[llvm] r289076 - [AMDGPU] Scalarization of global uniform loads.

[Michel Dänzer via llvm-commits]

Hi Alexander,


did you see my report? Are you able to reproduce the problem with llc?


Thanks,


On 09/12/16 05:09 PM, Michel Dänzer wrote:
> 
> Hi Alexander,
> 
> 
> On 09/12/16 02:28 AM, Alexander Timofeev via llvm-commits wrote:
>> Author: alex-t
>> Date: Thu Dec  8 11:28:47 2016
>> New Revision: 289076
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=289076&view=rev
>> Log:
>> [AMDGPU] Scalarization of global uniform loads.
> 
> This change broke a few OpenCL piglit tests for me with the Mesa OpenCL
> implementation on Kaveri, e.g.
> program at execute@store at store-double16-global:
> 
> LLVM ERROR: Cannot select: t77: v16i32,ch = load<LD64[%0(addrspace=1)(align=128)+64](align=64)(tbaa=<0x559592d62578>)> t11, t76, undef:i64
>   t76: i64 = add t344, Constant:i64<64>
>     t344: i64 = bitcast t343
>       t343: v2i32,ch = load<LD8[undef(addrspace=2)](nontemporal)(dereferenceable)(invariant)> t0, t8, undef:i64
>         t8: i64 = add t2, Constant:i64<8>
>           t2: i64,ch = CopyFromReg t0, Register:i64 %vreg1
>             t1: i64 = Register %vreg1
>           t7: i64 = Constant<8>
>         t4: i64 = undef
>     t75: i64 = Constant<64>
>   t4: i64 = undef
> In function: store_global
> 
> 
> You can reproduce it by feeding the attached LLVM IR to
> 
> llc -march=amdgcn
> 
> 


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