[PATCH] D63301: [AMDGPU] gfx1010 wave32 icmp/fcmp intrinsic changes for wave32

Fri Jun 14 02:14:03 PDT 2019

rampitec added a comment.

In D63301#1543215 <https://reviews.llvm.org/D63301#1543215>, @hakzsam wrote:

> HI,
>
> This change introduces a bunch of failures with CTS on GFX8/GFX9. Looks like the IR validation fail now, see below:
>
> $ RADV_DEBUG=checkir,nocache ./deqp-vk --deqp-case=dEQP-VK.glsl.derivate.dfdxfine.dynamic_loop.vec2_highp
>  Writing test log into TestResults.qpa
>  dEQP Core git-c2a38f956feeb2e141c76829c6ecb41c9f68d253 (0xc2a38f95) starting..
>
>   target implementation = 'Default'
>   
>
> Test case 'dEQP-VK.glsl.derivate.dfdxfine.dynamic_loop.vec2_highp'..
>  Intrinsic name not mangled correctly for type arguments! Should be: llvm.amdgcn.icmp.i64.i32
>  i64 (i32, i32, i32)* @llvm.amdgcn.icmp.i32
>  in function main
>  LLVM ERROR: Broken function found, compilation aborted!
>
> Can you investigate please?
>  Thanks!

Intrinsic mangling has changed. Normally llvm can handle it gracefully. How do you load a module? Can you attach a testcase please? I assume autoupgrade code may be needed, but we need to understand the workflow of RADV.

Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D63301/new/

https://reviews.llvm.org/D63301