrinsic?

In line 1157 of Verifier.cpp, there is this code:

      Assert1(!F->isIntrinsic() || (i == 0 &&
isa<CallInst>(I)),
              "Cannot take the address of an intrinsic!", &I);

This check appears to have a problem with this line:

        invoke void @llvm.memcpy.i32( i8* %._items.i.i, i8* 
%._items2.i.i, i32 ptrtoint (i32* getelementptr ([0 x i32]* null, i32 0, 
i32 4) to i32), i32 1 )
                        to label %UnifiedReturnBlock unwind label %failure

...in other words, it appears to be implying that it's not OK to use 
'invoke' on an intrinsic. Is that correct?

I should mention also that the code snipped above was produced by LLVM's 
optimizer, in my original code the call to memcpy was a call, not an 
invoke. So if it is indeed illegal to invoke intrinsics, the optimizer 
should be informed of this fact :)

-- Talin

On Apr 26, 2008, at 8:21 PM, Talin wrote:
> In line 1157 of Verifier.cpp, there is this code:
>
>      Assert1(!F->isIntrinsic() || (i == 0 &&
isa<CallInst>(I)),
>              "Cannot take the address of an intrinsic!", &I);
>
> This check appears to have a problem with this line:
>
>        invoke void @llvm.memcpy.i32( i8* %._items.i.i, i8*
> %._items2.i.i, i32 ptrtoint (i32* getelementptr ([0 x i32]* null,  
> i32 0,
> i32 4) to i32), i32 1 )
>                        to label %UnifiedReturnBlock unwind label  
> %failure
>
> ...in other words, it appears to be implying that it's not OK to use
> 'invoke' on an intrinsic. Is that correct?
It's correct.
> I should mention also that the code snipped above was produced by  
> LLVM's
> optimizer, in my original code the call to memcpy was a call, not an
> invoke. So if it is indeed illegal to invoke intrinsics, the optimizer
> should be informed of this fact :)
Make sure your front-end declares the intrinsic 'nounwind'.  If
you're
using the C++ API, use Intrinsic::getDeclaration(..) and it will  
automatically do this for you.

-Chris