The order in which instructions are generated by gen_insn() influences
assignment to tmp registers. During generation, tmp instructions (e.g.
generate_V6_vassign_tmp) use vreg_src_off() to determine what kind of
register to use as source. If some instruction (e.g.
generate_V6_vmpyowh_64_acc) uses a tmp register but is generated prior
to the corresponding tmp instruction, the vregs_updated_tmp bit map
isn't updated in time.
Exmple:
{ v14.tmp = v16; v25 = v14 } This works properly because
generate_V6_vassign_tmp is generated before generate_V6_vassign
and the bit map is updated.
{ v15:14.tmp = vcombine(v21, v16); v25:24 += vmpyo(v18.w,v14.h) }
This does not work properly because vmpyo is generated before
vcombine and therefore the bit map does not yet know that there's
a tmp register.
The parentheses in the decoding function were in the wrong place.
Moving them to the correct location makes shuffling of .tmp vector
registers work as expected.
Signed-off-by: Marco Liebel <quic_mliebel@quicinc.com>
Reviewed-by: Taylor Simpson <tsimpson@quicinc.com>
Tested-by: Taylor Simpson <tsimpson@quicinc.com>
Signed-off-by: Taylor Simpson <tsimpson@quicinc.com>
Reviewed-by: Brian Cain <bcain@quicinc.com>
Message-Id: <20230522174708.464197-1-quic_mliebel@quicinc.com>
3fd49e2217