While constructing very fast disasm engine which returns only instruction size (needed for an APIC emulator inside hypervisor), I studied CPU manuals again (AMD as well Intel) and I came across this thing. FASM version 1.71.22 generates correctly these opcodes:

mov [qword 0FEE00000h],eax -> A30000E0FE00000000
mov [dword 0FEE00000h],eax -> 67A30000E0FE

In CPU manuals I found a third way how to encode absolute addressing mode:
mod = 00, r/m = 100, SIB. base = 101 (none), SIB.index = 100 (none), SIB.scale = 0, 1, 2, 4
I was unable to tell FASM how to compile this way (is this possible?) so I encoded it manually as:
8904250000E0FE -> mov [FFFFFFFFFEE00000],eax
8904650000E0FE detto
8904A50000E0FE detto
8904E50000E0FE detto
note in this case CPU sign extends the dword address

Or someone may develop a macro for that using this formula:
; mod = 00, r/m = 100, SIB.base = 101, SIB.index = 100, SIB.scale=xx
db 89h, 100b, 00100101b, 00h, 00h, 0E0h, 0FEh
db 89h, 100b, 01100101b, 00h, 00h, 0E0h, 0FEh
db 89h, 100b, 10100101b, 00h, 00h, 0E0h, 0FEh
db 89h, 100b, 11100101b, 00h, 00h, 0E0h, 0FEh

There is nothing to solve, FASM generates fine code using first 2 choices. I just came across the third way which is very likely useless for FASM compiler, but should be taken into consideration when writing a disassembler/emulator etc.

I think it's what was discussed here: https://board.flatassembler.net/topic.php?p=211682#211682
Current versions of fasm assemble the examples given there the same way as the fasmg-based prototype.