[llvm] [AMDGPU][MC] Restrict op_sel in VOP3P dot instructions (PR #100485)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -389,345 +389,23 @@ v_dot8_i32_i4 v0, v1, v2, v3
 v_dot8_u32_u4 v0, v1, v2, v3
 
 //
-// Test op_sel/op_sel_hi.
+// Test op_sel/op_sel_hi: in VOP3P dot, op_sel must be 0, op_sel_hi cannot appear
 //
 
 // CHECK: encoding: [0x00,0x40,0xa3,0xd3,0x01,0x05,0x0e,0x1c]
 v_dot2_f32_f16 v0, v1, v2, v3 op_sel:[0,0]
-// CHECK: encoding: [0x00,0x50,0xa3,0xd3,0x01,0x05,0x0e,0x1c]
-v_dot2_f32_f16 v0, v1, v2, v3 op_sel:[0,1]
----------------
rampitec wrote:

Right, dot4 and dot8 do not have opsel. The difference is that these work on 4 and 8 component packed data and opsel can only select a word. Technically we are using V2F16 for v_dot2_f32_f16 in the profile and I32 for dot4 and dot8.

https://github.com/llvm/llvm-project/pull/100485