[PATCH] D157103: AMDGPU: Fix counting source modifiers as literal constants

[Joe Nash via Phabricator via llvm-commits]

Joe_Nash added a comment.

In D157103#4579608 <https://reviews.llvm.org/D157103#4579608>, @foad wrote:

>> 4b1702e87a2687569b197aea4721353f8b788182 <https://reviews.llvm.org/rG4b1702e87a2687569b197aea4721353f8b788182>
>
> This is causing:
>
>   $ cat r.ll
>   define amdgpu_ps <2 x half> @f(half %arg) {
>   bb:
>     %i = fmul contract half %arg, 0xH3A66
>     %i1 = fadd contract half %i, 0xH2E66
>     %i2 = fmul contract half %arg, 0xH0000
>     %i3 = fsub contract half 0xH3A66, %i2
>     %i4 = insertelement <2 x half> poison, half %i1, i32 0
>     %i5 = insertelement <2 x half> %i4, half %i3, i32 1
>     ret <2 x half> %i5
>   }
>   $ llc -march=amdgcn -mcpu=gfx1010 -o /dev/null r.ll 
>   error: Illegal instruction detected: VOP2/VOP3 instruction uses more than one literal
>   renamable $vgpr1 = contract nofpexcept V_FMAAK_F16 14950, $vgpr0, 11878, implicit $mode, implicit $exec
>
> Can you please fix or revert?

I think https://reviews.llvm.org/D157624 will fix it.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D157103/new/

https://reviews.llvm.org/D157103