[AMDGPU][MC][GFX9] Added op_sel support for v_mad_*16, v_fma_f16, v_div_fixup_f16

[dp]

Summary

This change implements features postponed in D35424 because of a dependency on D36322

[dp]

Needs refactoring after integration of D36674

[dp]

Updated to reflect D36674 changes. Refactored to improve readability.

[SamWot]

LGTM with small issue.

lib/Target/AMDGPU/VOP3Instructions.td

117

Better use HasModifiers or HasSrc0FloatMods from VOPProfile class.

lib/Target/AMDGPU/VOP3Instructions.td

class getVOP3ClampPat<VOPProfile P, SDPatternOperator node> {
   list<dag> ret1 = [(set P.DstVT:$vdst, (node P.Src0VT:$src0, i1:$clamp))];
   list<dag> ret = !if(!eq(P.NumSrcArgs, 3), ret3,
                   !if(!eq(P.NumSrcArgs, 2), ret2,
                   ret1));
 }
 
 class VOP3Inst<string OpName, VOPProfile P, SDPatternOperator node = null_frag, bit VOP3Only = 0> :
   VOP3_Pseudo<OpName, P,
+    !if(P.HasOpSel,
+        !if(isFloatType<P.Src0VT>.ret,
+            getVOP3OpSelModPat<P, node>.ret,
+            getVOP3OpSelPat<P, node>.ret),
         !if(P.HasModifiers,
             getVOP3ModPat<P, node>.ret,
             !if(P.HasIntClamp,
                 getVOP3ClampPat<P, node>.ret,
-            getVOP3Pat<P, node>.ret)),
-    VOP3Only> {
-  let IntClamp = P.HasIntClamp;
-}
-
-class VOP3OpSelInst<string OpName, VOPProfile P, SDPatternOperator node = null_frag> :
-  VOP3_Pseudo<OpName, P,
-    !if(isFloatType<P.Src0VT>.ret,
-        getVOP3OpSelModPat<P, node>.ret,
-        getVOP3OpSelPat<P, node>.ret),
-    1, 0, 1> {
+                getVOP3Pat<P, node>.ret))),
+    VOP3Only, 0, P.HasOpSel> {
 
-  let AsmMatchConverter = "cvtVOP3OpSel";
+  let IntClamp = P.HasIntClamp;
+  let AsmMatchConverter =
+    !if(P.HasOpSel,
+        "cvtVOP3OpSel",
+        !if(!or(P.HasModifiers, !or(P.HasOMod, P.HasIntClamp)),
+            "cvtVOP3",
+            ""));
 }
 
 // Special case for v_div_fmas_{f32|f64}, since it seems to be the
 // only VOP instruction that implicitly reads VCC.
 let Asm64 = " $vdst, $src0_modifiers, $src1_modifiers, $src2_modifiers$clamp$omod" in {
 def VOP_F32_F32_F32_F32_VCC : VOPProfile<[f32, f32, f32, f32]> {
   let Outs64 = (outs DstRC.RegClass:$vdst);
 }
 def VOP_F64_F64_F64_F64_VCC : VOPProfile<[f64, f64, f64, f64]> {
   let Outs64 = (outs DstRC.RegClass:$vdst);
 }
 }
 
 class getVOP3VCC<VOPProfile P, SDPatternOperator node> {
   list<dag> ret =
     [(set P.DstVT:$vdst,
       (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omod)),
             (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)),
             (P.Src2VT (VOP3Mods P.Src2VT:$src2, i32:$src2_modifiers)),
             (i1 VCC)))];
 }
 
-class VOP3_Profile<VOPProfile P> : VOPProfile<P.ArgVT> {
-  // FIXME: Hack to stop printing _e64
-  let Outs64 = (outs DstRC.RegClass:$vdst);
-  let Asm64 = " " # P.Asm64;
+class VOP3Features<bit Clamp, bit OpSel> {
+  bit HasClamp = Clamp;
+  bit HasOpSel = OpSel;
 }
 
-class VOP3Clamp_Profile<VOPProfile P> : VOPProfile<P.ArgVT> {
-  let HasClamp = 1;
+def VOP3_REGULAR : VOP3Features<0, 0>;
+def VOP3_CLAMP   : VOP3Features<1, 0>;
+def VOP3_OPSEL   : VOP3Features<1, 1>;
+
+class VOP3_Profile<VOPProfile P, VOP3Features Features = VOP3_REGULAR> : VOPProfile<P.ArgVT> {
+
+  let HasClamp = !if(Features.HasClamp, 1, P.HasClamp);
+  let HasOpSel = !if(Features.HasOpSel, 1, P.HasOpSel);
 
