UMLAL with rd_hi = rd_lo yields wrong result

[ladystarbreeze]

In mGBA 0.8.4, the UMLAL instruction yields incorrect results when rd_hi and rd_lo are the same register.
On hardware (3DS + open-agb-firm), UMLAL r4, r4, r4, r4 with r4 = 20001h sets r4 to 20005h. mGBA sets r4 to 60006h. Presumably, this is the correct order of operations:

a = do_unsigned_long_multiplication(rm, rs)
a += concat(rd_hi, rd_lo)
rd_lo = a_lo
rd_hi = a_hi

[endrift]

Can you please actually upload the test ROM so I don't have to make one myself?

[ladystarbreeze]

Sorry, here's the test ROM.

umlal.zip