AMDGPU: Make better use of op_sel with high components

Handle more general swizzles.

llvm-svn: 303296

1 files changed, 429 insertions, 2 deletions

diff --git a/llvm/test/CodeGen/AMDGPU/packed-op-sel.ll b/llvm/test/CodeGen/AMDGPU/packed-op-sel.ll
index 6ff0c54c33d..4970375d40d 100644
--- a/llvm/test/CodeGen/AMDGPU/packed-op-sel.ll
+++ b/llvm/test/CodeGen/AMDGPU/packed-op-sel.ll
@@ -181,8 +181,7 @@ bb:
 ; GCN-NOT: shl
 ; GCN-NOT: or
 
-; GCN: v_xor_b32_e32 [[NEG_SCALAR0:v[0-9]+]], 0x8000, [[SCALAR0]]
-; GCN-NEXT: v_pk_add_u16 v{{[0-9]+}}, [[VEC0]], [[NEG_SCALAR0]] op_sel_hi:[1,0]{{$}}
+; GCN: v_pk_add_u16 v{{[0-9]+}}, [[VEC0]], [[SCALAR0]] op_sel_hi:[1,0] neg_lo:[0,1] neg_hi:[0,1]{{$}}
 define amdgpu_kernel void @add_vector_neg_bitcast_scalar_lo(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(3)* %lds, half addrspace(3)* %arg2) #0 {
 bb:
   %vec0 = load volatile <2 x i16>, <2 x i16> addrspace(3)* %lds, align 4
@@ -260,6 +259,434 @@ bb:
   ret void
 }
 
