fleroviux / lunar Goto Github PK

Updating the computation results on each singular byte write is pretty inefficient.
This can improved if we implement specialized 16-bit and 32-bit MMIO accessors.

currently the ARM7 is treated like an ARM9 processor with an ARMv5TE ISA

I expect the following things need to be changed for proper ARM7 support:

• load instructions must not "branch and exchange" on r15 load
• adjust alignment edge-cases
• remove CP15 (throw undefined exception)
• add CP14 stub (doesn't do anything but prevents undefined exception)
• remove ARMv5 and ARMv6 opcodes
• handle LDM/STM edge-cases (done in ARM, only Thumb needs work)

The sequence of events is:

1. ARM9MemoryBus constructor
2. UpdateMemoryMap(0, 0x100000000ULL);
3. case 0x06: table[index] = VisitVRAMByAddress<GetUnsafePointerFunctor>(address);
4. GetUnsafePointerFunctor
5. GetUnsafePointer

At the first attempt to read an address above the 0x068A3FFF mark (VRAM 656KB limit) everything blows up.
Could fix it checking before calling VisitVRAMByAddress:

case 0x06: {
if ((address & 0xFFFFF) < 0xa4000) //Check
table[index] = VisitVRAMByAddress<GetUnsafePointerFunctor<u8>>(address);
break;
}

But for current lack of whole design complete understanding and fleroviux intentions, I'm just documenting here and not opening a pull request.

Tested on:

Platform: Win10
Compiler: Clang 11 under MSVCRT (MSVC can't handle the metaprogramming and won't build, at least in the current stable 14.29)
(logically should be an issue in other platforms, logically)

The sequence of events is:

1. Everything out of scope
2. ~Core
3. ~CoreImpl
4. ~Interconnect
5. ~APU

~APU tries to call close on its audio device, but apparently the check for null pointer doesn't help, because despite having an address it's invalid memory since sometime in the recent past(?). So the application crashes trying to access invalid memory.

Since Interconnect it's marked to be extinguished from existence, this rework could fix the destruction order and solve this issue.

Can't really suggest a solution, for currently lacking understanding of the whole program structure.

Tested on:

Platform: Win10
Compiler: Clang 11 under MSVCRT (Tests can only be performed after fixing #22, otherwise won't run)

Apply the same fix I did in NanoBoyAdvance for the switch between FIQ mode and modes other than User/System modes.

the ITCM and DTCM actually can be disabled via CP15 commands. we are not emulating this currently.

Wondering what your goal is, is it to be similar to Noo-DS but focus more on stable releases and/or otther things?

Curious I am, watching this and at least* one other repo to see which turns out better.

Hope all goes well

Culled polygon strips often commit more vertices to vertex RAM than they should.

The wrong code in question it this:
https://github.com/fleroviux/lunar/blob/master/src/lunar/src/nds/video_unit/gpu/primitive_assembly.cpp#L125-L141

• handle V-count coincidence IRQ more accurately (see the implementation in NanoboyAdvance).
• the setting of the H-blank flag is reportedly slightly delayed on NDS7
• NDS7 and NDS9 actually have separate DISPSTAT registers. (see d2c3a25)
• implement window bugs / glitches

Trying to set all IE bits gives FFFFFFFFh (DSi7) or FFFFFF7Fh (DSi9).

SIO/RCNT/RTC IRQ can only be used on the ARM7 core and trying to enable it on the ARM9 doesn't set the bit.