   // FIXME: Hack to stop printing _e64
   let Outs64 = (outs DstRC.RegClass:$vdst);
-  let Asm64 = " " # getAsm64<HasDst, NumSrcArgs, HasIntClamp, HasModifiers, HasOMod, DstVT>.ret;
-}
-
-class VOP3OpSel_Profile<VOPProfile P> : VOP3_Profile<P> {
-  let HasClamp = 1;
-  let HasOpSel = 1;
+  let Asm64 =
+    " " # !if(Features.HasOpSel,
+              getAsmVOP3OpSel<NumSrcArgs,
+                              HasIntClamp,
+                              HasSrc0FloatMods,
+                              HasSrc1FloatMods,
+                              HasSrc2FloatMods>.ret,
+              !if(Features.HasClamp,
+                  getAsm64<HasDst, NumSrcArgs, HasIntClamp, 
+                           HasModifiers, HasOMod, DstVT>.ret,
+                  P.Asm64));
 }
 
 class VOP3b_Profile<ValueType vt> : VOPProfile<[vt, vt, vt, vt]> {
   // v_div_scale_{f32|f64} do not support input modifiers.
   let HasModifiers = 0;
   let HasOMod = 0;
   let Outs64 = (outs DstRC:$vdst, SReg_64:$sdst);
   let Asm64 = " $vdst, $sdst, $src0, $src1, $src2";

 //===----------------------------------------------------------------------===//
 // VOP3 Instructions
 //===----------------------------------------------------------------------===//
 
 let isCommutable = 1 in {
 
 def V_MAD_LEGACY_F32 : VOP3Inst <"v_mad_legacy_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
 def V_MAD_F32 : VOP3Inst <"v_mad_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fmad>;
-def V_MAD_I32_I24 : VOP3Inst <"v_mad_i32_i24", VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
-def V_MAD_U32_U24 : VOP3Inst <"v_mad_u32_u24", VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
+def V_MAD_I32_I24 : VOP3Inst <"v_mad_i32_i24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
+def V_MAD_U32_U24 : VOP3Inst <"v_mad_u32_u24", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
 def V_FMA_F32 : VOP3Inst <"v_fma_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, fma>;
 def V_FMA_F64 : VOP3Inst <"v_fma_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, fma>;
 def V_LERP_U8 : VOP3Inst <"v_lerp_u8", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_lerp>;
 
 let SchedRW = [WriteDoubleAdd] in {
 def V_ADD_F64 : VOP3Inst <"v_add_f64", VOP3_Profile<VOP_F64_F64_F64>, fadd, 1>;
 def V_MUL_F64 : VOP3Inst <"v_mul_f64", VOP3_Profile<VOP_F64_F64_F64>, fmul, 1>;
 def V_MIN_F64 : VOP3Inst <"v_min_f64", VOP3_Profile<VOP_F64_F64_F64>, fminnum, 1>;

 def V_MIN3_I32 : VOP3Inst <"v_min3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmin3>;
 def V_MIN3_U32 : VOP3Inst <"v_min3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumin3>;
 def V_MAX3_F32 : VOP3Inst <"v_max3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmax3>;
 def V_MAX3_I32 : VOP3Inst <"v_max3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmax3>;
 def V_MAX3_U32 : VOP3Inst <"v_max3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumax3>;
 def V_MED3_F32 : VOP3Inst <"v_med3_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUfmed3>;
 def V_MED3_I32 : VOP3Inst <"v_med3_i32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUsmed3>;
 def V_MED3_U32 : VOP3Inst <"v_med3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>, AMDGPUumed3>;
-def V_SAD_U8 : VOP3Inst <"v_sad_u8",    VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
-def V_SAD_HI_U8 : VOP3Inst <"v_sad_hi_u8", VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
-def V_SAD_U16 : VOP3Inst <"v_sad_u16",   VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
-def V_SAD_U32 : VOP3Inst <"v_sad_u32",   VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
+def V_SAD_U8 : VOP3Inst <"v_sad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
+def V_SAD_HI_U8 : VOP3Inst <"v_sad_hi_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
+def V_SAD_U16 : VOP3Inst <"v_sad_u16", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
+def V_SAD_U32 : VOP3Inst <"v_sad_u32", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
 def V_CVT_PK_U8_F32 : VOP3Inst<"v_cvt_pk_u8_f32", VOP3_Profile<VOP_I32_F32_I32_I32>, int_amdgcn_cvt_pk_u8_f32>;
 def V_DIV_FIXUP_F32 : VOP3Inst <"v_div_fixup_f32", VOP3_Profile<VOP_F32_F32_F32_F32>, AMDGPUdiv_fixup>;
 
 let SchedRW = [WriteDoubleAdd] in {
 def V_DIV_FIXUP_F64 : VOP3Inst <"v_div_fixup_f64", VOP3_Profile<VOP_F64_F64_F64_F64>, AMDGPUdiv_fixup>;
 def V_LDEXP_F64 : VOP3Inst <"v_ldexp_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUldexp, 1>;
 } // End SchedRW = [WriteDoubleAdd]
 
 def V_DIV_SCALE_F32 : VOP3_Pseudo <"v_div_scale_f32", VOP3b_F32_I1_F32_F32_F32, [], 1> {
   let SchedRW = [WriteFloatFMA, WriteSALU];
   let hasExtraSrcRegAllocReq = 1;
   let AsmMatchConverter = "";
 }
 
 // Double precision division pre-scale.
 def V_DIV_SCALE_F64 : VOP3_Pseudo <"v_div_scale_f64", VOP3b_F64_I1_F64_F64_F64, [], 1> {
   let SchedRW = [WriteDouble, WriteSALU];
   let hasExtraSrcRegAllocReq = 1;
   let AsmMatchConverter = "";
 }
 
