[RISCV] vfslide1down+vmerge miscompile

[topperc]

This test case

define <vscale x 4 x float> @intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32(<vscale x 4 x float> %0, <vscale x 4 x float> %false, float %1, <vscale x 4 x i1> %mask) nounwind {
                                                          
entry:                                                                           
  %a = call <vscale x 4 x float> @llvm.riscv.vfslide1down.nxv4f32.f32(<vscale x 4 x float> undef, <vscale x 4 x float> %0, float %1, i64 4)
  %b = call <vscale x 4 x float> @llvm.riscv.vmerge.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> %false, <vscale x 4 x float> %a, <vscale x 4 x i1> %mask, i64 1)
  ret <vscale x 4 x float> %b                                                    
}

is compiled to

intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32: # @intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32
# %bb.0:                                # %entry
        vsetivli        zero, 1, e32, m2, ta, mu
        vfslide1down.vf v10, v8, fa0, v0.t
        vmv.v.v v8, v10
        ret

The vfslide1down's VL was changed from 4 to 1. This causes fa0 to be inserted into element 0 instead of element 3.

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: Craig Topper (topperc)

This test case

define &lt;vscale x 4 x float&gt; @<!-- -->intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32(&lt;vscale x 4 x float&gt; %0, &lt;vscale x 4 x float&gt; %false, float %1, &lt;vscale x 4 x i1&gt; %mask) nounwind {
                                                          
entry:                                                                           
  %a = call &lt;vscale x 4 x float&gt; @<!-- -->llvm.riscv.vfslide1down.nxv4f32.f32(&lt;vscale x 4 x float&gt; undef, &lt;vscale x 4 x float&gt; %0, float %1, i64 4)
  %b = call &lt;vscale x 4 x float&gt; @<!-- -->llvm.riscv.vmerge.nxv4f32(&lt;vscale x 4 x float&gt; undef, &lt;vscale x 4 x float&gt; %false, &lt;vscale x 4 x float&gt; %a, &lt;vscale x 4 x i1&gt; %mask, i64 1)
  ret &lt;vscale x 4 x float&gt; %b                                                    
}

is compiled to

intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32: # @<!-- -->intrinsic_vfslide1down_vf_nxv4f32_nxv4f32_f32
# %bb.0:                                # %entry
        vsetivli        zero, 1, e32, m2, ta, mu
        vfslide1down.vf v10, v8, fa0, v0.t
        vmv.v.v v8, v10
        ret

The vfslide1down's VL was changed from 4 to 1. This causes fa0 to be inserted into element 0 instead of element 3.

[lukel97]

I think we need to mark v[f]slide1down with ActiveElementsAffectResult since the index the element is inserted at depends on VL.