Hi,
As you can see, the structs does not have fields allocated from
different address space, but only pointers which points to certain
address space. If I leave out the __constant qualifiers from pointers
inside struct, it would mean that those pointers points to __private
address space, which would be plainly wrong and would be
against the specs to try to store adress of __constant address space
to __private pointer.
I didn't find anything specific about it from specs. It only says that
e.g. this cannot be done:
typedef struct {
private int data;
} invalid_struct;
constant invalid_struct = {0};
Since all the data of struct/union naturally must be stored in the
same address space.
I tested the cases also with 3.3 and both cases are still failing.
I'll try to get time to find out reduced test case for the instcombiner
problem as well.
- Mikael
On Fri, Jun 28, 2013 at 12:31 AM, Pekka Jääskeläinen <
[email protected]> wrote:
> On 06/27/2013 06:55 PM, Mikael Lepistö wrote:
> > I was able to limit the case where I got pocl-bzr to crash on osx. I
> > woudl appreciate if someone could reproduce the error.
>
> Quick guess: I recall there was some issue with LLVM 3.2 and
> structures with fields with address space qualifiers. They didn't work
> somehow. I cannot recall did it crash or silently miscompile.
> Is it even supported in OpenCL C? It sounds a bit creepy in a way.
> Like Erik suggested, please try to reproduce with LLVM 3.3.
> Or just try moving the __constant before the whole struct and
> remove it from the attributes.
>
>
> For the other kernel compiler instcombine problem that produced
> only with 'loops' and worked with 'repl' it'd be also useful to
> test LLVM 3.3. If it still breaks, open a bug report with as
> small test case as possible. Though, I cannot promise I can get
> to debugging it soon as I'm on vacations and AFK more than usual.
> So you might be able to fix it faster yourself by isolating the
> breaking piece of kernel code and then just dump some graphs in
> the WorkitemLoops.cc to see what's going on.
>
> --
> --Pekka
>
>
> ------------------------------------------------------------------------------
> This SF.net email is sponsored by Windows:
>
> Build for Windows Store.
>
> http://p.sf.net/sfu/windows-dev2dev
> _______________________________________________
> pocl-devel mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/pocl-devel
>
------------------------------------------------------------------------------
This SF.net email is sponsored by Windows:
Build for Windows Store.
http://p.sf.net/sfu/windows-dev2dev
_______________________________________________
pocl-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/pocl-devel