-def V_MSAD_U8 : VOP3Inst <"v_msad_u8", VOP3Clamp_Profile<VOP_I32_I32_I32_I32>>;
+def V_MSAD_U8 : VOP3Inst <"v_msad_u8", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_CLAMP>>;
 
 let Constraints = "@earlyclobber $vdst" in {
-def V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3Clamp_Profile<VOP_I64_I64_I32_I64>>;
+def V_MQSAD_PK_U16_U8 : VOP3Inst <"v_mqsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
 } // End Constraints = "@earlyclobber $vdst"
 
 def V_TRIG_PREOP_F64 : VOP3Inst <"v_trig_preop_f64", VOP3_Profile<VOP_F64_F64_I32>, AMDGPUtrig_preop> {
   let SchedRW = [WriteDouble];
 }
 
 // These instructions only exist on SI and CI
 let SubtargetPredicate = isSICI in {
 def V_LSHL_B64 : VOP3Inst <"v_lshl_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_LSHR_B64 : VOP3Inst <"v_lshr_b64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_ASHR_I64 : VOP3Inst <"v_ashr_i64", VOP3_Profile<VOP_I64_I64_I32>>;
 def V_MULLIT_F32 : VOP3Inst <"v_mullit_f32", VOP3_Profile<VOP_F32_F32_F32_F32>>;
 } // End SubtargetPredicate = isSICI
 
 let SubtargetPredicate = isVI in {
 def V_LSHLREV_B64 : VOP3Inst <"v_lshlrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_LSHRREV_B64 : VOP3Inst <"v_lshrrev_b64", VOP3_Profile<VOP_I64_I32_I64>>;
 def V_ASHRREV_I64 : VOP3Inst <"v_ashrrev_i64", VOP3_Profile<VOP_I64_I32_I64>>;
 } // End SubtargetPredicate = isVI
 
 
 let SubtargetPredicate = isCIVI in {
 
 let Constraints = "@earlyclobber $vdst" in {
-def V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3Clamp_Profile<VOP_I64_I64_I32_I64>>;
-def V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3Clamp_Profile<VOP_V4I32_I64_I32_V4I32>>;
+def V_QSAD_PK_U16_U8 : VOP3Inst <"v_qsad_pk_u16_u8", VOP3_Profile<VOP_I64_I64_I32_I64, VOP3_CLAMP>>;
+def V_MQSAD_U32_U8 : VOP3Inst <"v_mqsad_u32_u8", VOP3_Profile<VOP_V4I32_I64_I32_V4I32, VOP3_CLAMP>>;
 } // End Constraints = "@earlyclobber $vdst"
 
 let isCommutable = 1 in {
 def V_MAD_U64_U32 : VOP3Inst <"v_mad_u64_u32", VOP3b_I64_I1_I32_I32_I64>;
 def V_MAD_I64_I32 : VOP3Inst <"v_mad_i64_i32", VOP3b_I64_I1_I32_I32_I64>;
 } // End isCommutable = 1
 
 } // End SubtargetPredicate = isCIVI
 
 
 let SubtargetPredicate = Has16BitInsts in {
 
 let F16_ZFILL = 1 in {
 def V_DIV_FIXUP_F16 : VOP3Inst <"v_div_fixup_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, AMDGPUdiv_fixup>;
 }
 let SubtargetPredicate = isGFX9 in {
-def V_DIV_FIXUP_F16_gfx9   : VOP3Inst <"v_div_fixup_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16>>;
+def V_DIV_FIXUP_F16_gfx9 : VOP3Inst <"v_div_fixup_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
 }
 
 let isCommutable = 1 in {
 
 let F16_ZFILL = 1 in {
 def V_MAD_F16 : VOP3Inst <"v_mad_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, fmad>;
-def V_MAD_U16 : VOP3Inst <"v_mad_u16", VOP3Clamp_Profile<VOP_I16_I16_I16_I16>>;
-def V_MAD_I16 : VOP3Inst <"v_mad_i16", VOP3Clamp_Profile<VOP_I16_I16_I16_I16>>;
+def V_MAD_U16 : VOP3Inst <"v_mad_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>;
+def V_MAD_I16 : VOP3Inst <"v_mad_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_CLAMP>>;
 def V_FMA_F16 : VOP3Inst <"v_fma_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, fma>;
 }
 
 let SubtargetPredicate = isGFX9 in {
-def V_MAD_F16_gfx9   : VOP3Inst <"v_mad_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16>>;
-def V_MAD_U16_gfx9   : VOP3Inst <"v_mad_u16_gfx9", VOP3Clamp_Profile<VOP_I16_I16_I16_I16>>;
-def V_MAD_I16_gfx9   : VOP3Inst <"v_mad_i16_gfx9", VOP3Clamp_Profile<VOP_I16_I16_I16_I16>>;
-def V_FMA_F16_gfx9   : VOP3Inst <"v_fma_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16>>;
+def V_MAD_F16_gfx9   : VOP3Inst <"v_mad_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
+def V_MAD_U16_gfx9   : VOP3Inst <"v_mad_u16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>;
+def V_MAD_I16_gfx9   : VOP3Inst <"v_mad_i16_gfx9", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>>;
+def V_FMA_F16_gfx9   : VOP3Inst <"v_fma_f16_gfx9", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>>;
 } // End SubtargetPredicate = isGFX9
 
 def V_INTERP_P1LL_F16 : VOP3Interp <"v_interp_p1ll_f16", VOP3_INTERP16<[f32, f32, i32, untyped]>>;
 def V_INTERP_P1LV_F16 : VOP3Interp <"v_interp_p1lv_f16", VOP3_INTERP16<[f32, f32, i32, f16]>>;
 def V_INTERP_P2_F16   : VOP3Interp <"v_interp_p2_f16",   VOP3_INTERP16<[f16, f32, i32, f32]>>;
 
