Gives "value out of range" on that rva statement.
It seems to fail for any imagebase around 7FFF0000h, the actual value varying depending where the rva statement is used relative to start of code. Seems FASM is overflowing on signed dwords?

It's the bug caused by the fact, that during the calculations fasm extends all values to 64-bit, but while it uses sign-extension on the $ value, the code base value which is substracted from $ to get the RVA was zero-extended, and therefore the result was wrong, a large negative value that couldn't even fit in 32 bits. It's fixed in beta 1 of fasm 1.52.