On 6/19/20 12:06 PM, Peter Bergner wrote: > On 6/19/20 11:45 AM, Segher Boessenkool wrote: >>> +(define_insn_and_split "*mma_assemble_acc" >>> + [(set (match_operand:PXI 0 "fpr_reg_operand" "=d") >>> + (unspec:PXI [(match_operand:PXI 1 "mma_input_operand" "mwa") >>> + (match_operand:PXI 2 "mma_input_operand" "mwa") >>> + (match_operand:PXI 3 "mma_input_operand" "mwa") >>> + (match_operand:PXI 4 "mma_input_operand" "mwa")] >>> + UNSPEC_MMA_ASSEMBLE_ACC))] >> >> I would expect all those four last match_operand to be :V16QI, so why >> does it use this strange mode? > > Must be a cut/paste error and probably why we saw mode == PXImode > in the mma_input_operand predicate. I'll change that and the > predicate and retest. Thanks for pointing that out!
Yes, cut/paste error. I changed those to all V16QI and modified the mma_assemble_input_operand predicate to match (and made your other suggested changes) and retesting came back clean, so I pushed the patch with those changes. Thanks for your reviews!!! Peter