 }  // End isCommutable = 1
 } // End SubtargetPredicate = Has16BitInsts

 }
 
 defm: Ternary_i16_Pats<mul, add, V_MAD_U16, zext>;
 defm: Ternary_i16_Pats<mul, add, V_MAD_I16, sext>;
 
 } // End Predicates = [Has16BitInsts]
 
 let SubtargetPredicate = isGFX9 in {
-def V_PACK_B32_F16 : VOP3OpSelInst <"v_pack_b32_f16", VOP3OpSel_Profile<VOP_B32_F16_F16>>;
+def V_PACK_B32_F16 : VOP3Inst <"v_pack_b32_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
 def V_LSHL_ADD_U32 : VOP3Inst <"v_lshl_add_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 def V_ADD_LSHL_U32 : VOP3Inst <"v_add_lshl_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 def V_ADD3_U32 : VOP3Inst <"v_add3_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 def V_LSHL_OR_B32 : VOP3Inst <"v_lshl_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 def V_AND_OR_B32 : VOP3Inst <"v_and_or_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 def V_OR3_B32 : VOP3Inst <"v_or3_b32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 
 def V_XAD_U32 : VOP3Inst <"v_xad_u32", VOP3_Profile<VOP_I32_I32_I32_I32>>;
 
-def V_MED3_F16 : VOP3OpSelInst <"v_med3_f16", VOP3OpSel_Profile<VOP_F16_F16_F16_F16>, AMDGPUfmed3>;
-def V_MED3_I16 : VOP3OpSelInst <"v_med3_i16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUsmed3>;
-def V_MED3_U16 : VOP3OpSelInst <"v_med3_u16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUumed3>;
-
-def V_MIN3_F16 : VOP3OpSelInst <"v_min3_f16", VOP3OpSel_Profile<VOP_F16_F16_F16_F16>, AMDGPUfmin3>;
-def V_MIN3_I16 : VOP3OpSelInst <"v_min3_i16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUsmin3>;
-def V_MIN3_U16 : VOP3OpSelInst <"v_min3_u16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUumin3>;
-
-def V_MAX3_F16 : VOP3OpSelInst <"v_max3_f16", VOP3OpSel_Profile<VOP_F16_F16_F16_F16>, AMDGPUfmax3>;
-def V_MAX3_I16 : VOP3OpSelInst <"v_max3_i16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUsmax3>;
-def V_MAX3_U16 : VOP3OpSelInst <"v_max3_u16", VOP3OpSel_Profile<VOP_I16_I16_I16_I16>, AMDGPUumax3>;
+def V_MED3_F16 : VOP3Inst <"v_med3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmed3>;
+def V_MED3_I16 : VOP3Inst <"v_med3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmed3>;
+def V_MED3_U16 : VOP3Inst <"v_med3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumed3>;
+
+def V_MIN3_F16 : VOP3Inst <"v_min3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmin3>;
+def V_MIN3_I16 : VOP3Inst <"v_min3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmin3>;
+def V_MIN3_U16 : VOP3Inst <"v_min3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumin3>;
+
+def V_MAX3_F16 : VOP3Inst <"v_max3_f16", VOP3_Profile<VOP_F16_F16_F16_F16, VOP3_OPSEL>, AMDGPUfmax3>;
+def V_MAX3_I16 : VOP3Inst <"v_max3_i16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUsmax3>;
+def V_MAX3_U16 : VOP3Inst <"v_max3_u16", VOP3_Profile<VOP_I16_I16_I16_I16, VOP3_OPSEL>, AMDGPUumax3>;
 
-def V_ADD_I16 : VOP3OpSelInst <"v_add_i16", VOP3OpSel_Profile<VOP_I16_I16_I16>>;
-def V_SUB_I16 : VOP3OpSelInst <"v_sub_i16", VOP3OpSel_Profile<VOP_I16_I16_I16>>;
+def V_ADD_I16 : VOP3Inst <"v_add_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>;
+def V_SUB_I16 : VOP3Inst <"v_sub_i16", VOP3_Profile<VOP_I16_I16_I16, VOP3_OPSEL>>;
 
-def V_MAD_U32_U16 : VOP3OpSelInst <"v_mad_u32_u16", VOP3OpSel_Profile<VOP_I32_I16_I16_I32>>;
-def V_MAD_I32_I16 : VOP3OpSelInst <"v_mad_i32_i16", VOP3OpSel_Profile<VOP_I32_I16_I16_I32>>;
+def V_MAD_U32_U16 : VOP3Inst <"v_mad_u32_u16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
+def V_MAD_I32_I16 : VOP3Inst <"v_mad_i32_i16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
 
