[llvm] r324554 - AMDGPU: Fix incorrect reordering when inline asm defines LDS address

[Michel Dänzer via llvm-commits]

On 2018-02-10 12:58 PM, Kai Wasserbäch wrote:
> Hey Michel,
> Michel Dänzer via llvm-commits wrote on 09.02.2018 11:17:
>> On 2018-02-08 07:33 PM, Matt Arsenault wrote:
>>>> On Feb 8, 2018, at 12:35, Michel Dänzer via llvm-commits <llvm-commits at lists.llvm.org> wrote:
>>>> On 2018-02-08 02:56 AM, Matt Arsenault via llvm-commits wrote:
>>>>> Author: arsenm
>>>>> Date: Wed Feb  7 17:56:14 2018
>>>>> New Revision: 324554
>>>>>
>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=324554&view=rev
>>>>> Log:
>>>>> AMDGPU: Fix incorrect reordering when inline asm defines LDS address
>>>>>
>>>>> Defs of operands outside of the instruction's explicit defs need
>>>>> to be checked.
>>>>
>>>> This change broke a bunch of
>>>> spec at arb_gpu_shader_int64@execution at built-in-functions piglit tests with
>>>> radeonsi. I'm attaching good and bad code for one of them.
>>>
>>> I think this should be fixed by https://reviews.llvm.org/D42647
>>
>> Unfortunately, that doesn't apply to current trunk, so I can't confirm
>> right now.
> 
> if I didn't botch up too badly, the attached patch should work (the build for
> amd64 was successful and the build for i386 is still running, so that is
> promising at least... ;-) I haven't exercised this build yet though – clean LLVM
> builds always take a while for me). The patch applies cleanly on top of r324770
> and should also apply on top of SVN HEAD.

Thanks Kai, this patch indeed fixes the problem.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer