llvm-project

b2f77ace - [AMDGPU] Select the op_sel version of v_mad_u16 on GFX9 (#236042)

[AMDGPU] Select the op_sel version of v_mad_u16 on GFX9 (#236042) On GFX9, select v_mad_u16 (V_MAD_U16_gfx9) for i16 multiply-add instead of v_mad_legacy_u16, so that sources taken from the high half of a packed register can be read with op_sel, as on GFX10+. The legacy opcode is still selected when the result is zero-extended, since it zero-fills the high 16 bits of the destination. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>