I've been told that initializing the FPU is a monstrous task.

However, if one reads the August 2007 Intel Manual Volume 3A, Sections 9.1 & 9.2 it is actually a very simple task and it outlines it very well.

Here's the recommended initialization for a Pentium Processor or newer with a math coprocessor:



The two bits being set in the CR0 register are the MP bit (bit 2) and the NE bit (bit 6).

The MP bit tells whether or not the WAIT/FWAIT instructions react to the setting of the Task Switch (TS) flag/bit (bit 4 in CR0 register). If this bit is cleared, then the WAIT/FWAIT instructions will ignore the TS flag. Otherwise, if this bit is set and the TS flag is set, the processor will generate a device-not-avaiable exception.

The NE bit tells whether floating point exceptions are handled internally (bit set) or via an external interrupt (bit cleared).

Another bit which is vital is bit number 3, the EM bit. If this bit is set, it means that there is no floating-point coprocessor (x87). If it is cleared, then there is a floating-point coprocessor.

For more info, check the above reference where I got this info from. Also, in sections 9.7 and 9.8 it talks a great deal about mode switching into P-Mode and Real Mode. I'd suggest that people new to 32-bit OS programming would read it. It looks good.

Well, the complications of FPU init comes when you only want to save FPU task state per-thread if the FPU has actually been used... same goes for SSE registers.

You might also want to set up the FPU control word for rounding etc.

Where does it mention that in the Intel Manuals or how can I do it?

nobody knows
seriouslly, it's surelly in the system programming manual
section about all FPu, SSE.
but for sure it's not in task managment, i know this section by heart and i never saw any FPU, SSE reference with task switch

in fact, it's mostlly in the forums that you'll find how to do.

but in FPU and SSE sections, they give the manner to save and restore FPU and SSE states.
you just have to include this in a task switch
then task switch needs to be limited to the periodic style.

to switch with hardware, there are too many possibilities to make it easily reliable. jump, call, load a tss, interrupt,..., all these operations isues to a task switch

so, invalidate hardware task switch and make one software.
mans says that the software TS is faster, so why use the Hardware method?