-def V_CVT_PKNORM_I16_F16 : VOP3OpSelInst <"v_cvt_pknorm_i16_f16", VOP3OpSel_Profile<VOP_B32_F16_F16>>;
-def V_CVT_PKNORM_U16_F16 : VOP3OpSelInst <"v_cvt_pknorm_u16_f16", VOP3OpSel_Profile<VOP_B32_F16_F16>>;
+def V_CVT_PKNORM_I16_F16 : VOP3Inst <"v_cvt_pknorm_i16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
+def V_CVT_PKNORM_U16_F16 : VOP3Inst <"v_cvt_pknorm_u16_f16", VOP3_Profile<VOP_B32_F16_F16, VOP3_OPSEL>>;
 } // End SubtargetPredicate = isGFX9
 
 //===----------------------------------------------------------------------===//
 // Integer Clamp Patterns
 //===----------------------------------------------------------------------===//
 
 class getClampPat<VOPProfile P, SDPatternOperator node> {
   dag ret3 = (P.DstVT (node P.Src0VT:$src0, P.Src1VT:$src1, P.Src2VT:$src2));

 multiclass VOP3_F16_Real_gfx9<bits<10> op, string OpName, string AsmName> {
   def _vi : VOP3_Real<!cast<VOP3_Pseudo>(OpName), SIEncodingFamily.GFX9>,
             VOP3e_vi <op, !cast<VOP3_Pseudo>(OpName).Pfl> {
               VOP3_Pseudo ps = !cast<VOP3_Pseudo>(OpName);
               let AsmString = AsmName # ps.AsmOperands;
             }
 }
 
+multiclass VOP3OpSel_F16_Real_gfx9<bits<10> op, string AsmName> {
+  def _vi : VOP3_Real<!cast<VOP3_Pseudo>(NAME), SIEncodingFamily.GFX9>,
+            VOP3OpSel_gfx9 <op, !cast<VOP3_Pseudo>(NAME).Pfl> {
+              VOP3_Pseudo ps = !cast<VOP3_Pseudo>(NAME);
+              let AsmString = AsmName # ps.AsmOperands;
+            }
+}
+
 } // End AssemblerPredicates = [isGFX9], DecoderNamespace = "GFX9"
 
 defm V_MAD_U64_U32      : VOP3be_Real_vi <0x1E8>;
 defm V_MAD_I64_I32      : VOP3be_Real_vi <0x1E9>;
 
 defm V_MAD_LEGACY_F32   : VOP3_Real_vi <0x1c0>;
 defm V_MAD_F32          : VOP3_Real_vi <0x1c1>;
 defm V_MAD_I32_I24      : VOP3_Real_vi <0x1c2>;

 defm V_DIV_FIXUP_F16    : VOP3_F16_Real_vi <0x1ef>;
 
 defm V_MAD_LEGACY_F16       : VOP3_F16_Real_gfx9 <0x1ea, "V_MAD_F16",       "v_mad_legacy_f16">;
 defm V_MAD_LEGACY_U16       : VOP3_F16_Real_gfx9 <0x1eb, "V_MAD_U16",       "v_mad_legacy_u16">;
 defm V_MAD_LEGACY_I16       : VOP3_F16_Real_gfx9 <0x1ec, "V_MAD_I16",       "v_mad_legacy_i16">;
 defm V_FMA_LEGACY_F16       : VOP3_F16_Real_gfx9 <0x1ee, "V_FMA_F16",       "v_fma_legacy_f16">;
 defm V_DIV_FIXUP_LEGACY_F16 : VOP3_F16_Real_gfx9 <0x1ef, "V_DIV_FIXUP_F16", "v_div_fixup_legacy_f16">;
 
-defm V_MAD_F16_gfx9         : VOP3_F16_Real_gfx9 <0x203, "V_MAD_F16_gfx9",       "v_mad_f16">;
-defm V_MAD_U16_gfx9         : VOP3_F16_Real_gfx9 <0x204, "V_MAD_U16_gfx9",       "v_mad_u16">;
-defm V_MAD_I16_gfx9         : VOP3_F16_Real_gfx9 <0x205, "V_MAD_I16_gfx9",       "v_mad_i16">;
-defm V_FMA_F16_gfx9         : VOP3_F16_Real_gfx9 <0x206, "V_FMA_F16_gfx9",       "v_fma_f16">;
-defm V_DIV_FIXUP_F16_gfx9   : VOP3_F16_Real_gfx9 <0x207, "V_DIV_FIXUP_F16_gfx9", "v_div_fixup_f16">;
+defm V_MAD_F16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x203, "v_mad_f16">;
+defm V_MAD_U16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x204, "v_mad_u16">;
+defm V_MAD_I16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x205, "v_mad_i16">;
+defm V_FMA_F16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x206, "v_fma_f16">;
+defm V_DIV_FIXUP_F16_gfx9   : VOP3OpSel_F16_Real_gfx9 <0x207, "v_div_fixup_f16">;
 
 defm V_INTERP_P1_F32_e64  : VOP3Interp_Real_vi <0x270>;
 defm V_INTERP_P2_F32_e64  : VOP3Interp_Real_vi <0x271>;
 defm V_INTERP_MOV_F32_e64 : VOP3Interp_Real_vi <0x272>;
 
