[llvm] r239100 - [NVPTX] roll forward r239082

[Jingyue Wu]

Sorry that I missed your previous email.

I think it's a bug that should be fixed -- NVPTX should accept arguments
that are already in global. Clang always emit generic pointers for kernel
parameters, but we shouldn't restrict NVPTX to just Clang. Curiously, do
you use front-ends other than Clang? Xuetian recently did something similar
to alloca. Not sure if that affects you too.

I'll have a patch soon.

Jingyue

On Fri, Jun 26, 2015 at 5:45 AM, Tobias Grosser <tobias at grosser.es> wrote:

> On 06/22/2015 08:51 AM, Tobias Grosser wrote:
>
>> On 06/04/2015 11:28 PM, Jingyue Wu wrote:
>>
>>> Author: jingyue
>>> Date: Thu Jun  4 16:28:26 2015
>>> New Revision: 239100
>>>
>>> URL: http://llvm.org/viewvc/llvm-project?rev=239100&view=rev
>>> Log:
>>> [NVPTX] roll forward r239082
>>>
>>> NVPTXISelDAGToDAG translates "addrspacecast to param" to
>>> NVPTX::nvvm_ptr_gen_to_param
>>>
>>
>> This commit breaks kernels where the arguments are already in the global
>> address space (test case attached). Are such kernels invalid in some way?
>>
>
> Ping. Is this breakage intended.
>
> Tobias
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150626/e3d6a236/attachment.html>