Details of the Far JMP into Protected Mode

[Haroogan]

There are 2 different ways to achieve it. So I'm going to post 2 different snippets respectively:

#1

Code: Select all

BITS 16

%define LinearAddress(SegmentAddress,OffsetAddress) SegmentAddress * 0x10 + OffsetAddress

Code16:
	...
	JMP 8:LinearAddress(BootLoader2.SEGMENT_ADDRESS,Code32)

BITS 32

Code32:
	...

#2

Code: Select all

BITS 16

%define LinearAddress(SegmentAddress,OffsetAddress) SegmentAddress * 0x10 + OffsetAddress
%macro JMP32 2
	DB 0x66
	DB 0xEA
	DD %2;Address
	DW %1;Selector
%endmacro

Code16:
	...
	JMP32 8,LinearAddress(BootLoader2.SEGMENT_ADDRESS,Code32)

BITS 32

Code32:
	...

Both of them work just fine. But I've noticed that the 2nd one occupies more memory. So the 2nd one is really the 32 bit jump and the 1st one is not. But some people still recommend the 1st one and others - the 2nd one. Which one should I choose? And why? What's the essence of their difference?

[egos]

#2 of course if destination code segment is 32-bit. There is a probability that high word of eip is not equal to zero.

[Owen]

#2 of course if destination code segment is 32-bit. There is a probability that high word of eip is not equal to zero.

Utter junk. Both 32-bit and 64-bit jumps set the full width of EIP

[egos]

And what about 16-bit jump

[Owen]

It zeros out the upper bits of rIP. If the upper bits were not zeroed, you would violate the segment limit...

A good assembler will therefore detect when the address will fit in 16-bits and use the 16-bit version of the jump

[egos]

It zeros out the upper bits of rIP.

Maybe in emulators but not in real CPUs.

If the upper bits were not zeroed, you would violate the segment limit...

CPU ignores high word of eip while executing 16-bit code therefore an error doesn't occur.

A good assembler will therefore detect when the address will fit in 16-bits and use the 16-bit version of the jump

And this behaviour will be fatal if high word of eip during execution is not equal to zero. To jump from 16-bit code to 32-bit code you should use forced 32-bit far jump anywhere except if you sure that eip.31-16=0 (or if you want to jump to eip.31-16 shl 16 + 16-bit destination offset ).

[Owen]

If the operand-size attribute is 16, the upper two bytes of the EIP register are cleared.

This is from the Intel manuals, volume 2A, at time of writing page 3-554.

[Haroogan]

Anyway, I believe #2 is safer and more stable.

[egos]

If the operand-size attribute is 16, the upper two bytes of the EIP register are cleared.

This is from the Intel manuals, volume 2A, at time of writing page 3-554.

Good argument but you didn't take into account that some people make CPUs and another write manuals for them. Also I can assume that behaviour of CPUs can change from one model to another. I tested some Pentium processors and they worked precisely as I said above.

[Owen]

This is from the Intel manuals, volume 2A, at time of writing page 3-554.

Good argument but you didn't take into account that some people make CPUs and another write manuals for them. Also I can assume that behaviour of CPUs can change from one model to another. I tested some Pentium processors and they worked precisely as I said above.

The behavior of the instruction is architecturally defined. It cannot change from one processor model to another. Can you cite a reference on a processor having the bug you mentioned? I fail to believe that such a bug would not be documented anywhere on the internet, particularly when the Pentium dates back to the days of segmented software. I certainly can't find it in Intel's Pentium errata sheet.

[egos]

It was long time ago when I used 16-bit far jump to enter PM and recommended to other people to do so

Code: Select all

; fasm syntax
  if Startup32<10000h
  jmp KCODE:Startup32
  else
  jmp fword KCODE:Startup32
  end if

When I saw what happened when eip.31-16 != 0 on my testing computer (probably it was same one that I have now) I was shocked Now I have Intel Pentium MMX for testing. We can write testing code here and then test old and modern (I have Intel Core2 Quad Q9300 and Intel i7 920 at home too) CPUs with it.

[xenos]

I guess you really disliked that 32 bit jump when you wrote "fword" instead of "dword"...

[egos]

In fasm syntax fword (or pword) means six-byte pointer. "dword" in this case (in 16-bit code) is default and means four-byte pointer.

[xenos]

Ah, sorry, you're right - I haven't used fasm for a very long time... I was thinking about something like "ljmpl" in GNU as syntax, or the even more similar "jmp dword Seg:Off" in NASM syntax...