 defm V_INTERP_P1LL_F16  : VOP3Interp_Real_vi <0x274>;
 defm V_INTERP_P1LV_F16  : VOP3Interp_Real_vi <0x275>;
 defm V_INTERP_P2_F16    : VOP3Interp_Real_vi <0x276>;

test/MC/AMDGPU/vop3-gfx9.s

 // GFX9: v_fma_f16 v5, v1, -v2, v3 ; encoding: [0x05,0x00,0x06,0xd2,0x01,0x05,0x0e,0x44]
 
 v_fma_f16 v5, v1, v2, |v3|
 // GFX9: v_fma_f16 v5, v1, v2, |v3| ; encoding: [0x05,0x04,0x06,0xd2,0x01,0x05,0x0e,0x04]
 
 v_fma_f16 v5, v1, v2, v3 clamp
 // GFX9: v_fma_f16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x06,0xd2,0x01,0x05,0x0e,0x04]
 
+v_fma_f16 v5, v1, v2, v3 op_sel:[1,0,0,0]
+// GFX9: v_fma_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x06,0xd2,0x01,0x05,0x0e,0x04]
+
+v_fma_f16 v5, v1, v2, v3 op_sel:[0,1,0,0]
+// GFX9: v_fma_f16 v5, v1, v2, v3 op_sel:[0,1,0,0] ; encoding: [0x05,0x10,0x06,0xd2,0x01,0x05,0x0e,0x04]
+
+v_fma_f16 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX9: v_fma_f16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x06,0xd2,0x01,0x05,0x0e,0x04]
+
 v_fma_legacy_f16_e64 v5, v1, v2, v3
 // GFX9: v_fma_legacy_f16 v5, v1, v2, v3 ; encoding:  [0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x04]
 
 v_fma_legacy_f16 v5, -v1, v2, v3
 // GFX9: v_fma_legacy_f16 v5, -v1, v2, v3 ; encoding:  [0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x24]
 
 v_fma_legacy_f16 v5, v1, |v2|, v3
 // GFX9: v_fma_legacy_f16 v5, v1, |v2|, v3 ; encoding:  [0x05,0x02,0xee,0xd1,0x01,0x05,0x0e,0x04]

 // GFX9: v_div_fixup_f16 v5, -v1, v2, v3 ; encoding: [0x05,0x00,0x07,0xd2,0x01,0x05,0x0e,0x24]
 
 v_div_fixup_f16 v5, |v1|, v2, v3
 // GFX9: v_div_fixup_f16 v5, |v1|, v2, v3 ; encoding: [0x05,0x01,0x07,0xd2,0x01,0x05,0x0e,0x04]
 
 v_div_fixup_f16 v5, v1, v2, v3 clamp
 // GFX9: v_div_fixup_f16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x07,0xd2,0x01,0x05,0x0e,0x04]
 
+v_div_fixup_f16 v5, v1, v2, v3 op_sel:[1,0,0,0]
+// GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x07,0xd2,0x01,0x05,0x0e,0x04]
+
+v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,1,0]
+// GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,1,0] ; encoding: [0x05,0x20,0x07,0xd2,0x01,0x05,0x0e,0x04]
+
+v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,0,1]
+// GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x07,0xd2,0x01,0x05,0x0e,0x04]
+
 v_div_fixup_legacy_f16_e64 v5, 0.5, v2, v3
 // GFX9: v_div_fixup_legacy_f16 v5, 0.5, v2, v3 ; encoding: [0x05,0x00,0xef,0xd1,0xf0,0x04,0x0e,0x04]
 
 v_div_fixup_legacy_f16 v5, v1, 0.5, v3
 // GFX9: v_div_fixup_legacy_f16 v5, v1, 0.5, v3 ; encoding: [0x05,0x00,0xef,0xd1,0x01,0xe1,0x0d,0x04]
 
 v_div_fixup_legacy_f16 v5, v1, v2, 0.5
 // GFX9: v_div_fixup_legacy_f16 v5, v1, v2, 0.5 ; encoding: [0x05,0x00,0xef,0xd1,0x01,0x05,0xc2,0x03]

 // GFX9: v_mad_f16 v5, v1, v2, 0.5 ; encoding: [0x05,0x00,0x03,0xd2,0x01,0x05,0xc2,0x03]
 
 v_mad_f16 v5, v1, v2, -v3
 // GFX9: v_mad_f16 v5, v1, v2, -v3 ; encoding: [0x05,0x00,0x03,0xd2,0x01,0x05,0x0e,0x84]
 
 v_mad_f16 v5, v1, v2, |v3|
 // GFX9: v_mad_f16 v5, v1, v2, |v3| ; encoding: [0x05,0x04,0x03,0xd2,0x01,0x05,0x0e,0x04]
 
