llvm-project

dfba5a3e - [AMDGPU] Fold high-half extracts into op_sel for i16 mad (#231410)

[AMDGPU] Fold high-half extracts into op_sel for i16 mad (#231410) Use the VOP3OpSel complex pattern for the sources of the i16 mul+add patterns that select the op_sel-capable v_mad_u16 on GFX10+, so that a source taken from the high half of a packed register is read with op_sel instead of being extracted with a shift first. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>