+; GCN-LABEL: {{^}}fma_vector_vector_neg_vector_hi:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1] neg_lo:[0,0,1] neg_hi:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_neg_vector_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %vec2.fneg = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %vec2.fneg.elt1.broadcast = shufflevector <2 x half> %vec2.fneg, <2 x half> undef, <2 x i32> <i32 1, i32 1>
+
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %vec2.fneg.elt1.broadcast)
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_vector_neg_hi:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] neg_hi:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_vector_neg_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %vec2.elt1 = extractelement <2 x half> %vec2, i32 1
+  %neg.vec2.elt1 = fsub half -0.0, %vec2.elt1
+
+  %neg.vec2.elt1.insert = insertelement <2 x half> %vec2, half %neg.vec2.elt1, i32 1
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %neg.vec2.elt1.insert)
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}add_vector_scalar_hi:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_add_u16 v{{[0-9]+}}, [[VEC0]], [[VEC1]] op_sel:[0,1]{{$}}
+define amdgpu_kernel void @add_vector_scalar_hi(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(3)* %lds, i32 1
+
+  %vec0 = load volatile <2 x i16>, <2 x i16> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x i16>, <2 x i16> addrspace(3)* %lds.gep1, align 4
+
+  %vec1.elt1.broadcast = shufflevector <2 x i16> %vec1, <2 x i16> undef, <2 x i32> <i32 1, i32 1>
+  %result = add <2 x i16> %vec0, %vec1.elt1.broadcast
+
+  store <2 x i16> %result, <2 x i16> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_scalar_hi:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_scalar_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %vec2.elt1.broadcast = shufflevector <2 x half> %vec2, <2 x half> undef, <2 x i32> <i32 1, i32 1>
+
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %vec2.elt1.broadcast)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_neg_vector_lo_neg_hi:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]]{{$}}
+define amdgpu_kernel void @fma_vector_vector_neg_vector_lo_neg_hi(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %neg.vec2.elt1 = extractelement <2 x half> %neg.vec2, i32 1
+  %neg.neg.vec2.elt1 = fsub half -0.0, %neg.vec2.elt1
+  %neg.neg.vec2.elt1.insert = insertelement <2 x half> %vec2, half %neg.neg.vec2.elt1, i32 1
+
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %neg.neg.vec2.elt1.insert)
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_swap_vector:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1] op_sel_hi:[1,1,0]{{$}}
+define amdgpu_kernel void @fma_vector_vector_swap_vector(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %vec2.swap = shufflevector <2 x half> %vec2, <2 x half> undef, <2 x i32> <i32 1, i32 0>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %vec2.swap)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_swap_neg_vector:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+; GCN-NOT: xor
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1] op_sel_hi:[1,1,0] neg_lo:[0,0,1] neg_hi:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_swap_neg_vector(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+
+  %neg.vec2.swap = shufflevector <2 x half> %neg.vec2, <2 x half> undef, <2 x i32> <i32 1, i32 0>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %neg.vec2.swap)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_blend_vector_neg_vector_0:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+; GCN-NOT: xor
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1] op_sel_hi:[1,1,0] neg_lo:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_blend_vector_neg_vector_0(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %combined = shufflevector <2 x half> %vec2, <2 x half> %neg.vec2, <2 x i32> <i32 3, i32 0>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %combined)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_blend_vector_neg_vector_1:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+; GCN-NOT: xor
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] neg_lo:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_blend_vector_neg_vector_1(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %combined = shufflevector <2 x half> %vec2, <2 x half> %neg.vec2, <2 x i32> <i32 2, i32 1>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %combined)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_blend_vector_neg_vector_2:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+; GCN-NOT: xor
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] neg_hi:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_blend_vector_neg_vector_2(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %combined = shufflevector <2 x half> %vec2, <2 x half> %neg.vec2, <2 x i32> <i32 0, i32 3>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %combined)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}fma_vector_vector_blend_vector_neg_vector_3:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: or
+; GCN-NOT: xor
+
+; GCN: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[VEC2]] op_sel:[0,0,1] neg_lo:[0,0,1]{{$}}
+define amdgpu_kernel void @fma_vector_vector_blend_vector_neg_vector_3(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+  %neg.vec2 = fsub <2 x half> <half -0.0, half -0.0>, %vec2
+  %combined = shufflevector <2 x half> %vec2, <2 x half> %neg.vec2, <2 x i32> <i32 3, i32 1>
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %combined)
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}bitcast_fneg_f32:
+; GCN: v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+$}}
+define amdgpu_kernel void @bitcast_fneg_f32(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %f32 = load volatile float, float addrspace(3)* undef, align 4
+  %neg.f32 = fsub float -0.0, %f32
+  %bc = bitcast float %neg.f32 to <2 x half>
+  %result = fadd <2 x half> %vec0, %bc
+
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}shuffle_bitcast_fneg_f32:
+; GCN: v_pk_add_f16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} op_sel:[0,1] op_sel_hi:[1,0]{{$}}
+define amdgpu_kernel void @shuffle_bitcast_fneg_f32(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+
+  %f32 = load volatile float, float addrspace(3)* undef, align 4
+  %neg.f32 = fsub float -0.0, %f32
+  %bc = bitcast float %neg.f32 to <2 x half>
+  %shuf = shufflevector <2 x half> %bc, <2 x half> undef, <2 x i32> <i32 1, i32 0>
+  %result = fadd <2 x half> %vec0, %shuf
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}extract_from_i64:
+; GCN: v_lshl_or_b32
+; GCN: v_pk_add_u16 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+$}}
+define amdgpu_kernel void @extract_from_i64(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(3)* %lds) #0 {
+bb:
+  %vec0 = load volatile <2 x i16>, <2 x i16> addrspace(3)* %lds, align 4
+  %i64 = load volatile i64, i64 addrspace(1)* undef
+
+  %elt0 = trunc i64 %i64 to i16
+  %hi = lshr i64 %i64, 16
+  %elt1 = trunc i64 %hi to i16
+
+  %ins0 = insertelement <2 x i16> undef, i16 %elt1, i32 0
+  %ins1 = insertelement <2 x i16> %ins0, i16 %elt0, i32 1
+  %result = add <2 x i16> %vec0, %ins1
+  store <2 x i16> %result, <2 x i16> addrspace(1)* %out, align 4
+  ret void
+}
+
+
+; Bitcast is final obstacle to identifying same source register
+; GCN-LABEL: {{^}}bitcast_lo_elt_op_sel:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: _or
+
+; GCN: v_pk_add_f16 [[FADD:v[0-9]+]]
+; GCN-NEXT: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[FADD]] op_sel:[0,0,1] op_sel_hi:[1,1,0]{{$}}
+define amdgpu_kernel void @bitcast_lo_elt_op_sel(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %scalar0 = load volatile i16, i16 addrspace(1)* undef
+  %shl = shl i16 %scalar0, 1
+  %shl.bc = bitcast i16 %shl to half
+
+  %fadd = fadd <2 x half> %vec2, <half 2.0, half 2.0>
+  %shuffle = shufflevector <2 x half> %fadd, <2 x half> %vec2, <2 x i32> <i32 1, i32 0>
+
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %shuffle)
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
+
+; Bitcast is final obstacle to identifying same source register
+; GCN-LABEL: {{^}}mix_elt_types_op_sel:
+; GCN: ds_read_b32 [[VEC0:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC1:v[0-9]+]]
+; GCN: ds_read_b32 [[VEC2:v[0-9]+]]
+
+; GCN-NOT: pack
+; GCN-NOT: and
+; GCN-NOT: shl
+; GCN-NOT: _or
+
+; GCN: v_pk_add_f16 [[FADD:v[0-9]+]]
+; GCN-NEXT: v_pk_fma_f16 v{{[0-9]+}}, [[VEC0]], [[VEC1]], [[FADD]] op_sel:[0,0,1] op_sel_hi:[1,1,0]{{$}}
+define amdgpu_kernel void @mix_elt_types_op_sel(<2 x half> addrspace(1)* %out, <2 x half> addrspace(3)* %lds) #0 {
+bb:
+  %lds.gep1 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 1
+  %lds.gep2 = getelementptr inbounds <2 x half>, <2 x half> addrspace(3)* %lds, i32 2
+
+  %vec0 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds, align 4
+  %vec1 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep1, align 4
+  %vec2 = load volatile <2 x half>, <2 x half> addrspace(3)* %lds.gep2, align 4
+
+  %scalar0 = load volatile i16, i16 addrspace(1)* undef
+  %scalar1 = load volatile half, half addrspace(1)* undef
+  %shl = shl i16 %scalar0, 1
+  %shl.bc = bitcast i16 %shl to half
+
+  %insert0 = insertelement <2 x half> undef, half %shl.bc, i32 0
+
+  %fadd = fadd <2 x half> %vec2, <half 2.0, half 2.0>
+  %insert1 = shufflevector <2 x half> %fadd, <2 x half> %insert0, <2 x i32> <i32 1, i32 0>
+
+  %result = tail call <2 x half> @llvm.fma.v2f16(<2 x half> %vec0, <2 x half> %vec1, <2 x half> %insert1)
+  store <2 x half> %result, <2 x half> addrspace(1)* %out, align 4
+  ret void
+}
+
 declare <2 x half> @llvm.fma.v2f16(<2 x half>, <2 x half>, <2 x half>) #1
 
 attributes #0 = { nounwind }