AMDGPU: Mark global/scratch_load_d16 as D16Buf

[ruiling]

Summary

They behave the same way as other d16 buffer instructions.

[arsenm]

This is also missing from the DS instructions

[foad]

I am a bit surprised that this does not break the only existing use of this flag in SIInstrInfo::pseudoToMCOpcode.

The flag needs a better name if it is going to apply to all _d16/_d16_hi instructions.

llvm/lib/Target/AMDGPU/FLATInstructions.td

 defm GLOBAL_LOAD_SBYTE    : FLAT_Global_Load_Pseudo <"global_load_sbyte", VGPR_32>;
 defm GLOBAL_LOAD_USHORT   : FLAT_Global_Load_Pseudo <"global_load_ushort", VGPR_32>;
 defm GLOBAL_LOAD_SSHORT   : FLAT_Global_Load_Pseudo <"global_load_sshort", VGPR_32>;
 defm GLOBAL_LOAD_DWORD    : FLAT_Global_Load_Pseudo <"global_load_dword", VGPR_32>;
 defm GLOBAL_LOAD_DWORDX2  : FLAT_Global_Load_Pseudo <"global_load_dwordx2", VReg_64>;
 defm GLOBAL_LOAD_DWORDX3  : FLAT_Global_Load_Pseudo <"global_load_dwordx3", VReg_96>;
 defm GLOBAL_LOAD_DWORDX4  : FLAT_Global_Load_Pseudo <"global_load_dwordx4", VReg_128>;
 
+let D16Buf = 1 in {
 defm GLOBAL_LOAD_UBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16", VGPR_32, 1>;
 defm GLOBAL_LOAD_UBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_ubyte_d16_hi", VGPR_32, 1>;
 defm GLOBAL_LOAD_SBYTE_D16    : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16", VGPR_32, 1>;
 defm GLOBAL_LOAD_SBYTE_D16_HI : FLAT_Global_Load_Pseudo <"global_load_sbyte_d16_hi", VGPR_32, 1>;
 defm GLOBAL_LOAD_SHORT_D16    : FLAT_Global_Load_Pseudo <"global_load_short_d16", VGPR_32, 1>;
 defm GLOBAL_LOAD_SHORT_D16_HI : FLAT_Global_Load_Pseudo <"global_load_short_d16_hi", VGPR_32, 1>;
+}
+
 let OtherPredicates = [HasGFX10_BEncoding] in
 defm GLOBAL_LOAD_DWORD_ADDTID : FLAT_Global_Load_AddTid_Pseudo <"global_load_dword_addtid", VGPR_32>;
 
 defm GLOBAL_STORE_BYTE    : FLAT_Global_Store_Pseudo <"global_store_byte", VGPR_32>;
 defm GLOBAL_STORE_SHORT   : FLAT_Global_Store_Pseudo <"global_store_short", VGPR_32>;
 defm GLOBAL_STORE_DWORD   : FLAT_Global_Store_Pseudo <"global_store_dword", VGPR_32>;
 defm GLOBAL_STORE_DWORDX2 : FLAT_Global_Store_Pseudo <"global_store_dwordx2", VReg_64>;
 defm GLOBAL_STORE_DWORDX3 : FLAT_Global_Store_Pseudo <"global_store_dwordx3", VReg_96>;

 defm SCRATCH_LOAD_SBYTE    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte", VGPR_32>;
 defm SCRATCH_LOAD_USHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_ushort", VGPR_32>;
 defm SCRATCH_LOAD_SSHORT   : FLAT_Scratch_Load_Pseudo <"scratch_load_sshort", VGPR_32>;
 defm SCRATCH_LOAD_DWORD    : FLAT_Scratch_Load_Pseudo <"scratch_load_dword", VGPR_32>;
 defm SCRATCH_LOAD_DWORDX2  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx2", VReg_64>;
 defm SCRATCH_LOAD_DWORDX3  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx3", VReg_96>;
 defm SCRATCH_LOAD_DWORDX4  : FLAT_Scratch_Load_Pseudo <"scratch_load_dwordx4", VReg_128>;
 
+let D16Buf = 1 in {
 defm SCRATCH_LOAD_UBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16", VGPR_32, 1>;
 defm SCRATCH_LOAD_UBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_ubyte_d16_hi", VGPR_32, 1>;
 defm SCRATCH_LOAD_SBYTE_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16", VGPR_32, 1>;
 defm SCRATCH_LOAD_SBYTE_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_sbyte_d16_hi", VGPR_32, 1>;
 defm SCRATCH_LOAD_SHORT_D16    : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16", VGPR_32, 1>;
 defm SCRATCH_LOAD_SHORT_D16_HI : FLAT_Scratch_Load_Pseudo <"scratch_load_short_d16_hi", VGPR_32, 1>;
+}
 
 defm SCRATCH_STORE_BYTE    : FLAT_Scratch_Store_Pseudo <"scratch_store_byte", VGPR_32>;
 defm SCRATCH_STORE_SHORT   : FLAT_Scratch_Store_Pseudo <"scratch_store_short", VGPR_32>;
 defm SCRATCH_STORE_DWORD   : FLAT_Scratch_Store_Pseudo <"scratch_store_dword", VGPR_32>;
 defm SCRATCH_STORE_DWORDX2 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx2", VReg_64>;
 defm SCRATCH_STORE_DWORDX3 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx3", VReg_96>;
 defm SCRATCH_STORE_DWORDX4 : FLAT_Scratch_Store_Pseudo <"scratch_store_dwordx4", VReg_128>;