+v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,0,0]
+// GFX9: v_mad_f16 v5, v1, v2, v3 ; encoding: [0x05,0x00,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_f16 v5, v1, v2, v3 op_sel:[1,0,0,0]
+// GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_f16 v5, v1, v2, v3 op_sel:[0,1,0,0]
+// GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,1,0,0] ; encoding: [0x05,0x10,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,1,0]
+// GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,1,0] ; encoding: [0x05,0x20,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,0,1]
+// GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_f16 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x03,0xd2,0x01,0x05,0x0e,0x04]
+
 v_mad_f16 v5, v1, v2, v3 clamp
 // GFX9: v_mad_f16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x03,0xd2,0x01,0x05,0x0e,0x04]
 
 v_mad_i16_e64 v5, 0, v2, v3
 // GFX9: v_mad_i16 v5, 0, v2, v3 ; encoding: [0x05,0x00,0x05,0xd2,0x80,0x04,0x0e,0x04]
 
 v_mad_i16 v5, v1, -1, v3
 // GFX9: v_mad_i16 v5, v1, -1, v3 ; encoding: [0x05,0x00,0x05,0xd2,0x01,0x83,0x0d,0x04]
 
 v_mad_i16 v5, v1, v2, -4.0
 // GFX9: v_mad_i16 v5, v1, v2, -4.0 ; encoding: [0x05,0x00,0x05,0xd2,0x01,0x05,0xde,0x03]
 
 v_mad_i16 v5, v1, v2, v3 clamp
 // GFX9: v_mad_i16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x05,0xd2,0x01,0x05,0x0e,0x04]
 
+v_mad_i16 v5, v1, v2, v3 op_sel:[0,0,0,1]
+// GFX9: v_mad_i16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x05,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_i16 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX9: v_mad_i16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x05,0xd2,0x01,0x05,0x0e,0x04]
+
 v_mad_legacy_f16_e64 v5, 0.5, v2, v3
 // GFX9: v_mad_legacy_f16 v5, 0.5, v2, v3 ; encoding: [0x05,0x00,0xea,0xd1,0xf0,0x04,0x0e,0x04]
 
 v_mad_legacy_f16 v5, v1, 0.5, v3
 // GFX9: v_mad_legacy_f16 v5, v1, 0.5, v3 ; encoding: [0x05,0x00,0xea,0xd1,0x01,0xe1,0x0d,0x04]
 
 v_mad_legacy_f16 v5, v1, v2, 0.5
 // GFX9: v_mad_legacy_f16 v5, v1, v2, 0.5 ; encoding: [0x05,0x00,0xea,0xd1,0x01,0x05,0xc2,0x03]

 v_mad_u16 v5, v1, -1, v3
 // GFX9: v_mad_u16 v5, v1, -1, v3 ; encoding: [0x05,0x00,0x04,0xd2,0x01,0x83,0x0d,0x04]
 
 v_mad_u16 v5, v1, v2, -4.0
 // GFX9: v_mad_u16 v5, v1, v2, -4.0 ; encoding: [0x05,0x00,0x04,0xd2,0x01,0x05,0xde,0x03]
 
 v_mad_u16 v5, v1, v2, v3 clamp
 // GFX9: v_mad_u16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x04,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_u16 v5, v1, v2, v3 op_sel:[1,0,0,0]
+// GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x04,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_u16 v5, v1, v2, v3 op_sel:[0,0,0,1]
+// GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x04,0xd2,0x01,0x05,0x0e,0x04]
+
+v_mad_u16 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x04,0xd2,0x01,0x05,0x0e,0x04]

test/MC/Disassembler/AMDGPU/vop3_gfx9.txt

 # RUN: llvm-mc -arch=amdgcn -mcpu=gfx901 -disassemble -show-encoding < %s | FileCheck %s -check-prefix=GFX9
 
 # GFX9: v_fma_f16 v5, v1, v2, v3    ; encoding: [0x05,0x00,0x06,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x00,0x06,0xd2,0x01,0x05,0x0e,0x04
 
 # GFX9: v_fma_f16 v5, -v1, v2, v3    ; encoding: [0x05,0x00,0x06,0xd2,0x01,0x05,0x0e,0x24]
 0x05,0x00,0x06,0xd2,0x01,0x05,0x0e,0x24
 
 # GFX9: v_fma_f16 v5, v1, |v2|, v3    ; encoding: [0x05,0x02,0x06,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x02,0x06,0xd2,0x01,0x05,0x0e,0x04
 
+# GFX9: v_fma_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x06,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x08,0x06,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_fma_f16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x06,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x78,0x06,0xd2,0x01,0x05,0x0e,0x04
+
 # GFX9: v_fma_f16 v5, v1, v2, v3 clamp    ; encoding: [0x05,0x80,0x06,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x80,0x06,0xd2,0x01,0x05,0x0e,0x04
 
 # GFX9: v_fma_legacy_f16 v5, v1, v2, v3    ; encoding: [0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x04]
 0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x04
 
 # GFX9: v_fma_legacy_f16 v5, v1, v2, -v3    ; encoding: [0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x84]
 0x05,0x00,0xee,0xd1,0x01,0x05,0x0e,0x84

 0x05,0x00,0x07,0xd2,0x01,0x05,0xc2,0x03
 
 # GFX9: v_div_fixup_f16 v5, -v1, -v2, -v3    ; encoding: [0x05,0x00,0x07,0xd2,0x01,0x05,0x0e,0xe4]
 0x05,0x00,0x07,0xd2,0x01,0x05,0x0e,0xe4
 
