Support for original CL-GD5446 BIOSes

[Vort]

I tried to use several BIOSes for CL-GD5446 (5446BV.ZIP, stb nitro64v.zip, cl-gd5446_pci_vga_bios_version_1.31.zip) and all of them produce visual glitches, from black rectangles with some modes to wrong proportions and wrong colors with other modes:

1. Are there any original BIOS versions for CL-GD5446, which works correctly with Bochs?
2. Does it makes sense to improve support so BIOSes I mentioned above work correctly? With low priority, but anyway.

[Vort]

I decided to figure out why black rectangles appear and found these lines:
https://github.com/bochs-emu/VGABIOS/blob/cdb98a155773ff125f6632eee957f5fb43e2b830/vgabios/clext.c#L550
mov ax, #0xffff ; set to 0xff to keep win 2K happy

https://lists.archive.carbon60.com/xen/devel/83656
Thanks to Fabrice Bellard on the QEMU project who came up with a cleaner way of clearing the VGA memory on mode changes. This patch clears the memory in the BIOS call rather than when the Cirrus Logic register changes, which more closely matches what happens on the real hardware.

/* Win2K seems to assume that the VRAM is set to 0xff
 * whenever VGA/SVGA mode changes
 */

So what is really happening here? I see two possibilities:

1. Windows 2000 (and Windows XP) are not fully compatible with CL-GD5446 and display black rectangles on real hardware (with original VGA BIOS).
2. Windows 2k and XP are compatible with real hardware, but patch is wrong and FFs appear in VRAM via some different mechanism.

Does anybody have real hardware to check or knows for sure how real hardware behave in Win 2k & XP?

[Vort]

After further investigation I figured out this is just a bug.
No need to initialize RAM with FFs - there should be 00s fill like original BIOS does.
Driver will initialize this memory to FF by itself.
If bugs won't prevent it from happening.

First bug is located here:

Bochs/bochs/iodev/display/svga_cirrus.cc

Line 2125 in eb477d7

else if (((old_value & CIRRUS_BLT_START) == 0) &&

There should be no else there, because with else transition of value from 0x04 to 0x02 will fail to produce blt operation.
Both svga_reset_bitblt and svga_bitblt should be called in such case if I understand correctly.
But this is not all bugs: removing else makes Bochs crash.

I will try to track this problem further and create Issue about it (upd: done: #319).