Because neither Windows, nor Linux, nor Mac OS, BSD, BeOS or any other well conceived and "complete" OS ever made use of this ability, it has proven to be unreliable between CPU revisions and require many, many work-arounds to achieve, even on a completely new OS without any legacy software issues. The attempt to unify these into a single hardware API was called PAE. (
To view the content, you need to sign in or register
)
Now, it works, but is a huge pain in the backside, and when Windows attempts to utilise it (because of their "legacy software" restrictions) they can only do so to "
claw back" the (almost) 1 gigabyte of that 4Gig the OS usually gives over to making other things run smoother. (Memory mapped files and IO, BIOS extensions and fixed addressing kernel modules)
The best implementations of PAE (and all of the ones on 64-bit
capable processors) also include PAE-NX with enables hardware DEP in Windows. This system ensures that nasty things like the GFantasy (common in PT servers and many of the older clients) section addition are illegal at a hardware, CPU level. No memory block / section should be Writeable
and Executable.
If you have done the "over 2Gig patch" on the server PE and your executable includes such a section (ewww) then you should break the additions in it down into code, constants and variables, and give each kind of addition a section of it's own, with the correct permissions.
Additionally, the routines in the executable do not expect to be able to allocate, or address more than 2Gig, and many "loops" may expect to "wrap around" memory pointers which are incremented above 0x80000000. (By which I mean that they expect that 0x80000000 + 1 = 0x00) This will no longer happen. (0x80000000 + 1 will now equal 0x80000001) That
may greatly confuse them... or it may fix bugs in them if their developers (who probably couldn't find a machine with more than a Gig of RAM for the entire system back between 2000 and 2006) may not have been aware of this limitation to debug it. Many developers assumed that the user "hype" that a 32-bit CPU will address 4Gig was just simply "true" in all circumstances, and addresses would
not wrap around until you reached 0xFFFFFFFF. (a logical assumption, until you understand how the IBM compatible hardware and Windows OS (and Linux) use it. XD (If MacOS had remained on the PPC series processors and it's EFI hardware, I believe the expected result would be true on their 32-bit CPUs. I knew of this, because I was keen on the BeOS, before it got taken off desktops and turned into the newer incarnations of PalmOS.

They supported bot x86 and PPC and discussed the differences of the platforms from the OS perspective. ^_^)