 # GFX9: v_div_fixup_f16 v5, |v1|, |v2|, |v3|    ; encoding: [0x05,0x07,0x07,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x07,0x07,0xd2,0x01,0x05,0x0e,0x04
 
+# GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x07,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x08,0x07,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,1,0] ; encoding: [0x05,0x20,0x07,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x20,0x07,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_div_fixup_f16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x07,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x40,0x07,0xd2,0x01,0x05,0x0e,0x04
+
 # GFX9: v_div_fixup_f16 v5, v1, v2, v3 clamp    ; encoding: [0x05,0x80,0x07,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x80,0x07,0xd2,0x01,0x05,0x0e,0x04
 
 # GFX9: v_div_fixup_legacy_f16 v5, 0.5, v2, v3    ; encoding: [0x05,0x00,0xef,0xd1,0xf0,0x04,0x0e,0x04]
 0x05,0x00,0xef,0xd1,0xf0,0x04,0x0e,0x04
 
 # GFX9: v_div_fixup_legacy_f16 v5, v1, 0.5, v3    ; encoding: [0x05,0x00,0xef,0xd1,0x01,0xe1,0x0d,0x04]
 0x05,0x00,0xef,0xd1,0x01,0xe1,0x0d,0x04

 0x05,0x00,0x03,0xd2,0x01,0x05,0xc2,0x03
 
 # GFX9: v_mad_f16 v5, -v1, -v2, -v3    ; encoding: [0x05,0x00,0x03,0xd2,0x01,0x05,0x0e,0xe4]
 0x05,0x00,0x03,0xd2,0x01,0x05,0x0e,0xe4
 
 # GFX9: v_mad_f16 v5, |v1|, |v2|, |v3|    ; encoding: [0x05,0x07,0x03,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x07,0x03,0xd2,0x01,0x05,0x0e,0x04
 
+# GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x03,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x08,0x03,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,1,0,0] ; encoding: [0x05,0x10,0x03,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x10,0x03,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,1,0] ; encoding: [0x05,0x20,0x03,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x20,0x03,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x03,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x40,0x03,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_f16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x03,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x78,0x03,0xd2,0x01,0x05,0x0e,0x04
+
 # GFX9: v_mad_f16 v5, v1, v2, v3 clamp    ; encoding: [0x05,0x80,0x03,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x80,0x03,0xd2,0x01,0x05,0x0e,0x04
 
 # GFX9: v_mad_i16 v5, 0, v2, v3    ; encoding: [0x05,0x00,0x05,0xd2,0x80,0x04,0x0e,0x04]
 0x05,0x00,0x05,0xd2,0x80,0x04,0x0e,0x04
 
 # GFX9: v_mad_i16 v5, v1, -1, v3    ; encoding: [0x05,0x00,0x05,0xd2,0x01,0x83,0x0d,0x04]
 0x05,0x00,0x05,0xd2,0x01,0x83,0x0d,0x04
 
 # GFX9: v_mad_i16 v5, v1, v2, -4.0    ; encoding: [0x05,0x00,0x05,0xd2,0x01,0x05,0xde,0x03]
 0x05,0x00,0x05,0xd2,0x01,0x05,0xde,0x03
 
+# GFX9: v_mad_i16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x05,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x40,0x05,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_i16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x05,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x78,0x05,0xd2,0x01,0x05,0x0e,0x04
+
 # GFX9: v_mad_i16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x05,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x80,0x05,0xd2,0x01,0x05,0x0e,0x04
 
 # GFX9: v_mad_legacy_f16 v5, 0.5, v2, v3    ; encoding: [0x05,0x00,0xea,0xd1,0xf0,0x04,0x0e,0x04]
 0x05,0x00,0xea,0xd1,0xf0,0x04,0x0e,0x04
 
 # GFX9: v_mad_legacy_f16 v5, v1, 0.5, v3    ; encoding: [0x05,0x00,0xea,0xd1,0x01,0xe1,0x0d,0x04]
 0x05,0x00,0xea,0xd1,0x01,0xe1,0x0d,0x04

 0x05,0x00,0x04,0xd2,0x80,0x04,0x0e,0x04
 
 # GFX9: v_mad_u16 v5, v1, -1, v3    ; encoding: [0x05,0x00,0x04,0xd2,0x01,0x83,0x0d,0x04]
 0x05,0x00,0x04,0xd2,0x01,0x83,0x0d,0x04
 
 # GFX9: v_mad_u16 v5, v1, v2, -4.0    ; encoding: [0x05,0x00,0x04,0xd2,0x01,0x05,0xde,0x03]
 0x05,0x00,0x04,0xd2,0x01,0x05,0xde,0x03
 
+# GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x04,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x08,0x04,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[0,0,0,1] ; encoding: [0x05,0x40,0x04,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x40,0x04,0xd2,0x01,0x05,0x0e,0x04
+
+# GFX9: v_mad_u16 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x04,0xd2,0x01,0x05,0x0e,0x04]
+0x05,0x78,0x04,0xd2,0x01,0x05,0x0e,0x04
+
 # GFX9: v_mad_u16 v5, v1, v2, v3 clamp ; encoding: [0x05,0x80,0x04,0xd2,0x01,0x05,0x0e,0x04]
 0x05,0x80,0x04,0xd2,0x01,0x05,0x0e,0x04