[Bug] The REP reorganization broke NT 3.1 on 386DX

[Battler]

NT 3.1 now just shows a blue screen saying you might be using an early version of the 386DX.

It's probably this: set TF bit (0x0100) in eflags causes Debug exception (interrupt 0x01) only at completion of rep movsb (taken from here: https://www.geoffchappell.com/studies/w ... cation.htm ).

Edit: And there's also the fact REP with invalid instruction should just ignore and return pointing to the next instruction, something PCem did before but no longer does.

Also this:

Code: Select all

        if (x86_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32])
                return x86_opcodes_REPNE[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);
        return x86_opcodes[(fetchdat & 0xff) | cpu_state.op32](fetchdat >> 8);

Is this correct? Because every element of x86_opcodes_REPNE and x86_opcodes_REPE is xet to at least ILLEGAL, so it's never NULL, therefore that second return never gets executed.

[Greatpsycho]

This patch applies non-string instruction followed by REP instruction to be processed as real machine does. REP prefix should not occur exception #6 even if non-string instruction is followed.

[Battler]

Thank you very much, this fixes one half of the problem. The other half remains, which is REP not issuing the debug exception on every repeat when the trap flag is set, but only on command completion.

[SarahWalker]

Fixed the invalid opcode value in the REP tables in rev 814 - it should always have been NULL/0.

[Battler]

Actually, Greatpsycho did it correctly - it should be a simple ignore of the instruction, not a NULL which according to your code will just execute whatever the instruction there is, which is incorrect behavior.

[SarahWalker]

Rev 815 fixes the trap issue.

[SarahWalker]

Actually the behaviour with the REP prefix is undefined. However some stuff does use REP RET and expect a normal RET to be executed - the old REP code had specific handling for this.

[Battler]

- SarahWalker: Then add a specific handler for RET. Some Linuxes expect REP on invalid instruction to be ignored, and I think reenigne said that happens as well.

[SarahWalker]

Care to name one?

[Battler]

Gentoo was named in this thread: https://pcem-emulator.co.uk/phpBB3/view ... ?f=3&t=423 where I posted my original REP fix patch that you accepted.

Edit: OK, after doing some tests in VirtualBox, it seems that when REP is used with a non-string instruction, the REP prefix is ignored and the instruction is executed as if there was no REP prefix, which means, Sarah, that your fix was actually the correct one.

Edit #2: And it seems that's precisely what both my old patch and Greatpsycho's patch also do, except in slightly different ways.